Document - 2019-03-07T080258.777SPECTRALTEORIJA PDF

LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES'
Managing Editor: Professor J.W.S. Cassels, Department of Pure Mathematics

and Mathematical Statistics, 16 Mill Lane, Cambridge CB2 1SB, England
The books in the series listed below are available from booksellers, or, in
case of difficulty, from Cambridge University Press.
4 Algebraic topology, J.F.ADAMS

5 Commutative algebra, J.T.KNIGHT
8 Integration and harmonic analysis on compact groups, R.E.EDWARDS
11 New developments in topology, G.SEGAL (ed)
12 Symposium on complex analysis, J.CLUNIE & W.K.HAYMAN (eds)
13 Combinatorics, T.P.McDONOUGH & V.C.MAVRON (eds)
16 Topics in finite groups, T.M.GAGEN
17 Differential germs and catastrophes, Th.BROCKER & L.LANDER
18 A geometric approach to homology theory, S.BUONCRISTIANO, C.P.ROURKE
& B.J.SANDERSON
20 Sheaf theory, B.R.TENNISON
21 Automatic continuity of linear operators, A.M.SINCLAIR
23 Parallelisms of complete designs, P.J.CAMERON
24 The topology of Stiefel manifolds, I.M.JAMES
25 Lie groups and compact groups, J.F.PRICE
26 Transformation groups, C.KOSNIOWSKI (ed)
27 Skew field constructions, P.M.COHN
29 Pontryagin duality and the structure of LCA groups, S.A.MORRIS
30 Interaction models, N.L.BIGGS
31 Continuous crossed products and type III von Neumann algebras,A.VAN DAELE
32 Uniform algebras and Jensen measures, T.W.GAMELIN
34 Representation theory of Lie groups, M.F. ATIYAH et al.
35 Trace ideals and their applications, B.SIMON
36 Homological group theory, C.T.C.WALL (ed)
37 Partially ordered rings and semi-algebraic geometry, G.W.BRUMFIEL
38 Surveys in combinatorics, B.BOLLOBAS (ed)
39 Affine sets and affine groups, D.G.NORTHCOTT
40 Introduction to Hp spaces, P.J.KOOSIS
41 Theory and applications of Hopf bifurcation, B.D.HASSARD,
N.D.KAZARINOFF & Y-H.WAN
42 Topics in the theory of group presentations, D.L.JOHNSON
43 Graphs, codes and designs, P.J.CAMERON & J.H.VAN LINT
44 Z/2-homotopy theory, M.C.CRABB
45 Recursion theory: its generalisations and applications, F.R.DRAKE
& S.S.WAINER (eds)
46 p-adic analysis: a short course on recent work, N.KOBLITZ
47 Coding the Universe, A.BELLER, R.JENSEN & P.WELCH
48 Low-dimensional topology, R.BROWN & T.L.THICKSTUN (eds)
49 Finite geometries and designs,P.CAMERON, J.W.P.HIRSCHFELD & D.R.HUGHES (eds)
50 Commutator calculus and groups of homotopy classes, H.J.BAUES
51 Synthetic differential geometry, A.KOCK
52 Combinatorics, H.N.V.TEMPERLEY (ed)
54 Markov process and related problems of analysis, E.B.DYNKIN
55 Ordered permutation groups, A.M.W.GLASS
56 Journees arithmetiques, J.V.ARMITAGE (ed)
57 Techniques of geometric topology, R.A.FENN
58 Singularities of smooth functions and maps, J.A.MARTINET
59 Applicable differential geometry, M.CRAMPIN & F.A.E.PIRANI
60 Integrable systems, S.P.NOVIKOV et al
61 The core model, A.DODD
62 Economics for mathematicians, J.W.S.CASSELS
63 Continuous semigroups in Banach algebras, A.M.SINCLAIR
64 Basic concepts of enriched category theory, G.M.KELLY
65 Several complex variables and complex manifolds I, M.J.FIELD
66 Several complex variables and complex manifolds II, M.J.FIELD
67 Classification problems in ergodic theory, W.PARRY & S.TUNCEL
68 Complex algebraic surfaces, A.BEAUVILLE
69 Representation theory, I.M.GELFAND et al.
70 Stochastic differential equations on manifolds, K.D.ELWORTHY
71 Groups - St Andrews 1981, C.M.CAMPBELL & E.F.ROBERTSON (eds)
72 Commutative algebra: Durham 1981, R.Y.SHARP (ed)
73 Riemann surfaces: a view towards several complex variables,A.T.HUCKLEBERRY
74 Symmetric designs: an algebraic approach, E.S.LANDER
75 New geometric splittings of classical knots, L.SIEBENMANN & F.BONAHON
76 Spectral theory of linear differential operators and comparison
algebras, H.O.CORDES
77 Isolated singular points on complete intersections, E.J.N.LOOIJENGA
78 A primer on Riemann surfaces, A.F.BEARDON
79 Probability, statistics and analysis, J.F.C.KINGMAN & G.E.H.REUTER (eds)
80 Introduction to the representation theory of compact and locally
compact groups, A.ROBERT
81 Skew fields, P.K.DRAXL
82 Surveys in combinatorics, E.K.LLOYD (ed)
83 Homogeneous structures on Riemannian manifolds, F.TRICERRI & L.VANHECKE
84 Finite group algebras and their modules, P.LANDROCK
85 Solitons, P.G.DRAZIN
86 Topological topics, I.M.JAMES (ed)
87 Surveys in set theory, A.R.D.MATHIAS (ed)
88 FPF ring theory, C.FAITH & S.PAGE
89 An F-space sampler, N.J.KALTON, N.T.PECK & J.W.ROBERTS
90 Polytopes and symmetry, S.A.ROBERTSON
91 Classgroups of group rings, M.J.TAYLOR
92 Representation of rings over skew fields, A.H.SCHOFIELD
93 Aspects of topology, I.M.JAMES & E.H.KRONHEIMER (eds)
94 Representations of general linear groups, G.D.JAMES
95 Low-dimensional topology 1982, R.A.FENN (ed)
96 Diophantine equations over function fields, R.C.MASON
97 Varieties of constructive mathematics, D.S.BRIDGES & F.RICHMAN
98 Localization in Noetherian rings, A.V.JATEGAONKAR
99 Methods of differential geometry in algebraic topology,
M.KAROUBI & C.LERUSTE
100 Stopping time techniques for analysts and probabilists, L.EGGHE
101 Groups and geometry, ROGER C.LYNDON
102 Topology of the automorphism group of a free group, S.M.GERSTEN
103 Surveys in combinatorics 1985, I.ANDERSEN (ed)
104 Elliptic structures on 3-manifolds, C.B.THOMAS
105 A local spectral theory for closed operators, I.ERDELYI & WANG SHENGWANG
106 Syzygies, E.G.EVANS & P.GRIFFITH
107 Compactification of Siegel moduli schemes, C-L.CHAI
108 Some topics in graph theory, H.P.YAP
109 Diophantine analysis, J.LOXTON & A.VAN DER POORTEN (eds)
110 An introduction to surreal numbers, H.GONSHOR
111 Analytical and geometric aspects of hyperbolic space, D.B.A.EPSTEIN (ed)
112 Low dimensional topology and Kleinian groups, D.B.A.EPSTEIN (ed)
113 Lectures on the asymptotic theory of ideals, D.REES
114 Lectures on Bochner-Riesz means, K.M.DAVIS & Y-C.CHANG
115 An introduction to independence for analysts, H.G.DALES & W.H.WOODIN
116 Representations of algebras, P.J.WEBB (ed)
117 Homotopy theory, E.REES & J.D.S.JONES (eds)
118 Skew linear groups, M.SHIRVANI & B.WEHRFRITZ
London Mathematical Society Lecture Note Series. 76
Spectral Theory of
Linear Differential Operators and
Comparison Algebras
H. 0. CORDES
University of California, Berkeley
CAMBRIDGE UNIVERSITY PRESS

Cambridge
London New York New Rochelle
Melbourne Sydney
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo
Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521284431
© Cambridge University Press 1987
This publication is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 1987

Re-issued in this digitally printed version 2007
A catalogue record for this publication is available from the British Library
Library of Congress Cataloguing in Publication data
Cordes, H. O. (Heinz Otto), 1925-

Spectral theory of linear differential operators and
comparison algebras.
(London Mathematical Society lecture note series ; 76)

1. Differential operators. 2. Linear operators.
1. Title.
QA329.4.C67 1986 515.7'242 85-47935
ISBN 978-0-521-28443-1 paperback

P R E F A C E
The main purpose of this volume is to introduce the reader

to the concept of comparison algebra, defined as a type of C*-al-
gebra of singular integral operators, generally on a noncompact
manifold, generated by an elliptic second order differential ex-
pression, and certain classes of multipliers and 'Riesz-operators'
As for singular integral operators on In or on a compact manifold
the Fredholm properties of operators in such an algebra are gover-
ned by a symbol homomorphism. However, for noncompact manifolds
the symbol is of special interest at infinity. In particular the
structure of the symbol space over infinity is of interest, and
the fact, that the symbol no longer needs to be complex-valued
there.
The first attempts of the author to make a systematic pre-
sentation of this material happened at Berkeley (1966) and at
Lund (1970/71). Especially the second lecture exists in form of
(somewhat ragged) notes [CS] The cases of the Laplace compari-
.
son algebra of Rn and the half-space were presented in [C1] .
In the course of laying out theory of comparison algebras

we had to develop in details spectral theory of differential ope-
rators, as well as many of the basic properties of elliptic second
order differential operators. This was done in the first four
chapters. Comparison algebras (in L2-spaces and L2-Sobolev spaces)
are discussed in chapters V to IX Finally, in chapter X we
.
recall the basic facts of theory of Fredholm operators, partly

without proofs.
The material has been with the author for more than 20 years
and has been subject of innumerable discussions with students and
vi
associates. Accordingly it is almost impossible to recall in

detail the origin of the various concepts introduced. Especially
we are indebted to E. Herman, M. Breuer, E. Luft, M. Taylor,
R. McOwen, A. Erkip, D.Williams, H. Sohrab, in chronological, not
alphabetical order.
We are indebted to S.H.Doong, S.Melo, R.Rainsberger, M.Arse-
novic for help with proof reading. This volume originally was
planned under the title 'techniques of pseudodifferential opera-
tors', but then split into two parts, with the second yet to
appear. We are grateful to the publisher, Cambridge University
Press, for cooperation and patience in waiting for the manuscript.
Berkeley, September 1986 Heinz 0. Cordes

TABLE OF CONTENTS
Chapter 1. Abstract spectral theory in Hilbert spaces 1

1.1. Unbounded linear operators on Banach and
Hilbert spaces 1
1.2. Self-adjoint extensions of hermitian operators 6
1.3. On the spectral theorem for self-adjoint

operators. 12
1.4. Proof of the spectral theorem 17
1.5. A result on powers of positive operators 24
1.6. On HS-chains 29
Chapter 2. Spectral theory of differential operators 35
2.1. Linear differential operators on a subdomain of En 35
2.2. Generalized boundary problems;ordinary
differential expressions 41
2.3. Singular endpoints of a 2r-th order
Sturm-Liouville problem 49
2.4. The spectral theorem for a second order expression 53
Chapter 3. Second order elliptic expressions on manifolds 59
3.1. 2-nd order partial differential expressions
on manifolds;Weyl's lemma; Dirichlet operator 60
3.2. Boundary regularity for the Dirichiet realization 67
3.3. Compactness of the resolvent of the Friedrichs
extension 71
3.4. A Green's function for H and Hd, and a
mean value inequality 78
3.5. Harnack inequality; Dirichlet problem;
maximum principle 87
3.6. Change of dependent variable; normal forms;
positivity of the Green's function 93
Chapter 4. Essential self-adjointness
of the Minimal Operator 103
4.1. Essential self-adjointness of powers of H0 104
viii
4.2. Essential self-adjointness of H0 109

4.3. Proof of theorem 1.1 111
4.4. Proof of Frehse's theorem 118
4.5. More criteria for essential self-adjointness 123
Chapter 5. C -Comparison algebras 125
5.1. Comparison operators and comparison algebras 127
5.2. Differential expressions of order < 2 132
5.3. Compactness criteria for commutators 137
5.4. Comparison algebras with compact commutators 144
5.5. A discussion of one-dimensional problems 150
5.6. An expansion for expressions within reach
of an algebra C 155
Chapter 6. Minimal comparison algebra
and wave front space. 161
6.1. The local invariance of the minimal
comparison algebra 162
6.2. The wave front space 169
6.3. Differential expressions within reach
of the algebra J0 175
6.4. The Sobolev estimate for elliptic expressions
expressions on a compact 0 181
Chapter 7. The secondary symbol space 186
7.1. The symbol space of a general comparison algebra 189
7.2. The space M\W , and some examples 193
7.3. Stronger conditions and more detail on ffi\W . 200
7.4. More structure of 2s , and more on examples 210
Chapter 8. Comparison algebras with
non-compact commutators 218
8.1. An algebra invariant under a discrete
translation group 220
8.2. A C -algebra on a poly-cylinder. 228
8.3. Algebra surgery 239
8.4. Complete Riemannian manifolds with
cylindrical ends 246
Chapter 9. Hs-Algebras; higher order operators
within reach 250
9.1. Higher order Sobolev spaces and
Hs-comparison algebras 251
9.2. Closer analysis of some of the
conditions (1.) and (m.) 258
J 3
ix
9.3. Higher order differential expressions

within reach of C or C 262
s
9.4. Symbol calculus in Hs 267
9.5. Local properties of the Sobolev spaces Hs 272
9.6. Sobolev norms of integral order 279
9.7. Examples for higher order theory;
the secondary symbol 283
Chapter 10. Fredholm theory in comparison algebras 291
10.1. Fredholm theory in C C L(H) 292
10.2. Fredholm properties of operators within reach of C 296
10.3. Systems of operators, and operators acting on
vector bundles 303
10.4. Discussion of algebras with a two-link ideal chain 309
10.5. Fredholm theory and comparison technique in
Sobolev spaces 315
Appendix A. Auxiliary results concerning functions
on manifolds 320
Appendix B. Covariant derivatives and curvature 322
Appendix C. Summary of the conditions (xj) used 324
List of symbols used 327
References 328
Index 340
TO HILLGIA
CHAPTER 1. ABSTRACT SPECTRAL THEORY IN HILBERT SPACES.
In this chapter we give a short introduction into spectral

theory of abstract unbounded operators of a Hilbert space. In sec.
1 we give a discussion of general facts on unbounded operators. In
sec.2 we discuss the v.Neumann-Riesz theory of self-adjoint exten-
sion of hermitian operators. Sec.3 gives a general discussion of
the abstract spectral theorem for unbounded self-adjoint opera-
tors. We discuss a proof of the spectral theorem in sec.4. Also,
in sec.5 we discuss an extension of a result by Heinz and Loewner
useful in the following. Finally an abstract result on Fredholm
operators in a certain type of Frechet algebra related to a chain
of Hilbert spaces generated by powers of a self-adjoint positive
operator is discussed in sec.6. The typical 'HS-chain' is a chain
of L2-Sobolev spaces.
The chapter is self-contained and elementary, and only
requires some familiarity with general concepts of analysis and
functional analysis of bounded linear operators.
1. Unbounded linear operators on Banach and Hilbert spaces.
The term "(unbounded) linear operator" (between Banach spaces

X and Y) is commonly used to denote any linear map A:dom A -+ Y
from a dense linear subspace dom A of X to Y. The space dom A C X

then is called the domain of A. Here we distinguish between a li-
near map X + Y and a linear operator: A linear map X -F Y by defi-
nition has its domain equal to X .
The term "unbounded linear operator" will be used with the

meaning "not necessarily bounded linear operator", so that the
bounded linear operators are special unbounded operators. A boun-
ded linear operator, satisfying
(1.1) sup {UAuU/UuI : 0 9 u E dom A} < - ,

I.1. Unbounded operators 2
is necessasily continuous, hence admits a unique extension to X

in which dom A is dense. This is why we usually assume that a
bounded linear operator also is a linear map.
The class of all such unbounded linear operators between two
given spaces X and Y will be denoted by P(X,V). In particular the
class L(X,Y) of all continuous linear maps X -. Y then is a subset
of P(X,Y) .
Since unbounded linear operators are not linear maps from X

to V (but only from their individual domain to Y ) their sum and
,
product needs the following special interpretation: For A,BEP(X,Y)

and C E P(W,X) we define the sum A+B E P(X,Y) and the product AC
E P(W,Y) by setting
(1.2) dom (A+B) = dom A n dom B , (A+B)u = Au+Bu for u E dom(A+B),
and
(1.3) dom AC = {uEdom C : CuEdom Al , (AC)u=A(Cu) for u E dom AC
where it is assumed that dom(A+B) and dom AC are dense in X (or

else, we will say that dom(A+B) or dom AC is not defined ). Also
we define cA = with the identity operator 1 E L(X,X) .
A linear operator A E P(X,V) is uniquely characterized by

its graph, defined as the linear subspace
(1.4) graph A = {(u;Au) : u e dom Al
of the cartesian product XxY = {(u;v) : uEX , v(=-Y} , where (u;v)

denotes the ordered pair. Vice versa, if for any linear subspace
T C XxY the set of all first components is dense in X , and if T
does not contain elements of the form (O;u) other than (0;0), then
a unique unbounded operator A E P(X,Y) is defined by setting
(1.5) dom A = {uEX:(u;v)ET for some vEV} , Au=v , for u E dom A ,
and then we have graph A = T .
For two linear operators A,B E P(X,Y) we shall say that A

extends B (or that B is a restriction of A ) if graph A D graph B.
We then will write A D B (or B C A ) .
Notice that the cartesian product XxY of two Banach spaces is

a Banach space again, for example under the norm II (u; v) II = II uIl +II vII
Therefore it is meaningful to speak of a closed subspace of XxY
An unbounded operator A is defined to be closed if its graph is
I.1. Unbounded operators 3
a closed subspace of Xxy . The class of all closed operators in

P(X, Y) is denoted. by Q(X, Y) . It is clear that a continuous linear
map A E L( X, Y) is closed, since (uk;Auk) ->(u;v) implies v=lim Auk
= Au , hence (u;v) = (u;Au) E graph A.
An operator A E P(X,y) is called preclosed if the closure of
graph A is a graph again. Then Ac with graph Ac= (graph A)closure
is called the closure of A .
In the following we will be mainly interested in unbounded

linear operators A E P( H)= P( H, H) , where p is an infinite dimen-
sional separable Filbert space with inner product (u,v) and norm
Ilull = {(U, U) }1/2 . In that case the graph space HxH becomes a Fil-
bert space again, under the inner product and norm
(1.6) ((u;w),(v;z)) _ (u,v) + (w,z) , I1(u;w)II = (I1 ull2+I1w11

2)112
The following facts, regarding adjoint and closure all work

for general Banach spaces X Y with proper amendments. However
, ,
we will get restricted to the case X=Y=H in all of the follo- ,
wing. For an operator AEp(H) we will say that the (Hilbert space)
adjoint A E P(H) exists if the space TA = J(graph A) is a graph
again. Here J:HxH -. HxH denotes the map (u;w) (w;-u) and " L" - ,
denotes the orthogonal complement in the graph space (with respect

A"
to the inner product (1.6) Then we define the adjoint
. E P(H)
of A by setting
(1.7) graph A TA
Notice that A , if it exists, is closed, since all orthogonal

complements are necessarily closed and since J is inverted by -J
It is clear that A C B implies B C A assuming that A and B
exist. The proposition, below, translates the definition of
the adjoint into a more transparent form. (The proof is left to
the reader. )
Proposition 1.1. Assure that A E Q(H) exists, for some AE: P(H)
Then dom A* consists precisely of all u E: H for which there exists
an element v E it such that
(1.8) (u,Aw) = (v,w) , for all w E dom A.
Moreover the element v thus defined for each u E dom A is uniouely

determined, and we have A..u=v .
to
Proposition 1.2. An operator A_E P(H) admits an adjoint A if and
I.l. Unbounded operators 4
k
only if it is preclosed. Moreover, then A also admits an adjoint
C 'xx
A , and the closure A of A equals A
Proof. Let first A E P(H) have an adjoint Ask E Q(H). Then let T =
(graph A)clos* contain the element (O;z) It follows that there
.
exists a sequence wk E dom A with (wk;Awk) - (0;z) Substitute

.
w = wk in (1.8), and conclude that (u,z) = 0 for all u E dom A

since the inner products in (1.8) allow a passing to the limit.
Since dom A is dense, it follows that z=0, so that no element of
the form (O;z) is in T, except (0;0). Also, T J graph A, which im-
plies that the set of first components of elements in T is dense.
Therefore T indeed is a graph of some AcE P(H) and A is ,
preclosed.
Vice versa, let A be preclosed, and let again T be the clo-
sure of graph A
in TA=J(graph A)
1If(i.e.,
the set P of all first components of elements
of all second components of (graph A)
L
)
is not dense in H then there exists 0 9 z E H with ((0;z),(u;v))=

(z,v) = 0 for all (u;v)E(graph A) But this implies that (O;z)
.
A)H
E (graph (graph A)clos =T. However, for preclosedness of A
it is required that T does not contain such elements. Thus the set
V must be dense in H. On the other hand if (O;v) E TA = graph A
then (1.8) yields (v,w)=0 for all w E dom A , so that v=0 since ,
dom A is dense. This shows that then indeed A is a well defined

operator in P(H), hence in Q(H) q.e.d.
,
All continuous linear maps in L(H) are closed,hence have an

adjoint. Moreover L(H) is adjoint invariant, and, for an A E L(H),
the above adjoint coincides with the well known Hilbert space
adjoint A of the bounded operator A
Also, if A E P(H) is (pre-)closed, then yA+B is (pre-)closed,
for every B E L(H), 0 # y E M. Then (YA+B)o=YAo+B, (yA+B) * A*+B
in the sense of (1.2)
Our main interest, in the following, will focus on self-
adjoint unbounded operators. Here an operator A is called self-
adjoint if (A exists,and) A =A First of all a self-adjoint ope-
.
rator allows a spectral decomposition as a direct generaliza-

,
tion of the principal axis transformation of a symmetric matrix.

Second we will learn about important classes of differential ope-
rators which are unbounded self-adjoint operators, and therefore
allow such a spectral decomposition. Third, we will show how
results on unbounded non-selfadjoint differential operators can be
achieved by " comparing" them with certain standard self-adjoint
I.I. Unbounded operators 5
differential operators.
Note that a bounded operator (i.e., a continuous linear map)
A is self-adjoint if and only if it satisfies the relation
(1,9) (u,Av) _ (Au,v) for all u,v E dom A ,
(where dom A = H ).A general unbounded operator A satisfying (1.9)

needs not to be self-adjoint, because (1.9) just implies that
A D A , not that A = A . Such an operator is called hermitian
If the closure of a hermitian operator A is self-adjoint then we
speak of an essentially self-adjoint operator(i.e., A*=At ).
Note that a hermitian operator A indeed has an adjoint:If ukEdom A
Uk+O , Auk--w, then we may substitute u=uk into (1.9), for fixed v,
and pass to the limit, resulting in 0 = (w,v), for all v E dom A
*
It follows that w=0 , so that A is preclosed, and A exists, by
prop. 1.2. Comparing (1.8) and (1.9), it then follows at once that
A E P(H) is hermitian if and only if A C AF
If A is (essentially) self-adjoint, and B bounded hermitian
then yA+B is (essentially) self-adjoint for all 0#y E R .
More generally, two operators A, B E P(H) will be said to be

in adjoint relation if
(1.10) (u,Av) = (Bu,v), for all u E dom B , v E dom A .
The above conclusion, showing that hermitian operators have

adjoints, can be repeated to prove that operators A, B in adjoint
relation must have adjoints (both) and that,moreover, A and B are
*
in adjoint relation if and only if A C B (or if and only if
BCA* ).
One of the first major problems occurring in our discussion
of differential operators, in later sections, will be the con-
struction of all self-adjoint extensions of a given hermitian ope-
rator. It turns out that not all hermitian operators possess self-
adjoint extensions. On the other hand, the problem of characteri-
zing all self-adjoint extensions was solved by v. Neumann [vN11
and F. Riesz [Ri11. We will discuss the v. Neumann-Riesz theory in
section 2, together with some other constructions of self-adjoint
extensions.
1.2. Riesz-v.Neumann extension 6
2. Self-adjoint extensions of Hermitian operators.
In this section we discuss the v. Neumann-Riesz theory of

self-adjoint extensions of hermitian operators.
It is at once clear that a hermitian operator A has a hermi-
sY st sk ;:'e is
tian closure Ac = A , because A C A implies A C A Since a .
self-adjoint extension B = BC of A is necessarily closed, it also

must be an extension of the closure A Therefore, in looking for .
self-adjoint extensions of a given hermitian operator A we may

look for such extensions of the closure, and thus assume that A
is closed, without loss of generality.
Also if A C B = B then B = B C A
, , so that every self-adjoint
extension B of A is a restriction of A , as well: We have
( 2 . 1 ) A C B = B C A
Proposition 2.1. A hermitian operator A satisfies the identity
(2.2) l1(A-a)u02=11(A-Re A)u112+(Im A)2Hu112 , for all uEdom A , AEC
Proof.We have II(A-a)u11 2=((A-u-iv)u, (A-u-iv)u)=((A-p)u,(A-u)u)

+ v2(u,u) - 2Re ((A-u)u,ivu) where we have written u = Re A
,
v = Im A Here the last term vanishes, due to (Au,iu) + (iu,Au)

.
i((Au,u)-(u,Au)) = 0 , using that A is hermitian, q.e.d.

For a closed hermitian operator A it is implied by prop.2.1
that im (A-A) is closed for every nonreal A E E Indeed, consider .
a sequence uk E dom A such that (A-a)uk -> v It follows that .
(A-X)(uk-ul) } 0 , as k, 1 1 °° . But (2.2) implies the inequality
(2.3) U(A-A)ull > I Im AINO , u E dom A .
Substituting u=uk-ul into (2.3) yields Iluk-ulll - 0, since Im A 90,

by assumption. Hence uk u for some u E H and (uk;Auk) -. (u;v) ,
in HxH. But graph A is closed since A is closed. Thus it follows

that (u;v) E graph A , or, u E dom A , v=Au .
In the following we first consider the special case A = ±i

Since we have im (A±i) closed, by the above, we obtain a pair of
orthogonal direct decompositions
1
(2.4) H = im(A±i) $ V. , D.. = (im(A±i))
The two spaces D+=D+(A) are called the defect spaces of the closed
hermitian operator A , and their dimensions are called the defect

indices of A
(2.5)
. We write
def A = (dim(im(A+i))
i , dim(im(A-i))
L
(v+
)
JA-)
Note that the defect spaces D,. are just the eigenspaces of
the adjoint operator A to the eigenvalues ±i We have :
(2.6) D+ _ (im(Ati))1 = ker(A* i)

Indeed, f E D+ ,for example,amounts to 0 = (f,(A+i)u) for all u E
dom A . We write this as (f,Au) = (if,u) , u E dom A , and compare
with (1.8) , concluding that f E dom A* , A*f = if
As another consequence of prop.2.1 we note that (2.2) for
X = ±i implies
,
=
(2.7) n(Ati)ull = (IIAuN2+llull2)1/2 hl(u;Au)H , u E dom A .
Note that graph A. as a closed subspace of HxH is a Hilbert space

under the norm and inner product of HxH Moreover graph A is in .
linear 1-1-correspondence with dom A which may be used to trans-

fer that Hilbert space structure of graph A to dom A In other .
words, dom A is a Hilbert space under the (stronger) norm
(2.8) IIuIIA = (Ilull2+llAull2)1/2
with inner product
(2.9) (u,v)A = (u,v) + (Au,Av) , u,v E dom A
The latter is true for every closed operator B E 0(H) , not

only for hermitian operators. In particular we may apply it to the
adjoint B=A* of our closed hermitian operator A , obtaining a cor-
responding norm llull * and inner product (u,v) ,k , u,v E dom A
A A
In fact, we then get Pull Pull A ,for u E dom A C dom A* , and
A
*
dom A appears as a closed subspace of dom A under graph norm. ,
Note that (2.7) may be interpreted as follows: The two ope-

rators (A±i) are isometries dom A ; im(A±i) = D. In fact these .
isometries are 'onto'. Therefore V =(A+i)(A-i)71 im(A-i)-+im(A+i) :
defines an isometry between the two closed subspaces of H

Proposition 2.2. For a closed hermitian operator A we have
*
(2.10) dom A = dom A ® D +(A) ® D _(A)
as an orthogonal direct decomposition of the Hilbert space dom A*

under its norm and inner product.
Proof. We already noticed that dom A is a closed subspace of
* *
dom A , under graph norm. The eigenspaces D+ of A are closed sub-
spaces of H , as nulspaces of the closed operators A*'+i. Moreover,
on D+ we have nu II * _ V Hu II , so that D+ are also closed under

A -
graph norm. For f+ E D+ one confirms that (f+,f ) *= 0, using that
- A
*
A f+ _ -+if+. Also, for u E dom A , (u,f+) (Au,±if+) + (u,f+)
A
±i((A±i)u,f+) = 0, so that the three spaces in the decomposition
*
(2.10) are orthogonal. Suppose f E dom A satisfies 0 = (f,u) *
A
=(A*f,A*u)+(f,u) for all u E dom A. Comparing this with (1.8) it
is found that A*f E dom At, hence f E dom (A*)2, and (A*)2f+f = 0.
One also may write this as (A*+i)(A*-i)f = (A*-i)(A*+i)f = 0
* * * *
In particular, f E dom (A +i)(A -i) = dom (A -i)(A +i), in the
sense of (1.3). Now we write f = (A*+i)f/2i - (A*-i)f/2i = f++f
noting that (A +i)f+ = 0. This proves that every f orthogonal to
dom A , under graph inner product of A* is in D+®D- and thus , ,
completes the proof.
Corollary 2.3. A closed hermitian operator A is self-adjoint if

and only if its defect indices vanish (i.e. def A = 0 ). Or,
equivalently, if and only if im (A±i) = H for both, "+" and ,
:e *
Indeed, A = A
implies dom A = dom A , so that (2.10) gives
D+ = {0}, hence def A = 0. Vice versa, if def A = 0 , (2.10) gives
* * *
dom A = dom A , hence A =A since A DA , q.e.d.
,
We now state the v. Neumann-Riesz extension theorem.
Theorem 2.4. The closed hermitian extensions B of a given closed

hermitian operator A are in 1-1-correspondence with the extensions
W W- -> W+ of the isometry V = (A+i)(A-i)-1
: im(A-i) + im(A+i) :
as an isometry between the closed subspaces W. D im(A±i) This .
correspondence is established by assigning to B D A the operator

W = VB = (B+i)(B-i)-1 (which is an isometry extending V between
the spaces 0+ = im(B±i) D im(A±i)). Vice versa, given an isometry
W :W--t W+ extending the isometry V one must observe that W , as
,
an extension of V, is determined by its restriction W0=WIW°, where

W0= W n(im(A-D) = W nD is a subspace of the defect space D and ,
where W0 is just any isometry W0, W0 = W+nD+C V+. Then we have the
closed hermitian extension B given by
(2.11) dom B = dom A $ {W0- : E W0} ,
where the direct sum again is orthogonal in (.,.) * .
A
The proof is almost self-explanatory. It is clear from the
above that W = (B+i)(B-i)-1 is an isometric extension of V , for
every closed hermitian extension B of A . Vice versa, that W0
determines W , as described, follows from the well known fact that
isometries preserve orthogonality. Then, of course, the operator
B , if it exists,should satisfy
(2.12) W(B-i)u = (B+i)u , u E dom B = dom A $ ZO ,
with a certain subspace Z0 C D+$D-, because dom A C dom B C dom A*

and due to (2.10) . For u E Z0 let m = Bu-iu , X = Bu + iu = W0
It follows that u = (W04)-4))/2i, Bu = (WO4)+4))/2, in agreement with
(2.11). Now one simply must verify that the operator B of (2.11)
is closed and hermitian. The closedness follows if we show that
*
ZO = 4 E WO } is a closed space, under graph norm of A
{W0$-4
But we have
(2.13) 11W0 _O*2 = 11W04)_$112

+ 11W04)+4)112 = 211W04)112 + 2114)112 = 4114)112,
is, in fact
which shows that the map 4)-W04)-4), taking W0 onto Z0 ,
an isometry (up to the factor 4). Since W is closed, Z0 also is

closed. To verify that B of (2.11) is hermitian is only a calcu-
lation; since we know that Bldom A = A is hermitian one must show
that (AiYu,v)=(u,A*v) for all u,v E Z0, and for u E dom A, v E 20.
Both follow trivially, q.e.d.
Theorem 2.4 has the following important consequence.
Corollary 2.5. A closed hermitian operator A admits a self-adjoint

extension if and only if def A = (v,v), with v=0,1,2,...,- arbi-
trarily given. In other words, we must have
(2.14) codim im(A+i) = codim im(A-i) .
Then every self-adjoint extension B of A is obtained by picking

an arbitrary isometry W0 : D_ D+ between the two defect spaces
(2.6), and then defining B with (2.11) , and WO , with WO±=D±
The proof is evident.
Although the v.Neumann-Riesz theory completely clarifies
the problem of self-adjoint extensions, other criteria are useful,

of course, because the construction of isometries between defect
spaces is not always practical. In particular not every closed
hermitian operator A satisfies the condition (2.14), so that a
self-adjoint extension need not always to exist. There are two
well known general criteria giving existence of self-adjoint
extensions. Shortly, 'real' hermitian operators as well as 'semi-
bounded' hermitian operators always have self-adjoint extensions.
The concept of real operator refers to a given involution
u u of the Hilbert space H . In most applications we will have

H = L2(X,dp) with some measure space X and measure dp , and then
refer to the complex conjugation u(x) - u(x) of the complex-valued

function u(x)EH=L2. However, one may think of an abstract space
K and an involution map u -r u , with the properties
(U-)_ = u , (c1u+c2v) = clu +c2v

(2.15)
(u ,v ) _ (v,u) , for u,v E H , cjE T , j=1,2.
Then a real operator is defined as an operator A E P(H) satisfying
(2.16) (dom A) = dom A and (Au) = Au , for all u E dom A
Now, if a closed hermitian operator A is real with respect

to any such involution of H , then one confirms at once that
(2.17) D+- _ fu-: u E D+} = D- .
Indeed if f E D+, i.e., (f,(A+i)u)=0 for all u E dom A,then we get

,((A+i)u)_)=(f
0 = ((A+i)u,f)=(f ,'(A-i)u ), hence f E(im(A-i))

using (2.16). This conclusion may be reversed, so that (2.17) fol-
lows. Also it is clear that D+ and D+- have the same dimension.
We have proven:
Proposition 2.6. A closed hermitian operator A which is real with

respect to some involution of H has equal defect indices and hence
admits a self-adjoint extension.
A hermitian operator A E P(H) is called semi-bounded below,
if there exists a real constant c such that
(2.18) (Au,u) > c(u,u) for all u E dom A .
Similarly one speaks of semi-boundedness above if (2.18) holds

with ">" replaced by "<". In such case c is called an (upper or
lower) (semi-) bound of A .
Theorem 2.7. A semi-bounded hermitian operator (regardless whether

above or below) admits a self-adjoint extension.
We will prove thm.2.7 by constructing a distinguished self-ad-
joint extension B of a semi-bounded operator A , where B has the
same semi-bounds as A. This operator B will be referred to as the
Friedrichs extension of A (cf. K.O.Friedrichs [Fr2], and M.H.Stone
[St11). In the construction of B we assume that A is semi-bounded
below, and that c=1 is a lower bound. The general case may be re-
duced to this by considering the operator ±A + y with suitable
real y and sign ± .
In the case of c=1 we conclude from (2.18) that (u,v)..=(Au,v)

=(u,Av) defines a positive definite inner product in the space
dom A C H . One may complete dom A under the corresponding norm
(lull- = ((u,u)-)1'2, and obtain a new Hilbert space H containing
dom A as a dense subspace. It is easily confirmed that every class
of equivalent Cauchy sequences under 11.11., is contained in a uni-
que class of equivalent Cauchy sequences under 11.11 . This provides
an imbedding of H"" into- H , so that we get dom A C H- C H .
Now the Friedrichs extension B of A is defined as the

4e
restriction of A to the space (dom A )nH- . In other words,
(2.19) dom B = (dom A )CH" , Bu = A u u E dom B .
First of all, B still is hermitian: for u,vEdom B =H"'fldom A

there exist sequences uk , vk E dom A with uk'u
in H" ,
by construction of H-. Then (u,Bv) = (u,A v) = lim (uk,A v) _

= lim (Auk,v) = lim (uk,v)_ _ (u,v)-= lim (u,vk)-= (Bu,v)
Also it follows that
(2.20) (u,Bu) = (u,u). > (u,u) , for all u E dom B ,
so that B has the same lower bound 1 as A .

In order to show that B is self-adjoint refer to (1.8) and
let (f,Bu)=(g,u), for all uEdom B, and a given f,gEH Now con- .
sider the linear functional 1(u) = (g,u), for g as, above and all
u E dom A . We get 11(u) l < II glI II ull < II gll II ull _ , using (2.18). Hence
1(u) is a bounded linear functional over a dense subspace of H-,

and we may write 1(u) _ (g,u) _ (h,u)- , with some h E H-, by the
Frechet-Riesz theorem (or Hahn-Banach-theorem). As above, we find
that (g,u) _ (h,u)- _ (h,Au), for all u E dom A. Therefore it also
*, *
follows that h E dom A and that A h=g. Therefore, h E H- om A
= dom B , and Bh = g. We get (f,Bu) _ (g,u) = (Bh,u) = (h,Bu) ,
or (f-h,Bu) = 0, for all u E dom B. Note that we also have shown,

with the above conclusion, that im B = H, because the construction
of h works for every g E H. Thus it follows that f-h=0. Or f = h
E dom B and Bf = Bh = g. This proves that the Friedrichs exten-
sion B is self-adjoint, and completes the proof of thm.2.7 .
Note that, in the special case of the lower bound 1 consi-

B-1
dered we find that the operator is bounded and hermitian,
since we proved that im B = H , while (2.20) implies injectiveness
of B, so that B-1 exists. Also (Bu,v)=(u,Bv) u,v E dom B implies ,
(f,B-lg) = (B-lf,g) f,g E H Furthermore,

, .
(2.21) IIBuIl2= II(B-1)u+u112=11(B-1)u112+llu112+2((Bu,u)-(u,u)) > Ilull2
or, IlBul > Pull , u E dom B , which yields PB-1fII<Ilfll , f e H , or

B_1(=- L(H) , IIB-111 < 1 . The proposition, below, will be useful.
Proposition 2.8. Let C = B-1/2= (1/2)(B-1-l)0

be the unique
=
bounded positive self-adjoint square root of B-1 , satisfying C2

B-1
= . Then C:H -> H" is an isometry between H and H"
(2.22) H- = im C , and (lull = PCuF`, for all u E H .
We have C-1, with dom Cr -1H"' C H a self-adjoint operator in P(H),

with lower bound 1. Moreover, the restriction C-11dom A still is
essentially self-adjoint.
The proof is left to the reader (cf.also sec.4).
3. On the spectral theorem for self-adjoint operators.
The resolvent of a closed operator A E Q(H) is commonly

defined as the inverse R(a) _ (A-a)-1 similarly as for linear ,
maps. More precisely, the resolvent set Rs(A) is defined as the

set of all X E M such that
(3.1) im(A-A) = H , and P(A-A)ull > cllull , u E dom A ,

1.3. Spectral theorem, general discussion 13
with a positive constant c . It is clear that (3.1) holds if and

only if the linear map A-A between the spaces dom A and H has
an inverse (A-A)-1 : H -> dom A C H which constitutes a bounded
operator of H, with 11(A-A)-lII<c-l. Therefore R(A) = (A-A)-1 is
well defined in this sense,for all A E Rs(A) . The spectrum Sp(A)
of A is defined as the complement of Rs(A) : Sp(A) _ cC \ Rs(A)
It is easily seen that Rs(A) is an open set,so that Sp(A)
must be closed: Note that (3.1), for a=a0, and IA-A01<c/2 implies
(3.2) 11(A-A)u°I > II(A-A 0)uII_ja_a0j11u11 > (c/2)IIuII , u E dom A
so that the second condition holds for a neighbourhood of A0

Regarding the first condition we observe that
(3.3) (A-A)u = (1+(a0-A)R(A0))(A-A0)u , u E dom A
by a simple calculation. We know that im(A-A0) = H , and bounded-

ness of R(a0). For small JA_a0) the first factor at right of (3.3)
is of the form 1+ E , 11E11 < 1 , hence is invertible in L(H) , and
takes H onto H . Thus (3.3) shows that im(A-a)=H for all small
A-A01 , and Rs(A) is open.
Now we conclude from (3.3) that
(3.4) R(a) = R(A0)(l+(\0-A)R(A0))-1 ,

JA-A01 < c
where only bounded operators in L(H) occur. It is evident that

the right hand side of (3.4) provides a norm convergent power
series expansion of the operator R(a) in powers of (A-A0) ,
for A close to \0 . Therefore it follows that R(A) is an analytic

function from the resolvent set Rs(A) to L(H) .
Let us now return to self-adjoint operators.
Theorem 3.1. For a self-adjoint AE Q(H) we have Rs(A) J cC\ g, i.e.,

Sp(A)C 1, and
(3.5) IIR(A)fl c (or A<c) then Rs(A)J(-co,c) (or Rs(A)3(c,-)), and
(3.6) IIR(A)II < (dist(A,(c,oo)))-1 , (or IIR(A)II< (dist(A,-oo,c)))-1)
A hermitian AC=P(H) is essentially self-adjoint if(f) for some

A, Im A d 0, we have both im(A-A) im(A- 7) dense in H. If in addi-
,
tion A>c (or A<c) then A is essentially self-adjoint if(f) for some
real (y<c) (or y>c) we have im(A-y) dense in H .
Proof. The estimate (3.1), for a non-real is an immediate con-

).
sequence of prop.2.1, which also gives a value of the constant c

in agreement with (3.5). On the other hand, for X = ±i the first
condition (3.1) is a consequence of cor.2.3. In fact, it is noti-
ced that the conclusions leading to cor.2.3 can be repeated for
every pair of non-real complex conjugate numbers. Therefore we
indeed get the first statement and (3.5). On the other hand, for
a semi-bounded operator we must repeat the conclusions leading to
prop.2.8., particularly (2.21). Details are left to the reader.
The spectral theorem for unbounded self-adjoint operators
provides a Fourier type integral representation of every self-
adjoint operator by a so-called spectral family {E(X) a E 1} :
In this section we only will give a general discussion of the

spectral theorem, while the proof will be discussed in sec.4.
A bounded hermitian operator A semi-bounded below by 0 will
be called positive, and for two bounded hermitian operators A , B
we will write A A) if B-A is positive.
An orthogonal projection (here shortly 'projection') of the
Hilbert space H is a bounded hermitian operator P satisfying
P2 = P (i.e. an idempotent). One easily verifies
Proposition 3.2. The orthogonal projections P of H correspond to

the orthogonal direct decompositions of H :
(3.7) H = im P $ ker P .
For u E H we get u = v + w , corresponding to (3.7) , where v = Pu

and Pw = 0. For any direct decomposition H = M ® N with M , N ortho-
gonal the assignment u - v defines a bounded hermitian idempotent
operator P such that M = im P , N = ker P Pu=u in M , .
Moreover, for two orthogonal projections P,Q we have P<Q if

and only if im P C im Q and then Q-P is the orthogonal projec-
,
tion onto the complement (im Q)n(im P)1 . In particular this holds
if and only if
(3.8) PQ = QP = P .
Proof. For u E im P get u=Pv Pu=P2v=Pv=u

, Hence for w E ker P
.
get (u,w) = (Pv,w) _ (v,Pw) = 0, so that im P and ker P are ortho-

gonal. Also, for z E H write z=Pz+(l-P)z, where clearly Pz E im P
P(1-P)z = (P-P2)z=0 , hence (1-P)z E ker P so that we get the
,
desired orthogonal direct decomposition. Vice versa we confirm

easily that for an orthogonal decomposition H=M$N the map given
is a hermitian idempotent.
If P,Q are projections, and im PCim Q , then we get H=im P
1
® (im Q)f(im P) ® ker Q , an orthogonal direct decomposition,
where P = 1,0,0 , and Q = 1,1,0 in the three spaces. Hence Q-P=
,
0,1,0 , which shows that Q-P projects onto the second space. Also
we then clearly get (3.8).
Vice versa,for projections P,Q, let P<Q , which implies l-Q<1-P
so that Pu=u implies 11(1-Q)u112 = ((1-Q)u,(l-Q)u)=(u,(1-Q)2u)
_ (u,(l-Q)u) < (u,(l-P)u) = 0 , hence Qu=u , and im P C im Q
Then, of course Q-P projects as described, and we have QP=PQ=P.
Finally, if PQ=QP=P , one computes (Q-P) 2=Q2+P2-QP-PQ=Q-P
so that Q-P is a projection. If (Q-P)u=u then find Qu=Q(Q-P)u=u,
Pu=O , and vice versa, so that Q-P projects onto (im Q)n(ker P),
as stated, q.e.d.
A spectral family is defined to be an increasing one para-
meter family {E(A) :AE g} of orthogonal projections which is left
continuous, in strong operator convergence of L(H) and such that
(3.9) lime-SE(A) = 0 , limX-EM = 1 ,
also in strong convergence. In particular,
(3.10) E(A) < E(p) , as l 
and
(3.11) lima}u-OE(A) = E(u)
It is clear from prop.3.2 that the projection operators E(A)

of a spectral family all commute, and that HA = im E(A) forms an
increasing family of closed subspaces We have :
(3.12) HA C Hu , as A aE(p) , E(A-0) =
exist at each A E IR in strong operator convergence of H similarly

as for a real-valued function. In fact for p > A we have
(3.14) II(E(p)-E(A))u112 = (u,E(u)u) - (u,E(X)u) , uEH ,
hence the upper and lower limits will exist, just as they do for
the monotone real-valued functions (u,E(A)u), u E H. Using (3.14)
for an orthonormal basis of (the separable space) H one easily
concludes that E(a) is strongly continuous over all of I except ,
perhaps at certain 'jumps', and that there are at most countably

many jumps. Then the existence of the limits E(±-)=lim E(A)
in strong convergence may be derived from the monotony of E(a)
only, without the explicit requirements (3.9) .
For an interval A = (a',a"] of the real axis, left open

and right closed, let EA=E(a")-E(a'). If we have a partition
(3.15) a' = A0 < Al < . .. < AN-1 < aN = a" ,
of A into finitely many sub-intervals Aj = (aj_l,aj] , j=1,...,N,

then we get a corresponding orthogonal decomposition
(3.16) im EA = s T=1 (im EA ) .

J
For arbitrary u E H we get
(3.17) IIEAu1l2 = N_1 IIEA ul12

J
in view of the orthogonality. One may use this fact to define the
Riemann-Stieltjes integral J (A)dE(A), for a continuous function

A
(a), as the limit, in norm convergence, limS,0' Q , with Riemann sum
S and maximum interval length S (with points a.E[aj-l,x.}), where
Jj
(3.18) S = d = Max {(aj-aj_1) :j=l,...,N}
Here the existence of the limit for continuous functions is

easily derived from the observation that, for the Riemann sum S'
of any refinement of the partition (3.15),we have
(3.19) II(S-S.')ull < /_60EAull ,
by the same orthogonality. (Since any pair of partitions, with

sums S,S" , and maximal lengthes S,d" has a common refinement sum
S1, we get 11S-S1'11<,/_6+,/_6 , insuring convergence of the integral.)
Also,
(3.20) 11
A A A
with real-valued Riemann-Stieltjes integrals at right. For u E H

and a function ¢ E CU) one then may look for existence of the
improper Riemann-Stieltjes integral Iu = It is
found that Iu exists if and only if
(3.21) 111u112 =
(This follows by applying (3.20) to the difference IAu-IA,u , with
two large intervals A , A' , and IAu = Jo $(a)dE(A)u .)
In the theorem, below, we will say that an unbounded opera-

tor A and a bounded operator B commute if we have BA C AB , with ,
products defined by (1.3) . One confirms easily that this holds

true if and only if B commutes with the resolvent R(X)=(A-a)-1 for
every A E Rs(A) .
We now can state the spectral theorem.
Theorem 3.3. For every self-adjoint operator A E Q(H) there exists

a unique spectral family {E(A)} , with E(a) commuting with every
bounded hermitian operator B commuting with A such that
, ,
(3.22) A = J-XdE(X) ,
in the sense that
(3.23) dom A = {uEH :
and
(3.24) Au = J-AdE(X)u ,
with an improper Riemann-Stieltjes integral at right.

Vice versa, for every spectral family E(a) formula (3.22) ,
defines an unbounded operator A and E(A) commutes with every

,
bounded hermitian operator, commuting with A .
4. Proof of the spectral theorem.
In this section we discuss a proof of thm.3.3. Note that

there is a selection of proofs, mainly differing by the type of
construction used for the family E(A) . Here we define E(A) as

the orthogonal projection onto the null space of the bounded
operator
(4.1) SA = TA+(TA2)1/2 , Tx = (R(A+i)+R(A-i))/2
with the resolvent R(p) _ (A-u)-1 of the self-adjoint operator A.

More precisely, since the above would give a right (not left) con-
tinuous spectral family, the above family of projections is called
PX and then E(A) = lim u-,a,u<XPu = PX-0
, .
For a similar proof cf. Nagy [NSz1], [RN].

For a self-adjoint operator A the resolvent R(p) is defined
(and holomorphic) in all of U\1 (thm.3.1), and we have (A-p)'
_
= A-1i , hence R(p) Therefore the operator TA of (4.1)
= R(u) .
is bounded and self-adjoint, since we may write TA=(R(u)+R(u)")/2,

with u=A+i In fact we get IIT II<1 , using (3.5) . Accordingly we
.
a -
have TA2 bounded self-adjoint and O<TA2<1 and the binomial ,
series of prop.2.8 , i.e.,
U _ (T2)1/2 = 1--0(1/2)(TA2
(4.2) - 1)0
J j
defines a bounded self-adjoint operator Ux satisfying O<Ux<1, and
UX2=TA2. Indeed, the binomial series Y( 1/2)x0 converges absolutely

j
for x=1, as well known. We get -1<(Tx2-1)<0<1, hence IITx2-111<1

(using prop 4.3), so that the series (4.2) converges absolutely.
Thus the Cauchy-type reordering of the product
is correct, just as for a series of numbers, and we get U
(1+(TA2-1)) = The coefficients of the series are real
TA2 .
hence all terms are self-adjoint. Moreover, all terms but the
0-th power are <0 , and one estimates -Ii < 1 , so that UA=1+1i20.
It is clear that TA and Ux commute, by definition of Ux

(We have TAUA = UATA). Accordingly, 0=(TA2-UA2)=(TA+UA)(TX-UA).
For a uEim TA we get u=TAv=(TA+UA)v/2 + (TA-UA)v/2 = w+z , where
w = (TA+UA)v/2 E ker (TA-UA) z = (TA-UA)v/2 E ker (TA+UA)
,
Also, (w,z) = 4-1(v,(TA2-UA2)v) = 0 using that TA and Ux are ,
self-adjoint. This means that w and z are orthogonal.

Since TA is self-adjoint, we have
clo
(4.3) H = ker TA $ Um TA)s.
which is an orthogonal direct decomosition. Also, ker TA = ker U

If TAu=O, then UA2u = Tx2u = 0 hence IIUAuIi2 = (u,UA2u) = 0, so
,
that ker TIC ker UA and by symmetry we get equality. It follows

,
that ker(TA+UA) n ker (TA-UA) = ker TA = ker UA . Define

y
(4.4) Hx = ker(TA+UA) , Nx = ker TA , KA= ker(TA-UA) n (Nt ),
and conclude from the above that
(4.5) H = Hx ® KX ,
with an orthogonal direct decomposition. Also,
(u,Txu) = -(u,Uxu) <0 on HA = ker(TA+UA),

(4.6)
(u,Txu) = (u,UAu) >0 on K, C ker(TX-UA)
Finally, in this chain of arguments, note that
2Tu = R(A+i)u + R(A-i)u = R(A+i)((A-A+i)+(A-A-i))R(A-i)u
= 2R(A+i)(A-A)R(A-i)u , uEH ,
implying that
(4.7) TA = R(A-i) (A-A)R(A+i)
Combining this with (4.6) we get
(4.8) (R(A-i)u,(A-A)R (X-i)u) < 0 , as uEHA , >0 , as uEKx .
Furthermore we observe that R(v)R(p) = R(p)R(v), for every p,vET,

which also implies that R(p) commutes with TA and U . We conclude
from there that, for all p E ®\R we have ,
(4.9) R(p)HA = dom A n Hx , R(p)Kx = dom A fl K .
Indeed, for uEHx we get R(p)uE dom A , and (TA+UA)R(p)u =

R(p)(TA+UA)u = 0 , so that R(p)u E dom A n HA . For uE dom A fl HA
one may write u=R(p)v, v=(A-p)u, and then 0=(TA+UA)u=R(p)(TA+UX)v.
Since R(p) is 1-1, it follows that (TA+UA)v=O hence vEH, con- , ,
firming the first equation (4.9). Similarly one finds that

R(p)(ker(TA-UA)) = dom A n ker(TA-UA) . Also it is clear that
R(p)(ker TA)= ker T, = ker(A-A) C dom A , using (4.7) and (4.8),
so that also the second (4.9) follows.

One then finds that
dom A = (dom A n HA) (dom A n KA)
in the sense that every uE dom A allows a unique sum decomposition
u=v+w with v,w in the spaces at right. Such decomposition arises
if (A-p)u is subject to the decomposition (4.5), and then R(p) is
applied. Let R(4)X = R(u)IHA then it follows that
,
and im R(p)' is dense in HA , while R(p)A is 1-1, hence has an

inverse in Q(HA) One gets
. R(u)A and thus finds that ,
AX = (R(p)A)-1 + u
defines an unbounded self-adjoint operator in Q(HA) with domain
dom A n HA Moreover, AX is the restriction of A to that space,
.
and is independent of p. Similarly for the restriction of A to KX.

We summarize the above:
Theorem 4.1. We have
(4.10) ker(A-A) C HX , ker(A-A) n K. _ {0}
and
(4.11) (u,Au)>A(u,u), uEdom AnHA, and (u,Au)<A(u,u), uEdom AnKA.
Moreover, all the spaces H. KX for AEI

, are left invariant
, ,
by the resolvent R(p) and by A , where the latter means that

,
(4.12) A:(dom A n H ) - HA , A:(dom A n KA) 5

X
while dom A n H, , and dom A n KA are dense in HA and KA , respec-

tively, and the restrictions of A to these spaces are unbounded
self-adjoint operators of the Hilbert spaces HA and KA resp. ,
Thm.4.1 already holds the principal information of the

spectral theorem, at least as far as existence of the spectral
resolution is concerned. We complete this as follows.
Corollary 4.2. For every self-adjoint operator B E L(H), commuting

with every R(p), 4ET\l the restrictions BX=BIHA ,and BA=BIKA
,
are maps HA+HA and KA+KA , respectively .
In other words, every bounded self-adjoint operator B, com-

muting with the resolvent R(u) , is reduced by the direct decom-
position (4.5).
Indeed, such an operator B commutes with TA, hence with U,,
given as a power series in TA-1 As above we thus conclude that
.
(TA+UA)Bu=B(TA+UA)u=0 , as u E H , so that B: HX.HX . Similarly
B: KA+KA .
It is convenient now to introduce the orthogonal projections
PA by setting
(4.13) P Au=u on H, , P Xu=0 on K, .
Then, as a consequence of cor.4.2 we find that
(4.14) PAPX, = PA,PX , for all A, At EA .
Indeed, the operator PA is bounded and self-adjoint, and commutes

with R(4) The latter is an immediate consequence of thm.4.1:
.
For u E HA we get R(p)u E HA and uEKA implies R(u)uE KA , hence

,
PXR(p)u=R(p)u=R(p)Pu , uEHX , and (1-PX)R(p)u=R(p)u=R(p)(l-PX)u

uEKA , or also PR(p)u=R(p)PXu , for uEKX. Since H=HX®KX we get
PXR(p) = R(p)PX . Self-adjointness and boundedness of PA is a con-
sequence of the orthogonality of HA and KA .
Using cor.4.2 it thus follows that B is reduced by any of

the other direct decompositions (4.5) . Thus again, we get
(4.15) PA,PXu=PAu=PAPX,u, uEHV (1-PA,)PXu=PA(1-PA,)u, uEKX,,
implying (4.14).
Next let A',AER , a'<A . For uEH., write
(4.16) u = PAru = PA,PAu + PX,(l_PX)u = v+w .
The first term, at right, is EHA, and the second term wEHXtnKX
using the commutativity (4.14) . Thus the second term satisfies
(4.17) A(w,w) < (w,Aw) < A'(w,w) .
The latter is only possible if w=0 since A'<A , using (4.11).

,
Hence it follows that uEH' and we have proven that

,
(4.18) HA, C HA , as At <A .
Now we invoke prop.3.2, and conclude that JP X:AER} , is

a nondecreasing family; we have (3.10). As discussed in sec.3,
this implies existence of the monotone limits (3.9) and (3.13)
and these limits must be orthogonal projections, since the func-
tional equations P2=P and P = P are continuous. Thus we may set
E(A) = PX-0 , as postulated.
I.4. Proof of the spectral theorem 22
In particular E(--) and E(w) are the projections onto H_ _

nHx, and Hm (uHx), respectivey, with union and intersection over
I . For uEH__ we must have (u,Au) < A(u,u) for all AER , which is
only possible, if u=0. For uEK.= (H.)1 we must have (u,Au)>a(u,u)
for all AER , implying u=0 again. This yields H = (K.)-L= 0 ,
or, E(-=) = 1-E(w) = 0 confirming (3.9). Similarly we conclude

,
(3.11), by construction. Therefore E(A) indeed is a (left conti-

nuous) spectral family, while {PX} is right continuous.
Next consider an interval A=(a',a"] a, with finite
and a partition (3.15). With EA EA as in sec.3 observe that ,
EA = EX (1-EX). Thus (4.10) implies

J J J-1
(4.19) HE
Aj
ull2
-< (E Aj u,AE Aj u) -< aJ IIE Ajull2 , uEH
In particular one gets (4.19) first for uE dom A n H n K

aj aj_1
a space dense in im EA = Hx n H , by a direct sum argument,
J J j-l
as in the proof of thm.4.1. This implies boundedness of (the
restriction to im EA of) A in im EA , using prop.4.3, below.
J J
Since that restriction is a closed operator, it follows
dom A n im EA = im EA , hence im EA C dom A
J 7 J
From (4.19) we conclude that
(4.20) 0 < (EAJu,(A-x. )EAJu) < (Aj-aj_1)(u,EA7u) , uEH

3-1
Summing this over j one has
(4.21) 0 < (EAu,(A-S)EAu) < (u,EAu) ,
with S and , of (3.18) , for a = a. , and q(a) = X . Here we

j
used that (EA u,EA u)= (EA u,AEA u)=0 for j91
J 1 J 1
Now we use (4.21) and prop.4.3, below, and conclude that
(4.22) AEA = lA A dE(a)
with an operator norm convergent Riemann-Stieltjes integral.Then,

for u E dom A we get Au = lim , AE u , which equals the
improper Riemann-Stieltjes integral of (3.22). Vice versa, if that
integral exists, or, equivalently, if {-a2dIIE(A)u112 < , then

E6u}u , AEAu JIR AdE(A)u , so that uE H, and (3.24) holds, by
closedness of the operator A.

This completes the proof of thm.3.3, except for uniqueness
of E(a) .
Now, if QX is another spectral family such that QX commutes

with every self-adjoint BEL(H) commuting with A (i.e., with every
R(u)), then we get
(4.23) QxPU = PuQA ,
using that Pu commutes with R(v) by construction. Then the

,
argument leading to (4.14) may be repeated, showing that also
(4.24) in QX, C HX C im QA , as A'< A < A"
Then, by monotony of QX we get
(4.25) QX-O < PX < QX+O
resulting in equality QX=PX=E(X) at all points of continuity of

QX. Since both E(A) and QX are left continuous we get E(A)=Qx also
at the 'jumps' of QX using the fact that im(QX+O-QX-O)=ker(A-X).
,
This completes the proof of thm.3.3.
Proposition 4.3. Assume that a hermitian operator T satisfies
(4.26) l(u,Tu)j < c(u,u) , for all u E dom T ,
with some nonnegative constant c. Then T is bounded, and we have

11 T11 <c .
Proof. Apply (4.26) to u+v and u-v both, and subtract the two
inequalities, for
(4.27) -2c(11u11 2+11v112) < 4Re(u,Tv) <2c(Ilull2+Iivll2), u,v E dom T .
Here we introduce u=Tv/c (if necessary, let u run through a

sequence converging to Tv/c). It follows that
(4.28) (2/c)IITvll2 < (1/c)IITvll2 + clivll2 , vE dom T
which implies the statement.

I.S. Powers of positive operators 24
5. A result on powers of positive operators.
In the lemma,below, let H be a separable Hilbert space.
Lemma 5.1. Let B E L(H) be self-adjoint, and let A be an unboun-

ded self-adjoint operator with domain dom A . Assume A positive,
and B positive definite, so that B-1 exists as an unbounded self-
adjoint operator, and the positive powers As Bs , s>O are well
, ,
defined positive self-adjoint operators.

If AB is bounded, then also ASBS is bounded, for every 0<s<l
Moreover, we have
IIASBS N < IIAB U5 .
If AB is compact, then also ASBs is compact, for every O<s<l
Remark (a): The boundedness of ASBs ,and (5.1) is a well known

result, due to E.Heinz [Hzi],(cf. also [CQ], Dixmier [Dx21, Kato
[K3], Loewner [Lw1]). The compactness of ASBs seems to be new.
Remark (b): Note that the condition of boundedness of AB is equi-

valent to the condition dom A D im B and again equivalent to the
,
condition that the unbounded operator product AB has dense domain

and extends continuously to H Indeed, if dom A D in B , then
.
the product AB has domain H , while we get (AB)* D BA , due to

self-adjointness. Hence dom(AB) * is dense, which implies bounded-
ness of AB , by the closed graph theorem. If AB has dense domain
and a continuous extension to H , then for all uEH there
e Sc
exists uk E dom (AB) with uk-, u , ABuk+ v= (AB) u Since B is
.
bounded we also get Buk-> Bu , A(Buk) - v . The selfadjoint opera-

tor is closed. Thus we conclude that Bu E dom A , hence dom A
D in B . Thus indeed all 3 above conditions are equivalent.
Remark (c): Clearly boundedness of ASBs in lemma 5.1 again may be

interpreted in each of the ways of rem.(b). Also, clearly, since
compactness of AB implies boundedness of AB , all of remark
(b) applies in the case of AB E K(H) , where K(H) denotes the
(closed 2-sided) ideal of compact operators in L(H).
The proof of the lemma will require some preparations.
Let BEL(H) be self-adjoint, positive definite. Let H0 be the com-
pletion of H under the norm 9up0 = {(u,u)0}"2 , with inner pro-
duct (u,v)0 = (u,Bv) = (Bu,v) , u,v E H . From simple arguments it
follows that H may be regarded as a dense subspace of HO, Now let

us assume that A with domain dom A is an unbounded closed operator
of H which is hermitian with respect to the inner product (.,.) 0
That is, we have (u,Av)0 = (Au,v)0 ,for all u,v E dom A Also let .
the resolvent set Rs(A) of A contain at least one real point X0
Proposition 5.2. The linear map A:dom A } H0 defines an essen-

tially selfadjoint operator of HO If A0 denotes the closure of .
A in H0, then the resolvent set Rs(A0) contains Rs(A). Moreover,
(5.2) II(A0-a)-1110 < II(A-a)-111 for all A E Rs(A) .
(A-A0)-1
Proof. Let R = R(A0) _ E L(H) . We get ((A-X0)u,v)0 =

(u,(A-X0)v)0 for all u,v E dom A , which implies (f,Rg)0 = (Rf,g)0
for all f,g E H . In other words, the operator R E L(H) is hermi-
tian in H0 .
Proposition 5.3. We have R bounded in HO , and, IIR110 < IIRII
Proof. With a spectral family E(a) of the operator B let P. =
1 - E(E) , for e > 0 , and let S. = im PE . In the closed subspace

S of H the norms 11.11 and 11.11 are equivalent. We have Ellull2<
0
Bull2 IIBIIIlull2. The projection PE: H I SE is orthogonal with
respect to both inner products (.,.) and (.,.)0 . Define R. _
PEAIS£ . For u,v E SE we get (u,Rv) (u,(PRISE)v) _ (PEu,Rv)

_ (u,REv) , (u,REv)0 = (P6u,Rv)0 = (u,Rv)0 = (Ru,v)0 = (REU,v)0
Now it is easy to see that we must have IIREII0 < IIRE11 . Indeed, the
spectrum, and hence the spectral radius p of R. is independent of
the norm under which SE is considered. For a bounded hermitian
operator we get IIREII0 = P , while generally IIRE11 > P . Thus we
indeed have the above inequality.
It follows that
(5.3) sup{j(u,Ru)I/(u,u) 0#uESE} < IIP6RII < IIRR .
But the space of all u with u = P E u for some E > 0 is dense in ,
H since B was assumed positive definite. The proof is complete.

,
Continuing in the proof of prop.5.2 it now follows that

the linear map A:dom A - H0 is an essentially self-adjoint opera-
(A-A0)-1
tor, since is bounded and densely defined (cf. thm.3.l).

The closure of this operator (in H0) is self-adjoint. Its resol-
vent exists for all non-real points A and is a normal operator ,
wherever it exists. Also, at each point A of the resolvent set

Rs(AO) we find that A-X must be invertible as a linear map, at

(A0-a)-1
least, and the inverse is the restriction of to H . We

trivially have Rs(A)\l C Rs(A0) since Rs(A0) contains E\1
,
For every real A E Rs(A) prop.5.2 applies as well as to A0

Hence we also get RS(A)r C Rs(A0) thus Rs(A) C Rs(A0) , ,
as stated. Finally, for real A E Rs(A) prop.5.3 implies the

inequality (5.2). For non-real A E Rs(A) we know that A E
Rs(A0) and that (A0-a)-1 is normal. If P and p0 denote the
spectral radii of (A-A)-1 and (A0-X)-1 then again pO= II(A0-a)-111,
p < II(A-a)-111 and we must have p0 < p since p and PO are just
,
the reciprocal minimum distances from A to Sp(A) and Sp(AO) , and

since Sp(A) (Rs(A))comp J This comple-
(Rs(A0))comp=
_ Sp(AO) .
tes the proof of prop.5.2 as well.
Proposition 5.4. Under the assumptions of prop.5.2, if we have

A E K(H) then also A0 E K(H0) .
Proof. The resolvent set Rs(A) of an operator A E K(H) is known

9e
to be the complement in 1 = (E\{0} of an at most countable set of
eigenvalues, each of finite algebraic multiplicity, and with only
possible cluster point at zero. In particular, for each eigenvalue
A # 0 , the spectral projection
(5.4) P = i/(2Tr)Jr R(u)de , R(u) = (A-0-1

a
with a sufficiently small positively oriented circle rA of center

A not containing any other eigenvalue, is a bounded operator of
,
finite rank. The resolvent R(A) is a holomorphic map Rs(A)- L(H),

and the integral (5.4) exists in norm convergence of L(H)
Now we apply prop. 5.2 to conclude that Sp(A0) C Sp(A)
so that Sp(A0)\{0} is a collection of some of the eigenvalues A
(A0-u)-
of A For each such A we get R0(u) _

. as the continuous
extension of R(u) onto HO in view of (5.2), and uniformly so
,
over r . It follows that
(5.5) PO = i/(2,r)Jr R(u)du

O
u
is the continuous extension of PA to HO Since PX is of finite .
rank, this is true for PO as well It follows that all non-vani-.
shing points of Sp(A0) of the self-adjoint operator AO are isola-

ted eigenvalues of finite multiplicity. This proves that A0 is
compact and our proof is complete.
Proof of lemma 5.1. First we introduce the inner product (.,.) 0
1.5. Powers of positive operators 27
and space H0 with B of lemma 5.1 . Then if AB E L(H) , it may

be observed that (u,ABv) 0 = (u,BABv) _ (ABu,Bv) _ (ABu,v)0 for ,
all u,v E H so that AB satisfies the conditions of prop.5.3.

,
It follows that IIA1/2B1/2112= IIB1/2ABl/211 = IIABIIO< IIABII where we

have used the C -algebra property of L(H) , (i.e. II X112 = IIX"`XII ,
for X E L(H)), and the fact that B1/2 defines an isometry H - HO,
so that 11X110 = IIB1/2XB-1/211 for X E L(H0) Accordingly, we have
, .
proven (5.1) for s = 1/2 Similarly, if AB E K(H) , we may

.
apply prop.5.4 instead of prop.5.3 to conclude that B1/2AB1/2

(Al/2B1/2(A1/2B1/2)
E K(H) Since Bl/2AB1/2 = it then follows
that ASBS E K(H) for s=1/2 .
In a similar way one now proves the lemma for all rationals
of the form s = k/21 0 < k < 21 as follows. (We only indicate
, ,
how to proceed for s=1/4 and s=3/4 to make the method evident)
For s=1/4 we may just repeat the above, now using (u,v)O =
B1/2
(u,B1/2 v) and Al/2 as the operator A of prop.5.2 For s .
= 3/4 if AB E L(H) we already know that Al/2B1/2 E L(H)

, so ,
B1/2A3/2B = (A1/2B1/2)'ÀB
that C = E L(H) The operator C is .
hermitian with respect to (u,v)0 = (B1/2u,v) as easily checked. ,
Hence it also is bounded in the corresponding norm. With the iso-

metry B1/4.H->
O H we get that IIA3/4B3/4II2= IIB3/4A3/2B3/411 =
UC110 < UCH < IIA1/2B1/2IIIIABII Taking square roots, we

< IIABII3/2 .
thus get (5.1) for s=3/4 . If AB E K(H) we find that C E K(H)

so that prop.5.4 yields ASBs E K(H) for s=3/4 .
Note finally that boundedness of ASBS= E is equivalent to

s
the inequality
(5.6) IlAsull < IE 1111B-SuII for all u E dom B-s .

- s
If (5.6) with 11Es1I < 11E11IS is correct for all rationals r=p/q,
q=21 , 0<s<l , then it must be true for all 0<s<l Indeed, since .
B-s
B is bounded selfadjoint we have dom = imBs decreasing, as s
1.
increases. Let s . be a sequence of rationals with denominators 2 2
J
B- sj
B-s
with sji s, sj<s. Then u E dom C dom As implies u E dom
s s s.
C dom A J , and B- Ju -* B-su , A Ju -- Asu , as j ; . Using this
in (5.6) we conclude that this inequality is correct for all
0 < s < 1 if it is true for s=l .
Finally, regarding compactness of ASBS for general s we will
have to discuss the function F(z)=AZBz of the complex argument z .
1.5. Powers of positive operators 28
For u , v E H define the function (z) _ (u,AZBzv) , where it may

be noted that the general powers AZ and BZ are well defined
also for complex z The expression (z) is meaningful for 0
. <
Re z < 1 assuming that AB is bounded, and using our knowledge

,
on boundedness of ASBs for 0<s<l In fact, for z=s+it O<s<l . ,
z z
we get A B v = Ait(AsBs)(Bity) well defined, since the operators
Alt and Bit are unitary, of course. We get
(5.7) IIABII5IIuIIIIvII , 0<s<l , --<t<-
Note that t(z) also is analytic in the open strip 0<s<l , if we

assume that, for some O<c<l we have u e dom AE ,With a parti- .
tion 1 = X(X) + w(X) 0< X,w E C°°(R)

, X=l for )<l w=l A>2 , , ,
we write, with some 0<6<Min{c,Rez} ,
A)AbX(A))u,(AZ-dBz-d)(BSv))
fl(z) _ (((log +
(5.8)
((A6u),(AZ-dBZ-d)(Bdlog
((log A)w(A))u,(AZBZ)v) + B)v) .
All expressions at right are meaningful, in view of the selfad-

jointness of A and B using that B is bounded and positive defi-
,
nite, while A is possibly unbounded, but still positive. In parti-

cular, the function X(X)X6 log , extends bounded and continuously
to [0,-) , while w(a)log = O(TE) in [0,.) ,and adlog ) = 0(1)
on the bounded set Sp B C [0,o) .
Note that the restriction u E dom AE can be removed. For a

general u E H let the sequence ukE dom AE converge to u in H . We
get k(z), with uk instead of u, analytic in 0<s<l, and
(5.9) Wz)-Wz)I = I(u-uk,AZBZv)I < IIABIISIIvIIIIu-uk'
which implies uniform convergence Wz) -> (z) in the entire strip
Accordingly is bounded and analytic in 0< Re z <1
for any arbitrary u,v E H
For a given 0<s<l we may apply the Cauchy estimates on the
first derivative '(s) With n = Min{s,l-s} we get
.
(5.10) n-1 Max{IO(C)I:Ic-sI=n} < Ilunllvil Max{1,IIABH)/n.
From this estimate we conclude that
(5.11) (u,(ASBS-AtBt)v) _ sdTOu,v(T) = 0(Is-tIIIu1111viI)

t
with 0(I s-tI II ull 11 vll) < c(s,t) I s-tI II un 11 vll , the constant c(s,t)
independent of u,v and bounded in compact subintervals of (0,1).

Note that (5.11) implies the proposition,below.
Proposition S.S. Under the assumptions of lemma 5.1, if AB E

L(H) , then the operator family F(s) = AS BS , 0 < s< 1 , is norm
continuous (even Lipschitz continuous) . We have
(5.12) IIASBs - A t B t I < c(s,t)js-tl
Moreover, the corresponding is true for the more general family

F(z)=AZBz 0 < Re z < 1
, We then get (5.12) with s,t replaced
.
by z,z', with c(z,z') bounded in conpact subsets of the strip.

We will use only (5.12), not its complex extension.
Prop.5.5 also provides the final link in the proof of
lemma 5.1 If AB E K(H) then certainly AB E L(H)
. so that ,
prop.5.5 applies and ASBS is norm continuous in the open inter-

val (0,1) Since AS BS is compact in the dense subset of all
.
rationals of the form s=k/2 1 it must be compact for every s E

(0,1) , since the compact ideal is closed in norm topology.
Notice that A0B0=I of course, is not in K(H). Accordingly
,
it is clear that F(z) = AZBz cannot be norm continuous at z=0 .
6. On HS-chains.
In this section we shall discuss an abstract result on cer-

tain families of Hilbert spaces similar to the so-called Sobolev
spaces. For a concrete example cf. [C1],ch.III and IV, where
corresponding facts are derived for the Sobolev spaces over Rn
(we shall apply the abstract result in VI,3 and ch.IX below).
Let H be a (separable) Hilbert space and let H be a self-
adjoint operator of H with domain dom H . We assume H semi-bounded
below and that 1 be a lower semi-bound, i.e., H>l Also we assume .
H properly unbounded, i.e., that its spectrum extend to +_ Note .
that the complex powers HS, sEU are well defined closed operators.
For example we may use the spectral family E(X) of H and define
(with As = es log A
log A real-valued)
,
(6.1) HS = I'AsdE(A), dom HS ={uEH: Ix2Re(5)d11E(x12<.}
For a formal convenience, useful later on, we introduce the

inverse positive square root A = H-1'2 , and henceforth will also
Hs=A-s/2 .
write
1.6. HS-chains 30
For every sEg let H denote the Hilbert space obtained by

completing dom Hs/2 = dom A s C H under the norm
(6.2) ilulls = IIA- sull .
Evidently we get Hs = dom AS C H for s>0 and the norm 11.115 is ,
equivalent to the graph norm of the unbounded operator A-5. For

s<0 the norm 11.l1 s is weaker that the norm of H , whence Hs DH .
For example one may consider Hs , s <0, represented by the collec-

tion of all 1-parameter families such that E(p)u, _
up, as p<A , and that Ilull2 = JX2Re(s)dlluX112 <. . The same represen-
tation of uEHs as a function of ), also holds true for general sEiR.

In this way one finds that Hs is naturally imbedded in Ht
as s>t, so that {Hs:sE1R} is a decreasing family of spaces. We
still introduce the locally convex spaces
(6.3) H_ = n{HS:sER} , H__ = U{Hs:sElk} ,
where the union carries the inductive limit topology, while H

carries the Frechet topology induced by all the norms 11.0s sE1. :
A collection of spaces Hs, -=<s<0, as constructed above will

be referred to as an HS-chain.
It is natural to extend a projection E(a) of the spectral
family to an operator E(X)-_: H- -, H_ by defining E(X)__u=uA for
u={up>1}, as X>1 , and E(X) u=0 for X<l .
Proposition 6.1. The space H is dense in every Hs, IsL<= , under

its topology.
Proof. For u E H5 , s<0, we clearly have u, = E(a)_.u E H. and
(6.4) Qu-uX112 = J_p2Re sdllE(p)u112 + 0 , as Ai.
For general uEH- we must have uEHs for some 5E1. Then uA+u in Hs
as above, implying u1+u in H q.e.d. ,
We now define the order classes O(m).
Definition 6.2. For mE1 we denote by O(m) the class of all linear
operators A E L(H ) such that, for every sEl the induced unbounded
operator AEP(H H ) with dense domain H CH extends to a (conti-
s s-m s
nuous) operator AEL(Hs,Hs_m) (from Hs to Hs-m)
It is clear that an operator AEO(m) also admits a continuous
extension to H__ . In particular we must have Asu=Atu for u E
1.6. HS-chains 31
, r=Max(s,t) . Clearly every 0(m) is a linear space, and

HsnHt=Hr
(6.5) AB E 0(m+m') for all AEO(m) , BEO(m') .
Accordingly 0(0) and 0(o) = U0(m) are algebras and 0(-oo)=ff0(m)

is an ideal of both algebras.
In particular we have A-m E 0(m) (i.e., more precisely, the
restrictions A-mIH0, are contained in 0(m) , for every mEt ). More-
A-C
over the operators are in 0(Re ) for every GC . Also it is
(A-C)
found that s constitutes an isometry Hs-*Hs-
, E=Re C , for
every real s and complex 0 . Also E(a) (or rather E(a)IH.) is an
operator of 0(-W) and E(A). above is its canonical extension.
,
One may introduce a locally convex topology on 0(m), men,

using the collection of all operator norms
(6.6) RAN s,s-m= HA II s-m= sup{UA uU : uEH s, OUR <1) , 5E1 .

s s, s s-m s -
Theorem 6.3. The topology of (6.6) is a Frechet topology, and

0(m) is a Frechet space. In particular we have
UAU(t-r)/(t-s) IIAII(r-s)/(t-s)
(6.7) RAN < s < r < t
r,r-m - s,s-m t,t-m
Proof. It is sufficient to discuss the last statement. Moreover,
we may consider the case m=0 only, since RANpip-m=11AmAnpip= HAmAIIP.
This is a matter of Calderon interpolation (a straight application
of the Phragmen Lindeloef principle, cf.[Sel]): For fixed s<t and
and u,vE H_ define s(z)=(l-z)s+zt, zEM, and the function
(6.8) f(z) = (u,A-s(z)AAs(z)v) , z E T .
It was seen before that As is well defined for real and complex s.
We have 2d/dz(As(z)w)=-(t-s)As(z)log H w, with complex differen-
tion, for every wEHw. Thus f(z) is analytic in T. Write
Am-s(z) m+s(z
f=( Ps ,AA q) a P= A-mu s 1 =A-mv s Ps4 EH ms m=max{IsI > tI),
to see that f is bounded. For Re z=0, z=iy, get s(z)=s+iy(t-s),
hence,
If(z)I=I(Aiy(t-s) u,A-sAAsAiy(t-s)v)I<UAUsHui0UvH0
(6.9) ,
using that Alp is a unitary operator of HO = H , for all real p

and that IIAUs=UA-sAAs110 . Similarly, for Re z = 1,
(6.10) If(z)I < IIAUtHlui01IvI10 .

1.6. HS-chains 32
From the Phragmen-Lindeloef principle we thus conclude that

IIuIIDIIv11IIAllst-r)/(t-s)IIAlltr-s)/(t-s)
(6.11) lf((r-s)/(t-s)) <
Also note that f((r-s)/(t-s)) _ (u,A-rAArv) . Setting u=A-rAArv

and using that
(6.12) IIAIIr = IIA-rAArll = sup {IIA-rAAruq: uEH Hull 0<1} ,
we conclude (6.7), q.e.d.

An operator AEO(m) is said to have order m . Since A is
determined by any one of its extensions As we often will refer
to an operator A:H -*H and say that AEO() (or that A is of
r r-m
order m) whenever AIH, EO(n) .
Next let us discuss adjoints of operators in L(H ). For

s
A E L(Hs) the standard interpretation of adjoint would be that
of Hilbert space adjoint, an operator of L(H ) again. However,
s
for the present purposes it is more practical to regard H_s as
the adjoint space of Hs using the 'pairing'
,
(6.13) {u,v} - (u,v) (Asu,A_sv) , where u E Hs , v E H_s
Here we write As = (A-5)S for the isometry Hs}HO=H mentionned
above and (.,.) denotes the inner product of H=H0. A continuous

linear functional 1: Hs-}1 may be uniquely written in the form
1(u) = (f,u) = (A_sf,Asu) , with some f E H_s as an immediate ,
consequence of the Frechet-Riesz theorem. One then may regard

the adjoint Banach space Hs identified with the space H_s.
Accordingly for A E L(H ) we define the adjoint operator
s
A as an operator in L(H_s) , defined by the relation
(6.14) (u,Av) = (A*u,v) , u E H-s , v E Hs ,
with the pairing (.,.) of (6.13). Similarly, for AE L(H5,Ht)

there exists a unique operator A*EL(H_t,H_s) such that (6.14)
holds for uEH_t, v E Hs .
Proposition 6.4. For an operator AEO(m) there exists a unique

1:
operator A EO(m) such that
(6.15) (As )* _ (A*) m-s
Proof: For AEO(m) all the adjoints (As) exists and define conti-
nuous maps H -H Moreover, these maps coincide for s and t,
.
m-s -s
1.6. HS-chains 33
on uE H nH , due to
m-s m-t
(6.16) (As u-At u,v) _ (u,(A5-At)v)=0 , vEH
since H0 is dense in every H It follows that (A )*IHt = A*

r .
s
takes H_ to itself and is independent of s . Moreover, A* admits
A*s
continuous extensions defined by (6.15), and hence has all
et
norms IIAIIs m finite. This implies that A EL(H.) , q.e.d.
Now we come to the main result of this section, a result on
Fredholm operators in 0(m).
Recall that an operator AEL(Hs'Ht) is called Fredholm if
we have a(A)=dim ker A -, 6(A)=dim(Ht/im A)<-. A Fredholm opera-
tor always has closed image, and it follows that 6(A.)= dim ker A*.
Here we may interpret A as the above adjoint operator A :H-t+H-s.
The Fredholm index of A is defined as ind A = a(A)-S(A).
It is known that Fredholm property and index are invariant under
small bounded and arbitrary compact perturbations. For a more
detailed discussion cf.X,l. For fully detailed proofs cf.[C1],AI.
Definition 6.5. An operator AEO(0) is said to be K-invariant under

H-conjugation if we have
(6.17) AS AS A-1 - AO = K(s) E K(H) , s E R
with the above isometry As:Hs+H0 , defined as As=(A-s)s
Theorem 6.6. Let AEO(0) be K-invariant under H-conjugation, and

let A0 be Fredholm. Then all operators As5 sER, are Fredholm, and
we have ind A = const. independent of s Moreover, we even have.
s *
ker As and ker(A ) s independent of s , and both spaces are subspa-
ces of H- .
Proof. Since As and its inverse are isometries it is trivial that

As is Fredholm if and only if ASAA-1= AO+K(s) is Fredholm. The
latter is correct, since A0 is Fredholm, and K(s) is compact.
Similarly we get ind A = ind A A A-1 = ind(A +K(s)) = ind A
s s s s 0 0
In particular ind As is independent of s
On the other hand
(6.18) ind As = dim ker(A5) = dim ker As - dim ker(A )-s
using (6.15). Here both terms at right cannot increase, as s

increases: Indeed we must have ker As D ker At as s<t so that ,
e
a(As)=dim ker As is nonincreasing. Similarly, ker(A*)-sDker(A ) -t,
1.6. HS-chains 34
as s<t , so that S(As) = dim ker(A )-s is nondecreasing.

It follows that both dimensions at right of (6.18) also
must be constant. Theorefore the two null spaces do not depend on
s and must be contained in H. . This completes the proof.
Remark 6.7. Let P and Q denote the orthogonal projections in H=HO

onto ker AO and ker A respectively. Recalling that
0,
(6.19) Qu = I'Pj4j,u) , uEH ,
with orthonormal bases ¢.} and } of the finite dimensional

spaces ker A0 and ker A 0 respectively, it is clear at once
that P,QE 0(-o) , using that cj, 0j E H_ , and using the pairing
(6.13).
Later on we shall construct a distinguished Fredholm inverse
B0 of A0, such that B0A0 = 1-P , A0B0 = 1-Q It follows thus that
.
such operator B0 is an inverse even mod operators H_.-H. of finite

rank (called a Green inverse, later on).
Problem 6.8. Let F- denote the class of all operators T E 0(-o)
of the form T = If j)(gj , fj, gj E H_ , i.e.,
(6.20) Tu = If.(gj,u) , uE H-. , with (.,.) of (6.13) .
with given fj gj , and a finite sum. Show that the closure of

,
F. in the Frechet topology of 0(0) coincides with the class of

all AE 0(0) such that ASE K(HS) for all sE Ik
, .
Hint: Repeat the proof of [C1],p.144, lemma 3.2 in abstract form.

CHAPTER 2. SPECTRAL THEORY OF DIFFERENTIAL OPERATORS.
In the present chapter we will apply some of the abstract

principles developed in ch.l to the case of a differential opera-
tor. Here a differential operator appears as a realization of a
differential expression
(0.1) L = aa(x)D° , xEQ
where H is an open subset of Rn , and multi-index notation is

used (cf.sec.l). The coefficients as will be complex-valued
functions over 2 Later on we also will consider matrix-valued
.
coefficients, or differential expressions on manifolds, acting

on crosssections of vector bundles (cf. X,3). The term realiza-
tion is discussed in sec.l, below.
All of our discussion will be restricted to hypo-elliptic
expressions only. In fact, our main effort will be devoted to
elliptic expresions, and even only ordinary differential expres-
sions. In sec.l we discuss the v.Neumann-Riesz extension theory
for the minimal operator of a hypo-elliptic self-adjoint expres-
sion. In sec.2 we discuss generalized boundary conditions, in the
regular and singular case. In sec's 3 and 4 we focus on the
spectral theorem of an essentially self-adjoint realization of an
ordinary differential expression. Specifically a detailed proof
is discussed in the case of a second order expression.
In
1. Linear differential operators on a subdomain of
In the present section we will be concerned with differential

operators, as a special class of unbounded linear operators,
regarding their spectral theory. It is convenient to use multi-
index notation of In , as follows
aj = ax = 8/2x , Dj = -iaj > D = (D1,.... Dn) ' a = (°l....,an)

J J
II.1. Differential operators , generals 36
a a
Da = D1a1D2a2...D n , 8a = 81a1...an n
a
xa = x1a1...xn n , for x = (x1,...,xn) , etc.
For multi-indices a,6 it also will be convenient to abbreviate

lal 1 z
= al + ... + an , a! = al!...an! I (6) = (61)(62)... .
Suppose we are given a differential expression
(1.1) aa(x)Da
L = as E C(C) ,
a _
<N
over some open set C C In. For a differential expression a formal
adjoint L* is defined, also a differential expression,by
(1.2) (Lu,v) = (u,L*v) , u,v E CO(C)
with the inner product and norm of the Hilbert space H = L2(C)
(1.3) (u,v) = J,,u(x)v(x)dx , (lull = {J Iul2dx)1/2

Q
By a partial integration one finds that, for all u,v E CO
(Lu,v) = J dx a ()D uu v =J S2 dx laI<N (-1)lala D°uv

laI<N a a
(1.4) C
= la <NuDa(aav) = (u,w) , w = la <NDa(aav) = My
since all boundary terms cancel, for such functions. A differen-

tial expression is determined by the values Mu, for all u E CO
since they determine all aa Thus M is determined by all values of
(u,Mv), for u,v E CO, since C 0 is dense in H.Accordingly we have
(1.5) Lt = Daaa
laT<N
Using Leibnitz'formula
(1.6) P(D)(au) = J(D6a)(P(S)(D)u)/R! , PM(o _ (3Yp)( )
6
*
one may write L in the form (1.1) again, with new coefficients
as E CO. Note that as = as for IaI=N , so that the principal part

LN (i.e,the sum of highest order terms) of L is obtained by
conjugating the coefficients of LN
Generally, a differential operator will be a (linear) map
A:X--Y between spaces of functions or distributions X and Y such ,
that Au=Lu for all u E X , with some differential expression L.

,
Normally, unless stated otherwise, we shall assume the expression

L to have C'-coefficients, and that C'(D)C X C D'(D) , with the
space DI(D) of all distributions over H.
Conventionally spectral theory is considered only for ope-
rators defined by hypo-elliptic differential expressions.
An expression L (of the form (1.1)) is called hypo-elliptic if
the differential equation Lu=f admits only infinitely different-
iable solutions, whenever f is C' , for all open subsets Sl' C Q.
(Speaking precisely, every distribution u E D'(H') solving Lu = f
in H' necessarily is C°°(H'), and a classical solution of Lu = f.
In [C3] we will discuss criteria for hypo-ellipticity of a
differential expression (cf. also [Ho1] [Ho2]) . Here we note that
an elliptic differential expression necessarily is hypo-elliptic.
A differential expression L of the form (1.1) is called elliptic
if, for all x E H , and all 0 # E E In ,
(1.7) aa(x)Ea # 0
ja =N
Particularly an ordinary differential expression
(1.8) L = a.()D1 , a < x <

0<<N 1
with x=(x1) E il, and 0 = (a,8) C 1, an open interval, is elliptic
if and only if aN(x) # 0 , for all x E (a,g).
We shall require hypo-ellipticity only in the slightly wea-
kened form of lemma 1.2, below. Detailed proofs of such weaker
regularity property, in the cases of an ordinary expression L, or
a second order strongly elliptic expression are discussed in 11,2,
and III,1.
Presently we shall study unbounded operators of the Hilbert
space H above, in the sense of I,1, which are differential ope-
rators. For a given differential expression L one first intro-
duces the minimal operator L0 and the maximal operator L1 defi- ,
ned by setting Lju = Lu for u E dom L. , j=0,1 , where
(1.10) dom LO = C'(Q) C H , dom L1 = {u E C°°(H) : u, Lu E H}
Clearly dom L0 and dom L1 are dense in H , so that L0 and L1 are

unbounded linear operators in the sense of I,1. We have L0 C L1.
An operator A E P(H) with L0 C A C L1 will be called a realization
of the expression L. (In later chapters we will use the term
'realization' somewhat more generally for an operator AEP(H) with

C A C L1* . Actually this notation is in general use for
unbounded linear operators A of more general Hilbert or Banach
spaces in a similar relation to the (closures of) the minimal and
maximal operator. For the present chapter we appoint that L0 C A
C L1 is required for a realization.)
For a realization A we must have LOAD ADD L1 D (L)0
implying that dom Ax D dom (L") 0 is dense in H. Thus A is preclo-
sed. Note that realizations normally are not closed. Apart from
realizations also the (restrictions of) their closures and
(Hilbert pace) adjoints are differential operators in the distri-
bution sense, as easily verified.
,t
A differential expression L is called self-adjoint if L =L.
For a self-adjoint expression L the minimal operator clearly is
hermitian, but the maximal operator in general is not hermitian.
For spectral theory of differential operators one will be interes-
ted in the self-adjoint extensions of L0 As a restriction of L1
.
a realization cannot be self-adjoint, but at most will be essen-

tially self-adjoint. In view of the Riesz-v.Neumann extension
theory of 1,2 the question arises whether all self-adjoint exten-
sions of L0 are closures of essentially self-adjoint realizations
of the given expression L.
As a general assumption let us assume that all expressions
L-A, for A E M are hypo-elliptic. Then L will be called strongly
hypo-elliptic. If L is not of order 0 and is either elliptic or
of constant coefficients and hypo-elliptic, then it is strongly
hypo-ellitic. Similarly for the general formally hypo-elliptic
expressions fitting into the definitions of [C3],III, (having
a local parametrix of negative order). All this is discussed in
detail in [ C3] .
Lemma 1.1. Let the expression L be self-adjoint and strongly hypo-

elliptic. Then the defect spaces D+= (im(L0±i))L are contained in
dom L1 and are just the eigenspaces of L1 to the eigenvalues ±i
respectively. Moreover, every eigenvector of L0 also is an eigen-
vector of Ll.
Proof. It suffices to prove the last statement.Let L0u = Au, then
(1.11) (u,(L-A)v) = 0 , for all v E CA(D) = dom L0 .
Since L is self-adjoint and u E H C D'(D) , (1.11) implies that

(L-A)u=0 , xES2 , in the distribution sense. Thus u E C°'(c), since

by assumption L-A is hypo-elliptic and the function 0 is C°°(D).
Then Lu=au in the sense of standard partial derivatives. Since
uEH we get Lu=au E H, hence u E dom Ll which completes the proof.
The lemma, below, is a straight extension of lemma 1.1.
Lemma 1.2. Let L be hypo-elliptic, not necessarily self-adjoint.

Suppose for some f E H and g E C°'(c)nH we have
(1.12) (f,L*u) _ (g,u) , for all u E C0(D) .
Then we have f E C°°(D), and, moreover, f E dom L1 , Lf=g , x E n

with standard partial derivatives, possibly after changing f on a
set of measure zero.
Again the proof is immediate if we realize that (1.12) for
an L2-function f implies that f solves Lf=g in the distribution
sense. A function f E H satisfying (1.12) (for some gEH) is called
a weak L2-solution of the differential equation Lf=g. Evidently
such f is just a distribution solution which also belongs to H
Lemma 1.2, for the special case of a second order elliptic
expression, is generally known as Weyls Lemma. It was proven by
H.Weyl for the potential equation [We2) We will refer to the .
property of an expression to have at least every weak L2-solution

in C°° (for gEC°°) as (the property) of Weyl's lemma. This property
is weaker than hypo-ellipticity, but it will be sufficient for all
of our purposes. It will be more convenient to use since we do
not have to refer to distribution calculus.
For the remainder of sec.l let us assume that he expression
**
L is self-adjoint. For the closure L0 of the minimal operator (a
closed hermitian operator) we now have the well known direct decom
position (1.13), below. Moreover, it is orthogonal with respect to
the "graph inner product" (u,v)L (u,v)+(L*u,L*v) (cf.I,(2.9)).
0
**
(1.13) dom LO = dom LO ®D+ ® D_
Let us introduce the amended (slightly larger) minimal operator

LOS
LO as the restriction of to dom.,LO = C°° n dom LO. As a rest-
riction of the closure Lp* of L0 ,L0 can be reached from L0 by
closing That is, each u Edom L0 is the limit of a sequence un
.
E C' such that also Lun 3 LOu. We get (LOu,v)=(LO*u,v)=(u,LOv)

=(u,Lv)=(Lu,v) for u E dom LO , vECO(D), using (1.4). Since C-(D)
II.l. Differential operators , generals 40
is dense-in H, we get L0u = Lu for all u E dom L0. In other

,
words, LO is a realization of the expression L.

Notice that also L1 is the restriction of L0 to L0,
by the same conclusion: We get dom L1 C C- n dom LO , using (1.4).
Also, L0f=g, for some f E C', amounts to (1.14), below, using
(1.4) again.
(1.14) (g,u) _ (f,L0u) = (Lf,u) , for all u E C'O.
This implies L0f = Lf E H , so that f E dom L , hence dom L1 D

C-ndom LO . Thus L1 indeed is the stated restriction.
Theorem 1.3. For a self-adjoint, strongly hypo-elliptic expres-

sion L we have the algebraic decomposition
(1.15) dom L1 = dom L-0 + D+ + D_
where every pair of terms at right intersects at 0 only.

Every self-adjoint extension A of L0 (or L"'0) is the closure of a
unique essentially self-adjoint realization B determined as the
restriction of A to dom A n C'(D) = dom B.
Moreover, B is determined as a restriction of L1 to the space
(1.16) dom B = dom L0 + { (u-Wu) : u E D- } ,
where W: D- D+ is an isometry onto D+ , with D+ Hilbert spaces,

as closed subspaces of H. Also B is uniquely determined by W, and
(1.17) Bu = Lu on dom L0 , B(u-Wu) = -i(u+Wu) , u E D_
The decomposition (1.16) of dom B is orthogonal, with respect to

the inner product (u,v)L=(u,v)+(Lu,Lv), defined for u,v E dom L1.
The proof of (1.15) is a consequence of Lemma 1.1, and the
facts on L0 and L1 derived. The remainder of the theorem then
follows from the v.Neumann-Riesz theorem (cf. I, thm.2.4 ).
Theorem 1.3 amounts to a strong simplification of spectral
theory, for hypo-elliptic expressions, because all discussions
may be restricted to C"-functions, and no generalized derivatives,
Sobolev spaces, etc., have to be considered. This is under the
assumption that simple criteria for hypo-ellipticity can be
derived, which in turn do not use generalized derivatives. Note,
that the concept of distribution may be avoided, above, because
all conclusions already follow, if the statement of Weyl's Lemma
is known to be true.
Let us still mention the Carleman alternatives. Carleman
[Call, in essence, arrived at the conclusions of thm.1.3 in the
case of a Schroedinger operator L = H = -A+q , for Q=fin , with a
'potential' q(x) . He distinguishes between the definite case
where def L0= 0 , and L0 possesses precisely one self-adjoint
extension, and the indefinite case, where def L0 = (v,v)
v>0 Similarly, in the case of an ordinary differential
.
expression, H.Weyl speaks of the limit point case, and the limit
circle case, respectively, with a notation referring to his
special construction.
2.Generalized boundary problems;ordinary differential expressions.
In section 1 we derived a direct decomposition (1.15) of the

domain of the maximal operator, assuming that the expression (1.1)
is strongly hypo-elliptic and self-adjoint. In particular we con-
cluded that all self-adjoint extensions of the (hermitian) minimal
operator L0 are given as closures of certain restrictions of the
maximal operator L1, called e.s.a. realizations of L. In turn the
e.s.a. realizations are characterized by an isometry W:D- -+ D+
between the defect spaces D+ , similar as in the v. Neumann-Riesz
theorem (I, thm.2.4).
For a given e.s.a.-realization A of the expression L the
condition uEdom A ', imposed on a function uEdom L1, amounts to
'
a generalized boundary condition, because if u e dom L1 satisfies

u E dom A , then so does u + v = w , for all v E C0 , due to C0
C dom A. In other words, the condition u e dom A is not influen-
ced by the behaviour of u away from the boundary, it depends only
on the properties of u in some (arbitrarily small) neighbourhood
of the boundary of the domain D. In this sense we associate to
every e.s.a.-realization of L a (generalized) boundary problem
One will be tempted to ask for classes of e.s.a.-realizations
with boundary conditions of the conventional type. For an investi-
gation of this kind involving partial differential operators, and
more generally, dissipative and accretive boundary conditions cf.
[ CFO] .
In this and the following section we will get restricted to
n=1, i.e. , to the case of an ordinary differential operator.
Spectral theory of "ODE's" was completed to a high degree of per-
II.2. Boundary problems , ODE 42
fection, starting with the work of H.Weyl [We1]. From the large
number of contributions to this subject we mention the work of
Hilb [ H11] , K.Kodaira [ 1 < 0 1 ] , M.G.Krein [ Kr1] , N.Levinson [Lei] ,
E.C.Titchmarsh [Ti1] ,M.A.Naimark [Ne1], E.A.Coddington [Cd1],and
others (cf. also the monographs [ Ne2] , [Oh] , [ Bz1] , [ RS] , [ CdLi] )
We write
(2.1) L = J jN=0
a.(x)3 3 = d/dx a < x <
with extended real numbers a<S . For simplicity assume that a. E

C°°((a,s)) are real-valued. The adjoint of L is given by
(2.2) L _ 3=0(-1)030a.(x) _ (-1)NaN3N+ terms of lower order
Assume L*=L, and L elliptic (i.e., aN#0). Particularly aN=(-l)NaN

follows from (2.2), so that N must be even. Also assume NiO, so
that N=2,4,6,... .
We give a proof of Weyl's lemma (i.e., Lemma 1.2), for

this case.
Lemma 2.1. If (Let-X)f = g, for some g E HnC°°(S2), then fEdom L1.

Proof. We require a fundamental solution of L, that is, in distri-
bution terms, a solution e(x,t) of the differential equation
(2.3) (Lx-a)e(x,t) = 6(x-t) , a < x,t < S ,
with the'Delta-function' 6 . In simple terms this is a CN-2-func-

tion in the square of (2.3) being CN in each of the two triangular
regions a < x < t < g and a<t<x<S and solving (Lx_X)e=O there,
,
where writing Lx expresses that L is to be applied onto x (not t).

Moreover, at x=t we require the'jump condition'
(2.4) 3N-le(x+O,x) - 3N-le(x-O,x) = 1/aN(x) , a < x <
It then is easily verified that (2.3) holds.

Note that the general existence and uniqueness theorem for
ODE-s, and the known results on differential dependence of solu-
tions on initial data at once supply the existence of (an infinite
family of) fundamental solutions: One may prescribe arbitrary ini-
tial data 3Xe(t±O,t), subject to (2.4) and 3Xe(t+0,t)=3Xe(t-0,t),
O<j<N-2. For any such choice (in CN(O) of course) one obtains a
,
unique function e(x,t) satisfying the above requirements. If, in

addition, we require the data in C°° then also e will be C°° in the
11.2. Boundary problems , ODE 43
two triangles. The continuity of the derivatives then also gives

aN-le(x,x-0) aN-le(x,x+0)
(2.5) - = 1/aN() , a < x < 6 .
In the proof of Lemma 2.1 we may assume that a=0, due to

N L>2. Suppose we have LQf = g g E C,
, or, ,
(2.6) (f,Lu) _ (g,u) , for all u E CO
Let us focus on some point x0 E C _ (a,s) . We can find x E C0(C)
with X=l near x0, and ECOU) , X20 in R , = 0 , IxI?1,
and then let E(x) = Vx/E)/E . The function
(2.7) u(x) = u5(x) = X(x)J Yx0-y)e(x,y)dy , 0 < E
then will be in CO(C) and may be substituted into (2.6).
Indeed, we have v(x) = Ja E(x0-y)e(x,y)dy = Ja + JS , with both
integrands being Cam, so that v E Cam, hence u E C . Moreover (2.3)

suggests that
(2.8) Lv(x) =JOE(x0-y)6(x-y)dy
as also may be easily verified by a calculation, using (2.4).

Moreover, for sufficiently small E ,
(2.9) Lu(x) = Yx0-x) + Y={L x(X(x)e(x,y))}f
with the 'function part' {d}f of the distribution d (i.e., {d}f

omits the delta-function term arising from differentiating
at the discontinuity x=y). Clearly the integrand vanishes identi-
cally for small 1x-x01 using that Y = (Lxe(x,y))f=(6(x-y))f =
, 0
(due to E=1). Moreover, we have lx0-yI small in supp E(x0-y),

so that the possible singularity of Y(x-y) is outside supp E
unless Ix0-xl is small, in which case E=0 anyhow. Thus, for small
E>0 we have the integrand C for all x,yEC
Substituting into (2.6) and interchanging integrals we get
J E(x0-y)dyJX(x)g(x)e(x,y)dx = Jf(x)0E(x0-x)dx
(2.10)
1
Clearly (2.10) remains valid if x 0 is replaced by x close to x0,
where the functions X and may be kept the same. In the limit
e -> 0 , c > 0 , one obtains the equation
(2.11) f(xl) =/j e(x,xl)X(x)g(x)dx -J. dxf(x)y(x,xl)
for almost all x1 E D , with xl close enough to x0 so that ,
X(x)=l still holds near xl . Evaluation of these limits requires

a common technique known as regularization (or mollification)
(cf. [Cl1,I).
Note that the right hand side is C', as a function of xl
since the first term looks like v of (2.7) , while the kernel y
of the second term is C , as seen above. Thus by changing f on
a null set, keeping it within the equivalence class, representing
the given element of H , we find that fEC' , near x0 Since the .
construction is valid for every x0 E S2 the proof is complete.

,
The above proof is valid without the assumption of self-

adjointness for L or reality of coefficients, as long as L is
elliptic, as easily checked. Also, along similar arguments, one
easily confirms the following.
Corollary 2.2. The space dom L0 consists precisely of all u E

CN-1(0) with absolutely continuous N-1-st derivative such that
u(N
L u E H , where the N-th derivative is defined as derivative
almost everywhere of u(N-1) . Moreover, for fE dom L* Lgf=g we ,
still have (2.11) valid, for any C0'(c2)-functions X(x) equal to 1 ,
near x 1 E 0 , and the C'-function y(x,y)=Lx(X(x)e(x,y)), (ignoring

the delta-function arising from the discontinuity of the (N-1)-st
derivative of e(x,y) at x=y).
Indeed, we have proven (2.11) for all fE dom L0 with L0f=g,
where gEC°°nH, but notice that exactly the same derivation works
for fE dom L0, g=L0f, regardless whether or not g E C'. By diffe-
rentiating (2.11) it then follows at once that such f must be
CN-1(Q) and that its (N-1)-st derivative must be still absolu-
,
tely continuous. Also the differentiation implies that L"f=g E H

Vice versa, if fECN-1(0) has its (N-1)-st derivative absolutely
continuous, and if g=L*f , with f(N) existing only almost every-
where, is in H, then (2.6) follows by partial integration, so that
f E dom L0 and L0f=g q.e.d.
, ,
Note that, after Lemma 2.1, we now have complete control of

sec.1 , without having to rely on the hypo-ellipticity of L
(which in general is discussed in [C3). ).
If an endpoint a or a is finite and has the property that

all coefficients aj may be extended up to it, as Co-functions,
keeping aN90 also at that endpoint, then we shall speak of a
,
regular endpoint. Otherwise the endpoint is called singular.
Lemma 2.3. If a is a regular endpoint, then dom L C CN-1([a,g)).

1
Similarly for a regular endpoint S If both endpoints are regular
.
then dom L1 C CN-1([a,s]) .
Proof. If a is a regular endpoint, then all functions of the

defect spaces D± are C`,ta,s)), as solutions of the differential
equation (L±i)u=0. Indeed, it is well known that a solution of a
linear differential equation with C--coefficients stays C- in
the entire interval where the coefficients are defined, as long
as aN#0, also at the boundary (c.f. [CdL1] for example). In view
of the decomposition (1.13) it thus is sufficient to show that
dom L-0 C CN-1([a,s)) .
We again use a fundamental solution e(x,y). Since the solu-

tions of Lw=O extend smoothly into the left endpoint the same is
true for the fundamental solution e(x,y) we constructed. In fact,
we may assume e(x,y) defined for all x,yE [a',s) , with a'<a ,
a slightly larger interval. Let first fE CD , g=Lf . Then (2.11)

follows even for the more general case of a cut-off function X E
C-(1) , X=l near a =0 near a
, and is valid for xl near a
, The .
reason for this extension is that (2.6) now holds not only for u
with compact support, but for general u with LuEH , hence u of the
form (2.7), with our new function X . Of course we also must
define y with this new X , and then get yE C'([a',B)x[(X',S)), y=0
near x=$ .
The same formula then holds for general f E dom L0, because
a sequence fj E C' may be choosen with Lfj}g and setting ,
f=f. in (2.11), and passing to the limit j will give that

formula. Given the new representation of f near its left end-
,
CN-1
point, one may differentiate N-1 times to confirm that f E
near a. Similarly for the other endpoint. Q.e.d.
The extended formula (2.11) of the proof of lemma 2.3 also
gives the following.
Corollary 2.4. If a is regular then
(2.12) dom LO ={uEdom L1: u=u'=...=u(N-1)=0 at a, u satisfies BS},
where 8 S denotes a certain generalized boundary condition at s.

A corresponding formula holds, if S is regular. If both, a and

$ are regular, then dom L""0 consists precisely of all uE dom L1
,
with derivatives up to order N-1 including vanishing at a and S .
Proof. The statement may be expressed as follows: (i) a function

u E dom L0 must have all derivatives at a equal to zero, up to
order N-1 Moreover (ii) if v e dom L1 has that property near a
.
and vanishes near $ , then we get v E dom LO To prove this we .
apply our modified (2.11) (of the proof of lemma 2.3). For
u E dom LO , w=L0u we get
(2.13) u(a)=limaau(x)=JQe(y,a)X(y)w(y)dy-J.y(y,a)u(y)dy
Any such u is the limit of a sequence ujECo for which the right
hand side of (2.13) vanishes. Hence we also get the left hand
side zero. Similarly we may differentiate up to N-1-times, for
(2.14) u(1)(x) = K1(y,x)w(y)dy -J0X1(y,x)u(y)dy , 1<N-l,
with kernels K1, Al given as the 1-th x-derivatives of the kernels

in (2.13). These kernels still are L2 in the integration variable
y, and as L2-functions vary continuously with x as x is near a. ,
Therefore we also get u(1)(a)=O 1=1,...,N-l. This proves (i)

,
To verify (ii), let v E C' be 0 near S , and let

=limxyav(N-1)(x)
0 at a. (That is,
v=v'=...=v(N-1)=
=0 . ) Then we just show that v E dom L"'0, by constructing a

sequence vj E CD with vj + v , Lvj -r Lv . Let w E C-U) , w = 0
near a = 1 near R. Then vj(x) = v(x)w(a+j(x-a))

, j=1,2,..., ,
has the desired properties, as a calculation shows. Similarly

if s is regular or if both a and S are regular. Q.E.D.
Remark 2.5. Note that, for an elliptic ordinary differential

expression L a characterization of dom LO is possible along the
lines of cor.2.4: If the endpoint a is regular, then dom L0 is
given as the set of (2.12), with dom L1 replaced by dom L
Similar for a regular endpoint B or regular a and S , .
Let us first assume that both endpoints a and B of a given

expression L are regular. We introduce the sesqui-linear form
(2.15) Q(u,v) _ (u,Lv) - (Lu,v) , u,v E dom L1 .
Proposition 2.6. For uEdom L1 let us introduce the column vectors

(2.16) ua"=
(u(a),...,u(N-1)(a))
, us"'= (u(g),...,u(N-1)(6))
,
(well defined in view of lemma 2.3), and the (2N-column-) vector
(2.17) u'" _ (ua",u ) .
There exists a nonsingular skew-symmetric 2Nx2N-matrix
(2.18)
with NxN-matrices Q `, Q real coefficients such that, with

a
the inner product of T2N or or TN , respectively,
(2.19) Q(u,v)=(u-,Q-v"')=(us ,Q6-v)-(ua"',Qa-va-), u,vEdom L1.
Proof. Letting a < a' < g' < 13

, we get
(2.20) Q(u,v) = lima',R'-a,8 Ja,(av - L*uv)dx
Using a partial integration as in (1.4) we find the right hand

integral equal to a sum of boundary expressions of the form (2.19)
with a,13 replaced by a',S'. The matrices Qa , Qare composed
from derivatives of the coefficients a.. Only products u(j)v(1)

with j+l<N-1 can occur, so that qj1=0 for j+l>N-l. Moreover, for
j+l=N-1 the coefficient is either or ±aN(13 ). In the
limit a'ia, 6'-6 we get (2.19). the matrix coefficients remain the
same, except that the values of a(0) are taken at a,1 now. Hence
Qa and Q8 are nonsingular as triangle matrices with non-vanishing
diagonal terms, referring to the 'anti-diagonal' j+l=N-1. Finally,
(2.21) (Q-u-,v-) = v ,Q"u = -((u,Lv)-(Lu,v)) = -(u-,Q"v") ,
confirming that Q"' is skew-symmetric. (All 2N-vectors u",v" are

assumed, as u,v vary through dom L1.)
A linear subspace W of T2N will be called a maximal null
space of the matrix Q"' if (i) (u-,Q-u-)=0 for all u- E W, and (ii)
every subspace V with W C V C T2N and with property (i) coincides
with W.
Theorem 2.7. If both endpoints of an even order expression L with

real coefficients are regular, then the class of all e.s.a.-reali-
zations A of L is in 1-1-correspondence with the class of all

maximal null spaces W of the form Q-, as follows. For a maximal
null space W an e.s.a.-realization is defined by
(2.22) dom A = {u E dom L: u- E W} , Au = Lu for u E dom A .
Vice versa, for an e.s.a-realization A the space W is defined by
(2.23) W = {u-: there exists u E dom A with boundary values u-}.
A proof of Theorem 2.7 will be discussed in section 3. Here

let us get some insight into the nature of a maximal null space W.
The non-singular skew-symmetric matrix Q- has pure imaginary
eigenvalues. Since Q- is real there are precisely N eigenvalues
with positive and negative imaginary part each, and the corres-
ponding eigenvectors span spaces W+ with dim W+ = N We have the .
g2N=W+
orthogonal direct decomposition $ W_ We get positive.
definite sesqui-linear forms F+ defined by
(2.24) F+(u`,v-) = +iQ-(u-,v-) , u-,v- E W+
That is, the matrix Q- is reduced by the above direct decomposi-

tion, and, up to a factor ±i, F+ are the forms induced in W+.
Then the maximal null spaces of Q are given by
(2.25) W = {u-+W-u-: u E W-1 ,
where W- is any isometry W - W+, the latter two spaces considered

under norm and inner product given by the forms (2.24).
Note that Q also is reduced by the direct decomposition
2N = gN $ gN, as manifested by (2.18). Some (but not all of its)
maximal null spaces are of the form
(2.26) W = (u-=(u a-,u): ua- E Wa ,

u E Ws } I
with maximal null spaces Wa , WS of Q_ and QS', resp. The corres-

ponding boundary condition then will be called separated We get
(2.27) dom A = {u E don L1: ua-E Wa, uS-E WS } .
Note that (2.25) implies that all maximal null spaces 0 of Q

have dimension N, and all maximal null spaces of Qa, QS have
dimension N/2 (an integer). As easily seen, any N-dimensional null
Q,_'
space of Q is maximal. Similarly for Q
$_*
Theorem 2.7 then simply is a consequence of the fact that,

for regular endpoints a,$, every e.s.a.-realization has defect-
index (N,N), and the e.s.a.-realizations are characterized as the
hermitian extensions A , L0 C A C L1 , with dim (dom A/dom L0)=N.
This will be discussed in sec.3.
3. Singular endpoints of a 2r-th order Sturm-Liouville problem.
Let us next discuss the defect index of LO , under the

assumptions of section 2.
Lemma 3.1. Under the general assumptions of section 2 we have
(3.1) def LO = def L0'= (v,v) , 0 < v < N.
Moreover, if one endpoint is regular, we get r=N/2 < v < N . If

both endpoints are regular we have v = N .
Proof. Recall that the functions of D+ are Cm-solutions of Lu=±iu

in 0 . It is well known that there are precisely N linearly
independent such solutions of either equation. Clearly D+ will
consist of precisely the squared integrable solutions, hence must
have dimension between 0 and N. Both dimensions are equal since
D+ = D , using that L has real coefficients.
If both endpoints are regular then the solutions of Lu=±iu
are C'({a,s]) , as already observed. Hence all solutions are
in H, and v = N follows.
From (1.13) we get dim(dom L1/dom L0-)=2v. On the other hand,
if one endpoint, for example the point a, is regular, then we
have C'(La,s)) C dom L1 , while Corollary 2.4 implies that all
functions of dom L0-are CN-1([a,s)) and have their derivatives
up to order N-1 zero at a It follows that any two functions u,v
.
E dom L0` with different boundary values ua-9 va-are incongruent

modulo dom LO-5 so that 2v =dim(dom L1/dom li0) > dim UN=N q.e.d. ,
Now we discuss the proof of Theorem 2.7. First of all, for

any hermitian extension A LO-C A C L1, the space Ui={u-:uE dom A)
,
is a linear subspace of M2N and a null space of the form Q, by

,
definition of the hermitian operator. From (1.15) and lemma 3.1

it follows that dim (dom A/dom L0-) = N It follows from cor.2.4
.
that uEdom A is in dom L0- if and only if u-= 0. Accordingly, for

an e.s.a.-realization, we must have dim M = dim(dom A/dom L0-)=N.
so that Ul is a maximal null space. Therefore (2.23) indeed defines
11.3. Singular Sturm-Liouville problems 50
a maximal null space for every e.s.a.-realization.

Vice versa, for a maximal null space (d, let a linear opera-
tor A be defined as restriction of L1 with domain (2.22). Then A
is hermitian, in view of (2.21), since GJ is a null space of Q" .
Also, dim (dom A/dom L0")=dim (1 =N, so that A must be essentially

self-adjoint, again by lemma 3.1 and (1.15), q.e.d.
Let us consider the boundary form Q of (2.15), in case where
endpoint a is regular, while s may be singular. From (2.20)
we get
(3.2) Q(u,v) _ -(ua",Qa"ua") + limy,}s-0 (usr",QS,"us,")
where the limit at right exists, although it no longer be expres-

sed in the form of (2.19). Define
(3.3) Qa(u,v) = lims,-s(us,",QS,vs,") = Q(u,v) + (uaQava")
Note that QS(u,v) = 0 if either u or v are in LS = dom L0

{u E dom L1: u=0 near . Similar as in the proof of Lemma 3.1
we get dim (dom L11L S) = 2v-N = 2(v-r) = 2vs , r = N/2, 0 < vS <r.
One may refer to p( = dom L1/LS as to the (abstract) boundary
space at the right boundary point 5 Clearly QS induces a form
.
over M we get Qs(u,v)=(us,QS"vs") , with a certain skew-hermi-

tian operator QS of pas and the cosets uvs" E NS of u,v.
,
With this new meaning of us",vs",QS", (3.2) assumes the form

(2.10), and the construction of e.s.a.-realizations may be carried
M2N
out similar as in section 2, with replaced by ®N® Ms= N and ,
Qa"®QS"' One obtains the same 1-1-correspondence between e.s.a.

Q"
realizations and maximal null spaces of Q" as stated for the regu-
lar case in Theorem 2.6, with the only difference, that a maximal
null space now has dimension r+vs < N. Also the construction of
maximal null spaces from isometries remains unchanged from (2.25).
We may distinguish separated boundary conditions, etc. It is easy
to construct explicit e.s.a.-boundary conditions by first choosing
a basis in MS , a matrix for Q" and Qand then constructing
an explicit maximal null space.
In case of two singular endpoints it is convenient to first
consider the expression L in the two subintervals (a,y] and [y,s),
with some point y , a < y < a Let the restricted expressions be
.
denoted by La and LS respectively.

,
Proposition 3.2. We have, with La Ls as above, for La and Ls

,
respectively,
II.3. Singular Sturm-Liouville problems 51
(3.4) d = dom L1 /dom L- = dom La/L a ® don LS/LS = Pa .

V
Moreover, the defect index V = v+ = v- of LO now is given by

V = va + vs , with va and vs of La and LS the expressions with
,
one regular endpoint.

The simple proof is left to the reader.
Clearly the projections u ->ua- and u -u- may be defined, just as
above. Then, with two abstract boundary spaces M a and MS the con-
struction of e.s.a.-realizations from maximal null spaces assumes
the old form.
We now will state the spectral theorem for singular (even
order) Sturm-Liouville problems. (Note that, conventionally, the
eigenvalue problem Au = Au, for an e.s.a.-realization A of a self-
adjoint elliptic ordinary differential expression of even order,
and with real coefficients is called a Sturm-Liouville problem.
The special point will be that, for an ordinary expression
do*
L the spectral family of A
, can be linked to the finite dimen-
sional space of solutions of the differential equation Lu=au
For the remainder of this section we fix a fundamental set
(3.5) u1(x;A) , ... , uN(x;A)
(i.e.,a basis of the solution space) for the differential equation

Lu = Au by imposing the condition
(3.6) 33ul(Y,a) = 6j+l,l , j+l,l = 1,....N
where y is any point, a < y < S, kept fixed for all the discussion
From standard results on ODE's we conclude that uj E Cfor
fixed A E ® , and that 31u1(x;A) is an entire function of A , for
every fixed x E (a,B) and every j,l. For the following the condi-
tions (3.6) are not essential, but it is important that the u.
form a fundamental set with the above smoothness and analyticity.
Let o(A)=((a.l ))be a matrix-valued function,
j,1=1,...,N'
for A E I The coefficients are assumed
. real -valued, and a(A)
is symmetric, for every A. We shall call a non-decreasing if a(A)
_< a(U), as A < U, in the matrix sense (i.e., a. = a(U) - a(A)
satisfies jaijl 7cl > 0, for all E E ®N ). It is clear that the
diagonal elements ajj are real-valued, non-decreasing functions.

Moreover,
11.3. Singular Sturm-Liouville problems 52
1/2
(3.7) 10
X14 -< (a 4jj *'All) i/2(a
4jj
+ a
411
)
follows, so that all elements of a are of bounded variation, in

any compact subinterval of I . An inner product and a norm on
MN-valued C--functions may be defined by setting
(3.8) (u,v)a = J IN u.(A)v1()da.1(a) , [lull = (u,u)al/2,

J J
J,1=1
with an improper Riemann-Stieltjes integral, at right. The linear
space Ha of all Co-function with finite norm may be completed to a
Hilbert space Ha It is known that a Riesz-Fischer theorem holds,
.
so that H = L2(1,da) with a corresponding Lebesgue-Stieltjes inte-

gral. We shall not use this fact, however.
In all of the following we assume a continuous from the left.
Theorem 3.3. (Spectral theorem for an e.s.a.-realization A )
For every e.s.a.-realization A of L there exists a non-decreasing
matrix-valued function p(a), - < A < called the spectral den- ,
sity matrix, such that f defined by

1s
(3.9) $. U) = j uj(x;A)f(x)dx , j=1,...,N, f E
a
and f , defined by
(3.10) f(x) = J+ E Hp
- j,1=1
give a pair of mutually inverse unitary operators H H Hp, defined
by continuous extension of the maps (3.9) and (3.10). Moreover, if
U: H -* Hp denotes the operator (3.9), then UAU coincides with the
operator of multiplication by A in Hp . That is,
(3.11) UAU (A) = E U(dom A) , A E I
After choosing the fundamental set fu j} the spectral density

matrix is uniquely determined as a left continuous function,
up to an additive constant matrix, by the realization A .
A proof of theorem 3.3 will be discussed in section 4, (but
only for the simpler case of N=2 , in the interval [0,oo), with
0 a regular endpoint, and for the Dirichlet boundary condition
at x=0). Our proof is that of M.G.Krein, c.f. the book of Neumark
[Ne2]. An extension of the proof to general N would meet only
minor technical obstacles.
11.4. 2-nd order spectral theory 53
4. The spectral theorem for a second order expression.
Let us consider a second order self-adjoint expression L in

the interval [0,-), (i.e., with a regular endpoint at 0). From
the self-adjointness we conclude easily that
(4.1) L = -d/dx p(x) d/dx + q(x) , x E [0,o) ,
with real-valued p, q (-= C'00,-)), and p(x)# 0 Usually one assu- .
mes p>0. Expressions of this form are known as (second order)

Sturm-Liouville expressions.
From section 3 it follows that all realizations with sepa-
rated boundary conditions have domains of the form
(4.2) {u E dom L1 : u0 E WO , u: E W.
with certain maximal nulspaces WO W. We get u0 = (u(0),u'(0)),
, .
while W. C M = dom L1/L_. We have dim WO = 1 = N/2, so that the

condition 'u0 E R0 ' may be expressed as a conventional boundary
condition (with certain constants a,b jaj2+ lb12 = 1 ),
,
(4.3) au(0) + bu'(0) = 0
A calculation shows that the real boundary conditions, i.e., the

conditions (4.3) with real a,b , precisely characterize all maxi-
mal nul spaces at 0. On the other hand, at the singular endpoint
W the space W. has dimension v_ , where v. + 1 is the number of
,
linearly independent squared integrable solutions of the differen-

tial equation Lu = iu (or Lu = -iu). By lemma 3.1 and proposition
3.2 the defect index of L** must be v,+l > N/2 =1, so that there
are always either one or two linearly independent squared integra-
ble solutions of Lu=±iu each. In the first case dim Wm= dim M_
,
=0, so that u- E W. holds for all u E dom L1. That is, we need no
boundary condition at - in this case.(H.Weyl calls this the limit
point case.) In the other case (the limit circle caee,after Weyl),
we get dim W. = v_ = 1 , dim M = 2v_ = 2 , so that 'u_ E W_ '
amounts to one boundary condition at To express this condi-
.
tion explicitly requires a basis in the two-dimensional space M.

We refer to Neumark [Ne2] for details .
As a special case of the spectral theorem (thm.3.3) we for-

mulate and prove theorem 4.1, below, which refers to a second
order Sturm-Liouville expression in [0,oo), under a Dirichlet-boun-
dary condition at 0, i.e., the condition (4.3) with a=l, bzO . Or,
(4.4) dom A = {u E dom L1 : u(0) = 0 , u E
where the last condition is void if we have the limit point case.
For convenience we introduce the operator L2 , L2 C A C L1 , with
(4.5) dom L2 = {u E C0'([0,oo)) : u(0) = 0
Theorem 4.1. Let u(x;A) 0 < x < : A E B

, be the unique real- , ,
valued solution of the differential equation Lu = Au satisfying ,
(4.6) u(0;a) = 0 , u'(O,A) = 1/p(0) .
There exists a non-decreasing real-valued function p(A), -:<A<:,

such that the integral relations
(4.7) (A) = Ju(x;X)f(x)dx , f E dom L2

0
and
(4.8) f(x) = E C0(1)
define a pair of mutually inverse unitary operators H H HP by ,
continuous extension in H and H . Moreover, if U denotes the ope-

P
rator of (4.7), then we have
(4.9) UAU (A) = A (A) , for all E U(dom A)
The function p(a) is uniquely determined as a left continuous

function with p(O) = 0 , with the above properties.
Our proof of theorem 4.1 is based on the abstract spectral
theorem for unbounded self-adjoint operators (i.e., I, thm.3.3 ),
and some other preparations are required. The result was first
proven by H. Weyl [We11 in 1910 who did not use the spectral,
theorem for his proof. (This was, in fact, earlier than the
abstract theorem. ) The use of the spectral theorem in the proof
became customary with the extension of the result to general
order (i.e., thm.3.3).
For f E I+ _ [0,:), we introduce the entire function
(4.10) D(A) _ D(A;f) = l-u(x;a)f(x)dx

0
Proposition 4.2. We have

(4.11) VA;Af) _ V(A;Lf) = X'(X;f) , f E dom L

2
Moreover, if (D(A;f) = 0 for some AEI, then there exists w E dom L2

such that f = (L2-a)w .
Proof. (4.11) is a matter of partial integration, using that f=0

for large x, and that u and f both satisfy the Dirichlet condition
at 0. Introducing z(x) as the unique solution of Lz=Xz satisfying
z=l z'=0 at 0 , we define
x
(4.12) w(x) = u(x) u(y)f(y)dy
z(y)f(y)dy - z(x)
x 0
Note that the 'Wronskian expression' W = p(uz'-zu') is constant

(its derivative vanishes), so that Wsl , using the values of u and
z at 0. Thus one confirms by explicit differentiation that
(L-X)w=f. For large x the first integral in (4.12) vanishes while
the second equals '(X;f) (which also vanishes,by assumption),using
that f E dom L2 vanishes for large x. Hence w = 0 for large x.
Also w(0) = 0 since u(0)=0 . Hence w E dom L2
, q.e.d. ,
It is easy to find some p E CQU+) such that D(0;p)#0. Then

VO) = VX;ip) can have at most countably many zeros al, A2' ...,
and lim lajl = m, if the sequence is infinite. Applying propo-
sition 4.2 for a=a1 we get w=i1 E dom L2 such that _ (L-al)ipl
and (4.11) yields (D(A;ip) _ '(A;Lw)-'(a;w) = (X-X1)4(a,i1). Accor-
dingly has the same roots as except that Al has its
multiplicity reduced by 1 (or no longer is a zero). It is clear
that we may iterate this proceedure to arrive at the lemma, below.
Lemma 4.3. For any compact interval [a,b] C R there exists a

function p E C°([ 0,w)) such that f(A;i) 0 for all a E [a,b]
,
Lemma 4.4. Let E(A) be the left-continuous spectral family of the

self-adjoint operator B = A Let y be as in Lemma 4.3
. and let ,
A = [a,b). For any interval A'=[u1,42) C A , EA,=E(u2)-E(pl),

and any f E C'(]R+) we have
(4.13) EA,f = dE(A)V
Proof. It suffices to show, that, for v E
(4.14) (EA,f,v) = 1A' 1 (A)d(E(a)V,v) , y(7,)=
since the existence of the integral (4.13), in norm convergence

as a Riemann-Stieltjes integral, is well known. Note that (4.14)
amounts to K(u) = 0 , with

u
(4.15) K(u) = (d(E(X)f,v) - y(X)d(E(A)ip,v)) , u E A
a
We shall prove that the complex-valued function K is diffe-
rentiable and that its derivative vanishes. We have K(u)=(h(u),v),
u
h(u) _ (dE(X)f - y(X)dE(X)p) , and want to show that
a
(4.16) h(u+6) - h(p) = o(6) , as 6 - 0 .
For a moment let 6 > 0 , and let E = E(p+6) - E(p) . Write

u+6
(4.17) h(u+6) - h(u) = E(f-Y(u)$) - (y(a)-y(u))dE(a)i
u
u+6
In the second expression, called J2, we may write J
u+0
u+6
instead of 1
, since the integrand vanishes at p . This gives
u
u+6
(4.18)IIJ2n2= u+0 y
For the first expression, J1, in (4.17), let w = f - Y(V)*

Note that
(4.19) Vu;w) J ()(u;f) - Y(u)(D(u;ip) = 0 ,
by definition of y . Thus, by proposition 4.2 again, write w =

(A-u)w, with some w E dom L2. If u is an eigenvalue of A then one
may add a, multiple of the corresponding eigenfunction u(x;u) as
to obtain a revised w which is orthogonal to u(x,u) , and still
is in dom A . It follows that
u+6
(4.20) 11J112 = IIE(A-1)0 2 = (X-u)2dfE(A)wI12 = 0(62)
u+0
Similarly for 6 < 0 , q.e.d.
Lemma 4.5. We may write (4.13) in the form
(4.21) EA,f = Jo,D(A;f) dE(X)g , g = Jo
Here the function g E H is independent of the specific choice of

* , subject to the conditions of lemma 4.3.
Proof. It is clear that we may write (4.13) in the form of
(4.21). To show that g is independent of , let i"' be any function

with the properties of 'P . Applying lemma 4.4 with f = '_ we get
EA,*' = J6,0(a;*`)/D(X;'P)dE(a)' , hence g" = JA dE(A)V`/D(a;V')
This completes the proof.

We will write g = gO , since this element g E H evidently
is determined by the choice of the interval A . One finds that
(4.22) gp, = EA,gA ,

as A' C A .
Introduce a function g : I + H by setting
(4.23) g(a) = lime-++0,u+a+0 as A>0 , =-g[a,0) as 1 < 0.

g[e,u)
Then one confirms that gO, = g(u2) - g(ul). For f E CO we thus get
b
(4.24) (E(b) - E(a))f = (1) dg(a) , (X) = (D(a;f)
a
for all a<0 , b ? 0, with 'P as in (4.10). Also (4.23) implies
f
(4.25) IHEOf112 = J
l¢(X)12dP(X) , P(a) = Ilg(A)112
Letting a,b tend to we confirm ?arseval's relation
(4.26) 11fll2 = J 1'P(X)I2dp(1) , f E
and it is found that

f
(4.27) f = 'P(a)dg(A) , f E
with a strongly convergent improper Riemann-Stieltjes integral.

From (4.26) it is clear that the linear map f + 'P defines
an isometry in the dense subspace CO(l+) of H , which may be con-
tinuously extended to a map H - H Vice versa, let 'P E C0(I) .
P
be given, then we may define f by (4.27), and will find that fEH.
Then a sequence f. E C'(I+) may be constructed with f. + f in H.
Let $j be given by (4.7) with f = fj . We know that (4.26) holds
for f and 'P
. .It follows that
.
J J
(4.28) f-fj = J ('P-'j)dg(1)
Taking norms we get
(4.29) J Il2dp() = Of - fjll2
This proves that the above isometry maps onto HP , since C0(1+)
is dense in H
P
Next we prove the inversion formula (4.8). From (4.25)
we get
(4.30) (f,g) _ f,g E H
with , defined by (4.7) , with f and g Choose f E H , g = X,

the characteristic function of some interval A = [t,t+d] C 1+
Calculate that (X) = Therefore (4.30) yields
t+d t+d
(4.31) f(x)dx = j $Q)( ) u(x;X)dx)dp(A)
t - t
We assume E C U +) , and then may interchange the integrals at
0
right, then divide by S and let d -> 0. In the limit we get (4.8).
Finally the uniqueness of p is evident, because JI 2dp =
JI
412da for a pair of left-continuous nondecreasing real-valued
functions p , a , for all 4 E implies that p-a = const.

This completes the proof of thm. 4.1.
CHAPTER 3. SECOND ORDER ELLIPTIC EXPRESSIONS ON MANIFOLDS.
In the present chapter we turn to a discussion of a variety

of properties of second order self-adjoint strongly elliptic dif-
ferential expressions with real coefficients defined on a C--mani-
fold of dimension n. In local coordinates such an expression will
be of the form
-K-1a
(0.1) H = Khjk3 k + q ,
xj x
with function K,q>O , and a positive definite tensor ((hjk))
It then is self-adjoint with respect to the inner product (u,v)
= Juvdp , with the measure du=Kdx .
First we again discuss Weyl's lemma for expresssions of this

type, and a boundary extension of Weyl's lemma as well, similar to
that discussed for ordinary expressions in 11,2,3. Instead of the
fundamental solution used in 11,2 we here use a E.E.Lewy type
parametrix of the form (1.9), below. With this preparation we can
give a full discussion of the Dirichlet realization in case of
,
a regular boundary, as an extension of the minimal operator with

closure coinciding with the Friedrichs extension.
Next we will require criteria for the Friedrichs extension,
hence the closure of the Dirichlet realization, to be of compact
resolvent. Such criteria are developed in sec.2.
It is well known that differential expressions of the form
(0.1) possess a variety of important special properties, such as
the maximum principle, and the so-called Harnack inequality.
These facts will be required in the sequel, and will be discussed
in detail in sec.4. Also, in sec.3, we will discuss a detailed
proof of the existence of a (positive) Green's function for the
Friedrichs extension, again as required in the following.
The majority of the results mentioned are local results.
However, it is important to see them in this global surrounding,
111.0. Introduction 60
where they will have to be applied.

Finally we will be concerned with two normal forms for
expressions of the form (0.1). Both may be achieved by a transfor-
mation of dependent variable, but the second only if the minimal
operator of the expression H is positive definite. This is dis-
cussed in sec.6.
1. 2nd order PDE on manifolds; Weyls lemma; Dirichlet operator.
Let 0 denote a Co-manifold without boundary, of dimension

n > 1, not necessarily compact, but paracompact (and connected).
In fact, for convenience we assume existence of a countable atlas.
On 0 we assume given a positive Co-measure du locally of the ,
form du = Kdx , with K being C' in the local coordiates x. By H

,
we denote the Hilbert space L2(O,dp), with inner product and norm
(1.1) (u,v) = J2uv du , NO = (u,u)1/2 , u,v E H
Also we assume given on 0 a second order formally self-

adjoint strongly elliptic partial differential expression
(1.2) H = -K-18 hikK2 k + q

x] x
with Co-coefficients. In particular hjk denotes a symmetric posi-
tive definite contravariant tensor with real Cm-coefficients:
hjk hkj -k0 >.0

= = , , as E 9 0 ,
and q denotes a scalar real-valued C -function defined over 0

In the next sections we use the summation convention to
always sum from 1 to n over a pair of an upper and a lower index,
in a tensor, denoted by the same symbol. Also we will write local
coordinates with superscript indices x=(x1,x2,...,xn) E In not ,
subscripts, as for a subdomain 0 C In , according to convention.

As in II,1 the expression H induces minimal and maximal
differential operators of H . For the minimal operator H0 with
domain we get
(1.3) (u,H0v) = J0(hJku v Ik + quv)dp = H(u,v) , u,v E dom H0

Ix7 lx
Clearly H0 is hermitian and real, in the sense of I,prop.

hjk
2.6, since q and are real. Accordingly there exist self-ad-
joint extensions of H0 .
III.l. PDE on manifolds 61
In much of the following we will be interested in the case

where the sesqui-linear form H(u,v) of (1.3) is an inner product,
called (u,v)1 In fact if we assume q > 1 in G , then we get
.
(1.4) Ilulll > Hull , for all u E dom HO ,
with the norm Ilulll=((u,u)1)1/2 ,

since the tensor hjk is positive
definite. Condition (1.4) will be crucial in ch.5f, although the
condition q > 1 will be weakened later on (cf. prop.5.2).
For an expression H with (1.4) the Friedrichs extension of
H0 is well defined as restriction of H* to (dom H*) fl H1, with the
completion H1 of dom HO = C3(Q) under the norm Ilulll (cf. the proof
of I, thm.2.7 ). In fact, the norm Ilulll of (1.4) precisely coinci-
des with the norm null of that proof, and the completion H- defi-
ned there will be denoted by H1 here. We know that H1 is naturally

imbedded in H .
In ch.5 we will assume that (1.4) holds, and then denote the
Friedrichs extension by H again. This will be a self-adjoint ope-
rator,satisfying H > 1 We then shall speak of a comparison ope-
.
rator H The space H1 will be called the first Sobolev space of

.
the operator H Clearly H-1 E L(H) has a unique self-adjoint

.
yk=0(-1)k(lk2)(1-H-1)k
positive square root A = H-1/2 = , and
0 < A < 1 . We have (cf. I, prop.2.8)
(1.5) H1 = dom A-1 , (lull = IlAupl , u E H , Hvfl1 = RA-1vll, v E H1.
That is, A is an isometric isomorphism between the Hilbert spaces

H and H1 Generally the restriction of the self-adjoint operator
.
A-1 to dom H0 still is essentially self-adjoint, and A-1dom H is

dense in H , (while the corresponding is not generally true for
H dom H0 = im H0 ).
Note that H1 may be identified as the space of all functions
u E H such that a sequence um E dom H0 exists which is Cauchy in
the sense of H1 and converges to u in H It follows that there .
exists a unique covariant tensor (called the gradient of u , and

_ (u , even though the components are not proper derivatives)
Ix j)
which is p-measurable and satisfies
S2
h0k(u Ixj - um Ixj )(u Ixk - um Ixk )du -> 0 , m
III.1. PDE on manifolds 62
It is clear that u . is the local distribution derivative of u E

Ix]
L2(O,dp) C D'(O) while the above entitles us to speak of the

strong L2-derivative. One may introduce L2-norm and inner product
of gradients just as for covariant tensors in general by setting
(1.6) (pu,pv) hlku v dp , Ilpull = (pu,pu)1/2

Ix3 Ixlc
Returning to the general case, where (1.4) needs not to be

satisfied, we observe that the expression H is elliptic, since we
hjk
assumed the tensor positive definite : We have
(1.7) 0 , for all E 0
at each point x E 0 It was mentioned in II,1 that ellipticity

.
implies hypo-ellipticity, but this will be discussed in general

only in FC31. Since we need (strong) hypo-ellipticity (or at least
Weyl's lemma) in the following we shall offer a short independent
proof of Weyl's lemma here, which is quite similar to the proof
of II, lemma 2.2, in the case of an ODE .
It is clear why a proposition like Weyl's lemma is desira-

ble: Thm.l.3 of ch.II carries over literally, with the same
proofs, to the present case of an elliptic operator on a manifold.
We express this in thm.l.l, below, the proof of which will be
left to the reader, (except for our proof of Weyl's lemma, i.e.,
thm.1.2. )
Theorem 1..1. The defect spaces 4-(H ) are subspaces of

ao **0 4:'a e
HnC (Q) . Moreover, we have H0= Hl H1= H0 (i.e.,"weak" _ ,
"strong"), and the direct decomposition
(1.8) dom H1 = ((dom H** )nC'(0)) $ D+(H0) ® D-(H0)

,
which is orthogonal with respect to the inner product of graph H0.

The self-adjoint extensions A of H0 precisely are given as the
closures of the 'e.s.a.-realizations' A- obtained from an arbi-
trary isometry W: + D , using formulas II,(1.16) and II,(1.17)
Theorem 1.2. If (HO-A)f=g, for any fE dom H0 and gE HnC(2), ,
then we have f E C°10) and Hf-Xf = g , hence f E dom H1

, .
The proof depends on use of an (E.E.Levy-type) local

parametrix of the form
(p(x,x-y))2-n
e(x,y) _ as n > 2 ,
(1.9)
= log p(x,x-y) , as n=2
III.l. PDE on manifolds 63
with p=p(x,z) _ (hjk(x)zizk)1/2 (We may assume A=0, without loss

of generality. ) Note that e(x,y) is defined only
locally, in local coordinates, for x, y E WC 0 , with a
chart 0'. We note the proposition, below, which also will be
useful later on.
Proposition 1.3. For a function m E let
(1.10) v(x) = Je(x,y)$(y)dy
Then we have v E C-(Q') , and
(1.11) Hv(x) = J., y(x,y)$(y)dy , xEQ'
with a positive C'(Q')-function c(x) , and with a function

y(x,y) of the form
(1.12) y(x,y) = y0(x,Y,x-Y) + 2 i(y3(x,y,x-Y))

x
Here the functions y!(x,y,z) j=0,1,.... n, are C,(Q'xO'xjjn*)
,
and homogeneous of degree 2-n in the variable z , as n>2 In .
the case n=2 we have
(1.13) yi(x,y,z) = h1(x,y,z)log p(x,x-y) ,
where the h0 are and homogeneous of degree 0.

Proof. We only consider the case n>2 . The other case n=2 may be
treated similarly. Observe that the function e(x,y) may be written
in the general form
(1.14) e(x,y) = f(x,y,x-y) , f(x,y,z) E C,(O'xQ'x]Rn*) ,
where f is homogeneous in the variable z . (Presently we have
=(hjk(x)zjzk)1-n/2
(1.15) f(x,y,z) ,
independent of y and homogeneous in z of degree 2-n.

For any function g(x,y) = f(x,y,x-y) of the form (1.14) we
have
(1.16) (axi + a j)g(x,y) = fj(x,y,x-Y) n
y
where fj again has all the properties of (1.14) , with the same
homogeneity degree as f . This follows, because a function a(x-y)
is constant in the directions x. yj hence has (2 .+8 .)a=0.

x y
Now a single derivative 2 may be applied under the inte-
gral sign of (1.10), because the differentiated integrand still

is integrable. Applying (1.16), and a partial integration, we get
(1.17) vlx.(x) = (y)dy +
where gj(x,y) = fj(x,y,x-y) has exactly the properties (1.14)

again, (with homogeneity degree 2-n). In particular the partial
integration first may be performed on the integral over Ix-yI>e
where boundary terms may be explicitly evaluated. These boundary
terms tend to zero, as 6+0 , so that no special terms at x=y
from the partial integration appear in (1.17).
Accordingly the process may be iterated, it follows that
indeed all partial derivatives v(a)(x) exists , for x E Q' , and
(1.18) v(a)(x) = I
a<B,BAa
where all ea,$ are of the form (1.14) (degree 2-n) In particular .
we get vEC (Q'). Writing the local expression of H in the form
(1.19) H=Hw+Aw, Hw=-hjk(w)8 8 Aw=a.k(x,w)B 2 k+bj(x)2 +c(x)

xi x
k,
x x x
where wEO' is arbitrary, fixed, and a0k(x,w)=hOk(w)-hOk(x), it
follows that
(1.20)
x .(y)dy
with h0 , h0 again of the form (1.14) (degree 2-n) .
In (1.20) the first integral, at right, called Il, contains

the constant coefficient operator Hx (where x is only a parameter)
and the function e(x,y), depending on x-y only, as long as x is
kept fixed. One may make the linear substitution n=M(y-x) in the
integral Il, with M = ((hjk(x)))1/2, which brings it onto the form
(1.21) I1 = -det((1 Jk(x)))-1/2JInI2 nAPdn ,
with the Euclidean Laplace operator A , and Here,
(1.22) JIn 12-nAW(n)dn=lime+0 I =(2-n)limE+O (n)Inll-ndSn

ITIL In =e
An(InI2-n)=0
using partial integration, and that . Also that
2-n
lim JInI where dSn denotes the surface element
In =e
on the sphere InI=e, and 2v=-2 p, p=InI, the radial derivative.
From (1.22) we conclude that
(1.23) I1 = cn = (n-2)wn-1
In
with the area wn-1 of the sphere InI=l in . The above holds
for n>2. In case of n=2 one finds that cn = c2 = 27. Since the
determinant is C' and >0 this term accounts for the first term
in (1.11).
On the other hand another partial integration in the third
term of (1.20) will remove the derivative from and give an ,
integrand h3 (x,y)¢(y) instead, confirming (1.12). (In view of

Iv'
(1.16) we again may write h l +k with k(x,y,x-y) (degree
-hlYi=
2-n), etc.) This completes the proof of prop.1.3.

Now the proof of thm.1.2, in essence, is a repetition of
the argument used in the proof of II, 1.2.1. Using our parametrix
e(x,y) of (1.9) in place of the fundamental solution e(x,y), we
derive a local integral equation of the form 11,(2.11) for any
pair of functions f,gEH satisfying II,(2.6), with L=H, and with
the present Hilbert space H=L2(c) . Then, if gEC-, the right hand
side is continuous, using the nature of the kernels. Hence
f(x), after correction on a null set, also is continuous. Knowing
that f is continuous, one concludes that the right hand side even
is C1 , so that fEC1, etc. Continuing on, one concludes that fEC
Let us get such formula in detail. Let
(1.24) u(x) =
with a local cut-off function X(x)=l near some x0EO', XEC0(Q')

and a "regularizing kernel" c>0 , where ¢ECD(In).
Assume X=1 in 1"C2' , 0 open, and let xÈO" , and e so small that
supp 4e(x--.)C 0'. Then we may apply (1.11) for
(1.25) Hu(x) = c(x)me(x--x) +
with c(x) of (1.11), and 6(x,y) of the form (1.12) again. In fact,
6 is a sum of X(x)y(x,y) and terms of the form a(x)e(x,y),
b3(x)e (*x,y), with a,b3 E C' (c2') , where a and b3 vanish out-
Ix
side supp x . All these terms may .be written in the form (1.12).
Substituting u and Hu into the (present) equation II,(2.6),
and passing to the limit a+0, we get
(1.26) K(x")C(x")f(x") + JQ,S(x,x")f(x)dux = Jduxg(x)X(x)e(x,x-),
valid for x"E2". This relation is of the general type of II,(2.11)

However, the kernel 6 requires a somewhat refined proceedure.
If g E C_ , then the right hand side is a C--function h(x"),
by prop.1.3. The integral at left is at least Hoelder continuous,
since we find that the function x" + S(.,x") , with values in H
is Hoelder continuous: We get
(1.27) 2/lx"-x"Ie = 0(1) ,
L
for a suitable e>0 and x", x" E Q" , by a calculation. Hence
(1.26) shows that f is Hoelder-continuous as well. To show that
f is even C1 we write (1.26) as
(1.28) (KCf)(x) = h(x) - JS(y,x)f(y)dvy
and investigate the function
(1.29) e(x) = JS(y,x)f(y)dpy
The difference quotient ve = (9(x+e)-e(x))/e may be written as
(1.30) ye = Jvd(y)f(y)duy = Jvd(y)(f(y)-f(x))duy + f(x)Jvd(y)duy
where My) _ (S(y,x+c )-S(y,x))/e . Now we use (1.12) for
(1.31) Jvd(y)duY = J(V6 0K-V6JK My .

ly]
The right hand side limit exists under the integral sign, and
represents a Cm-function of x, since again the derivatives may
be taken over to the function K , using (1.16) The other inte- .
gral in (1.30) has a limit as well, since f is Hoelder conti-

nuous, so that the limit of the integrand is L1 . Thus we get
(1.32) 8 0(x) = k(x)f(x) + JSI (y,x)(f(y)-f(x))dpy ,
xj
confirming that fEC' . In particular, k(x) is a Cm-function.
Knowing that f is Cl we may integrate by parts in (1.32) , for
(1.33) (KCf)Ixi = hlx7 - kf + J S°(Y,x)flx.(y)dy ,

with a function d° of the general for (1.12) again.

It is clear now that this may be iterated, proving that
Kcf , hence f is C_ , q.e.d.
2. Boundary regularity, for the Dirichlet realization.
It will be of interest to note the following variant of

thm.1.2, applying in the case where g is only continuous, not
necessarily smooth.
Corollary 2.1. If g of thm.1.2 is only continuous, then we still

get f E C1(0) , and its first derivatives are Hoelder continuous.
Moreover, if g is Hoelder continuous, then also fEC2(0) follows,
and f is a classical solution of (H-X)f=g .
Indeed, it is clear that (1.26) is valid for general f,gEH

satisfying (HD-X)f=g . If g is only continuous then the right
hand side of (1.26), called V(x) still is C1(0)
, and p ,
IxJ
satisfies a Hoelder condition, by the above arguments, using an

estimate like (1.27) for the first derivatives of e(x,x-) .
Moreover, if g is Hoelder continuous, then the conclusion leading

to (1.32) may be applied to integrals representing i . It fol-
1.
lows then that also the derivatives 1 exist and are conti-
lxj x
nuous. This shows that, in the first case, f will be C1 , and
that (1.33) still holds. Also the first derivatives of f then
still will be Hoelder continuous. For Hoelder continuous g thus
we may iterate once more, getting that also fEC2 , q.e.d.
As in 11.2 we next try some boundary application of our
parametrix e(x,y) Let us investigate the Friedrichs extension
.
H of the minimal operator H0 in the special case where we have

only regular boundary points. Similarly as in the 1-dimensional case
of 11,2 it may occur that the manifold 0 is a subdomain with
smooth boundary 30 of another manifold R^ , and that the triple
{0,du,H} extends to a triple {0^,dp^,H^} on 0^ satisfying our
general assumptions. (That is, 30 is an (n-l)-dimensional Co-sub-
manifold of 0^ , and we have du = dpÎ0 , H = H-JQ .) In this
case we will say that 0 has the regular boundary 30 . Moreover,
if even OU30 is a compact subset of Q^ then we will speak of
,
an expression H (or a triple) with regular boundary, or we will

111.2. Boundary regularity 68
say that all boundary points of c are regular.

In the case of a triple with regular boundary one defines
the Dirichlet operator Hd J HO by setting
(2.1) dom Hd= {uEC'(SZUM): u=O on 8O}, Hdu=Hu , for uEdom Hd
Specifically the condition "u=0 at x(=-22" is referred to as the

Dirichlet (boundary) condition.
Theorem 2.2. The Dirichlet operator Hd , for an expression,H with

regular boundary, is an e.s.a.-realization of H , and its closure
**
H = Hd is the Friedrichs extension of the minimal operator HO
Proof. First we note that Hd indeed is a restriction of the Frie-
drichs extension H of HO For clearly we have HdCHd*
. (i.e.,
Hd is hermitian), since a partial integration confirms that
(2.2) (u,Hdv) _ (u,v)1 = (Hdu,v) , for all u,v E dom Hd
(The boundary terms vanish, since u and v vanish at 2D .) Also,

HOCHd , trivially, hence Hd*CHO* so that we get HdCH O*
,
For any u E dom Hd it is easy to construct a sequence ujE dom HO

with nu-u101-0 , as j- , since only the first derivatives of u
must be modified, while the 1-st Sobolev norm 11.111 contains only
first derivatives as well. In details, for a boundary chart c',
{xn>0} X(xn)EC,(J)
thought of as an open set in &+ _ , let , X?0
X=0 near xn=0 , X=l for xn>1 , and let X5(xn) = X(Jxn)
uj(x) = u(x)Xj(xn) . Then u-uj = u(1-X].) has support in O<xn<1/j.
If u has compact support in c' , then as 54 . The
same is true for IIO(u-uj)U , whenever u=0 at xn= 0, since we get
(u-uj)Ixk=(1-Xj)ulxk as k<n , while (u-uj)Ixn=(1-Xj)ulxn-uXjlxn
where vj=uXl n has support in O<xn<1/j, and is bounded in j and x,

X
due to u(x)Xjlxn = ju(x)X'(jxn) , where u = 0(xn)=0(1/j) in
supp X'(1xn) . Accordingly we conclude that I1u-uj1114-0 , since

the coefficients of H are bounded at the regular boundary. For
a general u E dom Hd one uses a partition of unity, corresponding
to a finite atlas of the compact manifold with boundary QU30 ,
for a similar result. Therefore indeed we get dom HdCHln(dom Hg),

the domain of the Friedrichs extension, and the above also shows
that Hdu=Hu for uEdom H .
Our theorem is proven if we show that Hd is essentially
self-adjoint. For then it is clear that its (self-adjoint) closure

**
Hd , as a restriction of the Friedrichs extension, must coincide
with the Friedrichs extension.
Since we know that Hd>1 , it suffices to show that im Hd is
dense. Thus, let fEfl satisfy
(2.3) (f,Hu) = 0 for all u E dom Hd .
By thm.1.2 we then get fEC_(O) . We want to show, that, in addi-

tion we get fEC°O()U90), and f=0 on M. If this can be established
dfp2=11f11i=(f,Hdf)=0,
then (2.2) for u=v=f implies that hence f=0,
so that indeed im Hd is dense.
Thus we again focus on a boundary chart n' as above. After
another change of coordinates one may assume that the co-normal
direction at xn=0 councides with the x n_direction. In other words,
this means that we have hnl(x)= 0 , j=1,...,n-1, as xn=0
Let us again write the expression H in the form
(2.4) H = hjk2 3 k + b32 + c

x3 x
x3
for simplicity, and let O" denote the 'doubled' chart 0' given by
(2.5) o" = 0' U O2, , O' _ x=(xl,...,xn-l,-xn)EO'} .
In 0" we define the doubled measure dp"=K"dx , with the even
extension K" of the function K to O" (i.e., K"(x)=K(x) , xE92" ).

We drop the notation K" , du" in favour of the old notation K, du.
Let an extension of H be defined by extending hnn , hjk , bi , c
for j,k<n, as even functions, but hnj = h:n , j<n , and bn as odd
functions. Also let the function f of (2.3) be extended as an
odd function, and assume u E C-(O") . We claim that then (2.3)
with integrals over 0" still holds for f an all such u . Indeed,
one may decompose
(2.6) u=ue+uo , ue(x)=(u(x)+u(x))/2 , u0(x)=(u(x)-u(x))/2 .
Note that the extended H maps even functions into even functions
and odd functions into odd functions. Since (the extended) f is
odd, we get (f,Hue)0 = J0" dpTHue = 0 , as an integral over an
odd function with symmetric range. On the other hand, (f,Hu0)0

= 2(f,Huo)0, = 0 as well, since (the restriction to 0' of) u0
is a function in dom Hd (A continuous odd function vanishes at

xn=0 ). Accordingly, the extended function f satisfies a relation
of the form II,(2.6) with the extended expression H .
Note that uEC0'(0") is not required, as long as u0 is conti-

nuous at zero, and u0In' is
Thus it appears that the proof of thm.1.2 will give fEC
on the part of an bordering the chart 0' Since f was defined as
.
odd function, it then follows that f=O on an , or, fe dom Hd .
However, it should be noticed that the assumptions of thm.

1.2 are not satisfied, since the coefficients of the extended H
and K ) are not C_ accross xn = 0 . Instead all leading coeffi-
cients hjk are continuous and C'(01) , since hJn=O at xn=0. It is
easy to make at least hnn E C1(H") , by the change of variable
y - _ (y 1 ,...,n-11
y _ (x,...,n-1
) x ) = x"
(2.7)
yn = xn(1 + xn.hnn(x-,0)/(4hnn n(x-,0))) ,
Ix
which takes H' onto a similar region in y-space, and makes the
normal derivative of hnn zero at xn=0 The lower order coeffi-
.
cients'c and b. , j<n also are continuous and piecewise Cl , while

7
bn has a jump accross xn = 0 Furthermore we have g = A = 0, in
.
the notation of thm.1.2.

Going over the proof of prop.l.3 we notice that v(x) of
(1.10) no longer is C-(0"), but will be C-(0') again, and conti-
nuous accross xn=O Therefore we again may substitute u of (1.24)
into (2.3) . We still get (1.25), as long as x- is not on an ,
for sufficiently small c and then (1.26) follows as well, with

,
g = 0 so that we have
,
(2.8) f(x) = jnit y(x,y,x-y)f(y)dy , for xn90 , xE 0-
with a sufficiently small neighbourhood 2- of a boundary point x0.

O"XQ"Xln*
Here y(x,y,z) is continuous and of compact support in ,
and homogeneous of degree 2-n in z , and its restrictions to xn>O

and to xn<0 are C_ . In this way we conclude that indeed f is
continuous, and has continuous one-sided derivatives of all orders
at xn=0 . Thus indeed f E dom Hd , and the proof is complete.
Note that a small modification of the above proof gives the
following.
Corollary 2.3. Under the assumptions of thm.2.2, if fECW(0U3c2) ,

relative to the imbedding 1-O^ , and if (H-X)u = (Hd -a)u = f

Hd
for some XEU , and u E dom H = dom then we have uE dom Hd
,
and (Hd-A)u=f
Indeed, under these assumptions the interior regularity of
u is a consequence of thm.1.2, while, at the boundary, the argu-
ment around (2.3) may be repeated, giving that also u E C_ at
a boundary point. Thus cor.2.3 follows.
3. Compactness of the resolvent of the Friedrichs extension.
This section discusses various compactness criteria, for

the resolvent of a comparison operator, as well as for the inverse
square root A=H-1"2 and also, more generally, for aA
, with a
function a(x)EC_(O) Its main applications will be found in V,3,
.
and V,4,6 , where compactness of commutators, and certain asym-

ptotic expansions are derived and applied. On the other hand, com-
pactness of the resolvent (H-a)-1 = R(a) will be essential in the
discussions of the present chapter, below, involving the Dirichlet
operator Hd .
In each case we consider the (self-adjoint) Friedrichs

extension of the minimal operator H0 for an expression of the ,
form (1.2). Note that compactness of the resolvent R(a) is an

immediate consequence of the compactness of A , since we get
R(a) _ (H-a)-1 = A2(1-XA2)-1, for points AERs(H), where (1-XA2)-1
E L(H) Since H=Hd for a manifold Q with regular boundary, and

.
0U30 compact, we get compactness of the resolvent of the Dirichlet

operator in cor.3.9.
The requirements on a for compactness of aA will be diffe-
rent near the various 'infinities of 0' For a clear formulation .
of these facts we introduce the concept of subextending triple,

below.
For technical reasons, for a while, we consider differential
expressions of the (formally) more general form
(3.1) L = q - K a Khjk3 , du = K dx
ks
xi x
with a C-(c2)-function $ > 0 and with all other coefficients q
,
K ,
hjk as in (1.2). Such an expression may be brought into the
form (1.2) : One has
(3.2) L = -K-1a K82h3k3 k + q + q0 , q0 = - a (h0kKs )

k
xi x 1xi ix
111.3. Compactness of the resolvent 72
Indeed, we get (u,Lv) (u,qv) + (4(Bu),V(Bv)) u v E dom HO , , ,
where we write V(Bu) = $Vu + u48 , and apply a partial integration
to the term (BVB,uVv+vVu) = (BOB,V(uv)), from the second inner
product, to free the term (uv) from its gradient. For more detail,
cf. the derivation of formula (6.6), below.
If we assume that
(3.3) (u,Lu) > (u,u) , u E
then the facts discussed for H in section 1 apply to L as well.

We get a well defined Friedrichs extension L of the minimal ope-
rator LO , and its inverse positive square root L-112 is a bounded
self-adjoint operator of H The space L-1/2C0(St) is dense in H
.
and L112 is essentially self-adjoint in dom HO

Let us adopt the convention to write
(3.4) limx y(x) = Y0 , (in S2 ) ,
where y(x) is a function over 0 , (with values in a space X) , and

YO E X if for every neighbourhood N of YO a compact set K C S2
,
can be found such that y(x) E N whenever x E 0 is outside K We .
are not excluding the case of a compact manifold ) Then condi- .
tion (3.4) is void: It is true for every function y(x) , limit y0°
and space X whatsoever. In particular condition (3.5), below, is
generally true whenever S2 is compact.
Theorem 3.1. Let the expression (3.1) satisfy (3.3) and
(3.5) limxq(x) = - , (in S2 ) .
Then L 112 is a compact operator of H = L2(c,dp)

The proof is rather technical,and will be prepared by a
series of propositions. Let w,X E C'(l) be such that
(3.6) 0<w,X<l , w2+X2=1 , w(t) =1 , t < 0 , w(t)=0 , t>l.
For 0 < A < - define the two functions
(3.7) XA(x) = w(q(x)-A) , uA(x) = X(q(x)-A)
It follows from (3.5) that aA has compact support.In fact, we have
(3.8) q < A+ 1 in supp aA q > A in supp

9 uA
Proposition 3.2. For every A > 0 there exists a constant cA with
(3.9) (u,Lu) > 7 IIUAull2 + IIV(AABu)112 - cAIIXABull , u E C0(52)
Proof. We get
(u,Lu) = J9<A+l AA2(hjk(Bu) (Bu) Ik + gIuI2)dp

IXJ
x
+ J9>A 11A2(hjk(Bu) (Bu)Ik + gIul2)dp

Ix] x
The second term at right is bounded below by AUPAuO2 . For the
first term we use (3.2), with B replaced by XA , to obtain, with
a=XA , u=VA ,
J82
JaA2h3k(Bu) (Bu) kdu = Uy(AA6u)112+ K(Khjka klul2dp.
Ix7 IxJ
Ix Ix
Here we estimate the last term at right by
cA' llABull Hull , cA' = supxES2 IS(Khjkx ) kl

Ixi Ix
Note that Hull = II(A2+P2)uU < IIX2ull + 1u2ull < Ilaull + Uuull,so that
(u,Lu) > AUP U112 + 114(aABu)U2 + J(ga2+62K(KhjkX ) k)IuI2du

A Ixi ix
with the last term estimated below by
-cA"IIX uI12 -°A1llasull2 _ CA'llxBullUuuU, cA" =sup{-q/B2:xEsupp AA
cA'= supp XA}
Using that HABuflluuf< Zlluull2+ 1/(2c)UXBuU2 we get (3.9),setting
e=A/cA , with cA = cA+ cA + cA2/(2A) . Q.E.D.

In the following let {S2i ) be a locally finite atlas of S2 , and
let w3.2 be a corresponding partition of unity, such that supp w,
C S2. , wi E C-(Q) , wj > 0 , w!2 = 1 , on all of S2 . We assume
without loss of generality that each S23 . is a subset of a larger

chart in which it has compact closure.
Proposition 3.3. For every A > 0 there exists a constant dA with
(3.10) (u,Lu) > 211uAull2+ Z V (XABwiu)U2- dA IIXABul2, u E C,(0).

Proof. In view of (3.9) it suffices to show that
(3.11) 11V(AARu)n2 > lj114(XARwju)H2 - cAIAARull2 u E C0(S2)
Write IIV(XASu)112 = IjHwjo(AASu)112 , and apply the conversion
(3.1) -> (3.2) , for S = wj , to bring wj under the gradient.Since

aA has compact support only finitely many of the summands do not
vanish,and the corresponding functions q0 of (3.2) are all boun-
ded. Hence (3.11) follows, q.e.d.
Q={x(=-]kn:O<xj<1,j=1,....
For a moment consider the cube n) ,
Ipvl2 = j=llaxjv12 ),
with the Riemmannian metric of In , (i.e.,
and the functions sinafrx = Rn sin 7rajxj , a E Zn As a consequence

1
of Parsevals relation for the Fourier sine and Fourier cosine
series we have
(3.12) JQIvj2dx = flval2 , JQIOv,2dx = j(a,a)jva12 , vEC0,(S2),
with the sums taken over all multi-indices a with aj > 1 , and
(a,a) = Zjaj2 , and with the Fourier coefficients
(3.13) va = 2n1Q v(x) sin anx dx
Proposition 3.4. For every C > 0 there exists an integer N = N(C),
such that v E C'(Q) , va = JQsinarx v(x)dx = 0 for jaj<N implies
JQIVv12
(3.14) dx > C JQIv12 dx .
Proof. It is clear that (3.14) is an immediate consequence of

(3.12) , setting N(C) = Max {Ial (a,a) < C ) : .
Returning to our general case we next prove the following.

Proposition 3.5. For every B > 0 there exists finitely many func-
tions zj E C_ (S2) , j = 1,...,M , M = M(B) , such that
u E C0'(S2) , (zj,u) = 0 1,...,M
(u,Lu) > Bllull 2

Proof. Apply proposition 3.3 , with A = 2B , and assume without

loss of generality that each of the compact sets supp wj relates
to a subset of the cube Q under the coordinates of the correspon-
ding chart Q. . In that compact set we also get a positive bound
below for the positive definite matrix function ((h3k(x))). Thus
AO(AAwju)112 > (B + dA)IIAAw.u112
f AAwj u sinalix dx = 0 , I a I < N. .
Since only finitely many indices j are involved (i.e.,

those with supp wj n supp aA 1 0 (3.18) translates into (3.15),
,
and (3.16) follows by combining (3.10) and (3.17), q.e.d.
L-1/2w
Proof of theorem 3.1. Setting u = in (3.15) , (3.16) we get
(3.19) IIL-1/2 w112<B-1]wV2, as wELl/2C'(c2), (L-1/2zj,w)=0, j=l,..,M.
Since L1/2C.(N) is dense in H, the estimate (3.19) follows for all

w E H orthogonal to L-1/2z. 1,...,M . Since B may be choosen
L-1/2,
arbitrarily large this implies compactness of by Rellich's
criterion ([C1], A1,4) : We get
IIL-1/2ITBI <
B-1/2 , TB = (L-1/2{zl,...,zM})1
(3.20)
where TB has finite codimension. This completes the proof.
Corollary 3.6. Under the assumptions of theorem 3.1, let " C 0

be a subdomain of Then the statement of the theorem holds for
.
the restriction of the differential expression L to Q" as well :
If L" is the Friedrichs extension of the minimal operator genera-

L"-1/2
ted by L in H" = L2(Q",du) then E K(H")
, Moreover, the .
same still holds for any expression
(3.21) L" = q"-(S/K)2 Kh"3k3

k
xj x
of the form (3.1) with C--coefficients, defined only in SZ" , and
satisfying
(3.22) q"> q , ((h" 3k)) > ((h 3k)) in c" .
Proof. For u E setting (.,.)Q" = 1 ..du , we get

(3.23) (u,L-u),,- > (u,Lu) > B(u,u) = B(u,u)0- ,
whenever (zj,u) = 0 , as a consequence of prop.3.5. Here we were

extending u to 0 by setting it zero outside Q- Notice that the .
conditions (zj,u)=0 may be written as (z-i,u),- = 0 , with z-i _

ziIsi` E H- In other words, we conclude that prop.3.5 is valid
.
also for L- in the subdomain Sl- C n, except that now the functions
z`3 . no longer are C'(c-), but only in C'(Q)1H` . This weakened
condition still allows its use for the argument of the proof of
thm.3.1 , because we still get (z-j,L--1/2w) _ (L--1/2z-j,w)
L--1/2z-.
where E H- q.e.d. ,
We now return to comparison triples. Let triples {c,du,H},

and {c^,dp^,H^} be given, both satisfying the basic assumptions
of sec.l. We write
{2,du,H} <c) {c^,du^,H^}
(or {0^,du^,H^} (c> {Q,dp,H} ) if
(i) 2 is an (open) subdomain of Q^ , (ii) du = dµÎc

(3.24)
(iii) ((hjk(x))) > ((hÔk(x))), q(x) > q^(x), for all x E Q.
In this case we will say that the triple {Q^,du^,H^} sub-extends

the triple {P,dp,H} .
The main point for introducing this more complicated notion

is the fact that many such singular problems display a very dif-
ferent behaviour near a point x0 E 2Q C 2^ than at infinity (of
0^ ) . For example, as shown by cor.3.6, we do not have to require
L-1/2
q - near such a point x0 to get compactness of
,
This will still become more evident in V,3, where we inves-

tigate compactness of commutators of the generators of a compa-
rison algebra C .
Here it may become important that a proper change of depen-

dent and independent variable may be necessary, before Q^ can be
defined (cf. the examples in V,4 and V,5).
Theorem 3.7. For a comparison triple {c1,du,H} let there exist a

triple {0^,dp^,H^} (c> {c,dp,H} satisfying
(3.25) q^(x) > 0 for all x e 2^\K where K C Y^ is compact .

Then, for every function a E C(Q) with a = i.e., with
(3.26) a2 = 0(q^) , and 0 , (in 2-
aH-1/2
the operator aA = is compact in H = L2(Q,dp).
Proof. Assume first a has a real extension 0<a^(=- and set
L=a-1Ha-1 a^-1H'a"-1
. Consider L^= , and note that L^+ Y0 , with
a sufficiently large positive constant 0 satisfies the assump-
tions of thm.3.1. Also the expression L+y satisfies the assump-
0
tions of cor.3.6. Hence (L+yo)-1/2, (L+y0)1 are compact, and L+y0
(therefore also L) has discrete spectrum. But A-1 = aH-1a is

well defined, hence 0 cannot be eigenvalue of L, and L-1 also must
be compact. But L-1= (aA)(aA)*, so that aA is compact, since an
operator A E L(H) is compact if and only if AAA is compact.
For a general aEC(2) we get UaAuU<UbAuf , where b = cJQ is
the restriction to 2 of O<cECT(Y^) constructed according to lemma
A.1, with suitably modified within K By lemma 3.10, .
below, it thus suffices to establish compactness of bA. This was

just done above. The proof is complete.
Corollary 3.8. Under the assumptions of thm.3.7 we also have

JaICAC E K(H) , for all O<e<l .
This corollary is a straight application of I,lemma 5.1.
Corollary 3.9. The operators Ae, O<e<l, are compact whenever 0 is

compact. Moreover, then H has discrete spectrum,i.e., its reso2vent
(H-a)-lis compact for all XERs(H)=T\Z, where Z=Sp(H) consists of
countably many eigenvalues of finite multiplicity without cluster.
The same holds if {Q,dp,H1<c){SZ;du^,H^), with 2U3SI compact in V.
Proof. Under these assumptions the conditions of thm.3.7 hold for
a=1 , evidently, q.e.d.
Lemma 3.10. Given two bounded operators A , B E L(H) . If A is

compact and B satisfies the inequality
(3.27) IIBuf < UAuU for all u E H ,
then B also is compact.

Proof. By Rellich's criterion (cf.[C1],App.AI,4) there exists a
closed subspace S£ C H with IIAISEU < e , for every e>0 . But then
(3.27) implies that
(3.28) IlBull < IlAull < e Iluh for all u E SE
Or, IBIS EI < e , so that B must be compact, q.e.d.
4. A Greens function for H and Hd, and a mean-value inequality.
For later application it will be useful to obtain a 'clean'

fundamental solution for our expression H , and, moreover, a
Greens function for our Friedrichs extension H as well as for
Dirichlet operator Hd of sec.2.
First we work locally, in a given fixed compact subset
B C 0' of a chart S1', with nonvoid connected Bint= , B.
Let n0 denote the (positive) distance (in In ) from B to W.

It is practical to consider the local expression H of the
form (1.2) redefined outside S in such a way that it equals 1-t in
n,compl , while still satisfying all other conditions of H . For
example one may choose a cutoff function X E 0<X<1

X=l near B , and then define the measure and expression
(4.1) dp'=K'dx , H' = q'-K-12 K'h'jk8 k e
xj x
over In , where
Xhjk+(1-X)Sjk
(4.2) K'=XK+(l-X) , q'=Xq+(1-X) , h'Jk =
For a while (until thm.4.4) we will entirely focus on this

expression over In . Hence we return to our old notation, and
will assume that H and du coincide with 1-0 and dx outside some
compact subset B of In
The 'parametrix function' e(x,y) of (1.9) again will play
a central part. Note that it now is defined for x,yEin. Moreover,
it is clear that e(x,y) _ (x-y12-n (or = log Ix-yl , for n=2)
as In fact, we again will depart from (1.26), which
xEScompl..
will have to be iterated. Let us reinterpret these facts as fol-

lows.
Proposition 4.1. For x,y E In define the function
(4.3) h=det ((hjk)),
with the constant cn of (1.23), and with X(x)?0

denoting a cutoff function, XEC0(IxI<2), X=l at IxI<l . There
111.4. Greens function 79
exist functions e3E C'(Inxinxin*) , j=O,...,n , with e3(x,y,z)

homogeneous of degree 2-n in z and support contained in the set
{(x,y)Einxin:1x-y1<2}xin* such that, for every solution
,
u of Hu=v , with u,vEC_(1n), we have the relation
(4.4) u(x) - J dye(x,y)u(y) = J g(x,y)v(y)by , x E In
with
(4.5) O(x,Y) = 00(x,Yex-Y) - (03(x,y,x-Y)) .
IY3
Here the functions e3 depend only on the choice of X
and the coefficients of H in In but not on the functions u, v.
,
For the proof one must look at the derivation of (1.26)

again, and just confirm that the functions O3 are as stated. A
pair of Co-functions u,v satisfying Hu=v may be put into the
places of f anf g of (1.26), of course, once we interpret II,(2.6)
with 2=St' =,n .
Proposition 4.2. Let 0 and G denote the integral operators on In,

with kernel e(x,y) and g(x,y) , respectively so that (4.4) assu- ,
mes the form u-Ou=Gv. Then the iterates Gk=OkG , k=l,...,n , and
0n+1,
all are well defined integral operators over In with kernels

gk(x,y) and On+l(x,y), respectively. In particular, 0n+l' gn-l'
gn are C(inxin), while F,k EC(inx]kn\{x=y}), for k=O,l,...,n-2.
Moreover, we have gk = 0((x-y)-n+k+2) , as O<k<n-2 , while gn-2
is 0((x-y)-n) , for all n>0 . If ri denotes the integral operator
with kernel gly.(x,y) , then the kernels of 0 , j,k=1,...,n,
are C(inxin\{x=y}) and 0((x-y)-n+k+l), as k<n-1, and 0((x-y)-n),

for all n>O, as k=n-1, and are C(inxin), for k=n. All the above
kernels have support in {lx-yl<2(n+l)} . In particular we may ite-
rate (4.4) n-times for (4.6), below, where again uEC-(in), v=Hu:
(4.6) u(x) _ (0n+lu)(x) + Ik=O(Gkv)(x) , for all xE8
The proof is an immediate consequence of prop.4.6, below.
Proposition 4.3. Given two kernels kj(x,y) continuous over ,
gnxRn\{x=y}
, and with support contained in lx-yl<c for some con-
stant c. Assume that, with constants vj O<vj<n , we have ,
_V.
(4.7) kj = O(lx-yl 3) , Kju(x) = Jkj(x,y)u(y)dy , uEC0(in)
Then the operator K3=K1K2 is an integral operator of the same form

-v
again, with kernel k3 = O(Ix-yl 3) , whenever v3=-n+v1+v2 > 0
while the kernel k3 is continuous at x=y if v3 < 0 and k3 still
is O(Ix-yj-n) , for every n>0 , if v3=0
Proof. If v3<0 then the integral
(4.8) k3(x,y) = Jk1(x,z)k2(z,y)dz
exists for all x,y including x=y . It is easily seen that the
,
product ax (z) = k1(x,z)k2(z,y) defines a continuous family of

functions in L1(Rn) , depending on x,y , hence k3 also is conti-
nuous in x and y . If v3>0 the same still is true for x#y , but
no longer at x=y . On the other hand, as v3>0 one may estimate
nlx-zl-vljz-yl-v2dz
(4.9) k3(x,y) < c f .
R
The right hand side integral I(x,y) exists, since the integrand is
O(Izl-vl-v2) for large IzI, with vl+v2>n , while the local sin-
gularities at x and y have order -v1.>-n . Moreover,
(4.10) I(x,y) = a=Ix-yl , e1=(1,0,...,0),
as shown by a substitution of integration variable involving a

suitable translation and rotation. Then another substitution
n-v -v
1
z = aw shows that I(x,y) = a 2I(0,eI) , where I(0,e1) is a
-v
constant. This confirms the estimate k3 = O(jx-yj3) . Finally,
if v3=0 we still get the estimate (4.9) with vl replaced by v1+n,
for sufficiently small n>0 (or, similarly, for v2, if more conve-
nient). Then we get (4.10) with 0=v3 replaced by v3+n=n>0 q.e.d. ,
Note that prop.4.2 suggests that the function
(4.11) g'(x,y) = lk=0 gk(x,Y) = g(x,y) + Ik=1 gk(x,y)
is an improved parametrix of the expression H , because the cor-

responding integral operator G- satisfies
(4.12) u - 0n+lu = G`Hu , for all u Ei C'(Rn) ,
which seems better than (4.4) , since the kernel 8n+l of 0n+1 is
continuous at x=y . In order to find a fundamental solution of H
we seek to add another term f- to g-, such that g '=g-+f- satisfies
(4.13) u = G°Hu , for all u E
Subtracting (4.12) from (4.13) we find that f` must satisfy
(4.14) On+lu=JOn+1(x,y)u(y)dy=(G°-G-)Hu=Jf-(x,y)Hu(y)dy, uECQ(In).
Let us require (4.14) not only for uECO , but even for all
uE dom H (the Friedrichs extension of H0) . Also introduce the
functions (of y) bx(y)=On+1(x,y)/K(y) , and ax(y)=f'(x,y)/K(y).
Observe that bx E C0(gn) C H = L2(gn) , and require that ax E H
Under these conditions (4.14) takes the form
(4.15) (bx,u) = (ax,Hu) , for all u E dom H ,
so that ax E dom H and Hax = bx follows, since the Friedrichs

extension is self-adjoint. In particular, of course, we also get
H*ax=bx , so that cor.2.1 applies, since bx E CO(In). Accordingly
ax at least is C1 , and has its first derivatives Hoelder contin-
H-1
uous. Moreover, since E L(H) , and bx E C(In,H) as x varies, ,
we also get ax E C(ln,H)

Now let us carry this back to the function g° = g-+f-
Introduce the function cx(y) = g4(x,y)/K(y) = g-(x,y)/K(y)+ax(y)
Clearly we still get cx E C(1n,H). Combining (4.12) and (4.15) get
In
(4.16) u(x) = (cx,Hu) , for all u E dom H and all x E
Applying (4.16) for u E C_(]n\{x}) , for fixed x , one

concludes that cx is a solution of H0 xcx=0, with the minimal
operator H of the expression H in the punctured In Thus .
O'x cxEC-(I
thm.1.2 implies that we have \{x}) , and Hcx=O , as y#x
with H acting on the y-variable.
Finally let us multiply (4.16) by v(x) E CD(Ln) and inte- ,
gn Note that the integration may be interchan-

grate dux over, .
ged with the inner product, since the family {cx:xEIn} is C(In,H).
It follows that
(4.17) (v,u) _ (Jv(x)cxdux,Hu) , for all vEC_ , uE dom H
Accordingly,
(4.18) w=Jv(x)cxdpx E dom H , and Hw = v , for all v E
But the Friedrichs extension H is self-adjoint and >1 , and

H-I
has a well defined bounded hermitian inverse E L(H) . Applying
H-1
to (4.18) we conclude that
(4.19) H-1v(y)=Jcx(y)v(x)dp =Jg (x,y)(K(x)/K(y))v(x)dx, vEC-(In).
Then the hermitian symmetry of H-1 implies that
(4.20) g (x,y)K(x) = g°(y,x)K(y) , x,yE In , x#y
We have proven the following result in the special case of

a triple {In,dp,H) , coinciding with the Laplace triple outside
a compact set. In case of a general manifold 0 we relate the
integral operator to the global measure du=Kdx again. Also we
denote the Green's kernel by g(x,y), not g°(x,y) ignoring the ,
earlier meaning of g(x,y) .
Theorem 4.4. For any triple {c2,du,H} satisfying (1.4) the Frie-
drichs extension has the form of an integral operator
(4.21) H-1u(x) = Jg(x,y)u(y)duy , u E H = L2(,,n)
with the 'Green's function' g(x,y) (E C-(HxQ\{x=y}) and with

g(x,.), g(.,x) E C(H,H). Also, g(x,y) = g(y,x), x,y E H, and, in
local coordinates, on a chart n', g has an expansion of the form
(4.22) g(x,y) = R7)e(x,y) + Ek=1gk(x,y) , x,yEH',
where e(x,y) denotes the parametrix of (1.9) and X(x) denotes ,
an even cutoff function (X(-x)=X(x) X(=-C0 , X=1 near 0

, X?0). ,
Also, gk(x,y) = 0((x-y)k-n+2), k=l,...,n-3, gn-2(x,y)=0((x-y)-n),
for all n>0 gn-1 E C(H'xH') , while gk , k=1,...,n-2, have

,
support contained in a ball Ix-yl<c0 Furthermore, the first

.
y-derivatives of the gk(x,y) exists and are continuous for y#x

The kernels gk (x,y) are 0((x-y)-n+k+1) as k=l,...,n-2 and ,
,
lyj
0((x-y)-n) for every n>0 for k=n-1 Finally, for any x0EH , and
.
a cutoff function X1EC'(H) 1=0 near x0 l-X1EC0_(H) , the func-

,
tion *(y) = X1(y)g(x0,y) is contained in C.fldom H

For a proof of thm.4.4 in the general case observe that the
function g(x,y) is uniquely determined by the relation
(4.23) u(x) = Jg(x,y)(Hu)(y)dpy , for all u E dom H
For xEH' , a chart of H , let 8 and H' be as initially in this

section. Let g'(x,y) be the Greens function of (the Friedrichs

extension of) H' in In , and let the cutoff function X have
support in 0' and be = 1 in B . Then we also get
(4.24) u(x) = Jg'(x,y)(H(Xu))(y)duy , uE C'(Q)
Subtracting (4.23) and (4.24) we get
(4.25)J(g(x,y)-X(y)g'(x,y))(Hu)(y)duy=J([X,H]g'(x,y))u(y)duy,uEC-0.
In particular we note that the commutator [H,X] acts on the y-var-

iable, and is a first order expression with support in 1'\B ,
so that the function [X,H]g'(x,y) = qx(y) is C-(c') , (as a func-

tion of y only) . Setting px(y) = g(x,y) - X(y)g'(x,y) , (4.25)
assumes the form
(4.26) (px,Hu) _ (qx,u) , for all u E C0'(S2)
We even request that (4.26) holds for all u E dom H . This deter-
mines a Cm-function p E dom H such that Hpx=qx in the sense of
x
ordinary differentiation. Then we define g(x,y)=px(y)+X(y)g'(x,y).
At least for all uE dom H with HuEC' we also get uEC_(G) , and
the right hand side of (4.25) may be written as (g'(x,.)[H,X]u)
so that (4.23) follows.
For the last statement observe that
(4.27) (i,Hv) _ (gx0,H(X1v)) + ([H,X1]gx0,v) , vGCCOfldom H ,
with gx (y) = g(x0,y) , where z = [H,X1]gx E HnC" , since the

0 0
commutator [H,X1] vanishes near x0 . Note that the first term at
right of (4.27) equals GH(X1v)(x0) = 0. Hence the selfadjointness
of H and essential self-adjointness of its restriction to
C'fldom H implies that Edom HnC' as stated. Therefore the proof
,
of thm.4.4 is complete.
Next we will investigate the Greens function g(x,y) near
its singularity. Given some fixed x0Ec let us ask for the surface
(4.28) Cx001={yEc:
g(x0,y)=6(x0)n2-n},
(or g=dlog n, as n=2) ,
with 6 = 6(x0) = cnVET-x T , as in (4.22) , where the constant n

0
is assumed sufficiently large. More precisely, if g in (4.28) is

replaced by g0 , then the surface (4.28) will be an ellipsoid, in
local coordinates. For g one expects a smooth surface C close
x0,n
to that ellipsoid, and with normal close to the normal of the

ellipsoid, as n is small. We leave open the possibility that other
points y exist in the set (4.28). These will be ignored, i.e., the
statement "y sufficiently close to x0 should be added in (4.28). if
Let us assume without loss of generality that, at the point

x0, we have hjk(x0)=6jk=0, as jtl, =1, as j=1, and that x0=0
so that the equation of (4.28) assumes the form
g(x0,y)=dn2-n
,
(4.29) IYI2-n + Y(Y) = .2-n
where we have
0(ly15/2-n)
(4.30) Y(y) = , VY(Y) =

o(IY13/2-n)
.
For n=2 the first term in (4.29) must be replaced by -loglyl, and
the second term is continuous with derivatives satisfying (4.30).
We again focus on the case n>2, with n=2 to be treated similarly.
It follows that for small IyI the first term at left of
equation (4.29) is large as compared to the second term. Along
each ray ty0 , with 1y01=l , 0<t<- , the equation (4.29) assumes
the form
n,
(4.31) ¢(t,y0)=t2 n+Y(ty0)=n2 +YO-VY(tyo)<O,
as t is small. In particular, also in the expression for atgt,y0)

0(t3/2-n),
the first term dominates for small t, since y00y(ty0) =
near t=0 , by (4.30). Let
a0(t)=sup{Tn-5/2IY(TYO)l, Tn-3/21yoVy(Ty0)l/(n-2):
(4.32) T<t}
and define a+(t) = t-1/2+a0(t). Then t1/2a+(t)-*l , as t-0, and
t5/2-na
(t) < (t,y0) < t5/2-nX+(t) ,
(4.33)
(n-2)t3/2-na+(t) (n-2)t3/2-nX-(t)
< .
Choose a t0>0 such that I1-t1/2a+(t)I<1/2 , for 0<t<t0

Then the function 0 decreases with t, for every fixed y0 . Hence
there will be precisely one root tE(O,t ) of (4.31) for each y
0
O<n<(2/3)1/(n-2)t0.
whenever n2-n>(3/2)t n, i.e., Moreover, then
(4.33) implies that (1/2)T2-n<

(t,y0) = n2-n < (3/2)t2-n . Or,
(1/2)1/(n-2)n (3/2)1/(n-2)n
(4.34) < t < .
This defines a surface Cn: t=tTI (y0), contained in the spherical

shell (4.34) . Combining (4.34) with the second (4.33) we get
.
(4.35) 2/9(n-2)nl-n
<
6(n-2)nl-n
This implies on CT, , so that Cn is smooth. Also we get Cn

inside C,,
n as n<n'. Notice that the 'angular derivatives' of
O(n3/2-n),
the function (y) are by (4.30), and (4.34), since the
first term is constant along the spheres jyl = const. By (4.35)
cnl-n
the radial derivative is bounded below by . Accordingly the
unit normal of C (which is parallel to the gradient) has its ra-
n
dial component of the form
(4.36) nr(y) _ ±1 + O(lyll/2) , as lyl-.0
Now we consider the integral
(4.37) In = JW (g(x0,y)-Sn2-n)Hu(y)duy
TI
for small n , and some Co-function u , where Wn denotes the inte-

rior of C . With a cutoff function V=l near x0 =0 near CT, ,
xOen
write u=ui+u(l-$)=ul+u2, and, correspondingly, In=ll,n+I2,n
Clearly we get Il,n = u(x ) - Sn2-nJqu du . On the other hand,

0 1
g(x0,y)-6n2-n
setting w(y) = , for a moment,
I = Jwqu du - Jdya (Khjku )w

(4.38) 2,n 2 yj 2lyk
= JHwu2du + JC dSyK(y)hjk(y)vk(y)w lju2(y)

n y
with the surface element dS and the exterior normal components V..
Noting that Hwu2 = -Sn2-nqu2 , and assuming that Hu=O , we con-
clude that
0 = In = u(x0) + JC dS p n u -6n2-nJW qudu ,

(4.39) n n
with p (y)=K(y)hJk(y)g (x0,y)vk(y) .
n IYj
Write the surface integral in (4.39) as an integral over the

Sn-l
unit sphere Sn-1={ly0I=1} , with the surface element dSl of
Iyln-1/nr(y)
noting that dS/dS1 = , with nr of (4.36) We get .
(4.40) u(x0)=6n2-nJWnqudu-JSn-lItn(y)In-lpn(tn(y))u(tn(y))/v rdS1

Izln-lpn(z)/nr(z)
Here we introduce the function an(z) = , and
observe that a(z) is bounded and bounded below by positive con-
stants independent of n and z Let us multiply (4.40) by nn-1
.
and integrate this do from 0 to ,

n0-(2/3)1/(n-2)t0
(4.41) n0u(x0) = P - J n-1dS1. Or. nn-ldn(au)(tn(y0))

S y
Ino
with P= S 0 n W qudu
T)
In the inner integral of (4.41) make the substitution

t = tn(y0) of integration variable. The derivative tn= 8ntn(Y0)
may be obtained by differentiating (4.31) for n . It follows that
(4.42) tn((2-n)t1 n + y04y(ty0)) _ (2-n)nl n

I
Clearly this implies that to is bounded and bounded below by

positive constants independent of t,n, and u . Therefore the
integral in (4.41) may be written in the form
(4.43) fW dys(y)u(y) , cl<$(y)<Cl

rb
with constants cl, C1 independent of n and u . On the other hand

the integral P above may be estimated by
(4.44) O(n2JW dyIu(y)I

rb
We have proven the following result.
Theorem 4.5. Let the 'local balls' B(x0,n) be defined by
(4.45) B(x0,n) = {xEO : d(x,x0)<n} ,
where d(x,y) denotes the geodesic distance from x to y, with res-

pect to ds2=hjkdxJdxk. For every compact subset K of 0 there exist
positive constants e0,c,C, which generally may depend on the
triple {c2,du,H}, and on K, but do not depend on u, such that, for
every nonnegative solution of the differential equation Hu=O , and
every e>O with e<e0 we have the inequality
(4.46) ce < u(x) < Ce-nJB(x,2e)udu , x E K .

nJB(x,e/2)udu
Indeed, our construction gives such inequality with the two

'balls' of integration replaced by the set Wn above. Moreover, it
follows from (4.34) that B(x,e/2) C Wn,x C B(x,2e) , for suitably
smaller co , which implies (4.46) . Also, it is evident from our
construction that each point x0En has a neighbourhood Nx such
0
that the constants c,C,e0 may be choosen independent of x, as long

as xEN x . In particular, our construction was carried out in a
0
special coordinate system, for each fixed x. Hence we must involve

a coordinate transform to establish the special coordinates, but
the coefficients involved then are locally continuous in x, and
all conditions are locally uniform. Since a compact set K may be
covered by finitely many such neighbourhoods Nx , we thus can
j
choose our constants to be valid in K,q.e.d.
5. Harnack inequality; Dirichlet problem; Maximum principle.
In all of this section let H denote a second order strongly

elliptic differential expression of the form (1.2), with real
coefficients, and q(x) of any sign, not necessarily satisfying
(1.4). Actually, formal self-adjointness is not required for thm.
5.1 nor thm.5.8. In fact, both results are known to be true under
still more general assumptions not important for us (cf. [GT1]) .
In preparation for some results of sec.ll we next will dis-

cuss a proof of the Harnack inequality, using the inequality of
thm.4.5.
Theorem 5.1. (Harnack's inequality). For every compact subset

KC(2 there exists a constant c=cK such that, for every nonnegative
solution of Hu=O we have
(5.1) sup{u: xEK} < c inf{u: xEK} .
Remark. It is known that the constant c may be choosen to depend

only on the dimension n, the ellipticity constant (i.e. the local
lower bounds of in a given fixed countable
atlas), on local bounds for all coefficients, and the choice of
0 and K (cf. Gilbarg-Trudinger [GT1]). We will not require (nor
III.5. Harnack inequality 88
prove) this much stronger result here, however.

Proof. First assume that K=B(x0,e) , as in (4.45) for some x0ES
and that B(x0,16e)C92 as well, where 8e<eo , with e0 as in thm.4.5.
A solution u of Hu=O assumes its maximum and its minimum at xM
and xm E K , respectively. Then we use (4.46) for
(5.2) u(xM) < Ce nJB(xM,2e)udu Cc-nJB(xm,4e)udu<8nC/c u(xm)

<
using that d(xm,xM) < 2e , hence B(xM,2e)CB(xm,4e) . Accordingly

(5.1) follows with the constant cK given by 8nC/c, where c and C
are the constants of thm.4.5.
Now consider a general compact subset KCQ . It is no loss
of generality to assume that Kint is connected, and (Kint)clos=K
since a larger K with these properties always can be found. The
constant cK will also be valid for the given K, since the sup
increases and the inf decreases, as K is enlarged.
Again the solution u assumes its maximum at xM and its mini-
mum at xm, and these points may be connected by a path r within
Kint
Then a finite sequence of points xjEr
.
j=1,...,N, may be
,
found such that xO=xM , xN=xm , and that the intersections

B(xj,e)fB(xj+1,e) , j=1,...,N-1 all are nonvoid, while e is choo-
sen to satisfy the above conditions. Accordingly, letting B
B(xj,e) , we conclude that
(5.3) u(xM) <supB u <y infB u <Y supB u <...<yN infB u <yNu(xm
0 0 1 r
with the constant y=8nC/c of (5.2). This completes the proof.
Proposition 5.2. Let 0' C 0 be a subdomain (an open connected

subset) of 0 . If u e C'(2') solves Hu = Au , x e 0', with A E I
and u is nonnegative in 0' then it must be either positive in all
of 0' or identically zero in 2' .
Proof. Follows from Harnack's inequality.

Next we will make some comments about the Dirichlet problem
of the differential expression H in a subdomain 2' C Q with regu-
lar boundary, in the sense of sec.2. That is, we assume Q'U3Q'
compact , and 8Q' a finite union of (connected) compact n-l-dimen-
sional submanifolds of Q We also allow the case of 32=O
. Note .
that the assumptions of cor.3.9 are satisfied, so that the

Friedrichs extension H, (i.e., the closure of the the Dirichlet
operator Hd) has compact resolvent. It follows that the spectrum
of H=Ha* is discrete. There exists an orthonormal basis
111.5. Harnack inequality 89
{,j of H'=L2(SZ') consisting of eigenvectors of H

such that H* = X. . , with corresponding eigenvalues
(5.4) A0 < Al < .... , aj i - , as j + . .
Proposition 5.3. All eigenvectors are in C'(S2'U3S2') . If AEM

A#A1, 1=0,1,2,..., and if 3S2'9 0 , then the Dirichiet problem
(5.5) u E C_ (S2' Ua c2') , Hu = Au , x E SZ' U8 S2' , u l 9 S2' = u 0
is uniquely solvable for all u0 E C0-(252')
Proposition 5.4. For the smallest eigenvalue A0 of Hd we have
12/]lull'2
(5.6) \0 =Min {J(u) :u E H1 \ {0} }, where J(u) = Pull
Here Ilull 1 and Hull' denote the norms IluO1 and (lull , for n' .
Proofs. Since H is of compact resolvent we find that A is not in

the spectrum of H , so that the resolvent (H-A)-1 exists, as a
bounded operator of H'. Hence the equation (H-A)u = f is uniquely
solvable, with u E dom H, for all fEH'. Again, if f E CW(SZ'U8S2'),
then we conclude that uEC(S2'U3S2') , using cor.2.2 In parti- .
cular this also implies that the eigenvectors are

For a given u0 E C'(8S2'), if 8S2' is nonvoid, it is easy to
construct v E C'(c') with vlac2' = u0 . Just extend u0 a little
distance inside of D', carrying it to zero,using a partition, etc.
then the problem (5.5) is equivalent to
(5.7) (H-A)w = - (H-A)v = f ,
where w = u - v , evidently. Since (Hd-a)w=f is uniquely solvable

for f E C'(S2) proposition 5.3 follows.
For prop.5.4 we first notice that, for u E C'(')
r
(5.8) (u,u)1 = (u,HOu)'=(u,Hu)'= j=0Xlaj12 aj =(,Pj,u),
Taking closure in H1 one confirms equality of the first and fourth

term in (5.8) for general u E H1 We know that H1 C H' . Hence
r
the functional J(u) of (5.6) is well defined over H1 , and
(5.9)
Eajla.I2/1lajl2
J(u) = > A0 = J(p0)
This completes the proof of prop.5.4.

I
Proposition 5.5. Any u E H1 \ {0} which minimizes the functional

J(u) is in dom Hd ,and is an eigen function of Hd to the minimal

eigenvalue a0 .
Proof. For a given function u with the assumed properties and any
w E dom H let 0(t) = J(u+tw) This is a well defined differentia-
ble function,since u+tw E Hl \ {0} for small t Using the fact
that u minimizes J we conclude that
(5.10) dJ/dt(0) = 2 Re ((u,w)/(u,u)' - (u,w)'/(u,u)'2) = 0
u
Or, replacing w by we with suitable real u , it is found that
'Re' in the last equation (5.10) may be omitted. It follows that
I
(5.11) (u,w)1 =(u,Hw)' = A(u,w)' A = J(u) ,for all w e dom H.
This implies that A is an eigenvalue of H hence of Hd ,by

prop.5.3. It is clear,from prop.5.4 that A is the smallest
eigenvalue of Hd , q.e.d.
Proposition 5.6. If u is a real-valued eigenfunction to the smal-

lest eigenvalue a0 of Hd ,then either u - 0 or u 9 0 in all of 0'.
Proof. In view of prop.5.2 it is sufficient to show that u can-
not assume both positive and negative values. If the latter were
true then the function v(x) = Max {u(x),0) vanishes identically
in some open subset of 0' but does not vanish on all of 0' We
claim that v is in Hl ,and that it minimizes J(u) This,if proven,
implies that v E dom Hd ,and Lv = a0v a contradiction because v
then will have to vanish identically,in view of prop.5.2.
We require the following special case of Sard's theorem
(cf. [ Mll] , [ GPl] ) .
Proposition 5.7. For any real-valued C'(0)-function f the set of

all critical values (i.e. values n E I such that n=f(x) at some
xEO with Of(x)=0) has Lebesgue measure 0 on the real line I .
Using prop.5.7 we conclude the existence of a decreasing

sequence {ej} ,ej + 0 such that the level sets u(x) = ej do not
contain a critical point for each j Indeed the nulset of all
.
critical values of u(x) cannot contain any interval (O,e) , e>0

It follows from the implicit function theorem that the sets
ail; = {x E 0': u(x) = e are finite unions of smooth compact
J
n-l-dimensional submanifolds of ill. In particular, ;Q naO'=o,

J
since u=09ej on aO' . (In case of infinitely many components there

III.5. Harnack inequality 91
would have to be a limit point in the interior of the compact

set O'U8O' , which would make E. a critical value.) It follows
1 1 J
that the set &IE U8S2E , with S2E _ is a finite dis-
J J J
joint union of compact manifolds with boundary. Accordingly we may

integrate by parts in these sets to obtain
JonhJku u kdu=-JOn(u-e1)K-1(KhJku )I kdu

(5.12) IxJ I x IxJ x
I
, S2" = S2
E1
Or, setting S2+ = US2E , we get

1
(5.13) (v,v)1 = (h Jku u k + gIul2)du = X0(v,v)' <

Jsl+
I xJ Ix
r
This indeed shows that v E H1 minimizes J ,q.e.d.

Finally, in this section we formulate and prove a maximum
principle for second order elliptic expressions, needed below.
Again we will not require the strongest known results (cf.GT11).
Theorem 5.8. (Maximum principle) Assume a differential expression

of the form (1.2), with q(x) > 0 . Let u E C2(O) be a solution of
the differential inequality
(5.14) Hu < 0 , x E S2
Then, if u assumes a positive maximum at some point of 0, or if
q(x)=0 , we have u = const.
Proof. Assume that uM=u(x0)>u(x) for all xEO, and some x0ES2. Let
um > 0 . Since we have q>0 , and since Vu = 0 for x=x0 , it
then follows from (5.14) that
(5.15) hjkulxjxk > 0 at x = x0 .
At a maximum we also have the 'Hessian matrix' ((u k)) = M

IxJx
negative semi-definite,as well known. On the other hand, it is
no loss of generality to assume that hJk(x0) = Sjk , by virtue of
a suitable linear coordinate transform. Then (5.15) amounts
to the condition that the trace of the matrix M is nonnegative.
One thus concludes that the Hessian matrix must be zero at x0 .
A contradiction results if we even have Auu>O at x0 , since then

it is implied that the trace of a negative semi-definite matrix is
positive.
A proof of the maximum principle results by proving that a

suitable modification of u indeed results in a function w with
Hw>O at its maximum. The construction, below, has been adopted
from Protter-Weinberger [PW1] p61f. ,
Suppose we have u(xl)<uM at some x1En . Connect x1 and x0

by a path r , and let x2 be the first point, coming from xl ,
where u=uM. Choose a chart for x2, and let B be a ball with center
x2 , and radius d, in that chart. On the part between x1 and x2
of r choose a point x3 with distance from x2 less than d/2. Let
B1 be the largest open ball with center at x3 such that u<uM in
B1 . Clearly B1 has radius <d/2 , hence is entirely in the given
chart. Also there must be a point E at the boundary of B1 where
u=uM . Finally let K be a ball inside of B1 tangent to B1 at
and with smaller radius, and let y be its center.
As a result of this construction we have a ball K (of radius
r and center y) such that u=uM at precisely one point of 3K,
while u<uM in all of KLA Also KU3K is contained in a ball
around y , twice its radius,
contained in a chart (Fig.5.1).
Let 0<r1<r , and let K1 denote
r II
the ball around C with radius r1.
Since uM>O we can choose ri so
small that u>O in K1 , so that
APu>0 in Ki. The boundary 3K1 is
the union of a closed arc C1' in
the set KU3K and an open arc
,
Cl" outside K Note that

.
max{u:xEC1'} = uM-C<uM , since

C1' is compact, and u<uM on C1' Fig.5.1
Now introduce the auxiliary function
2
e-a x-y - e-ar 2
(5.16) z(x) =
with a>O to be determined. Confirm that
(5.17) A z = e aI x-Y12{4a2hjk(x-y)i (x-Y) k 2a(h» +bJ(xj-yj ))},

11
where, for a moment we set 0 = hJk3 3 k+bj2 , and where the

u xi x x
coefficients hjk bi all are taken at x
, By choosing a large.
we can arrange for L z>O in KUBKi, using ellipticity of H

ar )- 1
Finally define w = u + ez , with Confirm 0<e<0(1-e-
.
that w has the following properties.

2
e(1-e-ar
() w<uM on C1', due to ez < ) < C .
(ii) w<uM on Cl" , since z<O there, and u<uM everywhere.

(iii) w=uM at E , since z=0 there, but u=uM .
It follows that there must be a maximum of w inside the ball K1,
while we conclude that Auw = Auu + Epuz>0 in K1. Thus we get the
contradiction mentioned above, and the theorem is established.
6. Change of dependent variable; Normal forms; Positivity of g.
Let us return to the configuration of section 1, of a triple

{Q,du,H} of a measure du = Kdx , and an elliptic self-adjoint
expression H of the form (1.2) on a manifold 0 The presence of .
hjk
the tensor on 2 suggests introduction of the Riemannian metric
((hjk))-l
(6.1) ds2 = hjkdxJdxk , ((hjk)) =
on . Then one also has the corresponding surface measure
(6.2) dS = 1 dx , h = det ((hjk)) (det ((hjk)) )-l
and the 'Beltrami-Laplace expression'
(6.3) 0
la Yi hJk2 k
X] x
which is self-adjoint with respect to the measure dS The choice .
of H = - 0 + 1 dp
, = dS , for a general Riemannian space 0 with
positive definite metric (6.1) will give a configuration meeting
(1.4). Then the Friedrichs extension H will be called the Laplace
comparison operator.
Two triples {l1,du,H} will be called equivalent if they can
be transformed into each other by a "transformation of dependent
variable". To explain this in more detail, if we set u=yu-, v=yv-,
with a positive CO(Q)-function y then the inner product (1.1)
assumes the form (u,v) = L u'v-y2du = (u-,v-)-. Also, the equation
Hu=v may be written as y-1Hyu" = v- . Accordingly we introduce

dp-
(6.4) = y2du , H- = y-1Hy ,
as a new measure and second order differential operator, and the

triple (2,du-,H-) will be called equivalent to (c2,du,H). (We then
111.6. Normal forms; positivity of g 94
write (H,dp,H)-(Q,dp",H-). ) It is clear then that the linear map

u' i yu- defines an isometry L2M,du-)-L2(c2,dp) . Moreover, we get
(6.5) (u,v)1= (u,Hv) _ (u-,y 2(y-1Hy)v-) = JHu-H-v-dp- _ (u`,v-)1,

C0-
for u,v E (0) , where (.,.)1 is the inner product (1.3) for H-.
H`-112
Since (A-liCO)** = A-1 , and the corresponding for A` = is
true, (6.5) implies that u- yu- also defines an isometry H1+H1.
-
Furthermore H-=y-1Hy , of course, is a second order diffe-

rential operator with principal part -hi13 a 1 . We get (H-u,v)"
xi x
(Y-1HYu,Y2v) = (Y2u,Y-1Hyv) = (u,H-v)` for u,v E CO(Q) ,
so that H- is formally self-adjoint,with respect to the new inner

product. In fact, H- again is of the form (1.2), with
(6.6) K- = Y 2 K , q- = Hy/y = q - (yK)-1(Khjky ) .

Ix3 Ixk
To verify this we write, for u,v E CO
(u,H'- v)` =(Y2u,y-1Hyv) =(yu,Hyv) = J(hjk(YU) (Yv) + gy2uv)dp

Ixi x
Ix
=JhjkuIxjvIxkY2dp + J(hjkYIxjYIxk + gy2)uvdu + Jhjkyy Ixj(uv) kdp

I x
A partial integration transforms the last integral into
- JK-1(yKhJky )
kuv du . Accordingly, by a calculation,
Jxj j x
(6.7) (u,H`v)` = J(h3kulxjvlxk + qùv)dp- , u,v e CO(Q)
with q- of (6.6) and dp- = K-dx

, K- of (6.6) Comparing this ,
with (1.3) one concludes that H` = y-1Hy indeed must have the
desired form.
It is natural to use the equivalence defined for introducing
a suitable normal form for comparison triples. We shall find two
specific normal forms convenient.
Proposition 6.1. Every comparison triple is equivalent to a unique

triple of the form
(6.8) (n ,dS,-A + q")
with surface measure dS and Laplace operator A of the given metric

ds2 , and with a C-(n)-function q" , explicitly given by
{K2/h}-1/4
(6.9) q = Hy/y , Y =
Proof. Note that y , as defined in (6.9) indeed is a Cam-function

on 0 because the transformation properties of the measure du and
hjk
the metric imply coordinate independence of q" and y of (6.9),
as easily checked. Then the proposition follows from the general
facts about the transformation u = yu- derived above, noting that
2
K_ = Y K = F ,q.e.d.
A triple (6.8) will be said to be in Sturm-Liouville nor-
mal form.
We note that the condition q>l imposed on comparison triples
so far has only one purpose It will guarantee that H is semi-
:
bounded below, and, moreover, that H > 1 , implying (1.4). On the

other hand, prop. 6.2, below, shows that for a triple with H>l
always another normal form can be introduced -i.e. an equivalent
triple - such that q" is a constant >1 In fact, if R is non- .
compact, we always can arrange for q-=l

In the following we often no longer require q > 1, but only
that (1.4) holds.
Proposition 6.2. Let (cl,du,H) satisfy all assumptions of early sec.

1, especially (1.4), i.e.,(6.10), below, but not necessarily q>l.
(6.10) (u,Hu) > (u,u) , u E C_(Q)
Then, if 0 is noncompact, there exists a positive C-(0)-function

y, solving the differential equation HY = y in all of 0 .
Using y in a transformation of dependent variable u = yu-,

as described, one arrives at an equivalent triple of the form
{0,du-,l-&u-} with
2
(6.11) -= K--18 K-h3k k , du = K-dx , K- = Y K
u xl x
in local coordinates.
On the other hand, if 0 is compact then the equation Hu=au,
for some real constant A > 1 admits a positive solution y which
may be used as above to obtain an equivalent triple with q-=A > 1.
This proposition is an immediate consequence of the facts
derived on transformations of dependent variable,as soon as we
establish existence of a positive solution y of the equation HY=y.
Such a solution will be constructed below. In fact it will become
evident that the condition H > 1 is necessary and sufficient for

existence of a positive solution.
The result, below is required for prop.6.2,and perhaps of
independent interest. It was shown by the author for an ODE and
a compact interval in [CEdI, p.65 (in 1957) (also [CS],IV,1.9.),
and for elliptic operators of varying generality by Allegretto
[Al1l, [Al21, [Al31, Piepenbrink [Ppl], [Pp2], Moss-Piepenbrink
[MP1l, Agmon [Agl]. We are indebted to M.Meier for the suggestion
to use the Harnack principle, which independently also was used
by Agmon (Agl] for the same result.
,
Theorem 6.3. Given a positive measure dp and differential

expression H as in (1.2) on a manifold Q, with real symmetric
positive definite hjk and real-valued q , all C` , but not
,
necessarily q > 1 .
(a) If D is noncompact, and the expression H-1 is positive, i.e.,
(6.12) j(hJku lx Iu lx k + (q-1)juj2)dp > 0 , for all u E Cp(D)
then there exists a positive solution u = y E C-(D) of the partial

differential equation Hu = u .
(b) If D is compact then the same is true under the additional

assumption that
(6.12a) inf {(Hu,u)/(u,u) : u E C0-(S2) , u 4 0) = 1
(c) Vice versa, regardless whether 0 is compact or not, if a posi-

tive C'-solution of Hu = u , defined over all of 0 , can be found
then H-1 is positive (i.e. (6.12) holds. If 0 is compact then
,
also (6.12a) follows.
Remark. Notice that a positive solution y of Hy = y may be used

for a transformation of dependent variable, as in section 1, to
take the triple (0t,dp,H) to an equivalent one {St,dp-,H-), with q-
=Hy/y=1, using (6.6). Since the new triple satisfies q"'=1>l, we
trivially get (1.4), coinciding with (6.12). Moreover, if 0 is
compact, then a positive solution of Hy = y trivially represents
an eigenfunction of H to the eigenvalue 1. Since we know that
(6.12) holds we then conclude that
(6.12b) 1=(Hy,y)/(y,y) < (Hu,u)/(u,u), for all u E C0_(S2)=C"(0).
This evidently implies (6.12a) . Accordingly we already have pro-

ven (c), using the above transformation of dependent variable.
Proposition 6.4. On a manifold n satisfying our general assump-

tions there exists a C'(O)-function 4) taking positive values on 0
such that limx_,°°4)(x) _ °D , in the sense of sec.3 (or app.A).
Proof. If Q is compact then the function 0 = 1 will satisfy all
assumptions. If Q is non-compact then set 4)(x) _ with
the partition of unity of app.A. Q.E.D.
Proposition 6.5. If n is noncompact then there exists an infinite

sequence
521C 522C ... C 52 , U SZj = 52 , 52j # 0 ,

(6.13)
0i = U1 nil , S2j1r1 njm = 0 , 19m ,
where each 0j1 is a nonvoid open subdomain of n with compact clo-

sure and nonvoid smooth non-self-intersecting boundary (i.e., 8S2 j1
is an n-l-dimensional submanifold of Q ). Every sum in (6.13) is
finite,and every inclusion in (6.13) is proper. Moreover, each
nilclos
is a proper subset of some unique SZj+l,m
Proof. Using proposition 5.7 on the function 4) of proposition 6.4
we conclude the existence of a strictly increasing sequence
1 < nl < n2 < ... < nj < ... nj + such that the "level set"
,
an3 = {x E 0 : 4)(x) = nil does not contain any point x with Of(x)
= 0 Indeed the set of all critical values cannot contain any
interval (n,°°) so that a sequence {nj) must exist. Then define
(6.14) SZj = {x E 0 : 4)(x) < nj} ,
so that an. is the set defined above, containing no critical

points. By the implicit function theorem we conclude that an
indeed is a compact n-l-dimensional manifold,not necessarily con-
nected. Each SZj is nonvoid and each Q.U3S2is compact, by con-
struction of 4) Thus each connected component is the interior of a
manifold with compact closure and smooth (compact boundary). There
can be at most finitely many components. (Otherwise a subsequence
of components must converge to a point x0E852., near which 8n.
is homeomorphic to a connected open piece while it is sup-
posed to contain an infinity of nonconnected boundaries of compo-
nents. ) Clearly each component 0 j1 has a nonvoid boundary, since
we assume 2 connected. Clearly we get 37 C 0j+l , which implies

the remaining statements, q.e.d.
Let us observe next,that prop.5.6 establishes theorem 6.3
for the case of a compact manifold Q .
Proposition 6.6. Let 2 be noncompact,and let be defined as in

2j
(6.13) Then if X3 is the smallest eigenvalue of Hd in Qj
the sequence {a0} is strictly decreasing and lim 0X > 1
0
Remark: The operator Hd for Qj is defined as the direct sum of the

corresponding operators for the components Qjl .
Proof. Notice that the minimal eigenvalue v of Hd is minimal eigen

value for at least one of the connected components Pjl. But Sljl C
S1 j+1'm , for some m , by prop.6.5, and this is a proper inclusion.
If i * 0 is a corresponding (positve) eigen function then define
(6.15) w = i in S1 jl , = 0 elsewehere in 0j+l,m
Then the conclusion of prop.5.5 may be repeated to show that

w E H1 - for 0j+l, and J(u) = v, so that the minimal eigenvalue
m,
v' of Hd in Qj+1 cannot be larger than v If v=v' .
then w will have to be an eigen function of Hd in Qj+l,m which is

impossible since w as defined in (6.15) vanishes in some open
set, by prop.5.2, thus will have to vanish identically. This
amounts to a contradiction unless v' > v Thus we indeed get a
.
strictly decreasing sequence,as j increases.Evidently there is

the lower bound 1 as well ,q.e.d.
For any j and some wjE C'M ,wj > 0 we now solve the
Dirichlet problem
(6.16) u E C-(Q j) , Hu = u in 2.U32j , uI92j = wj .
This is possible,since 1 is less than the smallest eigenvalue of

Hd for 0j (due to prop.6.6 and prop.5.3). The solution will
be called u.
Proposition 6.7. The function uj is positive in all of QjUaRj.

This proposition follows exactly as prop.5.6 .
Proof of theorem 6.3. To complete the proof in the noncompact

case let us now consider the sequence {vj = ajuj :j=1,2,...} with
the positive reals aj chosen to normalize vj according to
(6.17) sup {vj(x) : x E Q1 1 = 1 , j = 1,2,...

Using the Harnack inequality (i.e., thm.5.1) we get
sup {vj(x) : x E 03} < C inf {vj(x) : x E S23 }
(6.18)
< C inf {vj(x) : x E 01} < C sup {vj(x):x E 01) = C
where the constant C is independent of j , j = 4,5,... .
Now (6.18) at once implies a corresponding estimate
(6.19) J dp(glvjl2+hkly. k vj 1 < C' , C' independent of j

2 lx lx
Indeed, with a cutoff function X E C'(923) , X=l in 02 , we get
(6.20) (Xv,(q-l)v)0 _ (Xv,A v)0 =-Jhkl(Xv) kv 1dp

3 3 lx lx
where we set v=vj, for a moment. In the right hand side write
(Xv)lxk = Xvlxk + vXlxk , and, for the integral with the second
term, integrate by parts again, using that vvlx1 = (v2)Ix1/2 .
It follows that
JSt Xhkly
(6.21) J52 2
hkly
lxk vI xldp < 3 Ixkvl xl dp
=-(Xv,(q-1)v)0 3-(1/2)(A11 X,v2) Max{Iq-1IX+IOpXI/2:xE03}
Clearly this proves (6.19).

In turn, (6.19) implies existence of a subsequence {v. }
k
converging in L2(S21) to a positive solution v of Hv=v. Indeed,
with a new cutoff function X1ECO(S22) , X1=l in 0 we conclude
1
that wj= X2vj E Hl(S22), and that the norm I1wjII1 (with respect to
02) is uniformly bounded in j By cor.3.9 the imbedding.
H1-H (for 02) is compact. Thus wj must have a convergent subse-

quence in H(Q2) . Then the restrictions to S21 define a convergent
subsequence of vj in H(S21) . Again we get (H-1)v. =0 for x with
k
X1=1, i.e.,in a fixed neighbourhood 0" of SZl . It follows that
(v,(H-l)4)H1V= lim(vj for all tECD(0"), i.e. v is a
k
weak, hence a strong nonnegative solution of Hv=v Also Harnacks
-
inequality gives v.>l/C in S21, since sup v. = 1. Accordingly v=0
is impossible.
A similar argument may be carried out for each 0k , instead
of 921 Using the Cantor diagonal scheme one arrives at a sub-
sequence of vj converging in every nj . The limit v cannot vanish

identically anywhere and will be the desired positive solution
of Hv = v . This completes the proof of thm.6.3.
Finally, in this section we will show that (1.4) also
implies positivity of the Greens function of the Friedrichs exten-
sion H .
Theorem 6.8. Assume that condition (1.4) holds. then the Greens
function g(x,y) of the Friedrichs extension H of the minimal,
operator H0 , as constructed in thm.4.4, is positive for all

x,yESI , x9y
Proof. First we note that it is sufficient to consider the case
of q=const.>1, by virtue of thm.6.3, since we have (1.4).
Indeed, the Greens function of the transformed operator y-1Hy=H-
is given by
(6.22) g-(x,y) _ (y(y)/y(x))g(x,y) ,
which is positive if and only if g is positive.

Notice that we certainly have g(x,y)>0 as x and y are suf-
ficiently close together, by virtue of the expansion (4.22) .
Indeed, the first term in (4.22) is positive and is large in com-

parison to the other terms. Accordingly the positivity of g
follows from the maximum principle (thm.5.8), whenever 2 is com-
pact.
If G is noncompact, then we apply prop.6.5, constructing
an increasing sequence Q . with smooth boundary, and let H be
d,j
the Dirichlet operator of D. . Let gj(x,y) , x,yEGj be the Greens
function of Hd,j. From thm.4.4 we know that X(.)gj(x0,.)Edom Hd,j,
as x0EDj, so that cor.2.2 gives gj(x0,.) EC'(PjUe3j\{x0)) , and

gj(x0,y)=0 as yEaDj . Accordingly we again may use the maximum
principle to conclude that gj(x,y)>0 as x,yENj and x#y , and ,
then Harnack's inequality to show that even gj>0 .
Now we conclude the same for g by showing a type of weak con-

vergence gj-*g . Define the sequence of operators Gj`= GjXG , with
7
the characteristic function XN of Qj , where it is understood

3
that the function Gj-u , defined in G. is to be extended zero out-

side S1 Let vj=Gj-u-Gu. Note that vjE Hl, and that even Ilvjhl<c,
j,
with c independent of j . Indeed, we get GuEdom H , hence
(6.23) IMGull2 = (Gu,HGu) = (Gu,u) = IlAull2 < IluH2
Similarly one confirms that, with uj= ujStj ,
(6.24) IIGuH2 = IlGu112 < Hujll2 < Ilull2
so that Hvll1<211uhl=c
Next we note that, for uEHOC" , we have Hvj= HG.-u-HGu = 0
as x E Stj , and vjE C'(O,) as well. This implies that
(6.25) 0 , as
hence we also get
(6.26) limj. 0 , for all 4E H1 .
(This follows first for all 4EC0 and then for H1, since Ilvjlll
,
was seen bounded, and C- is dense in Hl
(6.27) G(u - Gu in weak operator convergence of L(H,H1 ) .
Moreover, we have strong operator convergence (in L(H,L2(K)) )

of the restrictions u + (G. u)IK (to (Gu)IK ) for each compact ,
subset K C 0
The first part of prop.6.9 was derived above. The second
part follows since the restriction map Hl -> L2(K) is compact,
by cor.3.9 , so that a weakly convergent sequence is turned
into a strongly convergent sequence,q.e.d.
We will apply only the weak convergence (6.27): Suppose
we have g(x0,y0)<0 , for some x0#y0 Then the same holds for .
xENx , yEN , with neighbourhoods Nx0 N of x0 and y0 We , .
0 y0 y0
may assume that N x0rN = 0 . For any SEC-(N
0 x0
) >0 , we conclude
y0
(G4)(y)<0 , as yE NY , and we get "<" whenever $ is not = 0

0
Accordingly we have (m,Gu) <0 whenever uECO(N y) , u>0, and both

0
u are not = 0 . On the other hand it is clear that ($,G. u)>0

Since the weak convergence in H1 implies weak convergence in H,
we get a contradiction. Accordingly we must have g(x,y)>0 , and

Harnack's inequality then implies g>0 . This completes the proof
of thm.6.8.
CHAPTER 4. ESSENTIAL SELF-ADJOINTNESS OF THE MINIMAL OPERATOR.
In this section we will discuss a variety of criteria, all

establishing essential self-adjointness of the minimal operator
HO for an expression H of the form III,(0.1), or one of its
,
powers. In other words, we seek to establish the definite case,

in the language of Carleman. This is a classical subject, with a
large list of contributions, too numerous to be discussed in
detail (cf. the bibliography in [F1] ).
Our selection, below, is guided by the requirements in
chapter's 5f. In particular we only consider the case of an
expression H bounded below.
All proofs offered are elementary. However we will require
the statement of Weyl's lemma for expressions H as well as for
their powers Hm , m=1,2..... . While the first was discussed
in III,1, the corresponding property for powers can be easily
derived with the same proof, using the explicit Green's function
we have established for the operators H . Clearly Hm has the m-th
iterate of the integral kernel g(x,y) as its Green's function, and
one easily analyzes the character of its singularity. Then the
proof of III,thm.1.2 may be repeated.
Actually, the latter argument is not required if distribu-
tion calculus is used: It is a trivial fact that products of
of hypo-elliptic operators are hypo-elliptic again. The proof
of III,thm.1.2 is easily rewritten to show that the expression
H there is hypo-elliptic. Hence Hm , for m = 1,2,3,... all ,
all must be hypo-elliptic. Hence Weyls Lemma holds for Hm

The only objection to this argument is that distribution calculus
must be used, which we have not introduced here (cf.[C1]).
Also, since the result is local, one may use a result in Rn,
for an expression H equal to 1-L outside a compact set. Such a
result is easily derived, even for general elliptic expressions
of arbitrary order, using a Green inverse of the form introduced
IV.l. Summary of results 104
in [C1],IV,3. This indicates a third alternative for proving

Weyl's lemma for Hm . Accordingly, and since the result is gene-
rally well known and contained in standard texts on elliptic par-
tial differential equations, we will not discuss a detailed proof.
1. Essential self-adjointness of powers of H0.
In this section we turn to the question about criteria to

insure Carleman's definite case for an expression H of the form
III,(1.2). We will not only look for H itself, but also for its
powers which will be needed in chapter 5. It was shown by Gaffney
[Gfl], and Roelcke [Rol] that the Laplace operator A on a complete
Riemannian manifold always is in the definite case.
This was generalized to all powers Am, m=1,2,..., by the
author [CWS], together with a more general result on expressions
H of the form (1.2) (cf. thm.l.l, below). Chernoff [Ch] gave a
very elegant proof for the powers Am on a complete manifold, using
as an essential ingredient that the first order hyperbolic system
8u/3t = (d+6)u , with the operators d and S acting on the diffe-
rential form u , has finite propagation speed, so that a solution
of the Cauchy problem with CD-initial-values will be C' for all t.
We shall discuss Chernoff's proof in [C3], in connection with
theory of hyperbolic equations.
On the other hand, for the special case of S2=Yn, dp=dx, m=1,
there is a multitude of results. We mainly will be interested in
the case of a semi-bounded HO , and, moreover, q0. The most gene-
ral set of results perhaps are those of J.Frehse [F1], some of
which we will discuss in sec.4 (cf. thm.1.9). Notice that the
case of n=1, under these assumptions, does not require any addi-
tional assumptions (cf.thm.1.10). For many other contributions
to this subject we refer to the bibliography in [F1], c.f. also
Devinatz [ Dv1] .
As in III,l we work with the expression
(1.1) H =-AU + q , Au = 1/K3 jKhJk3 k , dp = Kdx

x x
Theorem 1.1. Let us assume that there exists a non-negative

Lipschitz continuous function a defined over Q and satisfying
the following conditions
(i) a(x0) = 0 for some given fixed x0 E S2
IV.1. Summary of results 105
(ii) limx a(x)=- , (i.e., for N>0 there exists a compact set
KN such that a(x) > N as x E n\KN);
,
IVal2
(iii) < q x E 0 \ K , with some compact set K C Q
,
(iv) lVol < Cena x E 0 , with constants n,C > 0

,
Then there exists a nonincreasing sequence nm , m=1,2,...

such that (i),(ii),(iii),and (iv)TI implies essential self-adjoint
m
ness of H0m = (Hm) 0 (i.e., the definite case for the expression
Hm.) Moreover, nl may be choosen as any number with 0 < nl < 1
Remark 1.2. Notice that the conditions (i), (ii), (iii), (iv)
of thm.l.l are equivalent to the following alternate conditions:
There exists a positive Lipschitz-continuous function T(x)
xE0 , such that
(i') T(x0)=l for some fixed x0E 0
(ii') limx T(x) = 0 ;
(iii') VT(x) is bounded on 0

(iv')n q(x) > n-2IVT(x)12/T2(x) , for all x E CZ\K , with
some compact set K .
Indeed, (i), (ii), (iii), (iv)T1 translate into (1'), (ii'),
(iii'), (iv')n, if we set T(x)=e-na(x), as shown by a calculation.
Remark 1.3. Notice that the conditions of the thm. imply semi-
boundedness of H0 , because we will have q > c0 in the compact set
K , and hence q bounded below in all of 0 , since (iii) implies q
> 0 outside K .
Our proof of thm.l.l is elementary, but quite technical.

It will be discussed in the following sections. Presently let us
discuss the theorem and some of its consequences.
Proposition 1.4. For a semi-bounded H0 , if (H9+c0) > 1 , and

(H0+c0)m is e.s.a., for some real c0, then (H0+c)k is e.s.a., for
all c E I , and k=l,...,m .
Proof. Let C'= (H0+c0)m , B'=H0+c0, for a moment, and let C be the
unique self-adjoint extension of C'. Clearly B'>1 implies Hull2
< (u,B'u) < llulIfB'ull, or, IIB'ull > Hull, u EC-, so that 0u112< IIB'u112
(u,B'2u) < (B'u,B'2u) =(u,B'3u) <IIB'2ull2= (u,B'4u)<...<(u,C'u) ,
so that C'> 1. It follows that C must be the Friedrichs extension

of C', so we get C > 1 as well. As in the proof of I, prop.2.7 we
C-1
find that is bounded, and 0<C<l, implying uC-1Q<1 , im C =H .
Let B be the Friedrichs extension of B' , then Bm is self-adjoint,

since it is hermitian, and one easily concludes in Bm=H . Also,

BmDB,m=C', so that C=Bm Now for f E H there exists
. fj=C'uj=
B,(B,m-1u j) -* f, with u.E CO = dom B'. Note that vj=B'm-lu. E CO,

and v.- v, due to lB'z >Ilzll . It follows that im B' is dense in If,
and that B'o = B , so that B' is e.s.a. Similarly we may write f.
B'k(B'm-k u.) above,and conclude that B'k is e.s.a.
Now we are finished if we also can derive that (B'+n)k is
e.s.a., for every real n , because clearly H0+c = B'+n , n=c-n
For this we note that, for a given n there exist positive
constants X. yl, y2 such that
(1.2) y1IIB'ku]2< II(B1+n)kull2+allull2< y2flB'ku12, for uEdom (BI )k
Indeed, the second inequality is evident, and the first

follows from
(1.3) B'i < CB, k+ cE , j=l,...,k-1 ,
valid for every e > 0 , and sufficiently large c6>0 , as easily

derived. Now (1.2) at once implies that (B'+n)k =(B+n) , which
is self-adjoint, q.e.d.
Theorem 1.5. Let 2 be the interior of a complete Riemannian

manifold Q0 with compact boundary 30 , and assume that
(1.4) q(x) > y/(distance(x,3c2))2 , x E cA K
There exists a nondecreasing sequence ym , m=1,2,.... such that ,
for y > ym , and H = -A+q , we have H0m = (Hm) 0 essentially self-

adjoint. Here condition (1.4) will be considered void (i.e.,
always satisfied) if the boundary 83 is void. Also, A is the
Laplace operator of the Riemannian metric of 0 .
Proof. Let p(x) be the Riemannian distance from x to 83 and let ,
K(x) be the distance from any given fixed point x0 E Q Then .
T(x) =Max { e-nK(x) ,

P(x)}
where we set p = - if 8G = 0 , so that then T = K , may be seen

to satisfy all conditions ('). Indeed, the first three
conditions are trivially satisfied, using that VK and Vp are
bounded, and that p and K tend to - , of the same order, as x
on 0 in view of the completeness of Q0 (cf. [KG1]).
,
In view of remark 1.3 and prop.l.4 we may consider H+c0,

IV.1. Summary of results 107
for large co instead of H

, Note that x = e-hk
. near 3C , and
T = p near - (of 00 ). We get
y/p2>p-2Vp2/p2
(1.6) q+c0> VK near - of NO , and q+c0> near 3C ,
for sufficiently large y, depending on n. Thus the theorem follows
Corollary 1.6. Under the assumptions of thm.1.5, if we have
limx;3 (q(x)(dist(x,aC))2 = + ,
then all powers H0m are e.s.a.
Corollary 1.7. For a complete Riemannian manifold without boundary

all powers of the Laplace operator have their minimal operator
essentially self-adjoint.
The proofs are evident, in view of thm.1.5 and prop.l.4.
We shall discuss Chernoff's proof of cor.1.7 in [C3],
since it depends on the basic properties of the solution operator
of the hyperbolic equation Btu = Du to preserve the compact
,
support of its initial values, as t propagates (i.e.,finite propa-

gation speed). We note that his proof also applies to the case of
the Laplacian acting on differential forms, which we do not
consider here.
The theorem,below,is am immediate consequence of thm.1.5,
and III,thm.6.3, regarding a normal form of H>l. It appears as an
interesting generalization of a conjecture by F.Rellich [Re l1 for
Schroedinger type operators in Rn, first proven by Wienholtz [Wil]
(cf. also [Chi, sec.4).
Theorem 1.8. Let C be a complete Riemannian manifold (with coun-

table atlas), and let a differential expressions H of the form
hjk
(1.1) be given, with the metric tensor of the space, but a
general positive measure dp . Then, if the minimal operator H0 is
semi-bounded below, all its powers H0m are essentially self-
adjoint in the Hilbert space H = L2(C,dp).
Indeed, we conclude from III, prop.6.2, that the dependent
variable may be changed such that q=y = constant. In view of I.
prop.2.4 we may consider H+c0 instead of H and thus may assume
that y=l . Then thm.1.5 may be applied with v(x)=dist(x,x0)
again, with a fixed point xO Thus thm.1.8 follows.
.
Finally let us mention some results of J.Frehse [F11 concer-

ning the first power of a second order operator of the form (1.1)
on 0 = In. Actually we are stating weaker results, since we will

get restricted to C--coefficients, as we always do here. Also, in
[F1] one finds two more applications of refined nonlinear techni-
ques to problems of essential self-adjointness of the minimal ope-
rator which we do not include.
On In we of course can get restricted to a fixed global
coordinate system x = (x1,...,xn) (written with subscripts again)
Then, with fixed coordinates, let A(x) and a(x) be the largest,
and the smallest eigenvalue of the nxn-matrix ((hjk)) J,k=l,...,n '
Let Br (x) be the closed ball with center x and radius r and let ,
Br = Br(0) Also we get restricted to the case Kel

. .
Theorem 1.9. Given the differential expression
( 1 . 8 ) H = -Xax hJkax + q , x E In , dp = dx .
J k
where hjk q E C,(ln), and the matrix ((hjk(x))) is real,symme-
,
tric and positive definite. Then each of the following conditions

is sufficient for essential self-adjointness of the minimal opera-
tor HO .
1) q bounded, below, and
(1.9) sup{A(x):xEB2R\BR}/inf{X(x):xEB2R\Br} = O(R2+1) , as R-+°
2) q bounded, below, and
(1.10) sup{A(x):xEBr(y)}/inf{A(x):xEBr(y)} = 0(1), with some r>0,
The proof will be discussed in sec.4.

In the case of n=1, and a q bounded below it turns out that
(1.9) and (1.10) are superfluous:
Theorem 1.10. For the singular Sturm-Liouville expression
(1.11) H = -axp(x)ax + q(x) , -°°<x<°° , q(x)>0 ,
in H=L2(1) the minimal operator H0 is essentially self-adjoint.

Proof. We may assume that q > 0 , since H0 may be replaced by
H0+c. Since H is hypo-elliptic, it suffices to show that Hu=O
uEC0'(I)f1H implies u=O . Now, since uEL2(1) there must exist
sequences such that u(Ej )-+0 , u(Oj )-+0 , as j-+co.
However, since q>0 we have the maximum principle (III, thm.5.8)
valid, so that Max{lu(x)I: C <x-"J.) + 0 , i.e., uSO
j- q.e.d. ,
IV.2. Limit. point case for H 109
2. Essential self-adjointness of HO.
It is practical to discuss the proof of thm.1.1 for m=1

first. This will be a matter of an established elementary
technique, applied to manifolds by W. Roelcke. The lemma, below,
is crucial.
Let G,du, H=-Am+q be as in (1.1). For any compact set ECG
we define the 'local inner products'
(2.1) (u,v)E = JEu(x)v(x)du , (Vu,Vv)E = JEh3ku IxjuIkdu

x
Lemma 2.1. Let a(x) be a nonnegative Lipschitz continuous func-

tion over 0, satisfying o(x0) = 0 for some given x0. For some
R>0 let the set ER = {xES2 : a(x)<R} be compact. Then we have
R
(2.2) J ((Auu,v)E +(Vu,Vv) E )dp J: IVoIIVuIIvIdu
0 p p R
Proof. For given fixed e,p > 0 , 0 < p < R introduce the
continuous, piecewise linear function n = np'E on with
n(t)=0 on [p,°) , n(t)=l on (--,p-e] This function is Lipschitz .
continuous, hence the composition fi(x) = n(a(x)) defines a Lip-

schitz continuous function over S2 and we get V = n'Va , almost ,
everywhere. For u,v E CO(Q) we get Cv absolutely continuous, and

is locally bounded. Thus Green's formula may be applied for
(2.3) 0 = JE (DUu vC + hjku Ix.(cv)

k)du
R I x
It follows that
(2.4) JE (Auu v + hjku JE In'IIValIVulIvldu

R R
Note that (2.4) holds for all p , 0 0
We may integrate (2.4) dp from 0 to R , observing that the only

term at right depending on p is the term n' = dnpeE/dt(a(x))
Hence we get
R
(2.5) J v +h3kul jvI k) <JE duIVoIIVulIvl dpIn'I
0 x x R
As e -* 0 the function converges to the characteristic function
of Ep , boundedly, and almost everywhere, while JRIn'Idp converges
to 1 , as easily calculated. Therefore (2.2) follows, q.e.d.

IV.2. Limit point case for H 110
We now can prove thm.1.1 for m= 1 , as follows. Suppose

a function a exists with the properties of the thm. since a--
as x-, all the sets ER of lemma 2.1, formed with this function
a, are compact. Now the operator H0 is >1 , (if necessary we may
replace H by H+c0, using rem.1.3). Thus it is sufficient to prove
that im H0 is dense (I, prop.2.7). Or, we may show that (f,H0u)=0
for all u E dom H0= C0 implies u=0. In view of Weyls lemma (i.e.,
III, prop.l.2) any such f must be C'(O), and then must solve Hf=O.
Thus we are finished if we can prove the lemma, below.
Lemma 2.2. Under the assumptions of thm.l.l , for m=1 , if

f E H12W(O) solves Hf=O , then we get f=0 .
Proof. Define the functions $(p) , I(p) , A(p) , by setting
(2.6) VR) = JRIIflValdg dp , *(R) = JRIVfII2 dp, 0<R<-

0 p 0 p
Note that , j are integrals of nondecreasing functions over
[0,") each. They are continuous and have a left and right first
derivativative at each point. Moreover, the first derivative
exists at all points except at the (at most countably many) jumps
of the integrand in (2.6). Moreover, 4 and L are convex. The same
is true for A . We apply lemma 2.1, now valid for all R > 0 , and
use that Hf=O, i.e., Apf = of , to get
R
(2.7) dp(IIV-qfl2 +HVf12 < [IValfllE [Vf6E , 0 < R <
0 p p R R
With condition (iii) of thm.l.1 and the above functions we have
(2.8) A(R) < (4'(R+0)tp'(R+0))1/2 < A'(R+0)/2 , 0 C R < - .
If A does not vanish identi-

Notice that A , A ' are > 0 , by ( 2 . 6 ) .
cally then we get A > 0 , as p > pl , for some p1 > 0 Clearly .
A' is absolutely continuous. (2.8) may be integrated in [p1,")

for A(p) > c e2p , c= A(p1)e-2p' . Using (iv) of thm.1.1 we get
(2.9) Vol (< ena < CenP < C(A/c)n/2 = c1An/2, as x E Ep, P > P1.
Now we go back to (2.6), and conclude that

< c2angfll2 = c3an , since IIfU < . This may be taken into
c4/X-ran/2
(2.8), for A < .
Or
c5A2-n
(2.10) A' > , as p > pl
with a positive constant c5 . Now it is well known (and easily

IV.3. Proof of theorem 1.1 111
derived) that the differential inequality (2.10) cannot have a

positive solution in an interval of infinite length, whenever n<1.
Therefore, for n < 1 we get a contradiction, and aa0 , i.e. f=0
follows, q.e.d.
3. Proof of theorem 1.1.
We now will discuss the proof of thm.l.l for general m=2,

3,... The basic tool will be lemma 2.1 again, which will yield a
system of differential inequalities instead of the single inequa-
lity (2.8). It then is a rather complicated matter to show that
the system obtained admits only the identically zero solution.
We start by arguing as before that the proof is accomplished
if we can show that the only solution f E Hf1C'(Q) of the differen-
tial equation Hmf = 0 is the function f a 0 (Here we first have
added a constant to q to insure that q > 1, and H m > 1, which may
0
be done, in view of prop.l.4). This works just like in sec.2
and will not be discussed again.
For an f with above properties define g. = Hm-jf, j=0,±l,±2,
Clearly gj = 0 for j<0 , and we interpret Hk as a differential
operator applying to a C"-function, for j=0,...,m . The negative
powers Hm-j , j > m , will be interpreted as the (inverse) powers
of the Friedrichs extension H , a self-adjoint operator (and the
closure of HO , since our assumptions (and n < 1) already give
the e.s.a.-property for H0 itself.
It follows that gjE H , for all j E Z . Using Weyl's lemma
a finite number of times one also concludes that all gj are CW(SZ),
and we generally have H1gj = gj-1 , with derivatives in the ordi-
nary sense.
Similarly as in sec.2. we now define the functions
(3.1) (R) = 1RII,Vo,gjflrdr , i.(R) = JRIIVg.II dr, j=0,±1,...,

0 0
where we use the abbreviation 1.112 = II.12 (cf.(2.1)).

r r
Clearly 4j, ,j have similar properties as , $ of (2.6). Specifi-
cally they are absolutely continuous and the left and right deri-
tives exist at each point and coincide, except possibly at coun-
tably many jumps. The first derivatives exist except perhaps at
the jumps. In the estimates below we write '(R) for the right
derivative of 4 , i.e., q'(R) = and
similarly for VP'(R) Clearly than j4jt,V,j, j' are nondecreasing

.
and nonnegative, and j , V,j are convex. We have the fundamental

R
theorem of calculus valid in the form (R)-4(S) J
'(p)dp with
S
a well existing Riemann integral. For j<O both Q,j and V,j vanish
identically. Moreover, using (iii) of thm.l.l we get
1IVoIgj1I2+UVgj112
(3.2) < II/gj112+11Vgj112= (gj,Hgj) =(gj,gj_1) < °° ,
where the partial integration without boundary terms is justified

by the essential self-adjointness of H0, already proven (One may
carry it out for a sequence g in Cp with - gj, Hg - Hgj g
Notice that (3.2) yields boundedness of the derivatives 4j', i,j
for all j .
Proposition 3.1. The functions 0j , j satisfy the system (3.3)

of differential inequalities, where the series at right represent
finite sums, since 0j Vj vanish for j<0 and where the argument
, ,
at right is R+0 , not R (just as in (2.8)).
0 j+Vj < (0'IVj)l/2 +
erlr
(3.3)
j < C
j = 1,2,...., 0 < r < .o .
Proof. This follows by repeated application of lemma 2.1 to the

inequalities
+Ipj
< 0 (UVgj112 + IIVgjU2)dr
(3.4)
R
j < enR IIgjII dr
0
which follow from (iii) and (iv)n of thm.1.5. For example, the
first inequality (3.4) implies
R R
+V' < 0 (gj,(q-Au)g.)rdr + j ((gj,ougj)r+ (Vgj,Vgj)r)dr
0
1R
< (gj'gj-1)rdr + IIIVaIgjIIRUVgjUR
0
=
R
((q-Au)gj+l,gj_1)rdr + 4jV,j .
0
The first integral may be "integrated by parts" again, using lemma

2.1, for
_ jj + ((gj+l,(q-DU)gj-1)rdr
0
- (R
J ((Vgj+l'Vgj-1)r+ ('ugj+l,gj_1)r)dr
0
R
+ JO ((Vgj+1'Vgj-1)z,+ (gj+l,Apgj_1)r)dr
< m' j'7j + Zj +lvj-1 + /V_-,.+, l; j-1 + 0 ((q-Ap)gj+2'gj-2)rdr
It is clear now that this process may be iterated until an

index j-l<0 is reached, at which time the last term vanishes. This
results in the first formula (3.3). For the second inequality
one starts with
R 2 R
0 11gj11rdr = 0 ((q-ou)gj+l,gj)rdr
and again uses lemma 2.1 to increase the index of the first
factor, and decrease the index of the second factor, until again
the inner product vanishes. Then the second (3.4) implies (3.3),
completing the proof of prop.3.1.
Only the first 2m inequalities (3.3) will be used, but we
leave the system infinite for technical reasons.
We now reduce the system (3.3) as follows. First introduce
new variables by setting
(3.5) aj = q.+ 1Pj , IT 0j- Ij .
Using that Zm1pk by Schwarz' inequality, we get

$km + < kX
(3.6)
X2 j'2-uj'2 1
+11=0 j+1+lj-1-1
2(X.+ PI +.-1+1X1
uj)
Note that Xj , uj satisfy the following conditions.

(A): Both aj , pi are absolutely continuous and their right
derivatives aj', uj' exist everywhere. The functions XI are convex
and Xj, X it are nondecreasing and nonnegative. The aj' are bounded
the uj' of bounded variation. Furthermore, Xj(0) = pj(0) = 0, and
(3.7) 1 uj < aj , I uj'I < aj as 0 < r < W .

Proposition 3.2. The inequalities (3.6) and conditions (A) imply
(3.8) A < 3V7 X13/211(,7e nr/2x J+l + 2rajaj+l+laj-1-1)

j
for j=1,2....., and 0<r<-

Proof. If A.'-O the corresponding j-th estimate is trivial, since
aj(0)=0 then implies also aj=0 Since aj' is nondecreasing, we .
may have X.'=O only in some initial interval [O,p], but then will
get aj'>0 in (p,co), (unless aj=0 for all r). We derive the j-th
estimate (3.8) from the j-th pair of estimates (3.6). Note that
all j-th estimates involve the Ak k9j only in form of their ,
derivatives. We thus may translate left by p and replace Ak by

changing the j-th pair (3.6). If the j-th
Ak-ak(p) , without
(3.8) can be derived for these new A. then the same follows for
J i
the old aj . Thus, we may assume that X .>O for r>0

From (3.6) we obtain
/----
(3.9) a
2' vlx
+ 1l A +l+laT 1
We plan to integrate (3.9). Note that
(3.10) 2 JRajdr = JR(aj/aj)2d(aj)
with well existing Stieltjes integrals. In (3.10) we may let e-}0,

R
since A. exists, and the integral at right (called IE) increa-
0
ses, as a decreases. If Ie: , then a/A , hence X'/A?< C in

J 7 7
[0,1] . Integrate this to get
1/aj(6) - 1/aj(r) < C(r-6) , 0<6<r .
As 6-'0 the left hand side goes to - since aj(0)=0. Thus a con- ,
tradiction results. Not only must the integral IO exists but also
we must have A./A *0 , as r-'0 It follows that .
R
(3.11) A.I(2aj) < aj(r)dr
0
Here we substitute aj from (3.6), and use the inequality

R
(3.12) Cdr < ab , as a,bEC1([0,R]) , a,a',b,b'>O
0
From Schwarz' inequality we get

R
(3.13) V777 dr < (Ja'dr Jb'dr)1/2<((a(R)-a(0))(b(R)-b(0))1/2
0
confirming (3.12).] Thus we get
(3.14) X?/(2X )
J J -'f
<
J J J 1
a.J+1+1a.J_1_1
Here we substitute (X +uj)/2 from the second (3.6), for

r
(3.15) j1( cenraj47j+l+1X3-1 +

)j+l+laj-1-1
A /siimilar proceedure is now applied again to (3.15). Multi-

ply by X. , and integrate from 0 to R . Clearly (3.11) implies
0 , as r-0 , so that the integral at left exists. We
apply a partial integration as before:
JR(X,)-3/2d( =XjXj_3/2IR_ JRa3d(A' 3/2

(3.16) 3 JRX
As E->0 , the last integral, JE, is monotone, hence either conver-

ges or diverges to -- , as e->0. The latter proves impossible,
because then a3/kr3/2 = (X2/X.)3/2 , while we showed above
J J J J
that it tends to 0 as c-0 . It follows that
, -3/2(R) R
(3.17) < aj dr
3XJX j 0
Combining (3.17) and the integrated (3.14) we g e

IXX'-3/2(R)
<
1{TenR/2 JR jaj aj+ly
--
tt
+laj 1 dr
( 3 . 1 8 )
R
dr}
+ J
0
j+l+1X. 1-1
Using (3.12) twice, the first integral at right of (3.18) is
estimated
< )1/4
(3.19) < (X2/2J aj+l+laj-l-ldr)1/2
Similarly the second integral is estimated by
(3.20) < (Rajaj+l+lxj-1-1)1/2
using the monotony of ak

Now we may substitute (3.19) and (3.20) into (3.18) to
obtain the desired estimate (3.8), q.e.d.
For the derivation of thm.l.l from prop.3.2 we introduce
the following two order relations, defined for the set of nonnega-
tive functions a, b, defined over [0,-) We write a <p) b , if .
a(r) < cb(r) , with a constant c, and we write a <q) b if a(r)
< c(log(2+K(r))sb(r) , where K(r)=h= a.(r), with above functions

1 J
aj Clearly both relations are transitive and reflexive, and are

.
compatible with addition and multiplication of functions, and with

taking positive powers. Moreover, a <p) b implies a <q) b
Notice that K,K' are >0 and nondecreasing, and
I
(3.21) X .<K j=l,...,m, X <p) r

, X <K' , , a <p) 1 , as j>m
J - J J J
as a consequence of cdn.(A) , above.
Proposition 3.3. If K is not -0 , then we have eer <p) K+l

for some e>0
Proof. We sum over j in the first estimate (3.6), replacing u.
J
by 0 and using (3.21). Thus it follows that K <p) K' + Rr <p)

K' + y , for every y>0 . This amounts to the differential inequa-
lity ca < at , for a=K=y an some e>0 . Suppose K is not -0 , then
K(p)>0 for some p>0. For small y>0 we get a(r)>0 as r>p. Then we ,
get e>a'/a=(log a). This may be integrated for

a(r)>a(p)ee(r-p)
or, K(r)>ceer , as r>p, with c >0, implying the statement.
Corollary 3.4. We have either K-0 or r<q) 1 .
Proof. Take logarithm's.

For the reminder of this section we assume that K is not -0.
This will lead us into a contradiction, thus completing the proof
of thm.l.l. Define the functions am+1=0, aj - lk=jak m.
K?1/2{K5/12a1/12 +
K1/3a 1/6 + K5/12 + K1/3
X. <q)
(3.22)
+ K5/12+n/6ea1/12 + K5/12+n/6e}
J+l
Proof. User prop's 3.2 and 3.3. Note that F <q) 1 , by cor.3.4.
and that X <q) K'. For the two terms S. and T. in (3.8) we get
S.e-nr/2 <q) K5/4(xk_oaj+k+1)1/4 <q) K5/4(aJ+1)1/4

(3.23)
Ti <q) K (Xk-iaj+k-1)1/2 <q) K(aj+1)1/2
Taking the cube root we thus get (3.22) , also using prop.3.4.
The proof is complete.
Proposition 3.6. If n<nm is sufficiently small, then, for every

j=1,2,...,m, there exist finitely many positive numbers a3,k, Sj,k

Yi,k, dj,k satisfying ,,3'k + Bj,k < 1, yj,k+d0'k < 1, such that
JK,aj,kKsJ,k ' <q) JK'YjekKdjek

(3.24) a. <q)
k k
Proof. We use induction and first note that am+l
= 0 am = Am , '
so that (3.22) give the assertion for j=m. Assume it true for a
certain j<m . Then substitute aj from (3.24) into (3.22) for
j-1. It is clear then that
K5/12aj/12 <q) J{K5/12+gj,k/12 K,1/2+aj'k/12}

(3.25)
which again is a sum of terms of the form in view of 5/12

+1/2 + (Sj'k+aJk)/12 = 11/12 + ($J,k+aJ'k)/12 < 1. Similarly for
the other terms in (3.22) This shows that a.-1 is dominated in
.
the described way, and we have a.-1 =a.-1

+a. , of course, q.e.d.
Note that K = a1 now is dominated by a sum of the form
R
(3.24). Also we get K(R) < K'dr < RK'(R) , by monotony of K
0
Thus we get K < rK', and cor.3.4 implies K <q) K'. This may be
substituted into the right hand side of (3.24) , for j=1. It fol-
lows that
(3.26) K <q) K'E , for some 0<E<l .
Finally we observe that (3.26) implies
(3.27) K <p) K,(1+)/2
because we have
(3.28) (log(2+K))s <p) KE , for every c>0
since we know that K- as r-

, , in view of prop.3.4.
Now it is well know that the differential inequality (3.27)
(or in details
(3.29) K < c K'Y ,
with constants c, y >0 , y < 1 , does not have a nonnegative solu-

tion, defined in [0,-), except K(r)=0 . Indeed, write (3.29)
in the form
(3.30) -(l/KT)' =
TK'/KT+1 > c- ,
t = 1/Y - 1 > 0 , c_>O .
IV.4. Frehse's theorem 118
where r must be choosen large enough so that K90 . Integrating

(3.30) we get
(3.31) K-T(P) > K-T(P) - K_T(r) > c`(r-P) , as r>P ,
where the right hand side tends to - , as r- , while the left

hand side remains constant. This results in a contradiction
so that we must have K = 0 for all R. This, in turn, implies
that f = 0 .
This completes the proof of thm.l.l.
4. Proof of Frehse's theorem.
In this section we finally want to discuss a proof of thm.

1.9. As in the discussion of thm.l.l we only must show that
every C'-solution of the differential equation Hu=O in H=L2(In)
vanishes identically, under the assumptions of the thm., such as
(1.9). Here we may assume that us is real-valued, since the coef-
ficients of H are real-valued. Let us assume q?0 again, without
loss of generality.
In the following we consider open balls Br with center x0
kept fixed for a while, with radius r. For uEC(1n) solving Hu=O,
and EEC_(Br) we have
(4.1) fB 0
r
In (4.1) we introduce $ = X2IulPsgn(u), with sgn u = ±1 and 0, as
u>0 , u<0 , and u=0, respectively, and with a suitable Lipschitz
continuous cut-off function XEC0(Br) , to be chosen later on.
Clearly such is not C_ but still Lipschitz-continuous. Never-

,
theless the relation remains valid, by an approximation argument,

as long as p?l (For example one may introduce a mollified me
.
$E* , as in the proof of II,lemma.2.1, and then pass to the limit

E-0 .) From (4.1) we conclude that
(4.2) JB hOkuli (X2IulPsgn u)Ikdx < 0

r
since the second term is nonnegative.
Let us write the integrand J in (4.2) in the form
(4.3) J = PX2IuIP-lhjkulJulk + 2XIuIP(sgn u)hJkulHXlk .

In (4.3) introduce
u1jXIuI(p-1)/2=
2/(p+l) (CIA-nj)
( 4.4)
C =
Xlul(p+l)/2sgn u ,
nj = XIj IuI(p+l)/2sgn u
A calculation yields
(4.5) J = 4(p+l)-2(phJkCI.clk- hOknj nk -(p-l)hJkCIJnk)
Integrating (4.5) and applying (4.2) we thus get
(4.6) 0 > JBrh3k'lj'lkdx - p(JBrhOknjnkdx
Or, with A+=A+(r)= sup{A(x):xEBr}, a-=a-(r)= inf{X(x):xEBr}, with

the largest and smallest eigenvalue A(x) and ),(x) of the matrix
((hjk(x))), respectively,
(4.7) JB dxIVCI2 < pA+/a-JB dxIVXI2Iulp+l

r r
We now require a special case of the Sobolev estimate, as
stated in prop.4.1, below.
Proposition 4.1. For every Lipschitz continuous uEC0(Rn) , and

every p satisfying 1<p<n we have the estimate
(4.8) Hull np/(n-p) < n-1/2p(n-1)/(n-p) IlVullp ,
n)
with the Lp(.n)-norms ( IVuI=(flu1J)1/2 , integrals over
(4.9) llullp = {Jlulpdp}l/p ,

IlVullp = {JIVulp}1/p
The proof of prop. 4.1. will be discussed at the end of

this section.
Let us assume n>2 , so that the assumptions of prop.4.1
hold for p=2. (The case of n=2 will be discussed later on.)
Then we may combine (4.8), for u=r , and (4.7) to obtain
(4.10) IlXIuI(p+l)/2b2n/(n-2) IIIVXIIul(p+l)/2fl2

< Cl
with
2n-1/2(n-l)/(n-2)(A+/a-)1/2
(4.11) C1 =
Finally we replace p by p-l in (4.10), assuming p>2, and

take the 2/p-th power, and arrange for La-norms of u :
(4.12)
&X2/Pupgn/(n-2)
< (PC2)1/PIIIVX,2/Pulip , for all 2<p<-.
The above is correct for every X meeting the above require-

ments, and it should be noted that the constant C1 depends on the
choice of X only insofar as A+ and a- might change, according to
the support of X . Also C1 is independent of p.
Next let us choose a sequence of cutoff functions Xj, j=1,2,
... such that XjE CO(BR), and that X3.>0, and VXj= Xj-1VXj
, for ,
all j. We explicitly choose Xj(x)=Pj(Ix-x0I/R) with tj=(2j)-+1/2,
(4.13) .(r) =1 in r<tj+1, =0 in r>tj, =(r-tj)/(tj+l-tj) else
Then clearly Xj is Lipschitz continuous, and we get
(4.14) IVXjI < 2j(j+l)/R = dj-l
In (4.12) let p=pj=2TJ , with T=n/(n-2), and let X = Xj+l'

Using (4.14) and VXjXj-1=VXj, it follows that
2/p. 1/g 21 p
(4.15) IIXj+1 Jullp < JIIXj J-lullp j=0,1, ...
,
J+1 J
This relation may be iterated for

1/pJullp
. 1/pkjullp
(4.16) IIXj+1 < IIk=O(dkpkC2)
J+1 0
Notice that the product in (4.16) is bounded; we get

II(6kpkC2)1/Pk<e2Elog6K/Pk((T)l/p0)Ek/Tk((pOC2)1/po)ET-k
(4.17)
with infinite series, at right, taken from 0 to . All 3 series

CT-k/2
converge, since T>l , and (log 6k)/pk < , for large k
Accordingly we may replace the product in (4.16) by the constant
C2 at right of (4.17), and then pass to the limit j;co Note that .
2/pJ l
(4.18) HullL_(B ) = lim pX uh
R/2 J R/2 J J pi.
Therefore, expressing the at left by the sup, we get
(4.19) sup {Iu(x)I: xEBR/2} < C2RUHL2(B

R)
In particular we used that p0=2. The dependence of the constant C
on R, A+ and a- is easily evaluated: Only the first infinite
series in (4.17) involves the radius R. Using (4.14) we get
(4.20) 21(log 6k)/pk=IT-k log 2k(k+l) -IT-k log R =C3-n/21og R ,

with C3 independent of R . Similarly only the third term at right
of (4.17) involves the quotient A+/a-=a. With IT-k=(1-(n-2)/n)-1
=n/2, as used above, and with p0=2, the term becomes Cln/2= C4Qn/4
with C4 independent of a . In conclusion we get
(4.21) C2(R,A+,a-) = C5 R-n/2(A+(R)/a-(R))n/4
with C5 depending on n only.

The above iteration technique was first applied by J.Moser
[Msl), cf. also [GT1].
We have proven the following result in the case n>2.
Proposition 4.2. If n>2, and q>O , then every C(I)-solution u
of Hu=O satisfies the estimate (4.18) with C2 of the form (4.19)
where C5 depends only on n , while the center and radius of the
ball BR are completely arbitrary.
Exactly the same proof works in the case of n=2, with the
following modification. Hoelder's inequality implies that
(4.22) Ilullp < V(2-p)/2pnull2 , with V = 1B dx = nR2 , 1<p<2

R
Here we set and then apply prop.4.1 with p = r , for
(tR2)(2-p)/2pllVCll2
(4.23) 11CI2p/(2-p) < , p < 2
Now we combine (4.23) with (4.7), replacing p with p-1:
(4.24) IIXIuIP/2112p/(2-p) < cpR(2-p)/peal/2fIVXIIuIP/2112
with cp=c YT (2Y)-1 , Y=y(p)=p/(2-p) , c' a constant independent

of R,a,p . Or,
(4.25) IIX2/pupYP < (cpRl/Y(po)1/2)2/PIIIVXI2/PuIIp , P>2 .
-k
Notice that 0Y =p/2(p-1). Thus, for pj =2.Y3 the iteration lea-
ding to (4.16) yields (4.16) again, but now with the constant
(4.26)
c" depends on n only, and 6=6(p)=p/(p-1) > 2=n , 6(p)-2,

as p;2 It is clear that this is somewhat weaker than required
.
in thm.1.9, at least under estimate (1.9), but will be enough for

an estimate (1.9) with '0(1+R2)' replaced by '0(1+R2-E)', e>0 .
However, for n=2, a simpler proceedure, not using the above tech-
nique, still will give the same result as for general n, as will
not be discussed here.
Proof of theorem 1.9. In view of thm.1.10 we may assume that
n>2 , so that prop.4.2 holds. Let us first focus on assumption
(1.10). For R=r/2 , with the r>0 of (1.10) we conclude that
(4.27) sup{1u(x)I:ixl=n} < KJ n_r<Ixl <n+r lu(x)I2dx - 0, as n-,
since uEL2(Rn) , where K denotes a constant independent of n .
Indeed, for each x0 with lx0H=n we get (4.19) with the ball BR
around x0 and a constant independent of x0 and n However the .
maximum principle (III, thm.5.8) holds, since q>0 . Therefore

un=Max{Ju(x)I:IxI<n} is assumed on lx!=n , and we indeed must have
limn}_un = 0, or, u=0 in all of 1n, showing that H0 is essen-
tially self-adjoint.
Next let us assume (1.9). Notice then that the sphere
aB3n/2={x:Ixl=3n/2} , for large n, may be covered by balls Bn/2
with center on 3B 3n/2 and contained in the spherical shell
,
B2n\Bn The corresponding concentric balls Bn/4 still cover

.
9B
3q/2
, and the constants n-n/2 (A+/a_)n/4 are uniformly bounded,
in view of (1.9). Accordingly, again using the maximum principle,
we get
(4.28) Ju(x)J < c1n<JxJ<2n lu(x)I2dx + 0 , as n,- .
This again implies u=0 , hence H0 again is essentially self-ad-

joint, and theorem 1.9 is established.
Finally, in this section, we discuss a proof of prop.4.1.
(cf.[GT1]). We shall require Hoelder's inequality for n functions:
(4.29) Julu2...ukdx < flk 111u1Ip , u E L p3(,n)
valid if pl, ... , pk E R satisfy the relation
(4.30) 1/pl + ... + 1/pk = 1
Also we use that, for k nonnegative numbers al, ..., ak the

'geometric mean' (H%-1a is never larger than the arithmetic
mean k-1yj=1ak . The latter is well known. Also, (4.29) is
easily derived from the well known case of k=2, by induction.

For uEC0(&n) write

X.
(4.31) Iu(x)I J Iluljldxj < luljIdxj
Taking the product of these n inequalities, and the power 1/(n-1),

we get
(4.32)
Iu(x)In/(n-1)
< =1(Jluljldxj)l(n-1)
H
The inequality (4.32) may be integrated with respect to each xj,

j=1,2,...,n. After each integration Hoelder's inequality is
applied, with k=n-1 , P1-P2=. For example, the first pn-1=n-1.
factor, at right of (4.32) is independent of xi. Hence there are
n-1 unknown functions under the integral Jdxl. We get
(4.33)
Jdx1Iu(x)ln/(n-1)
<
{Jdxllulll 1Jdxldxjlul. }1/(n-1)
fl
For each subsequent integration one of the n factors at right is

a constant, and (4.29) may be applied with n-1 factors. We get
(4.34) Ilulll < (III=llluljll)1/n < n-lylluljlll < n-1/2UIVul1
Notice that (4.34) amounts to (4.8) for p=1. In particular

the inequality may be obtained for Lipschitz continuous u E C0(1n)
by an approximation argument.(For example one may use a mollifier
as in the proof of III, thm 1.2).
The general inequality (4.8) then follows by applying
(4.34) to the function v=lulY , with y = p(n-1)/(n-p) We get .
IVIujYI = yIujY-ljVuI Substituting this into (4.34) and using

.
Hoelder's inequality we conclude (4.8) for general p.
5. More criteria for essential self-adjointness.
In this section we discuss an extension of thm.l.5, useful

in ch.8. Let us assume that the expression H , defined over S2 ,
has an essentially self-adjoint minimal operator HO . Let U C ci

be an open subdomain of St with compact smooth boundary 3U, where
3U is assumed to bean n-l-dimensional submanifold of S2 . Assume the
potential q of H bounded below on U , and let q'EC'(U) denote a
nonnegative potential satisfying (1.4) in some neighbourhood of
3U , with a suitable constant y . Assume that q'=O outside another
IV.5. More essential self-adjointness 124
neighbourhood of 3U .
Theorem 5.1. Under the above assumptions the minimal operator H'0
of the expression H'=H+q' , in the Hilbert space L2(U,dp), is
essentially self-adjoint, whenever y>yl , where yl is a positive
constant.
Proof. One may repeat the proof of thm.1.5 in the following amen-
ded form. Let T(x)=dist(x,3U) and define
,
(5.1) O(x) = n-llog(E/T) , as T(X)<E , = 0 , as T(X)>E ,
for suitable constants E,n>0 . Then use this redefined function

o(x) as in the proofs of lemma 2.1 and lemma 2.2, to get essential
self-adjointness of H' .
In particular the 'spheres' ER again are defined by the

inequality o(x)<R . These, in general, are noncompact sets, so
that (2.3), (2.4), (2.5) are correct only for u,vEC'(R) Now, if .
fEC-(U)nL2(U,du) satisfies H'f=0 in U , then we get wf E dom H ,

for a suitable cutoff function, zero near 3U =1 outside a neigh- ,
bourhood of 3U in UU3U . Since H0 (in 0) is essentially self-ad-

joint it follows that there exists a sequence zkEC'(U) such that ,
zk-wf , k-* , in L2(O,dp) Using this, with a suitable

.
limit k- in (2.3), (2.4), (2.5) , one shows that the 'partial
integration' , taking JH'rfd14 into J(,Vf,2+qIfI2)dp still will
give no boundary terms at infinity of U , although there will be

such terms at the boundary of ER , of course.
In other words, one finds that inequality (2.7), for q+q'
instead of q , remains intact. Therefore the remainder of the
proof in sec.2 may be repeated, and one finds that f=0 . This
proves thm.5.1.
Remark 5.2. It appears that a corresponding statement, regarding

essential selfadjointness of powers H' 0m , can be proven by repeat-
ing the chain of arguments in sec.3, with similar amendments.
However, we have not checked this in detail.
CHAPTER 5. C -COMPARISON ALGEBRAS.
In this chapter we open the discussion of one of our main

objectives: Comparison algebras are certain C -algebras of linear
operators on an L2-space. In essence their operators may be expli-
citly represented either as Cauchy-type singular integral opera-
tors or else as pseudo-differential operators of order 0. However,
we never use (or even discuss) such explicit representations.
Except for a short excursion into Fredholm results for dif-
ferential operators on compact manifolds in VI,3 and VI,4, in ch's
V,VI and VII we will entirely focus on the discussion of the struc-
ture of such C*-algebras. The reader should keep in mind, however,
that the results obtained will be used later on (ch.X) to discuss
the properties of differential operators. For certain differential
tial expressions (said to be 'within reach' of a comparison alge-
bra C ) a realization will be defined (cf. def.6.2 below) In ch.X
we will focus on the spectral (and Fredholm) properties of such
realizations.
While we have discussed abstract spectral theory in ch.I,
we refer to [C1],app.A2, regarding an abstract theory of Fredholm
operators. Also we will use some results of abstract spectral
theory not discussed in detail in earlier chapters, such as the
discussion of the essential spectrum of a differential operator.
(Note that at least a survey of some such abstract results is made
in ch.X.)
We work in the Hilbert space H= L2(O,dp) , on a Cm-manifold
0 , with a positive Co-measure du, as introduced in ch.3. On Q we
also have a comparison operator H defined, where H is the (self-
adjoint) Friedrichs extension of the minimal operator of a diffe-
rential expression H, as in III,(0.1) In [Cl] we considered the
.
special noncompact manifold In . There we introduced a class of

C*-algebras generated by multiplication operators and convolu-
V.0. Introduction 126
tion operators. Especially the singular integral operators

S. = DjA A=(l-A)-1/2 , were used as such convolution generators.
,
In the present approach, on a general compact or non-compact

manifold 0, we replace the generators Sj by operators of the form
DH-1/2
DA = , with certain first order differential operators D
defined on Sl , and with A = H-1/2 , with the above self-adjoint
positive definite operator H
Accordingly, a comparison algebra C is a C-algebra genera-
ted as norm closure from generators
(0.1) a , DA , aE A# , DE
with certain classes A# , D# , of bounded C--functions and first

order linear partial differential expressions (folpde's), respec-
tively.
We work with general assumptions on A# , D which always
imply that CDK(H) , with the ideal K(H) of all compact operators
on H The commutator ideal E of C may or may not be K(H) as well.
.
At any rate, for all examples considered there is a chain
(0.2) C D E D K(H) ,
where the second inclusion may be proper or improper, but such

that both CIE, and E/K(H) are isometrically isomorphic to an
algebra of continuous functions. Then the Fredholm property of
an operator AEC is governed by one or two symbol functions, ari-
sing naturally from the above ideal chain. For more details and
simple explicit examples we then refer to [C1], where both kinds
of ideal chains are obtained.
It should be noted that algebras with two-link ideal chains
also have been studied elsewhere (cf. Dynin [Dyl], Upmeyer, [Upl])
although, as it seems, not in connection with noncompact manifolds
In applications it is found that the two-inversion symbol calculus
has its disadvantages, insofar as the second symbol usually is
available only after constructing an explicit E-inverse, not
available from the first symbol alone. In some later examples we
have remedied this by extending the second symbol to the entire
algebra C Then usually the second symbol takes values in some
.
algebra of singular integral operators over a lower dimensonal

space. This will be discussed in CCMe1I, [CPo] .
In the present chapter we only discuss some basic facts.

In particular it is proven that every comparison algebra C con-
V.1. Comparison algebras 127
tains the compact ideal, and that K(H) even is contained in the
commutator ideal E of C (lemma 1.1). We explicitly discuss L2-
continuity of DA , and AMA , for certain first and second order
expressions D and M, and derive explicit formulas for commutators
(sec.2). In sec.3 we derive sufficient conditions for compactness
of commutators in a comparison algebra. In sec.4 and 5 we discuss
a variety of examples. Specifically sec.5 considers the one-dimen-
sional case.
Finally, in sec.6, we prepare the later application of our
theory by introducing the concept of an expression within reach
of a comparison algebra. Specifically, a Taylor-type expansion
(with integral remainder) is discussed for "H-compatible expres-
sions", having all H-commutators within reach. Such expansions
prove useful as an efficient substitute for the calculus of pseu-
dodifferential operators, which we do not assume here.
1. Comparison operators and comparison algebras.
Let ci denote a paracompact Co-manifold without boundary,

with a countable atlas, as in III,1, and again consider a measure
dp=Kdx, and a second order elliptic expression
H = -K-12 hjkK3 k + q
x3 x
which is self-adjoint with respect to the inner product of the
Hilbert space H = L2(D,du) As in III,1 we assume that H has real
.
C°°-coefficients, that ((hjk))>O , and that HO > 1 , i.e.
(1.2) (u,Hu) > (u,u) , for all u E CD(D)
Then we introduce the first Sobolev norm and inner product by
(1.3) (u,H0v) = JD(hJku IxjvIk + quv)dp = (u,v)1 , u,v E dom H

x
The Friedrichs extension H is a well defined self-adjoint
realization of H We get H > 1
. Clearly H-1 E L(H) has a self-
.
H-1/2 = 1k_0(_1)k(1k2)(1-H-1)k
adjoint positive square root A =
and 0<A<l . A is an isometric isomorphism between H and H1. We get
(1.4) H1 = dom A-1 , (lull = IlAulll , u E H , Ovlll = IIA-lvll , v E Hl.
The restriction A-'Idom H0 is essentially self-adjoint, and

A-ldom HO is dense in H , while the corresponding is not always

true for H dom HO = im HO. For details cf.III,l. In fact, we
already discussed this extensively. The operator H0 will have to
be essentially self-adjoint for this.
We also note the possibility of transforming the dependent
variable,to get from the triple {c,du,H} to an 'equivalent triple'
{c,du-,H"} with possibly better properties,such as a normal form
for H etc. (cf.III,6).
,
If 1,dp,and H are given as described we shall speak of the

comparison triple {S2,du,H}, and of the comparison operator H on n.
Also H1 will be called the corresponding first Sobolev space.
Let a comparison triple be given, and let
(1.5) D = bl2 +p
xj
denote a first order differential expression on S2 , where (bi) and
p denote a contra-variant tensor field and a complex-valued func-
tion on 0 , both assumed C
Then the linear operator DA:CD()*
.
L2loc(St) is well defined, since the square root A of H-1 maps onto
the first Sobolev space Hl (cf.III,(1.6)). In fact, assuming that
IIDIIH = II(IbI2 +IpI2/q)1/2p - ,

L_(Q) <
(with IbI={hjkFibk}1/2 ) , it follows that
(1.7) IIDuf2 < IIDIIH(u,Hu) _ PDIIHIIuII
provided that q > 0 on 2 Using (1.4) , it then follows that
(1.8) IIDAvu < IIDIIHIIvII ,vE im(A-lIC0-(S2))
where the space im(A-ljC-(c)) is dense in H. Accordingly, by con-

tinuous extension of D0A , we get a bounded operator A E L(H)
and it is clear that A = DA , with A as a map H - H1 , and a dif-
ferential operator D:H1 } H , defined with the strong L2-gradient
of III,(1.6) .
The adjoint operator A* of A E L(H) is in L(H) again, of

course, and we have
*
(1.9) A u = AD*u , u E C0'(52) , D* _ -K-18 jK50 +
x
In chapter X we will study the Fredholm properties of
certain C-subalgebras
* of the algebra L(H), called H-comparison
algebras. Here by an H-comparison algebra we mean a C*-algebra C

C L(H) generated as operator norm closure of the smallest algebra
CO containing all multiplication operators u -* au , and all opera-
tors A , A* of the above form A = DA , where a E A# D E V#. Here
,
A# , D# denote a class of bounded continuous complex-valued func-

tions, and a class of first order linear partial differential ex-
pressions (folpde's) on Q. of the form (1.5), with IDBH < - .
We will write C = C(A#,D#) , CO = CO(A#,D#) omitting the argu-
,
ments if no confusion can arise.

We require the following general assumptions:
Condition (a0): A# contains C'(c1), and its complex conjugates.
Condition (a1): A# contains the constant functions
Condition (d0): D# contains its complex conjugates T7=5j2 j+p ,

x
and contains all C'-folpde's with compact support.
Here (a0) and (d0) are always imposed, and (a1) is required
unless explicitly stated otherwise. Especially there are two
exceptions to this last rule: The minimal algebra JO defined
,
below, and the class A# 0 of VIII,3 normally will not contain all
constant functions.
In these cases we have the following condition instead.
Condition (al'): For every D E D# there exists an a E A# such
that Va has compact support, and aD = Da = D
For every comparison triple we have a minimal comparison
algebra, denoted by J , obtained by choosing AT as the class of
functions a=a0+ c , a0 E C , and D# as the class
of folpdes D = b3a j+p with b3E C' , p E C'(D) . The C -algebra
x
(not necessarily with unit) generated by norm closing the algebra
generated by a and DA for all, a and C'-folpde's
with compact support, will be denoted by JO. Clearly J is obtained
from J0 by adjoining the identity 1. We shall see in VI,l that 1
and J0 have compact commutators. J0 is unital if and only if D is
compact. Both J and J0 will be called minimal comparison algebras.
Lemma 1.1. The minimal algebra JO , and hence every comparison

algebra C , always contains the entire ideal K(H) of compact ope-
rators. Moreover, K(H) even is contained in the commutator ideal
E = E(C) of any comparison algebra C , and we have E(J0)=K(H)
In lemma 1.1. the commutator ideal E(C) of C denotes the
smallest closed 2-sided ideal containing all the commutators of
operators in C .
Proof. As in [C1],IV,thm.l.l, we use the fact that an irreducible

C*-subalgebra A of L(H) containing at least one nontrivial com-
,
pact operator, must contain the entire compact ideal (cf.Dixmier

[Dxl], cor.4.1.10). Here 'irreducible' means that the space Au is
dense in H for every 0 # u E H
, .
It is easy to construct nontrivial compact operators in a

comparison algebra. For example, for any function a E C'(Q)
not vanishing identically we have aA=aH-1"2EK(H), by III,thm.3.7,
below, while trivially aA 9 0 . In fact, also the commutator
[a,bA] , for suitable a,b E C'(Q) is compact and # 0 . Hence it is
sufficient to show that Cu is dense in H , for every 0 # u E H ,
(Clearly then the construction of [C1],IV,lemma 3.3 may be repea-

ted to show that K C JO , and, starting with the given compact
commutator # 0 that K C E holds as well. Indeed, one first will
,
construct an operator of rank 1 in C or E , of the form )( ,

with ,1p c H ,,p 9 0
, Then the operator
.
comes arbitrarily close to every dyad f)(g , f,gE H , as A,BEL(H).

Accordingly C and E contain all such dyads, and hence K(H) , the
closure of the class of operators of finite rank.)
H-1
Note that the operator may be written as an integral
operator
(1.10) H-Iu(x) = JQG(x,x')u(x')dp
where the 'Green's function' G(x,x') is positive on all of Q .
Indeed, from III,thm.6.2 we know that H - H-= 1 - Ap_ , after a

suitable change of dependent variable. Such an operator satisfies
the conditions of the maximum principle, which insures that the
Green's function is positive, since G solves HG(.,x')=0 , and is
clx-x'12-n (or c loglx-x'j ) near its singularity x=x' , and is
limit of the sequence of Green's functions Gk of the Dirichlet
problem on an exhausting sequence Qk of subdomains of n with
smooth boundary Clearly the Gk are positive, since their boun-
dary values are.Hence G is positive. For details cf. III, thm.6.8.
For any OiuEH we have 0 # H-ncu continuous for a suitable
C0(c1)-function c, by the Sobolev imbedding lemma, (VI,1.4.2), or
since the iterated Green's function has the corresponding smoothe-
ning property). Hence we may choose aECO(1)CA# such that aH-ncu =
aA2ncu does not change sign, and is not always zero. Assume it >0,
without loss of generality. Then w=A2aA2ncu is positive on 0,
and every v E C' can be written as v = bA2aA2ncu , where

b = v/w E C0(O) . Choose bkE with Ilb-bkll + 0 , then
let Ak = bkA2aA2nc and note that IIv - Akull } 0

, Furthermore, if .
0 < k 1 , kE C'(O) , k(x) } 1 for all x E 0 then A ,
A2n
hence (A4.)2, A2, (A'. )2n in strong operator convergence.
-
,
Thus we also get IIv - Bkull - 0 , for Bk = )2nc E J0,

k k
with a suitable sequence {jk} of integers jk This shows
, .
that indeed J0u is dense in H q.e.d.,
For a comparison algebra C we focus on the commutator ideal

E , i.e., the closed 2-sided ideal generated by the commutators.
Note that C normally has a unit (the function al). The quotient
C/E, as a unital C*-algebra, is isometrically isomorphic to a func-
tion algebra C(N), with a certain compact space N, called the sym-
bol space, by the Gelfand-Naimark theorem. The homomorphism a:C;
C/E;C(N) assigns to AEC a function aA' called symbol of A. In sec.
5 we will prove that N always contains a subset W homeomorphic to
the bundle of unit spheres in the cotangent space T'O of 0 . We
call W the wave front space (of the algebra C). The continuous
function aA:M - U corresponding to the coset mod E of an operator
A E C coincides with the principal symbol of D on W if A=DA , with
a differential operator D .
In the simplest cases we will get E=K(H) equal to the ideal

of compact operators on H We already mentionned that this is
.
true for the minimal algebras J and J0 The space W mentionned

.
above coincides with the maximal ideal space of J0/K On the .
other hand the set N \ W can be described as a certain set

located "over - of the manifold 0 ". Then an operator A E C
will be Fredholm if and only if aA # 0 on all of N However, in .
some more general cases we have a proper inclusion E D K(H) Then .
we shall tend to obtain a similar "function structure" for the

quotient E/K. The Fredholm theory then will be somewhat more com-
plicated.
The techniques to be described in the following can be
extended to other types of differential operators as comparison
operators, not necessarily of second order, or elliptic, or self-
adjoint. However, the fact that so many properties of these
'Schroedinger operators' are known makes the use of them particu-
larly attractive. In particular the e.s.a.-properties discussed
in ch.5 will be important, in this connection.
V.I. Comparison algebras 132
For many parts of this chapter it will be crucial to have

Carleman's definite case of II,1 , for the operator Ho in the ,
Hilbert space H In other words, no boundary conditions at infin-

.
ity will be required ; the Friedrichs extension is the only self-

adjoint extension of HO , and it coincides with HO , the closure
of Ho .
We shall use this condition in weaker or stronger form, as

follows:
Condition (w) : H0 is essentially self-adjoint.
Condition (s) : For all m = 1,2,.., the minimal operator (Hm)0=H0m
of the m-th power Hm is essentially self-adjoint.
Condition (sk): (where k = 1,2,.... o) : (Hk)0 = H0k is essentially
self-adjoint.
It follows easily that (sk) implies (sl for 1<1<k . Accord-
ingly (s1) equals (w) and (s,) equals (s) (cf.IV, prop.l.4).
Various results asserting the validity of (w) or (s) have
been discussed in ch.4. In particular (s) holds whenever the
manifold 0 is complete under the metric (1.7). On the other hand
we never will have (w) nor (s) if 0 is the interior of a manifold
with (nonvoid) boundary onto which the triple may be extended.
2. Differential expressions of order not larger than two.
For a comparison triple {O,dp,H} , as in section 1, let us

consider more general differential expressions on the manifold 0
We speak of a differential expression L of order N on 0 if a lin-
ear operator L:C'(O) ; C_(O) is defined such that locally(within a
chart with coordinates x) Lu is given by a differential expression
of the form II,(1.1). Using the measure du = Kdx a general (<)
second order differential operator M on 0 always may be written in
the form
(2.1) M = -1 3 Kmjk 2 + bj3 + P

K xj xk xj
with contravariant C--tensor fields m5k , b3 , of second and first
degree, and with a C--function p over 0 . This follows because for
an N-th order differential operator on 0 the local principal part
defines a contravariant tensor of N indices. For N = 2 , and any
tensor field a jk one finds that -K-1
3 Kajk 8 k= A defines a glo-
xi x
V.2. 2-nd order differential expressions 133
bal second order differential operator with principal part -ajk.

Indeed,for u,v with support in a chart one gets (u,Av) _
fa3ku kdV which is coordinate invariant, by the well known

(x7vI
x
transformation properties of tensors. For a general second order
operator M one may subtract this operator, formed with its princi-
pal part ajk = mjk to obtain an operator of order one or zero.
But every first order global differential operator is of the form
(1.5). Therefore (2.1) gives the general 2-nd order operator. For
additional symmetry and later convenience let us replace (2.1) by
(2.2) M = - Km3k2 + b38 - K-13 Kc3 + p

K xj xk xj xj
which is only formally more general, of course.
Now the comparison operator H , or rather its inverse square
root A of (1.4) may be used to convert a general first or second
order operator into an L2-bounded operator. For D of the form
(1.5) and M of the form (2.1) we formally consider the operators
(2.3) A = DA , B = AMA .
It is not hard to set up conditions insuring that A and B of (2.3)

define bounded operators of H In section 1 we already found that
.
(2.4) HAD = IIDAII < IIDIIH
where A:H + Hl , and D:H1 + H , with the strong L2-gradient in D ,
and with UDIIH of (1.6). Similarly one may look at
(2.5) (u,Mv) = f(mJku v k+ubiv +cjuv +puv)dp, u,v E C'(SZ) ,

j Ix Ixj Ixj
Ix
by a partial integration. For the integrand m of (2.5) one gets
(2.6) Im(u,v)I2 < ci h(u)h(v) , h(u) = hjku u

k
+ glul2
Ix Ix
assuming finiteness of (2.7) , with the supremum over all x E SZ
(2.7)cl= sup{(hijhklmikmjl+h]kFibk/Iql+hjkc3ck/IqI+IPI2/IgI2)1/2
by a calculation. From (2.5),(2.6),(2.7), and (1.4) we get
(2.8) I(u,Mv)I < c1Pull 111vI11 = c11IA-luUUA-1vU
which yields
(2.9) I(u,Bv)I < c1Uuffvf , u,v E A-Idom HO

Recall that A-1dom H0 is dense in H , so that (2.9) implies
(2.10) BBull < IIMIH (lull , IIMIIH = c1 (with Cl of (2.7)) ,
for all u of a dense subspace. It follows that B admits a contin-

uous extension to H , with the operator bound of (2.10).
Finally it may be observed that we may improve on the esti-
mates (2.4) and (2.10) by replacing IID0H and IIMIIH with
(2.11) inf{IID"IIH-} , int{IIM-11 H-.} ,
where the infimum is taken over all operators y-1Dy =D"', y-1My
= M` , y-1Hy = H- y E C_(S2)
, y>0 , representing D M and H
, ,
after a transformation of dependent variable, preserving the con-

dition q > 0 In fact, only one such norm needs to be finite, in
order to get A or B in L(H) .
In the next few sections we mainly shall focus on first or-

der expresssions D with finite fDIIH, so that A=DA is bounded in H.
Our aim will be the investigation of algebras generated by classes
of operators DA together with classes of multiplication operators
Au(x) = a(M)u(x) = a(x)u(x). Of crucial interest will be the com-
mutators of two generating operators [A,B] with A and B of either
form DA , or a(M) In that respect we look at commutators of the
.
forms [ a,D] , [ D, F] , [ H,a] , and [ H,D] , for a = a(M), and first

order expressions
(2.12) D = b33.
x + p , F = claxj + r
One calculates that
(2.13) [D,F] = (bjcklx. - cjbklxjxk + (b1rlxj - c3plx.)
which is another first order expression.

For the commutator [H,D] it is useful to introduce covariant
derivatives with respect to the Riemannian connection of the
rhetric tensor hjk . With the Christoffel symbols (cf.app.B)
(2.14) rji k =
hil(h .1
k + hkl j - hjk 1
Ix Ix Ix
write
bi lx. , bjIk = bjllhlk

(2.15) - bkrk.
biij =

(2.16) [H,D] -K axjK(b3lk + bklJ)8xk _ D + D b3q

p I
xj
where we define
(2.17) Da = hJka j2 k , a E C1(S2)

Ix x
and where p is defined by
(2.18) D* Kax.Ks3 + p = _S3 ax3 + p , p = P -..(Kb0) Ixj
The proof is a somewhat lengthy calculation. For D0=b3a

x3
All = K-18 jKhJk8 k write [H,D] _ -[A ,DO]-[&u, p]+[q,D0], where
x x
[q,DO] _ -b3q j . For u,v E C'(Q) we get
Ix
-(u,[Au,p]v) _ (Vu,V(pv)) - (V(pu),Vv) = (Vu,(Vp)v)
-(u,(Vp).Vv) _ (DPu,v) - (u,Dpv)

with Da as in (2.17). Accordingly,
(2.19) -[A p] = DP* - Dp
Similarly ,
-(u,[AuDOjv) _ (Vu,VDOv) - (V(DO u),Vv)
using that D0 -bO - c , c = K-1(Kb3) j

Therefore we get
IX
-(u,[DU,DO]v) (Vu,VDOv) + (V150u,Vv) + (V(cu),Vv)
(hikb3(u iv + hlku ib3 + hikb3 iu v k)dp

k Ix3 kv
Ix Ix Ix Ix Ix3 Ix Ix 3 lx
+ (Vu,cVv) + (u,Dcv)
hjkbi hik
= [(h'jbk Ix3 + - b3)u v du + (u,D v)
Ix3 Ixj Ix1 Ixk c
Now we use (2.15) to express the derivatives b3 k by covariant

Ix
derivatives. In particular we may differentiate h13hkj
= 6k
for x1 and solve for him 1 to obtain
Ix
him Ix1 = _hkmrikl - hijrm3l

(2.20)
using (2.14) . With (2.20) we get
h]Jbk +
hjkbi - hik bj
(2.21) lxJ lxJ 1X3
h17bk + hJkb1 = bilk + bkll

l7 l7
Accordingly,
(2.22) - -K-13 iK(bilk + bkli)8 k + Dc

x x
and formula (2.16) follows. Q.e.d.
Since we introduced covariant derivatives for the commutator
[H,D] we may as well use them in (2.13). This gives the formula
(2.23) [D,F] _ (bJcklj clbklj)3 k + (b7rlj - c7Plj) ,

-
x
noting that first covariant derivatives of scalars are ordinary
derivatives. The symmetry r k=tkjwas also used for (2.23).
A simple calculation also gives
*
(2.24) [D,a] = b7alj , [H,a] = Da - Da
with D of (2.17).
a
Let us note, finally, that (2.23) is valid with any conve-
nient affine connection, while (2.16) only holds for the Rieman-
nian connection of the metric tensor hik
It is easy now to set up conditions for boundedness in H of
the operators
(2.25) [D,a] , [H,a]A , [D,FIA , and AIH,D]A ,
by just arranging for finiteness of the corresponding norms (1.6)

or (2.10).
Proposition 2.2. Let the tensors bJ and cJ be bounded , together

with their first (Riemannian) co-variant derivatives. Assume that
p and r are and that Vp , Vr Vp* are ,
0 , and that Vq = 0(q) . Also let a , Va , and Aua be

bounded Then all operators (2.25) are bounded in H
.
The proof is evident, after the above calculations.

The result will have to be refined, before it becomes useful.
This we shall attempt in the sections, below.
V.3. Compactness of Commutators 137
3. Compactness criteria for commutators.
Throughout this section we shall assume a triple {Q,du,H}

given with H = q - Au of (1.1) satisfying (1.2) , and such that
condition (w) holds: the minimal operator H0 is essentially self-
adjoint (which implies that im H0 = H<(52)) is dense in H ). For
most results we also will refer to a sub-extending {S2",du",H"} (c>
{c,du,H} (which may be {52,du,H} itself) , and assume here in addi-
hjk h"jk
tion to III,(3.24) that the tensors and coincide on S2 ,
and that q" >,l outside a compact set K C 52" (we may achieve K=0
by (1.2) and III,prop.6.2, changing variablEs).In particular the
norm of a tensor is unambiguously defined, since the two metric
hjk
tensors and h"jk coincide. Generally (w) will be required for
{O,du,H} , but not for {52",du",H"} . In view of IV,thm.1.5 it is
naturalt expect q to tend to - near 30 C 0" whenever (w) holds.
Accordingly thm.3.1, lemma 3.2 and lemma 3.3 often will have the
best use for 0 = 02" , although they are formulated for general
O C 0" (Indeed the condition Va=o0(ge/2) is weaker than Va"=
oQ"(gne/2) near a point x0 E 30 , whenever q -* - as x -r x0 , since
a" and q" must stay continuous near such a point. ) On the other
hand, thm.3.4 and the lemmata following it strongly depend on
condition (3.13), below, which cannot be satisfied for q near a
point x0 E 30 C 0" unless q stays bounded there (which is impossib-
le if (w) is to hold). (This reflection assumes 0 to look like
a manifold with smooth boundary near x0 ) Note that near such
.
point x0 the conditions (3.16) still allow some degeneration of

the coefficients bi , p of a differential expression D , but weak-
er than near infinity (of 0") .
The sequence of results, below, discusses conditions on the

generating sets A# and D# implying compactness of commutators in
the comparison algebra C(A#,D#). Thm.3.l deals with commutators of
the form [a,DA], while thm.3.4 looks at [DA,FA], a E A#, D,F E D#.
We use the nctation oO(.) , etc., of app.A.
Theorem 3.1. Let aEC_(O) satisfy the condition Va=o0"(q"£/2)(in 01)
for some 0<s<l , and let the first order expression D (over 0),
of the form (2.2) satisfy b = 0(1) , p = Then we have
(3.1) [a,DA] E K(H) .
More precisely, the unbounded operator [a,DA] = aDA - DAa is

densely defined, and extends continuously to a compact operator.

Proof. For u E im HO C CD(0) write [a,DA]u = -b3a .Au+D[a,A]u
Ix
Here the first term is of the form Cu , with C E K(H), due to our
conditions on a and b (we get b1al j=on^(ne/) 0)),
x
by III,thm.3.7. For the second term we will need the two lemmata,
below.
Lemma 3.2. Let 0 < c < 1 be given, and let the function a E CW(R)
satisfy
(3.2) Va = o.^((q^)E/2)
Then the commutator [a,A2] = aA2 - A2a, defined on im HO = HC-(Q),

extends to an operator in K(H) . Moreover, there exists compact
operators C. E K(H), j=1,2, such that
AC2A2-E
(3.3) [a,A2] = A2-eC1A +
Proof. Use App.A, L.A.l to construct 0 < y E C'(0^), y=o,^(qÊ/2),
Va = 0(y) and define 6 = yl/E , so that Va = 0(Se), 6=oQ^(/

,
SEAS
i.e., E K(H), by III,cor.3.8. For u E im H0 we have u, A2u
E CO . Therefore, using (2.26), we get
aA2u - A2au = A2HaA2u - A2aHA2u = A2[H,a]A2u = A2Da*A2u - A2DaA2u
= AU-ED-A) SEA')A2-eu - A2-E(AEse)(S-EDaA)Au
Since SEAS E K(H) , and a EDa has bounded coefficients, so that

(2.6) implies 6-ED aA E L(H) , we conclude that (3.3) holds.
It is trivial then,that also [a,A2] E K(H), q.e.d.
Lemma 3.3. Let 0 < s < 2, and let the constants E,r,p,t,T satisfy
(3.4) 0 < e,r,t < 1 , 0 < p,T < 2-s , r+p ,

t+T < l+s-c .
Then, if we have (3.2) with c of (3.4) for some a E C-(0) , we get
(3.5) [a,A5] = aAs - Asa = APCIAr + AtC2AT
where Cl. C2 are operators in K(H) .
Proof. The self-adjoint operator A2 has its spectrum on [0,1], and

X)-1
its resolvent R(a) _ (A2 - is a holomorphic function from
U \ [0,1] to L(H) . Accordingly ,
(3.6) As = i/2n Jr R(A) As/2dX , s 0
with As/2 = es/2(loglal + i arg A) ,

larg al < n and the positi-
vely oriented circle r = {la - 11 = 1 }.
The integral in (3.6) exists as an improper Riemann integral

in norm convergence of L(H) Its integrand is continuous, except
.
at A = 0, where we have R(A) s/2 = O(IAI-l+s/2) as A E r . We get
[a,R(A)] = R(A)(A2-A)aR(A) - R(A)a(A2-A)R(X) = R(A)[A2,a]R(X),
as follows for u E im(H0-X) = (H-X)C-, which is also dense in H,

for all A in the resolvent set, since H0 is essentially self-ad-
joint (cf.I,thm.3.1). Then the same follows for uEH, by continuous
extension.Next we apply (3.3) with c as in (3.2), and e,r,t,p,T
satisfying (3.4) . It follows that
As/2[a,R(A)]
(3.7) = APC3(A)Ar + AtC4(A)AT I
with
C3(A) = _As/2A2-E-PR(A)C1A1-rR(A)
(3.8)
C4(A) = -as/2A1-tR(A)C2A2 E-TR(A)
Clearly C3 and C4 are compact for all A E r . Also we have
, 0 < s < 2
(3.9) AsR(A) = 0(lal-l+s/2) , A E r
from which we get
fC3H=O(lal-1+(l+s-e-r-P)/2) IIC411=0(lal-1+(l+s-e-t-T)/2).
(3.10) ,
It follows that the integrals (3.11) exists as improper Riemann

integrals in L(H) and that ,
(3.11) J Cj(A)da = Cj E K(H) , j = 3,4
with the ideal K(H) of compact operators of H. Thus (3.5) follows

if we substitute (3.6) into the commutator, and use the above.
This completes the proof of lemma 3.3. Also we note that lemma
3.3 ,for s=1, p=t=1, r=T=O, and any e<l=1+s-r-p=l+s-t-T gives
(3.12) D[a,A]u = DA(C1+C2)u = Cu , C E K(H) ,

which completes the proof of thm.3.1 as well.

Let us emphasize that the function a in thm.3.1 does not
have to be bounded. Only the derivative is restricted by (3.2).
This will be useful in thm.3.4 below for commutators of the form
/2
[qe ,A] , using the restriction (3.13) . On the other hand in
sec.4 we will discuss a class A# of function generators giving
compact commutators, using thm.3.1.
While we note that no restrictions at all on the comparison
operator H were needed for the commutator of thm.3.1, if one wants
compactness of commutators of two generators DA of the second kind
then it will be practical to require the condition
(3.13) Vq" = on"(q") .
on the 'potential' q" . No further condition on q" will be impo-

sed, and none on q as well, if {O,du,H} _ On the
{St",du",H"}
other hand, if D is a proper subdomain of Sl", then at the boundary

an of 0 the function q must be singular, in order to get (w) ,
while q" is smooth there. (Otherwise the Dirichlet and Neumann

operators are different self-adjoint extensions of H0 ; hence H0
cannot be essentially self-adjoint.) For that case we will
need a condition on the expression b3g1j, included in (3.16).
x
Note that (3.13) holds trivially for the Laplace operator.
The comparison operator 1-A not only has q=l , but also C=l,
with the function = h = det((hjk))-l. Note that E C_(0),
due to the invariance properties of K and T . Similarly we define

for the triple {Q",du",H"} For an expression D of the form
.
(2.2) the adjoint D is given by (2.20). If we want D and D to

satisfy the same type of estimates we must look at the term
(3.14) K-1(Kb3)Ixj = (-1(b0C)lj ,
with Riemannian covariant derivatives. We will require that the

coefficientsbi p of D satisfy the conditions
,
(3.15) (b3) = 0(1) , p = C-1(b3C)lj = 0(1q) (in 0)
for D of (2.2) , just to insure boundedness of D and D , i.e.,

finiteness of c1 in (2.6) for D and D* . For H=1-A the term (3.14)
equals the 'divergence' b3lj . Note that a change of dependent
variable can make either q=const or C=l, as was seen in 111,6.

For a result on compactness of commutators between two gene-
rators DA D E D# we also introduce the following conditions:
,
b01k + bk13 =
Vp = b3q = for some 0 < c < 1,

.
lx
Theorem 3.4. Let the comparison operator H" satisfy (3.13), and
let the expression D of (2.2) satisfy (3.15) and (3.16). Then the
*
operators DA , D A E L(H) (by (2.6),(3.15)) satisfy
(3.17) (DA)* - D*A E K(H) , DA - (AD)** E K(H)
where in particular AD, as an unbounded operator with domain H1

admits a continuous extension (AD)** . Moreover, for D and F of
(2.12) both satisfying (3.15) and (3.16), we get
(3.18) [DA,FA] E K(H) .
The proof again requires some preparations.
Lemma 3.5. Let D satisfy (3.16) for some 0<e<1 Then the .
commutator DA2 - A2D (defined in extends to a bounded

operator in H , which is compact. Moreover, we get
AC2A1-e
(3.19) [D,A2] = Al-cCIA = , C. = Cu(e) E K(H) , j= 1,2.
Proof. Again we get
(3.20) DA2u - A2Du = A2[H,D]A2u u E im H0 = HCO ,
where we may express [H,D] by (2.16). Accordingly, we must discuss

the operators A2XA2 with X being one of the four terms of (2.16).
First consider the term A2M0A2 , with
(3.21) MO=-K-13 K M k, mOk=b31k+bk10=9mik,

xi x
where B>0 denotes a C'(D^)-function with S=qÊ/2, xEc^\K. Write
Ml=-K-1a
(3.22) M0= 8M1- k
, Kmlka k , 8 = log a
x x x
and observe that (3.13) implies
(3.23) ck= mikO1j = e/2 mikq^lj/q^ (for xED^\K) = o,.(1) (on D).
For the second term in (3.22) we get A2-E(AEB)c33 x AA

=A2-ECA , of the proper form for (3.19). For the first term apply
lemma 3.3 with s = 1 , p = t = 0 , r = T = 1 , for a = B , noting
that VB = BV6 =o,,,.(q"E/2), i.e., (3.2) holds. We get [A,B]=CA,
with CEK(H) , hence A2M1A2 = Al-e(AEB)(AM1A)A + AC(AM1A)A ,
where the second term is of the form ACA , by (2.10) The first .
E
term is of the form Al-ECA Indeed, we get A B E L(H), by (3.2)
.
and I,lemma 5.1, and AM1A E K(H) can be proven as follows: For
6 > 0 write mil = wmil+ Xmil , where w E C'(O') is chosen such
that the tensor Xmil, with X=1-w, has norm <6. Correspondingly
write M1= M2+M3, and note that IIAM3A<6. On the other hand, M2=
wM1+ D2 , where the differential expression D2 is of the form yD4,
with D 4 A bounded and y E C;(O') It follows that AM2A = CAM1A +
.
w(AM1A)+AD2A = C+wAM1A. Accordingly, (AEB)(AM1A)=E6+AEBw(AM1A)+C,
where E6= (AEB)(AM3A) has norm = 0(6) , while the other two terms
are compact. Since 6 may be chosen arbitrarily, we get
(AEB)(AM1A) E K(H) , using that K(H) is norm closed. Thus we
have shown that A2M0A2= Al-ECA is in the proper form for (3.19).
Next let us focus on the term A2bOgI.A2=J. Using (3.16) we
may write J = A(AB)y(BA)A, where 0< B E C'(O") B = ,
is constructed according to lemma A.1 such that y=O(1) This .
shows that J=ACA , with compact C , using III, thm.3.7

Finally, for the two terms A2D A2 and A2DP A2, it is suffi-
p
cient to look at J'= A2DpA2 only, for symmetry reasons. With B as
in the discussion of A2M0A2 and, similarly, y=q1/2 in O\K, O<y E

C'(S2) we write Dp= ByD3 , where D3 has coefficients o,,.(l) , by
the third formula (3.16). Then, J'= A2ByD3A2 = ABAyD3A2+ ACAyD3A2,
=Al-E(AEB)(AyDaA)A + Al-E(AEB)C(AyD3A)A = Al-ECA , using the same
techniques over again. In particular we used lemma 3.3 on B again.
Note finally that we also may push the factor B to the right :
BYD3 = yD3B -c/2 y(D36)B, with the function 6= log q" again. Thus
we also get J'= ACA1-e. This completes the proof of lemma 3.5.
Lemma 3.6. Under the assumptions of lemma 3.5 we have
(3.24) [DA,As] = ArCAP , 0 < s < 2 , C = Cr02sE K(H)
for every pair r,p of reals satisfying

(3.25) r+p+c<l+s, and either 0<r+E<l, 0<p<2, or O<r<l, 0<p+E<2
Proof. We substitute (3.6) into the commutator of (3.24) for
[DA,As] = i/2Tr )
Xs/2R()[A2,D]AR(a) dA
r
Now (3.19) may be used in the integrand I(X) for ,
I(X) _ X, /2 A1-ER(A)C1 A2R(a) = ArC3(a)AO , with
C3(X) _ Xs/2A1-r-ER(A)C1R(a)A2-p = Xs/2A1-rR(a)C2R(a)A2-p-E
Using (3.25) and the appropriate (3.19) we get C3(a) E K(H)

I(X) = O(Iail+(1+s-r-p-E)/2)
on r Thus (3.24) follows as
.
earlier (3.5) The proof is complete.

.
Lemma 3.7. For some 0<E<1, let the two differential expressions
D,F , as in (2.12), satisfy the conditions
(3.26) (qE/2) c5plj-oQ^(q(1+E)/2),

bJcklj - cObkWOO- , bOrlj -
Then the operator A[D,F]A is bounded in H, and may be written as
(3.27) A[D,F]A = Al-EC1
with a compact operator C1 .
Proof. Use lemma A.1 to construct a function 0<6 E C°°(Y^) with

6 = such that $-1[D,F]A is bounded, using (2.23), (2.6)
and (3.26). Therefore A[D,F]A = A1-E(AES)(S-1[D,F]A) =Al-EC1, with
C E K(H), using III, cor.3.8.
1
Proof of theorem 3.4. For D satisfying (3.15),(3.16) we apply
lemma 3.6 with s=p=l , r=0 , for [DA,A] = DA2 - ADA = CA . This
implies DAu = ADu + Cu , u E H1= im A . In particular AD with
domain H , must have a continuous extension (AD) E L(H) and ,
DA = (AD)**+ C follows. Taking adjoints we get (DA)* = DA + C

so that (3.17) follows. For (3.18) use (3.27) for E=l and the
**
second (3.17) for DDFAu = (AD) FAu + Cu = Cu + ADFAu = Cu + AFDAu
+ C'u = C"u + (AF)DAu = FADAu + C"'u for all u c HC' , with C, ,
C', C", C"' E K(H) , completing the proof of the theorem.

V.4. Compact commutator algebras 144
4. Comparison algebras with compact commutators.
In this section we will discuss some concrete cases of com-

parison algebras with compact commutators. Since all the discus-
sion in sec.3 is based on the validity of condition (w) we first
will engage one of the results of ch.4.
First assume that the assumptions of IV, thm.1.5 hold for
m=l. In details, let O^ be a smooth complete Riemannian manifold
without boundary, with a positive Cm-measure dm^=K^dx and let ,
the CW-function q^ be > 1 outside some compact set K , and satisfy

condition (3.13) , i.e. (with the limit in c^),
q^ k/q^2=lima
(4.1) h^3k(log q^) (log q^) k=0.
Ixj Ix IxJ Ix
We assume that O^ has a countable locally finite atlas.

Our manifold 0 is a subdomain of 0^ , and its boundary 30 is
assumed to be a smooth compact n-l-dimensional submanifold of 0^
On 0 we have the restricted measure du = duÎO = Kdx , and the
restriction to 0 of the (contravariant) metric tensor is called
hjk
A comparison operator H is introduced by
.
(4.2) H = -K-12 Khjk3 k + q

X] x
with a real-valued Cm-function q>q^. Then condition (w) for the
triple {0,dp,H} results if we require condition IV, (1.4), for q
(4.3) q(x) > yl/dist(x,30)2 , for all x E Q\K' ,
with a suitable compact set K'CO, where yl is a certain positive

constant defined in IV,thm.l.5.
Now the results of sec.3 suggest introduction of the follo-
wing generating sets A# and D# .
A# is defined as the class of all bounded functions a E

c
C,(c2), such that Va = o0(ge 2) , with some 0 < e < 1 (That is, .
the C-()-function q e/2hOka a k is bounded over O\K , and has

Ixj Ix
limit zero, as x- (in 0) ) .
D# is defined as the set of all first order partial diffe-

rential expressions D=b03 j+p with Cam-coefficients, defined over
x
0 satisfying (3.15),(3.16) with some 0<e<l
, . That is, in detail,
the 3 complex-valued C-(O)-functions
(4.4) h. jbk , p/Vq , K-1(Kb3) j/i

Ix
are bounded over S2 , and also, the 4 C-(S2)-functions (with C
q"-6(h..,h ,(bjlk+bklj)(bj'lk'+bk1Ij1) , bi(log q)

(4.5) » kk Ixj
4 1q" ehikPlx-Plxk
also are bounded over 0 (or at least R\K ), and have limit zero
as x- (in 12") .
It is trivial that A# and D# satisfy the general assump-

tions (a0), (a1), (d0) of sec.l. In addition D# contains its
formal Hilbert space adjoints.
Now thm.4.1, below, is an immediate consequence of thm.3.1,
thm.3.4, and IV,thm.1.5. (We use thm.3.1 for S2=S2", but thm.4.4
with S2" as above.)
Theorem 4.1. Under the above assumptions the comparison algebra

C ,D#) of section 1, formed with the above classes A c D
C c c
contains the class K(H) , and has compact commutators. Moreover,
C is generated by a E A# , DA , D E D# alone, i.e., the generators
(DA) may be omitted.
Let us shortly focus on some special cases onto which thm.
4.1 applies. In particular, for technical reasons, one often looks
at the algebras of smaller generating sets. The above sets A# D#
may be regarded as the largest practical such sets, allowing for
compact commutators.
Example (A): Q=In, dp=dx, H=1-A, j2, (i.e., the Euclidean

x
Laplace comparison triple of In) . This case has been studied
extensively in [Cl], ch.III,IV We have condition (s), since In
.
is a complete space (IV, cor.1.7). Since q(x)El is a constant, we

also get q=q"=qE=qêEl , so that the number e becomes redundant.
AO consists of all bounded Co-functions with gradient vanishing at
infinity. The largest comparison algebra of [C1],IV,l , called
there, uses as D# the set of all folpde's with coefficients
r'V
in A A calculation confirms that the class D# properly con-
tains D# .
In [C1] we obtain precise knowledge about the 'symbol space'

(i.e.,the maximal ideal space of (/K(H) ).
Problem 4.2. Use the techniques of [C1] to characterize the symbol

space of C(A#,Do)
Example (B): The Laplace comparison algebra of the Euclidean Rn
proves particularly simple because this space O=Rn is a locally
compact abelian group (under translation). In [C1] we essentially
use the Fourier transform for most of the proofs. Of course, the
Fourier transform appears as the Plancherel transform of the
translation group, which explains the simplicity of the case.
Other cases of manifolds which are translation groups, are
Mjk
just as simple. For example, let p = = RjxTk, j+k=n, with the
circle T = 1/(2rZ) For the simplest case of j=k=l, n=2 we obtain
.
the 2-dimensional straight circular cylinder.

One may introduce global coordinates x=(xl,...,xn) , appoin-
ting that functions on M3k must be 2r-periodic in the last k var-
iables. then a comparison triple again is given by the same dp=dx,
H=1-A as under (A)
, .
In [CS],V,2 we investigate the comparison algebra C(A#,D#),

where D# again is a proper subset of D (defined as under (A)).
Clearly our present triple satisfies (s), as the Laplace compari-
son triple of a complete Riemannian manifold, by IV,cor.1.7 .
There is an important difference between the symbol spaces for

R2 (as under (A)) and the cylinder M1'1 , which will have to be
discussed later on (cf.VII ,2, VIII,2.)
Example (C): 0=Rn, du=dx, with the expression (where a,6 E R, a>O)
2R3
(4.6) H = (x) 2a - div( x) 2Sgrad = (x) 2a j(x) j
x x
(x) =(l+Ixl2)1/2. We have hjk = (x) 286jk , hence hjk= (x) -2a6jk
Proposition 4.3. The manifold O=Rn , with metric tensor
(x)-26
hjk = 6 is complete if and only if S<1 .
jk
Proof. For any smooth curve {x(t):tEJ) we obtain as arc length
(4.7) 1= J((x(t))-2SIx,(t)I2)l/2dt = J(x)-slx*(t)Idt .

J J
For a moment define the real-valued function E(t)=Ix(t)I over the
interval J We have
.
(4.8) Ix'(t)I
Accordingly, if J=(p,p') , we get
EV(PP
(4.9) < JJ( <11
This proves that the geodesic distance between two point x , x- is
bounded below by the integral p=lxl, p_=Ix-J. This

P
distance tends to - as x-->- if and only if J_(C)-Bd

0
This happens precisely for B<1 . Also (4.8) implies that

iE-(t)i _ ix.(t)l whenever the curve runs in a radial direction.
Accordingly, if 0>1 then indeed all curves
, {tx0 O<t<co}} :
have finite length, so that then 0 cannot be complete. Q.E.D.

Considering the expression (4.6) we thus have cdn.(s) when-
ever 6<1 , by IV, cor.1.7. On the other hand, for a>1 we will
change the dependent variable by setting with X>O to
be chosen. This transformation yields a new expression H"' with
potential given by III,(6.6). That is, with B"=B-l ,
q` = q - (x) Adiv ((x) 2 B grad (x))

(4.10)
x) 2a + X(n+2B"-X)(x)2B -
Let us first consider the case B>a+l Then the second term .
at right is leading, as lxi is large. Its coefficient becomes

largest if we choose X=$-+n/2 = B-l+n/2 . Now, although we no
longer have a complete Riemannian manifold we still may apply
IV, thm.l.l.
The proper choice of the function a will be a(x)=T log(x)
with suitable T>0 . We then get
T2lxl2(x)28--2<
(4.11) hjko a k= q-, as T=("-E+n/2), any c>O,

x Ix
(at least outside a large compact set, which is sufficient). We
trivially have a(0)=0 , and (ii) of thm.l.l holds, since T>0
Checking on (iv)n , we find that
(4.12) e- naiVal =
0((x)B--n(B E+n/2)) = 0(1)
whenever B""<n0"'+n/2) , for sufficiently small E>0 . In other

words, we must require that
(4.13) n > B"'/(B"+n/2) .
Notice that the right hand side of (4.13) is <1 , for all B"">0 ,
while that expression tends to zero, as B'"+ 0 and $->0 . Thus we

always can satisfy (iv)n for n<l enough for (w), for every
,
choice of 8 whatsoever.
So far we have looked at the case a<$- only. If a=B-, then
in (4.11) we may set T=((6"-e+n/2)2+1)1/2 . Similarly, for a>$- we
can allow any choice of T . Correspondingly (4.13) improves to
(4.14) n>S-/(1+(0-+n/2)2)l/2, as a=8 , n>O , as a>S`
We have proven the result,below.
Theorem 4.4. For a comparison operator of the form (4.6) we always

have cdn.(w) satisfied, regardless of the choice of a and 8.
Moreover, we have cdn.(s) for all S<1 , regardless of the choice
of a . If 0<a<0-1, we at least have cdn.(w). As $ decreases from
to +1 , we will have cdn.'s(sk) for larger and larger k , such
that for every integer k0>0 there exists Sk >1 such that (sk )
0 0
holds for every S<Sk

0
For a concrete example we now set q"=1, and then, with A#,
D# as above, introduce the classes A#=A# , and
(4.15) D#={DEV#: (x)-Sb7EA#, P=O((x)a'), p j=o((x)-a+a))

1x
The symbol spaces for some such algebras have indeed been
worked out by Sohrab [S3] . We shall discuss this in VII,thm.4.8.
Example (D): This example, and example (E), below, only point to
special cases of thm.4.1 where the various conditions can be
translated into a simpler (or different) form. None of the alge-
bras have been investigated in generality, regarding their symbol
or symbol space.
Let 52 be an open subdomain of In with compact connected
smooth boundary 30 , where 2S2 is an n-l-dimensional smooth sub-
manifold of In Let H = q - A , with the Euclidean Laplace opera-
.
tor A , above. Here the real-valued Co-potential q>l must satisfy

(4.3), above, with the Euclidean distance. Set S2" = I dV = du" ,
= dx, H^ = 1-A, to get all conditions satisfied, including (4.1).

In other words, we have 0 either the interior or the exte-
rior of a compact subdomain of In
Now the class A# consists of all bounded C'(c2)-functions a
c
such that, for suitable e , 0<e<l and all j=l,...,n, we have
,
alxj = o(qe/2) in S2 That is, we have

. <6 outside
lalxj/qe/21
a suitable compact subset K6 of 0 , for every 6>0 .
Also the class D# consists of all (globally defined) D=b33

xj
+ p , with b3 , j=l,..,n , p E C'(0) and with the functions
b3, p/v, bjI j// bounded over 0 , j=l,...,n , and with

x
k
b3
Ixk +b , b3q , p Ixj/,q , (bl l) Ixj/V-q , all, = 0(1),
Ixj Ixj/q I x
for j,k=l,...,n, in the sense described for A#

As a consequence of thm.4.1 we then get a comparison algebra
with compact commutators, containing K(N) . The symbol space of
such an algebra seems to be unexplored.
Example (E): Let 0 be the interior of a complete Riemannian
manifold QU3M with compact boundary 30 It is known that then .
0U3M may be imbedded, in the sense required for thm.5.1, into a

manifold Q^ which is a doubling of 0 (i.e., in two copies of OLGO
one identifies corresponding points of M. using as chart of a
boundary point the union of the chart of OU30 in the half space
{x(=-In: xn>0} together with its reflection into {xn<0} ).
Assuming the metric tensor hjk E C-(0U3O) , we also may

double h3k i.e., reflect it at the boundary. However, the
,
resulting extension in general is no longer continuous at 30

On the other hand, it is easy to arrange for a smooth transition,
accross 30 , by modifying the reflected h3k in a strip along O
"below O " (cf. also the proof of III,thm.2.2) .
Similarly for a measure dp on 0 Furthermore, if a poten- .
tial q is given on 0 (normally singular on 30), it also may be re-

flected and suitably modified in a (two-sided) neighbourhood of
30 in the doubled manifold, called 0^ , to obtain a q^ E C-(92^)
In this way, if a triple {0,dp,H} is given, on the above 0
it is possible to satisfy all assumptions of thm.4.1, regarding
the imbedding into 0^ , with a sub-extending triple {O^,dp",H^}.
We must require that q satisfies (4.3), and that Vq/q = o(l) in
Q\K , with a compact neighbourhood of O in cu50 in order to ,
insure that (4.1) can be satisfied.

This shows that the case of a manifold with boundary is
naturally included in our discussion.
Example (F): Finally we point to the case of 0 _ In, with du=dx,
and a (globally defined) comparison operator H, under the assump-
tions of IV, thm.1.9 .
V.S. One dimensional problems 150
In this case we will assume S2" = S2 = In , and q"=l . Then we

again have all assumptions of thm.4.1 satisfied, for the classes
A# , D# . Again, the structure of the corresponding symbol space
seems to be unexplored.
5. A discussion of one-dimensional problems.
In this section we get restricted to the case of a 1-dimen-

sional manifold 0. That is, our comparison triple assumes the form
(5.1) {O,dp,H} _ {I,Kdx,K-1(-d/dx p d/dx + q)}
with a finite or infinite open interval I = {a < x < g} # 0

and globally defined Cam-functions K(x) p(x) q(x) , x E I , , ,
K(x) , p(x) > 0 , as x E I .

In this case the two normal forms of III, prop.'s 6.1 and
6.2 may be further simplified by also introducing a special inde-
pendent variable . A simple calculation,left to the reader, gives
the following result.
Proposition 5.1. a)If new dependent and independent variables

(v,y) instead of (u,x) are introduced by setting
(5.2)u=yv, Y=(pK)-1/4 y(x)=JX dt, a-=limx;ay(x), R-=limx}6y(x)

0
with any given x0 E I then the triple (5.1) assumes the form
(5.3) {S2-,du-,H'} _ {1-,dy,-d 2/dy2 + q-}
with I` _ (a`,B-), and q"' of the form, with Y-=Y-1=(pK)1/4(x(y)),
(5.4) q-(y) = (q/K)(x(y)) - (d2Y /dy2/Y )(y) = Hy/y(x(y)) .
b) Let y E C'(I) be a positive solution of the differential

equation Hy y (which exists according to III, prop.6.2, since I is
noncompact, and the minimal operator H0 has the lower bound 1)
Then, if new dependent and independent variables are introduced
by setting
(5.5) u = Yv , y(x) = JX K(t)y2(t)dt

0
then the triple (5.1) assumes the form
(5.6) {I-,dy,H-1 _ {(a-,s-),dy,-d/dyp-d/dy+1 } ,

with
(5.7) p'(y) = (Kpy4)(x(y)) , a- = limx+ay(x) limx+sy(x).
Generally a transformation of dependent and independent

variable of the form
(5.8) u = yv , y = fi(x) , 4'(x) > 0 ,
with a positive Cm-function y and an invertible C'-map :I + I-

may be applied to the triple (5.1) as well as to any first order
differential operator
(5.9) F = bD + c , b,c E C_(I) , D = d/dx ,
and the Hilbert spaces H and H1 .
In view of proposition 5.1 it is no loss of generality to

depart from a triple in Sturm-Liouville normal form :
(5.10) I = (a,8) , du = dx , H D2 + q , Dx= d/dx
The proposed change of variables will carry (5.10) into

,
I- _ (a",$-) a- = 4(a+0) , S- _ (6-0) ,
dp-
= K-dy , K_(Y) =
(5.11) H"' -K--1 DyK-p-DY + q- , F` = b-Dy + c- , Dy = d/dy ,
p"'(y) _ q-(y) =
b_(y) = c-(y) _ (c + by'/y)(4(y)) .
All this assumes that the transformation amounts to an isometry

H « H`=L2(I`,dp"'), carrying the equations Hu=f Fv=g into H-u'=f"',
F-v"'=g"' , with the transformed functions u-,V",f-,g- of u,v,f,g,
respectively.
While we will work with the triple in the form (5.10) we
also note that our conditions on the coefficients of H and F will
not remain invariant under a transformation (5.8). Hence we shall
k
tend to repeat the construction of a C -algebra with compact com-
mutator of the preceeding sections,while trying to optimize these
V.5. One dimensional problems 152
conditions, or else trying to bring them onto a simpler form.

First the requirement of condition (w) is invariant under
(5.8) as well as that of (s) or (sk because (5.8) amounts to
a unitary transformation U: H + H- taking the operator H into
,
H- = U-1HU, and C_(I-) . An application of IV, thm.1.8

to (5.10) yields the following
Proposition 5.2. a) If a = -- , $ = -, then the assumed semi-boun-

dedness of H alone implies that (s) = (so) holds.
b) If a or S (or both) are finite then there exists a nonde-
creasing sequence nk , k = 1,2,... , such that (sk) holds whenever
(5.12) q(x) > nk(x - a)-2 near a, or q(x) > nk(x - S)-2, near B ,
respectively. (Both for 2 finite ends). Specifically (w)=(s1)
holds for q with (5.12), for nl = 12 , and (s) _ (s,) holds for
(5.13) (x-a) 2/q(x)=o(1) , as x+a , or, (x-a) 2/q(x)=o(1) . as x-5,
respectively. (Both conditions, if a and $ both are finite, but if

only one end is finite then the infinite end never requires a spe-
cial condition,except the semi-boundedness of the operator).
From (5.11) and (1.7) we know that the operator FA , with
A = H-1/2 is bounded whenever, for some y>0, y E C-(I)
,
(5.14) b = 0(1) , c + bY'/y = O(/q--ywTy) .
This condition may be simplified by introducing the function
(5.15) 6 = Y'/Y , 6' + 62 = Y"/Y , Y = Y(x0) exp(JX 6(t)dt)

0
Proposition 5.3. The operator A = (FA)** is in L(H) whenever there

exists a real-valued function 6 E C'() such that
(5.16) b = 0(1) , c + b6 = O(V`q-77 67)
The proof is immediate. Note that change of coordinates does

not give an improvement of conditions for b. To evaluate the
second condition (5.16) we assume that the operator bound, hence
the 0(.) -constant in (5.16) is 1 , as always can be achieved by
multiplication with a positive constant. Then (5.16)2 is equiva-
lent to the Riccati type differential inequality
6' + 26 Re(cb) + (l+Ibl2)62 < q - Icl2 .

(5.17)
For example, consider the case of b-0 , where (5.17) yields
(5.18) 0 < icl2 < q - 6' - 62 .
For 6 = 0 (5.18) gives the trivial condition IcI <
A more general 6 will improve this only if
(5.19) 6' + 62 < 0
over the interval I or at least near one of the endpoints.

,
(or in a sequence of intervals clustering at an endpoint), in view

of the fact that a C_(I) -coefficient b or c always gives a boun-
ded operator FA In fact, q0= -(6+62) also is the improvement of
.
the potential q under the change of variable (with 6=Y'/Y)

Notice that a real-valued 6 satisfying (5.19) in some open
interval J must be a decreasing function of x which can vanish at
most once.At its only possible zero we have 6' < 0 On the other .
hand for all x E J with 6 9 0 , (5.19) amounts to
(5.20) y' = (1/6)' = - 6'/62 > 1 , for y(x) = 1/6(x) .
Consider a right endpoint B=co , for example. If 6(x0)<0 ,
for some x0 , then 1/6 , having slope > 1 , by (5.20), must get
zero after a finite x-increase, leading to a contradiction. Hence
6>0 for all x>x0, x0 sufficiently large. Then (5.20) yields
1/6 - 1/60 > x-x0 for all x>x0 , with 60=6(x0)
, Hence, .
(5.21) 0 < 6(x) < 60/(1+60(x-x0)) , for all x > x0 ,
so that 6-0 , as x- . One then concludes that also

x+h
(5.22) (6'+62)dti < 60/(1+60(x-x0))+h60/(1+60(x-x0))2 , h>0.
X
Accordingly, while the possible improvement of q may be
large at some specific point, it will become small in the average,
over a finite interval (x,x+h) , as x-* , and can be significant
only if q+0 , as x-*0 . Since we get q_l in suitable coordinates,
it follows that q, in the average cannot be too different from 1
Accordingly the improvement of the boundedness estimates by coor-
dinate changes may be of little use, if the interval is (--,+°)
On the other hand, if the right end B of the interval I=S2
is finite, then one may think of possibly significant changes of
q by change of coordinates.
V.5. One dimensional problems 154
Example 5.4. The choice 6(x) = (B-x)A , 0<A<l , near B -i.e.,
(5.23) y(x) = exp(-(B-x) X+1/(X+1)) , near B x < B ,
gives the additive improvement
(B-X) 2X
(5.24) q0(x) = X(B-x) X-1 - = A(B-x) A-1 x.B
Similarly 6(x) = -(B-x)X , -1<X<0 , near B , resulting in

the even faster increasing
(B-x)2a
(5.25) q0(x) _ IXI(R-x)X-1 + = IAI(B-x)A , xy(B-x)-2,

which again makes the possible gain by either (5.24) or (5.25)
appear insignificant.
Coming to some more explicit examples in 1 dimension, let us
examine examples (A) and (C) of sec. 4 in this connection. Both
,
give some typical 1-dimensional examples. While (A) already is

in Sturm-Liouville normal form, (C) is not. Its Sturm-Liouville
transform is very similar to that used in the proof of thm.4.4, in
the case of B>1 We will take up the discussion of the correspon-
.
ding algebras C in VII,'+. Also there we will discuss certain more

general examples of Sohrab, around the quantum mechanical harmonic
and anharmonic oscillator equation in one dimension.
Finally, in this section, consider the following:
Example (G): Define
(5.26) du=dx , H=1-Dx2 , Dx=d/dx
Let A# be the class of all C'(1)-functions
(5.27) a=b+c, b,cEC'(1n), where bECS(1), and c is 2ir-periodic
Here CS(I) denotes the class of continuous functions over I with

limits at +- , and at -- , where the limits need not to be equal.
Also, D# is the class of all folpde's with coefficients in A#
Here we want to show that the corresponding algebra C(A#,D#)
does not have compact commutators. Indeed, consider the commutator
(5.28) [ elx,DA] _ [ elx,s(D)] , s(x) = x/(x)

F-
If we conjugate with the inverse Fourier transform we get
(5.29) Fs(D)F-1 = s(x) ,

FeixF 1 = e-iD = T-1
with the translation operator T-1u(x) = u(x-1) . Therefore,
(5.30) [e1X,DA] = (V-1s)(D)e1X , V-1s(x)=s(x-l)-s(x)
eix
This is not a compact operator, since multiplication by is
a unitary map of H=L2(1) , while the function V-Is(x) does not
vanish identically, so that V-1s(D) cannot be compact.
Similarly we get (for j=0,±l,±2,... , with X(x)=(x)-
(5.31) [e'j',DA] = V-js(D)e1jx , [e13X,A] =V-jX(D)e13X .
where Vjb(x)=b(x+j)-b(x) . On the other hand, for aECS(E we get
(5.32) [a(x),DA] E K(H) , [a(x),A] E K(H) ,
(cf.[C1] , ch.III , or also by application of sec.3, above.)
Proposition 5.5. The commutator ideal E of the present algebra C

is the norm closure of the algebra of all operators of the form
X X-()lj=-Na-.(D)e13X + K, where
(5.33)
ajE CO(s), K E K(H), X1 for ±x?y »1, =0 for ±x<y-1.
Proof. It is clear that all commutators of generators of C are of

the form (5.33), by the above. Any product of an operator (5.33)
with a generator of the algebra again is of that form. Especially,
a product a(D)b(x) or b(x)a(D) for a,b E CO(I) , is compact, by
,
[C1],III,L.8.1, or III, thm.3.7. By the Stone-Weierstrass theorem

one concludes that every operator a(D)e1JX is in E , for arbitrary
aECO. Similarly E contains the entire compact ideal, by lemma 1.1.
In VIII,l we will look at this algebra again, trying to
determine its symbol space and ideal chain.
6. An expansion for expressions within reach of an algebra C.
In the present section we shall prepare relating a differen-

tial expression L to a comparison algebra in a manner which will
allow us to connect an invertibility or Fredholm statement of a
certain realization of L to a corresponding statement involving
a bounded operator in the algebra. The general technique appears
to be common in theory of singular integral operators (cf. Calde-
V.6. Expressions within reach 156
ron and Zygmund [CZ2], for example).

Given a comparison triple {D,dp,H) on a manifold D , and
any subalgebra A of L(H), H=L2(D,dp). Let us assume cdn.(s) for
all of the present section.
Definition 6.1. A differential expression L of order < N (for

some integer N=0,1,...,) is said to be within comparison reach
(or shortly 'within reach') of the algebra A if the unbounded
A=H-1/2
operator L0AN , with the minimal operator of L and
has a continuous extension A to H, where A E A .
More precisely we should speak of an 'N-reach' , since the

condition obviously may depend on N. We will omit this distinc-
tion, however, since the choice of N always will be clear.
In particular we note that cdn.(s) implies that the space
dom LOAN = A-NC-(G) is dense in H Indeed, for even N=2k we get
.
A-NCO
= HkC' = im H dense in H , since H is > 1 and essentially
self-adjoint, using I,thm.3.1. For odd N=2k-1, if fEH, (f,A-N0=0

for all uECO, then we set f=A-lg , using that A-1 is self-adjoint
and >1, hence im A-1=H . It follows that (A-1g,A-Nu)=(g,A-N-1 u)=
=(g,H0u)= 0 for all uE dom H0 , hence again g=0 and f=0, so that
A-NCO
again is dense.
A-NCO
Since all spaces are dense it follows that L of order
N is within reach of L(H) whenever we get, with a constant c,
(6.1) HLup < cIIA-Nul for all u E C_(G).
It is trivial that all 0-order expressions (i.e. multipli-

cations) aE A# and all folpdes DED# are within reach of the com-
parison algebra C= C(A#,D#), by definition of C. In VI,3 we shall
see that every differential expression of compact support is wit-
hin reach of the minimal comparison algebra. The general question
for expressions within reach of a comparison algebra C is vital
for our applications and will be studied in IX,3 and X,2.
For a differential expression L of order N within reach of
a comparison algebra C there is a natural realization Z defined:
Definition 6.2. Let AEC be the continuous extension of L0AN to H

requested in def.6.1. Then we set
(6.2) dom Z = im AN = dom A_N , Zu = AA-Nu , u E dom Z
i.e., Z=AA-N, in the sense of I,(1.3).

Remark 6.3. The space im AN = dom A-N= HN is commonly referred to

as the (L2-) Sobolev space of order N of the triple {D,du,H}
(assuming cdn.(s), as we presently do). Here HN is a Hilbert space
under the graph norm I,(2.9) of the (closed) self-adjoint operator
A_N . Since A-N>l it follows that the graph norm is equivalent to
(6.3) IIuIIN = 11A-Nun , u E HN ,
called the Sobolev norm of order N . With this norm the operator
A-N:HN. H becomes an isometry between the Hilbert spaces H and H
Then the unbounded operator Z E P(H) may just as well be regarded
as a bounded operator of L(HN,H) There exists an inverse Z-1 E
.
L(H,HN) if and only if A-1 exists in L(H) Similarly the unboun-

.
ded operator Z admits a bounded closed inverse if and only if

A E L(H) is invertible. We also set HZnHN and note that all L and
ANtake H.->H.. For more details cf.VI,3 and X,2.
In order to facilitate a later discussion of expressions
within reach and their realization Z we next will discuss a cer-
tain Taylor type expansion with integral remainder, valid for
certain expressions within reach.
Definition 6.4. A (C--)differential expression L of order <N is

said to be H-compatible (or said to satisfy cdn.(p)) if for all
j=0,1,2,... the iterated commutators
(6.4) (ad H)0L=L, (ad H)OL = [H,[H,...[H,L]..] , with j commutators,

are within reach of L(H) (as expressions of order N+j).
Proposition 6.5. Let again R(a) _ (A2-a)-lE L(H), and let L be

H-compatible. Then the operator A=LA NE L(H) satisfies the identity
R MA - AR(A) a3-'((ad H)OL)A23+N+2Rj+1(A)

(6.5)
+ XMA2R(a)((ad H)M+1L)A2M+N+2RN+1(a) , x E r ,
for all M=0,1,2,.... .
Proof. We get
(6.6) R(a)Lu - LR(A)u = A2R(X)[H,L]A2R(a)u , u E im(l-AH0)
as used before (cf. sec.3) , and may extend this to u E H. .
Note that (6.6) amounts to (6.5) for N=O. Suppose (6.5) holds
for N>O then write A2R(A)=1+AR(A) , and LN+l = (ad H)N+1L

for a moment. For u E im(1-AH0) we get,
A2R(X)u = (1+XR(X))LN+lu
= LN+l(1+R(A))u
(6.7)
+ X(R(a)LN+lu-LN+lR(X)u) = LN+lA2R(X)u+AA2R(A)LN+2A2R(A)u ,
and again we may extend this to H. . Substituting (6.7) into

(6.5) yields (6.5) for N+1 q.e.d. ,
Proposition 6.6. In the notation of lemma 3.3 we have
(6.8) Jr A2JR]+1(A),s/2da = 2rti(-l)j+l(s,2)As , s>0, j=0,1,...,
with binomial coefficients Here the integral converges in

norm convergence of L(H) as an improper Riemann integral.
,
Proof. Integrate by parts in formula (3.6), using that

Rk+l(a) .
d/dXRk(A) = k
Proposition 6.7. With the notations of prop.6.5 we have
(6.9) AsA =
(-1)3(s/2+j-1) ((ad H)OL)As+2j+N + RM
0
where
(6.10) RM = i/27r JrA2R(a)((ad H)M+lL)A2M+2+NRM+1(X)Xs/2+MdA
Proof. Multiply (6.5) by i/(2n)as/2 , and integrate over r

Note that the integral may be interchanged with the operators
(ad H)3L = Lj , which are preclosed. Then use (6.8) to evaluate
some of the integrals.
Theorem 6.8. For all (real or complex) s with -2M-2<Re s < M+l
M=1,2,..., and A D LAN , with an H-compatible expression L of
order <N we have the expansion
M=1(-1)3(s/2]3-1)(LjAN+20)
(6.11) AsAA-s - A = + SM,s
where Lj = (ad HA E LN+j, and the 'remainder' SM's is given by
-2'rtiS M s = fl, XC-1dXSl(a)(ALM+aAN+M+l)S2 s(a)

>
(6.12) >
SM s(a) = AM+1/2+s/2-eAM+l-sRN+l() , 0<e<l.

Here the two L(H)-valued functions Sj(A) of A , A , (and s) are

bounded and regular analytic for A E r\{0} -2M-2+2e < s < M+l ,
as e is kept fixed. The integral in (6.12) converges in L(H),

as an improper Riemann integral of a continuous integrand, even
if the constant operator AAM+lAN+M+l is replaced by an arbitrary
operator VEL(H) . Moreover, we even have the functions
(6.13) SM s(a)A s-M-1

2
, as s>0 ,
2
and SM s()A 2e-M-1 as s<2
> >
bounded and analytic on the (A,s)-set specified.

We should emphasize that, by writing (6.11) we intend to
ASAA-s
imply that the (product of unbounded) operator(s) is
bounded, and that its continuous extension to H should be taken
in its place.
Remark 6.9. It is an immediate consequence of thm.6.7 that the

remainder SM s is a holomorphic function of s (with values in
L(H) ), defined in the strip -2M-2 < Re s < M+1 and that even ,
the operators SM's AS-M-1, for 0<s<M+l and SMM's

S A
AE-M-1 ,
, for
-2M-2+e<s<e are in L(H) . Moreover, it should be observed
,
that the two factors S1(A) and S2 (A) both are functions of A
hence commute with every function of A
The proof of thm.6.8 is a matter of analytic continuation
of the identities (6.9), (6.10) from the positive real s-axis
onto the s-strip specified. First we get (6.11) for 0<s<M+1, by
multiplying (6.9) from the right by A-S Next we note that all
.
terms at right are entire functions of s except the remainder.

On the other hand, the remainder may formally be written, as
specified in (6.12), where the function Si are analytic. Also,
boundedness of the functions S! , near their singularity a=0 is
a consequence of estimate (3.9). Thus one concludes that indeed
the remainder integral admits an analytic extension into the
strip -2M-2< Re s < M+l Finally, we at least have the complex-
.
valued function ((ASAA-S - A)u,u) _ 4(s) analytic in the strip,

for every fixed uEH. . One thus concludes that (6.11),(6.12)
holds in the same inner product sense. This, in turn, implies
ASAA-s
H-boundedness of the pre-closed operator for s in the
strip. The remaining estimates follow as part of the above proof,
or are trivial consequences.
Remark 6.3'. Note that the above definition of Sobolev space is
meaningful not only for integers NO but also for general s>0 .
Similarly, for s<0 one defines the Sobolev space H as the comple-
s
tion of H under the norm Ilups IIA-sup (note that now A- ELM).
=
For a more detailed investigation of the spaces Hs cf.IX,l.

Here we only note that the spaces {Hs:seU} define an HS-chain in
the sense of 1,6. Then thm.6.8 above has the following consequence
Corollary 6.10. The operator LOAN, for an H-compatible differen-

tial expression of order <N, not only admits an extension A to
L(H) but the restriction AIH. extends to an operator As E L(Hs).
In other words, A (or AIH_) is of order 0 in the sense of 1,6 .
Also, for every s,N there exists a constant c and an integer

M>0 such that the operator norm in L(Hs) of As is bounded by
(6.14) c I IILkAN+2k II , Lk = (ad H)kL

0<k<M
for all H-compatible expressions L of order <N
The proof is an immediate consequence of thm.6.8: We must
show that IIL0ANups = IIA-sL0ANup < cllups = cIIA-sull or, with ,
A-s-NCO
v=A-su , IIA-s(LOAN)Asvll < cllvp for v E , That, exactly,
follows from our expansion (6.11) if only we choose M large
,
enough to include the real number s in the strip of analyticity,

defined in the thm. In particular we find that L.AN+j E L(H)
and the remainder S as well, in view of the properties stated.
Ms
s
CHAPTER 6. MINIMAL COMPARISON ALGEBRA AND WAVE FRONT SPACE;
SOBOLEV SPACES ON COMPACT MANIFOLDS.
Starting a general investigation of the symbol space of a

comparison algebra, (i.e., the maximal ideal space of the commu-
tative algebra C/E) , we first will focus on the minimal compari-
son algebra JO , introduced in V,1 (It is generated by the classes
of a and D with compact support). We know that it always has the
commutator ideal K(H) . Its symbol space will be called the wave
front space , and be denoted by W . We first will show that the
wave front space is independent of the choice of H and dµ (sec.l).
In fact, it proves to be homeomorphic to the bundle of spheres
of radius infinity compactifying the cotangent space T Ox at a point
x GO (sec.2). For simplicity one commonly identifies these spheres
with the corresponding unit spheres
(0.1) { (x,F) : 1 } C T*Ox
Thus W then is called the 'co-sphere bundle' of 0 (cf.[AS],[Se2]).
The wave front space is found naturally imbedded in the
symbol space of every comparison algebra C meeting our general
assumptions (cf. VII,l). The result described above is due to
Seeley [Se21 in the case of a compact manifold, and to Gohberg
n
[Gbl], in the case of It (cf. also [CI], Herman [H1]).
For a compact manifold the minimal comparison algebra is the
only comparison algebra, evidently. For noncompact manifolds one
will ask the question about points of the symbol space M , not
contained in W This will be the focus of our attention in
.
ch.VII . In fact, it will be found that such points are linked to

the essential spectrum of the comparison operator H .
In sec.3 we show that all differential expressions of com-

pact support are within reach of the minimal comparison algebra
JO This leads to a result about the Fredholm property of the
.
realization of V,6 (thm.3.4) on compact manifolds. In fact, we

even get a Green inverse, i.e., an inverse modulo 0(-co) (thm.3.7).
VI.l. Local invariance 162
This fact is exploited in sec.4 to obtain the familiar Sobolev

estimates on compact manifolds.
1. The local invariance of the minimal comparison algebra.
In this section we focus on the minimal comparison algebra

JO introduced in V,l Let Am = C0 (U) and Dm = {D=b38 j+p: b3 and
.
x
p are C' , and of compact support } be the corresponding sets of
functions and folpde's. Essential self-adjointness of H0 is not
required in this section.
Theorem 1.1. The algebra J0 remains unchanged if the measure

dp , and the comparison operator H are modified over a compact
subset U C 0 . That is, if {O,du,H} and {Q,dv,K} are two triples
and if there exists a compact set U c 0 such that H=K and dp=dv
outside U , then both triples have the same minimal comparison
algebra J0
Remark. Notice that the two Hilbert spaces L2(O,dp), and L2(Q,dv)
coincide as sets, and have equivalent norms, but in general have
different inner products. Accordingly there also are two different
adjoints. Our proposition states that the two corresponding alge-
bras J0 of bounded operators coincide. They necessarily are C -
algebras with respect to either adjoint.
The proof of thm.l.l will require some preparations.
In order to use methods similar to those developed in V,3
it will be necessary to have condition (w), or even (sk) for ,
larger k, available. Since we focus on a and D with compact sup-

port in this section, it will be convenient to relate the results
to a modified comparison operator H° = H+q° , where the nonnega-
tive function q° is chosen such that {O,dv,H°} satisfies (sk)
for suitable k .
First of all, by III, prop.6.2 we may change the dependent

variable, and then obtain a constant nonnegative potential
q= const >1 . In this representation define
(1.1) q° = X(nkT)-2h1kTlxjTlxk
, T = jj=12-3ejwj
with our partition {w.} of app.A, where e.}0, as j}-, and
3 1 3
0 < £j < (Max{l, sup{IVwj(x)I x E nj}})
: Also, nk denotes
.
the positive real number specified in IV, thm.l.l, and X denotes

any C-(Q)-function with 0<X<1 , X=l outside some compact set.

In particular x may be made to vanish in some given compact set,
so that H=H° there. Then a calculation confirms that the expres-
sion H° satisfies the conditions of IV, thm.l.l, for m=k , so that
we have (sk) satisfied for H° . Finally we may switch back to the
old dependent variable, which does not disturb (sk), and leaves
the function q° unchanged as well. We have proven:
Proposition 1.2. For every comparison operator H and k=1,2,...

there exists a nonnegative function q° E C'(c) such that He=H+qs
satisfies condition (sk) of V,1 . Moreover, q° may be chosen
to vanish in any given compact set of D .
Remark. It is easy to also obtain a q° such that H° satisfies (s),

For example, one just might replace the constant nk in (1.1) by
a positive C°(c2)-function n(x) satisfying limx+.n(x)=0
A°=H°-1/2
Proposition 1.3. Let A=H-1/2 and . For every function
a E Am and folpde D E Dm we have
(1.2) aA - aA° = AC = C'A , DA - DA' = C" , C,C'5C" E K(H) .
Proof. For u E C-(SI) write
(1.3) D(H°+A)u = (H+A)Du + Dq°u + [D,H]u = (H+A)Du + Mu
with a second order expression M of compact support. For any A >0

(H+A)-1
write v = (H°+A)u , and multiply (1.3) by , to get
(1.4) R A)Dv=DS(A)v+R(X)MS(A)v=DS(A)v+(A-1R (X))AMA(A-1A°)A°-1S(A)v
(H+a)-1
with R(a) _ , S(a) = (H°+a)-1. By (w) the space (H°+A)C'
is dense, so that (1.4) may be extended continuously to v E H .
In particular A-1R(a) and A'IS(A) are bounded and 0((l+a)-1/2)

as easily seen, while IIA-1A°II < 1 follows from the inequality
(1.5) nA-1u112 = (u,Hu) < (u,H°u) = IIA°-1u112 , u E C'(cl)

Now a resolvent integral may be employed again. Let us use
the resolvent of the operator H , not the resolvent of A2 , for
a change. We have the formula
H-s S-1
(1.6) = m-lsin rs /(M-1 J'(R(A))mam-s-lda, s>0 , m=1,2,...
0
with binomial coefficients (m_i) , where the denominator may have

a simple zero, for certain integers s , but then the numerator
sin its vanishes also, and (1.6) holds with the limit of the
quotient, as s approaches such points. The proper and improper
Riemann integrals exist in norm convergence, as usual.
Formula (1.6) is a direct consequence of V,(6.8). One
simply must introduce the new resolvent and make a substitution
p=-1/A of the integration variable. Also the integration path
must be deformed into the positive real axis, chosen as slit of
Am-s-1
the branch of , and then the integrand must be evaluated
along the slit. Finally an analytic continuation in s must be
made, using the spectral theorem for the self-adjoint H
We use (1.6) on H and H° for s=1/2 to obtain
,
A = 1/7 J-R(A)A-1/2dA , A° = 1/7r J-S(A)A-1/2dX
0 0
Integrating (1.4) it follows that
(1.7) (AD) = DA' + 1/Tr N(A)dW

0
with
(1.8) C(A)=A-1/2(A-1R(A))(AMA)(A-lA°)(A°-1S(A))=O(A-1/2(l+A)-1),
as is derived, using the fact that A-1R(A) = 0((l+A)-1/2) , and

AA-1S(A) =
O((l+A)-1/2).
Also C(A) is compact for all A > 0 and

norm continuous in A again. (One may write (A-1R(A))(AMA) =
(A-1-ER(A))(A1+EMA) , where the first term is bounded,the second
is compact, due to the compact support of M .) It follows that the
integral in (1.7) is a compact operator. Taking adjoints in (1.7)
we get DA = (A°D) + C , C E K(H) This may be combined with the
second formula V,(3.17), applied to H° and A° for the second ,
relation (1.2). For the first relation we assume the expression D

of order 0 (i.e., bi = 0) Then M will be of first order only,
.
and we may repeat the argument with AMA replaced by AM or MA

while the remaining factor A (or A°) may be taken out of the
integral, either left or right. This completes the proof.
Proposition 1.4. The minimal comparison algebras J0 of {D,du,H},

and {O,du,H°} are identical.
This proposition is an immediate consequence of V,lemma 1.1,
and the second formula (1.2) .
We will need (1.2) in a slightly more general setting,

expressed by the corollary,below.
Corollary 1.5. Formula (1.2) also holds if H and H° are two gene-
ral comparison operators of the form V,(1.1) satisfying V,(1.2),
perhaps with respect to different measures dp and dUA , as long as
the two Hilbert spaces over 0 coincide (i.e., the function * =
du/dpA is bounded and bounded below), and as long as H and H° have
the same principal part (i.e., H-H° is a first order expression).
Also (w) must hold for H° not necessarily for H , and (1.5) must
,
hold, at least with an additional multiplicative constant at right

Proof. The important point is that (1.3) is still valid, i.e.,
D(H°+a) = (H+X)D + M , with a second order expression M of compact
support. Also the resolvent integrals are still well-defined, and
satisfy the estimates required, so that the proof just may be
repeated, q.e.d.
It is clear now that, for the proof of thm.1.1 it may be
assumed that both triples {52,dp,H} and {S2,dv,K} satisfy (sk), for
an arbitrary k (or even (s)=(sue) ). Also we may assume that dp=dv.
Indeed, let c = dii/dv . Note that is a positive C-(S2)-function
equal to 1 outside the compact set U . We get
K=-(CK)-12 KhÔkB k+q=K°+D° K°=-K-18 jKh^jkB k+q+y

,
xj x x x
with DA _ -C a k - y having compactly supported prin-
Jxj x
cipal part, while the constant y may be chosen such that K°
satisfies V, (1.2) Clearly H and K° now are self-adjoint with
.
respect to the same inner product, while K and K° generate the

same minimal algebra JO , in view of cor.1.5. If necessary, prop.
1.4 must be used again, to get (s) or (sk) for K° .
Now theorem 1.1, will be a consequence of the following

result, first published in [CS] , ch.V,lemma 4.10.
Theorem 1.6. Let condition (s) hold for H and K , and assume that
dU = dv , so that we only have one adjoint. Then the operators
Vt=(HtK-t)**
are in L(H) , for every t E I . If Ut= Vt+K(H)
denotes the coset mod K(H) of Vt, then {Ut tEg} is a 1-parameter
abelian subgroup of positive self-adjoint elements of the linear
group GL(L(H)/K(H)) in the C*-algebra L(H)/K(H).
Let us first show that thm.l.6 implies thm.l.l. For a moment
write J0(H) and J0(K) for the minimal algebras of H and K. We con-
clude that V1-1 = HK-1-1 =(H-K)K-lE J0(K), because G=H-K has com-
pact support, and may be written as a finite sum of products DF ,

DK-1/2FK-1/2 DK-1/2CK 1/2
with folpde's D,F E Dm. (Then DFK 1 = +
E J follows from V, lemma 3.6.) Thus U E J /K , and for the

1- 0
coset U-1/2 of V1/2=(H-1/2K1/2)** we also must have U_1/2-lEJ0/K
Indeed, U-1/2 as positive hermitian element of LIK must be the
,
unique inverse positive square root of U1, since the group pro-
perty of Ut gives U1 = (U-1/2)-2 This means that U_1/2 may be .
obtained as a resolvent integral
Jr(U1-a)-1X_1/2dA
(1.9) U-1/2 = i/2,
with a curve r surrounding Sp(U1) c (0,co), and staying in the

right half-plane Re X > 0 In particular a1/2 denotes the branch
.
of the root assuming the value 1 at X=1 It is important that .
0 is not in Sp(U1) since U1 is invertible (its inverse is U-1

,
by thm.1.6). Write U1=1+J J E J0/K. One may select r such that

,
1 is inside (it will belong to Sp U1 anyway). Then
(1-a)-1)X-1/2dX
(1.10) U_1/2-1 = i/2n Jr ((J+l-x)-1 -
by the residue theorem, where the integral converges in norm of

L/H . Clearly,
(J+1-a)-1-(1-a)-1
(1.11) = -J(1-a)-1(J+l-a)-1 E J0(K)/K, X E r
Since J0/K is closed it follows that U-1/2-1 E JO(K)/K. Thus also
(1.12) V-1/2 = (H-1/2K1/2):;* = 1+J^ , J^ E J0(K)
DH-1/2
Now from the generators a E Am , , D E Dm , of JO(H)
the first type trivially is in J0(K) as well. On the other hand,
DK-1/2(1+J^x)
DH_ 1/2 = DK-1/2(K-1/2H1/2) =
(1.13) E JOCK)
as well, by (1.12) so that J0(H) C J0(K) . For reason of symmetry

we then get the opposite inclusion, hence J0(H) = J0(K) , q.e.d.
Notice that the full strength of thm.1.6 was not used in the
proof of thm.l.l. In particular, only the classes U+1 and U+1/2
were ever used, together with the group property (U-1/2)-2 = U1
Thus, while stating thm.1.6 in its elegant form, we only discuss
the following weaker result (for full proof of thm.1.6 cf.IX,S)
Lemma 1.7. Assume only (s2), for H and K Then we have Vt e L(H) .
for all ltl<1. The cosets Ut mod K are hermitian and positive in
VI.1. Local invariance 167
L/K , and we have (Us/2)2 = Us , UsU-s=U0=l for all lsl<1 .
Proof. We already proved that HK-lE K(H), so that U1= HK-1+K(H)

is well defined. Similarly KH-1E L(H) , thus also
(1.14) V-1 = (H 1K)** _ (KH 1) E L(H)
so that U_1 is well-defined. Moreover, using I,lemma 5.1, we now

get HtK t and KtH-t E L(H) , 0<t<l , and hence get Ut well
defined for all ltl<l . Also we get
(1.15) HK-1 - (HK-1)* = GK-1 - (GK 1) , G = H - K
The compactly supported second order expression G is a sum of

products of < 2 folpde's. We get DFK-1 = DK-1/2FK 1/2 + C , CEK
as already used above. Accordingly DFK-1 - (DFK-1)* E K, by V,thm.
3.6, and U1 = U1 follows Similarly U-1 = U-1
.
Next we need a commutator relation.
Proposition 1.8. For 0<s<1/2 we have
(1.16) Hs[H-s,K-s]Hs E K(H) , and H2s[H-s, Ks] E K(H)
Proof. With the resolvents R(X)=(H+a)-1, S(A)=(K+A)1, and inte-

grals as in (1.6) (and either (a,T)=(s,t), or (a,T)=(2s,0)) write
(1.17) Ha[H-s,K-s]HT = (sin 2Trs)/Tr2 1-x-s dAU-sdpHa[R(A),S(p)]HT,

1-
0 0
(cf. (1.6) for more detail ) where
(1.18) [R(A),S(p)]v = R(A)S(p)[K,H]S(u)R(A)v , v E im(H0+X)(K0+p),
by a simple calculation.
We now need condition (s2) , at least, to insure that
im(H0+A)(K0+u) is dense for all A,u > 0 (cf. prop.1.9, below).
.
The expression J = [K,H) is of order 3 , and has compact support.

Accordingly we may write J as a finite sum of products of at most
3 folpde's with compact support. Accordingly, for evaluation of
the term at right of (1.17) we will substitute [K,H] in (1.18)
by a product DEF of 3 such folpde's , to indicate the procedure.
We then will have to commute S(U) with D :
(1.19) VD = [S(p),D] _
(K1/2S(ir))(K-1/2[ DFK] K-1/2)(K1/2S(u)) ,
(using (w)) , where the middle bracket is bounded (and compact,

K- E
if multiplied by , with arbitrarily small e>0 ) , while the
two outer brackets decay with p , so that [S(p),D] is compact, and
(1.20) VD = [S(p)5D] = 0((l+u)-1) .
Moreover, the commutator VD not only is bounded, but even GVD

with an arbitrary folpde G of compact support, is bounded.
Therefore we now can write
R(X)S(p)DEFS(p)R(X)v = R(X)DS(p)ES(p)FR(X)v
(1.21)
+ R(X)VDES(p)FR(X)v - R(X)DS(p)ESF,R(X)v - R(X)VDEVFR(X)v ,
where the right-hand side contains only bounded operators, so that

a continuous extension to v E H is possible.
The remainder of the proof consists of just balancing all
the various effects occurring in the integrand
X-sp-SHaR(X)S(p)DEFS(p)R(X)HT
(1.22)
where we focus on the first term at right of (1.21) . The integral

(1+X)s-1-d(1+11)s-1-d
needs a power with 6>0 to converge. That
is, in view of , KVS(p)=0((1+p)v-1)
HVR(X)=0((1+X)v-1)
Hs-g+l
we can absorb a total power (left and right) , using the
two factors R(X) For 0<s<1/2 this means that both a=T=s and
.
v=2s, T=0 give a converging integral of compact operators, as long

as two of the folpde's D,E,F can be absorbed by the factors S(p).
That can be done, because it still leaves a factor (l+p)-1 , hence
a total power p-s(l+p)-l to generate a convergent integral. For
s=1/2, on the other hand, one may split K 1/4 from the second S(p)
and H-1/4 from the last R(X) for a term K- 1/4 FH- 1/4=
(K-1/4H1/4)(FH-1/2+[H-1/4,F]H-1/4)
E L(H) by V, lemma 3.6, while ,
0(X-1/2(1+X)-3/411-1/2(l+p)-3/4).
we still have the integrand The
other three terms in (1.21) are similar, although easier to treat.
This completes the proof of prop.l.8. Then relation (1.16)1
(K-sHs)*- HSK-s
gives E K so that U ,is hermitian, 0<s<1/2
From the second (1.16) we get (HSK s)2H2sK 2s+H2s[H-s,K-s]HsK-s =
H2sK-2s+C ,
CEK, so that Us2=U2s This also gives Ut=(Ut/2)2
positive hermitian, for all 0<t<l Interchanging H and K then .
yields the same relations for -1<t<0 Finally it is trivial that .
UtU-t = U-tUt = U0 = 1 This completes the proof of lemma 1.7.

.
Proposition 1.9. If condition (s2) holds for K then it also holds

for H , and we have im(H0+A)(K0+u) dense in H , for all 0<A,p<° .

Proof. Let fEH , and (f,(H+A)(K+p)u)=0 , for all u E CE (O) . Write
this as (f,(K+p)2u) = (f,((K-H)+(u-X))(K+p)u) , u E C_(O). Observe
that (H+X)(K+p) is an elliptic, hence hypo-elliptic expression.
Thus the distribution solution f of (K+X)(H+A)f=0 is C_(O) Hence .
we can write (f,(K+p)2u)=(((K-H)+(p-a))f,(K+p)u). Due to self-

adjointness we may write g=(K-H)f+(u-a)f=(K+p)h, with hEdom(K+u).
Conclusion: (f-h,im(K+p)2)=0 i.e., f=h, (K+p)f-(H+X)f=(K+p)f, or
,
(K+p)f=O , f=0 , so that indeed im(H0+a)(K0+u) is dense, q.e.d.
Remark: A proof of prop.l.9 avoiding hypo-ellipticity is possible,

using the same arguments as in the proof of IX, prop.5.1.
2. The wave front space.
In this section we will discuss the structure of the symbol

space W of the minimal comparison algebra J0 The space W is
.
defined as the maximal ideal space of the commutative CC-algebra

J0/K(H) (note prop.2.1, below). This space will be seen to be
the bundle of unit spheres in the cotangent space of 0 , while the
continuous function aA:W*U assigned to a generator A=a (or A=DA)
turns out to be the formal principal symbol of the (ze'o-or first-
order) expression D or a .
Especially, for compact manifolds 0, this result is well

known (cf. Seeley [Se2], also Gohberg [Gb], Cordes [CI], in case
of the (non-compact) Rn). (Note that for a compact 0 the algebra
J0, hence the space W, is entirely independent of the operator
H or measure dp used, within the general restrictions made. This
is an immediate consequence of thm.l.l.)
Proposition 2.1. The algebra J0 contains the compact ideal K(H)

and has compact commutators.
Proof. It is sufficient to discuss compactness of commutators,
since J0 3 K(H) was seen in V, lemma 1.1. Moreover, we may apply
prop.l.2 again, to conclude that, for the commutator of two gene-
rators with compactly supported a or D one may assume H replaced
,
by H° satisfying M. Then,of course, V,thm.3 .1 and V,thm.3.4 give

the desired compactness of commutators, q.e.d.
The theorem,below, gives complete control of the wave front
space , as we shall call the maximal ideal space W of the commuta-
tive J0/K(H), henceforth. We use the notation 'symbol'
VI.2. The wave front space 170
(or symbol function) for the continuous function over W associated

to the coset A+K(H) of an operator A E JO by the Gel'fand map.
Theorem 2.2. The space 1 is homeomorphic to the co-sphere bundle

S S2 of Q. Here S S2 is the collection of all (x,>;) in the cotangent
bundle T*S2 of S2 (i.e. x E =Q , E n ) with
hjk(x)EjEk =1
Moreover, in this interpretation of W the symbol functions of the
algebra generators a E A# and A = DA , for D = b3 + p E D#
X3j
are explicitly given by the formulas
(2.2) aa(x,l;) = a(x) , ib0(x)i;j , E S*S2 = w
Proof. First we observe that JO contains the commutative C*-subal-

gebra CO(D) of all continuous functions a(x) with limx,a(x) =0 ,
which is obtained as closure of the algebra Am C JO The maximal .
ideal space of CO(D) is equal to S2 itself . The imbedding CO(D) -

JO generates a corresponding 'associate dual map' i:W - D , which
must be surjective (cf. [C1], App.AII,5) (In particular we use
.
that a multiplication operator u-au is never compact unless asO

so that also an injection CO(D)-JO/K is induced.) Later on we will
*
recognize i as the bundle projection in S D as defined. ,
Let Slj , j=1,2,..., be a locally finite atlas of D. Assume

QJlos
that every S2 has compact closure C :2-i , where {c'' } is an
atlas again. Define J. as the closed 2-sided ideal of JO contai-
ning all operators with symbol = 0 outside of i-1Q.
Proposition 2.3. The ideal Jj is generated by K(H) and all a, DA

with a E Am , D E Dm having compact support, contained in Q.
Proof. Evidently all the above generators are contained in Jj
since for a E C-(D.) we have o a () = a(i(m))=0 , as m E W\1-152
using the inherent properties of the dual map i Similarly, for .
D with compact support, contained in D. a function X E C;(52.)

may be found with X = 1 on the support of D , so that DA = XDA
aDA = aX.aDA We have aX(m) = X(i(m)) again, so that aDA=O
.
for all m with i(m) E D\D. , follows as well.

Vice versa, let A E Jj. We choose functions X1E C0 (Q 1=1,
2,..., with 0<X1<l , Xl=l in K1 , where K1 denotes an increasing
sequence of compact sets with U K1 = 52j . Such a sequence may be
constructed, since Stj may be regarded as an open subset of In with
compact closure. Since aA(m) is continuous and =0 outside 1-1QJ

we conclude that
Ia(1-Xl)A(m)I=(1-Xl(t(m)))IaA(m)I<sup{IoA(m)I:1(m)EOj\Kl}.0, 1--,
in view of the compactness of aQj , and uniform continuity of aA

It follows that the coset mod K of A1=(l-X1)A converges to
0 as 1}- . Or, there exists a sequence {Ci } of compact operators
such that IIA-X1A-C1H - 0 , 1}°° . However, it is clear that X1A+C1
E J , since Ai can be approximated by a finite sum C + JP with
CEK , and P = D1AD2A.... DMA , D.E Dm , so that

7
(2.3) X1P = XlD1AX1+1D2A...Xl+N-1DNA + C', C'E K ,
where Xl+k-1Dk have support in . For (2.3) we used prop.l.2,

0.
and Xl = X1X1..1.. .X1+N-1 ,
of course. The proof is complete.
Note that a partition of unity 1 = wi , supp wi C Qj , may be

J
used to verify that J = +.J3 . . That is, every A E J may be obtai-
ned as a limit (for m-'") of finite sums IjAm = Am, where Am E J.
Again it is clear that each Jj is a C -algebra without unit

(except in case c._0 compact) containing K, and that J./K has
1-l0j
maximal ideal space ffi C W With this identification the
functions in CO(ffi7) correspond to the cosets A' = {A+ K} of Ji /K
for A E Ji . Clearly we get W =U.ffiY due to 0 = U,Oj
For the description of W it therefore suffices to discuss ffi0
and the homomorphisms J. -* J-1K i Since O^.D njlos is a
chart, we have global coordinates xl,...,xn defined by a homeo-
morphism Q^ O' C In , mapping fj onto OTC Oi An explicit .
set of generators mod K of Ji then is given by
(2.4) { xkA ,k=l,...,n ,: f E C'(O4) } .
(Here we break with our usual convention to silently identify Q

In
with the open subset 03 . of , for reason of clarity.
Proposition 2.4. The space ffi3 . is homeomorphic to the set
(2.5) {(x,E) E 0i X In , 1 } ,
equipped with its natural topology as a subset of I 2n . Moreover,

under this homeomorphism, we have
(2.6) Q _ fi(x) ' QA,(x'E) = (x)Ej, for Ai _ -i4a A, J=l,...,n.

I xi
Proof. For an m E ffi let x = t(m) , k = QA (m) , where E C'(c2.)
k
(real-valued) satisfies fi(x)=l near x=t(m) , 1=1,...,n. We claim
that this defines the desired homeomorphism. Indeed, first of all
we note that (x,C) is well-defined if $ and both are = 1 near :
x = t(m) and if B. = -i'Pa

, then we get A7-B. = (1-X)(Aj-Bj ),
,
x
with a C0--function X , equal to 1 near t(m)= x , hence aA (m)
3
aB (m), showing that is well defined. Also is real, due to
Cj = oA *(m), where C = -i$3 A + C', C,C' E K,

I x xi
using (1.2) and V, thm.3.4 , so that 7l = l . Furthermore, with

as above we note that h3k¢ E Am For u,v E AC' we then get
.
(u,A(-a k)Av) =
x x
- (u,A(D 3 + q¢3)Av) .
Since ACD is dense we conclude that
$3+C
(2.7) Ai A(D 3+0 3)A = , C E K
using (1.2), V. lemma 3.3, V, lemma 3.6, and that AHAu=u5 uEACO
Taking symbols one obtains the relation
(2.8) hJk(x)EiEk = 1 , as x = t(m) .
With the above we have defined a map u:ffii -> Mi , with
(2.9) M. _ {(x,E) : x E 0i , E E n 1 } .
We shall prove that u defines a homeomorphism onto Mj. Assuming

this to be true let us observe that the proper transformation
law is valid,since they will inherit the co-variant tensor
property of the 8 .
x]
To verify that u is 1-1 we recall that the maximal ideal mEffi.
is characterized by its corresponding homomorphism J -' M , given
by A - aA(m) , A E J J. Let m, m' be such that i(m) = t (ml) = x, and

as above. Note that this implies a4)(m)=a4)(m'), and aA (m)=aA (m')
J J
for all E C0-(S2j), regardless whether 4)(x) = 1 , or not. Since a

is a homomorphism, and since the m and Aj listed form a set of
generators of J., modulo K , it follows that aA(m) = aA(m') for ,
all A E,J Thus the homomorphisms corresponding to m and m' are

identical, which means that m = m' Thus the map u is injective.
.
Finally we focus on surjectivity of u . It is clear that,

above every x GO there will be at least one point (x,l;) E u(ffij).
j
For the map i maps onto I so that some mE M with x=t(m) can be
,
found. For this m we then get the l;j by our above construction.
To show that every (x,n) n is an image, we
, for general n E
construct an automorphism of Jj , leaving K invariant, which
carries (x,F) into (x,n) , in form of a coordinate transform.
Let M=((m.k)) be a constant real nxn-matrix with M>;=n, M*J2M=J2,
where J=((hJk(x)))1/2 (i.e., JMJ-1 is orthogonal). Such a matrix
exists for every i;,n E gn and one even may choose det M = 1
, ,
so that JMJ-1 is a rotation, and can be smoothly deformed into the

identity matrix. Denoting x=x0 , for a moment we design a diffeo-
morphism e:Q.- 52. as follows: Let R(T), 0<T<-, be a Coo-family of
rotations such that R(T)=JMJ for O<T<E , R(T)=l for T>2e. Define
6(x)=x0+JRT(IJ(x-x0)I)J-1(x-x0), x E Q.. By construction 0 maps
J
each (Euclidean) ellipsoid T2=IJ-1(x-x0)I2=h.k(x0)(x3-xj)(xk-xk)
J
onto itself, and is equal to the identity as lJ(x-x0)1>2e There- .
fore, for O<e sufficiently small, 0 defines a diffeomorphism of

9j onto itself, and 0 is linear near x0 and equal to the identi-
,
ty near a12j . In particular we may extend 0 to a diffeomorphism

a-0 equal to the identity outside S2j .
Using 6 we define the unitary operator 0 : H -y H by
(2.10) Ou(x) _ (det(a0/3x))1/2u(0(x)) .
Then A-A°=0-1A0 is a *-automorphism J.+ Jj leaving K(H) invariant.

Indeed this automorphism trivially leaves K invariant, and
a =0-1a0 = ao 6-1, where ao 0- E CD(c.) if a E C0_(52j ). Similarly,
(2.11) 0-1u = (u/v)oe-1 , v = (30/ax)1"2
and, for b E CD(S2j) Dr = ba Xr , we get
(2.12) D A u =0-1b3 r®u =(boo-1)((.(v 6-1)a lu),

xr
r/v)oe-1)u+(01
r x
x Ix
which again is an expression with compact support, contained in

s0j. Accordingly, the automorphism 0 leaves the generating sets
of J. invariant.
J
In fact, near x=xO , the expression Dr=-i43 r goes into
x
Dr° = -i4mr]ax. = mr3Dj , modulo a zero order term.
We now must find out how the operator A changes under conju-
ation by 0 Clearly the differential expression H goes into
.
HA = 01H0 =-K°-1a K°hAjka k + q° , du° = K°dX ,
(2.13) xi x
K° = Koe-1 , h'jk=(hlmOjIx10klxm)oe-1 q°=(Hv/v)oe-l

,
0=0-1
Indeed, since 0 is unitary (i.e., ), we get for u E CD(S2),
(2.14) (u,HAu) =J S2
(hJk(v(uoO))
IXj (v(uoe)) Ixk +IgoOI2v2)du
For the first term we use the formula
(2.15) (v(uoO)) = v (uoe) + v(u 1oe)e1

Ixj Ixj IX IXj
and a partial integration similar to that of 111.(6.7), to get
(2.16) (u,H°u)=J(hlmej ek Ixm ((u u k )oe)+(Hv/v)Iuoel2)v2d u

1 Ixj
Ix Ix
Using that v2=1ae/ax) we may transform the integral backward. Then
the hermitian form may be polarized to to get an expression for
(u,H°v) . The second order expression ft is uniquely determined
by the values of (u,H4v) , for all u,v E C , and (2.13) results.
Clearly H° = H and du = dp° outside a small compact neigh-
bourhood of the point x0 . Accordingly we may apply thm.l.l and
{Q,dp°,H°}
the triples {1,dv,H} and generate the same JD . In

,
fact, by definition of A and A° as inverse positive square roots

of H and H° = 0-1HO (where now H and H° mean the self-adjoint
closures of the minimal operators) it is clear that A° = 0-1A0
We thus get
0-1a0 = ao O-1 = a° , O-1DAO = D° A° ,

(2.17)
D° = 0-1DO = ( (b1ej 1)o 0-1)a + (Dv/v)o e-1 .

Ix xi
Now consider the associate dual map m + m° = 0-1m0 of the
automorphism A -)- A° By construction we get i(m) = i(m ) , when-
.
ever t(m)=x0 , since a°(x0) = a(x0) for all a . Therefore for our
(x0,E) we have m° E 1-1(x0) as well, and we now will
show that u(m ) _ (x0,n) . Indeed, aA (m') = aA e(m) , by the pro-
r r
perties of the associate dual map m-m° . We have seen that
(2.18) Ar° = Dr A° = mr1DIA' + C = mr1A1(A-1A°)+ C .
1
Accordingly, a Ar e(m) = m r 1a (A-s
lA ) (m) = nra (A-,
lA ) (m) . Thus
we are only left with showing that a A (m) = 1 . However,from

(A-lA )
lemma 1.8 we know that the coset U-112 of A-1AA is the inverse
H°H-1 Zhr1ArA1 + C , near x0. Thus, over x0

square root of U1 = =
we get aU = 1 . This completes the proof.
3. Differential expressions within reach of the algebra J.
In chapter X we shall exploit the information on the symbol

space of the general comparison algebra to its full extent. Pre-
sently let us discuss differential expressions within reach of
the minimal comparison algebra. For the special case of a compact
manifold D we then will discuss the implications of our theory,
such as existence of a Fredholm and Green inverse, the Sobolev
estimates etc.
First let 0 be general, not necessarily compact. Consider
a given triple {G,du,H} on G , with its Hilbert space H=L2(Q,dp).
Let J0 be the corresponding minimal comparison algebra. Using
prop.l.2 (and the remark following it) and prop.l.4 we may modify
the potential q of H such that cdn.(s) holds, without changing J.
Moreover, using thm.l.l we also may modify H and du on compact
subsets of G (In fact, applying later results of algebra surgery
(cf.VIII,thm.3.1) it will follow that J0 remains unchanged as long
as the measure du, hence the Hilbert space H is kept the same out-
side some compact set).In particular, in the following, it is no
loss of generality to assume that cdn.(s) holds.
Specifically for a compact manifold there is only one Hil-
bert space H, and the choice of H and dV will not affect J0 at
all. Then we have cdn.(s) for just any H, regardless of the poten-
tial q , by IV,thm.l.l.
VI.3. Reach of the minimal algebra 176
Proposition 3.1. Let L be an N-th order differential expression

with C"-coefficients and of compact support on 0. Then the unboun-
ded linear operator LAN with domain A-NC-(Q) admits a continuous
extension A to all of the Hilbert space H , and we have AEC=JO
In other words, every differential expression with compact support
is within reach of the minimal comparison algebra JO .
Proof. For NO and N=l the statement is trivial, since the opera-
tors aA0 = a and DA , for a multiplier a or a folpde D belong to
the generators of C=JO For the general case it is essential that
.
every L of order N and of compact support can be obtained as a

finite sum of (not more than N) first order expressions. Indeed,
refer to some finite open chart cover of the compact set supp L
a subordinate partition of unity, and explicit form of L in the
local coordinates
(3.1) supp L NS2,, 1=1w, on supp L, L=L3= a33(x)Dxa on 52j,

JaT<N a
and write
(3.2) L = 1w.L. _ . (w.a3)IIn IIa

v=1 p=1
(w. D
3vp xv
7 3 3 a'f<N 3 a
using the auxiliary functions w,vpE C0-(cr.), =1 near supp wjCcr..
Here the expressions wjvpD may be interpreted as folpdes in Vm
v
x
and, for jaI=N, we may multiply the factor (mia3a) into the last
such folpde. Accordingly each summand in (3.2) is a product of at
most N folpdes. Trivially we may assume exactly N folpdes in each
product. Accordingly is is sufficient to prove the statement for
products
( 3 . 3 ) L = ...... , Dj
D E Dm .
Notice also that A-NC0(S2) is dense in H for every N, as

implied by cdn.(s), as seen in V,6, for example.
Therefore it is sufficient to show that LANu for uEA-NC0, ,
can be written as a sum of terms A1...ARu , with generators

Ai of J 0 .
For an induction proof consider K = D1D2 , M = D3...DN.
Assuming that an operator CJKA2 , CEL(H) exists, we get
A-NCO
(3.4) KMANu = CMAN-2u + CI H,M] A Nu , uE (11)
(Confirm in details that CHMANu=KMANu and CMAN-2u=CMHANu.)

Note that the commutator is an expression of order <N-1 , so that

[H,M]AN=[H,M]AN-lA E L(H), by induction hypothesis. Thus we only
must prove the statement for N=2 to make the induction complete.
On the other hand, for a twofold product DF , with folpdes
D,F, we write (with exactly the same reasoning)
(3.5) DFA2u = DA2Fu + DA2[H,F]A2u , u E A-2C'(O)
Again the first term DA2F = (DA)(F*A)* is a product of generators,

while [H,F] is of order <2 , hence a sum of products of at most
two folpdes. It follows that the second terms operator is a finite
sum of terms
(3.6) (DA)(AD')(F'A)A E C ,
with folpdes D', F', where F'=1 may occur as well. This completes
the proof.
Next let us ask for the symbol of AJLAN , as in prop.3.1.
We find that A D Dl..... DNAN has symbol
(3.7) aA = RJ=IaA , A = DMA

J
Indeed the second terms in both (3.4) and (3.5) carry an

extra power of A which may be written as XA , mod K(H), with XE
C- since there always is a expression with compact support in the
same term. Recall that XAEK(H), by III,thm.3.7. Also we get
DA-AD E K(H) by V,thm.3.4, so that aDA
= aAD
Now we may combine (3.7) and the product representation
(3.2) of an expression L and apply thm.2.2 to obtain the symbol
of A D LAN. The proof of prop.3.2 below is a formal verification
left to the reader:
Proposition 3.2. Let L be as in prop. 3.1. Assume that, in local

coordinates of O'C 0 L has the form
(3.8) L = a(x)DX
a , x E 0'
with coordinates x of the chart O'C 2 Then the symbol of the

.
operator LAN C A E C on the set i-1(0') C W _ M = SCO

, is ,
explicitly given by the formula
(3.9) a 'AN 'O = I

ac=N a E 1-1(D')
Notice that the expression at right of (3.9) is commonly

referred to as the principal part polynomial (or the principal

symbol) of the expression L Since we know that oA
. for AC-C, is ,
a continuous function over the subset S*c of the cotangent space,

it follows that also the right-hand side of (3.9) is independent
of the coordinates used, as also easily confirmed directly. In
particular it sometimes is more convenient to use tensor notation
instead of multi-index notation: We may write
a-il...3NEj
(3.10) aa(x)i;a = ...Ej
Ja=N 1 N
with summation convention and a unique symmetric contravariant
tensor a- of N variable indices.
Remark 3.3. Observe that the expression L is elliptic over c'ai if

and only if the symbol of A J LAN (with respect to the algebra J0
does never vanish on the part of the wave front space t-1(0') .
From now on consider the case of a compact manifold Q only.

Then the algebra J0 contains the identity operator of H and its
wave front space W is compact. As a consequence of prop.3.1 every
differential operator on 0 with C"-coefficients is within reach
of the unique (minimal) comparison algebra C=JO. Then let ADLOAN
and Z=AA-N be the operator of V,def.6.1 and realization of V,def.
6.2, respectively.
Theorem 3.4. The operator A D LAN has an inverse modulo the class
F(H) C L(H) of all operators of finite rank if and only if L is
elliptic on 0 .In details, there exists an operator B E L(H) with
(3.11) AB = 1+F , BA = 1+G , F, G E F(H) ,
if and only if L is uniformly elliptic. Moreover, we then get
(3.12) dim ker A < - , dim H/im A < - ,
and B may be chosen as operator in JO such that -F and -G coincide

with the orthogonal projections onto the finite dimensional spaces
*
ker A and ker A , respectively. On the other hand, we instead may
choose BEJ0 such that for an arbitrarly given e>0 we get
(3.13) IBH<Ill/Q II
+e, a (x,E)= a (x)Ea in local coord's.
A A a =N a
Proof. For a bounded operator A an inverse modulo F(H) exists if
and only if A has an inverse modulo K(H). Indeed, since F(H)CK(H),
we only must show that (3.11) for a BEL(H) with F,GEK(H) instead
of F(H) implies (3.11) for some other B=C. By Rellich's criterion

we get orthogonal projections P , Q with finite rank such that
(3.14) IIF(1-P) II < 1/2 , IIG(l-Q) II < 1/2

Then we may write
(3.15) AB = (1+F(1-P)) + FP , BA = (1+0(1-Q)) + GQ
where the first terms at right are invertible (an inverse is
given by the norm convergent infinite series J_(£(P-1))3 , in

j=0
the first case, for example). Thus we have
(3.16) A(B(1+F(1-P))-1) = 1 + F' , F' = FP(1+F(1-P))-1 E F(H),
showing that B' = B(1+F(1-P))-1 is a right inverse mod F of A .
Similarly we use the second (3.15) to obtain a left inverse. It

is known that existence of both a left and a right inverse implies
that both, left and right inverse are inverses, so that indeed
A is invertible mod F .
This shows that we only must discuss invertibility of A D

LAN mod K(H) . But K(H) is a closed 2-sided *-ideal of the C*-
algebra L(H) . Moreover, C = JD is a C*_subalgebra of L(H) con-
taining K(H) . It therefore follows that L' = L(H)/K(H) and C"
C/K(H) are C*-algebras, and that C"C L', using a well known result
on C-algebras (cf. [RlI or any general book on C-algebras).
*
Moreover, using another standard result on C -algebras, C" is
inverse closed in L". That is, A"= A+K(H) E L"is invertible in C"
if and only if it is invertible in L" (cf.[C11,AII,L.7.15).
It follows that A D LAN is invertible mod F(H) if and only
if its coset A A+K(H) has an inverse in CV . Here we finally
use the fact that the commutator ideal E of C=J0 equals K(H), and
that C" is isometrically isomorphic to C(M). Since the cosphere
*
bundle ID=S 0 is compact, a function aEC(W) is invertible if and
only if it never vanishes. Hence A D LAN is invertible mod F(H)
if and only if L is elliptic, by remark 3.3 above. This completes
the proof of the first statement.
Observe that ker A C ker(l+G) and ker A C ker(l+F ), where
both right hand sides are finite dimensional. Hence dim ker A <-,
dim ker A <co . Notice that in A is closed: For Auk+v , k+. get
(l+G)uk+Bv . If wk is the component of uk orthogonal to ker(l+G)
then we must have wk-+w and BAw = By . Hence v=Aw + z , zEker B

Then let pk=uk-wk (Ecer(1+G). We get Apk;z . If qk denotes the
component of pk othogonal to ker A C ker(l+G) then we also have
Aqk+z , hence qk->q, since A is invertible in the finite dimensio-
nal space ker(l+G)n(ker A)1 Then v=A(w+q), so that im A is closed
hence im A = (ker As*)-L. Thus dim(H/im A) = dim ker A*<co , and we
have (3.12) confirmed.
It then is clear that A0 = Aim A constitutes an inver-
tible map im A ; im A. Let its inverse be called B and extend
0
B0:im A ; in A to an operator B':H;H by setting B'=0 on ker A
(im A)'L. It follows that B'EL(H) and that AB' abd B'A are the
orthogonal projections onto in A and im A , respectively. Thus
1-AB' and 1-B'A project onto ker A* and ker A, resp. Since B' is
an inverse mod K(H) of AEJ0 we conclude that also B'EJ0, using
again that inverses in Lv of elements in C' must be in C' .
Finally it is found that '1/2' in (3.14) may be replaced by

any arbitrary positive number 6 Then the conclusion leading to
.
(3.16) gives an inverse mod F(H), called B", for a moment such
that IIB"II<IIBII+6 On the other hand the isometry of the Gelfand
Naimark isomorphism implies that
(3.17) 1111a 0 = inf{IIB+KII: KEK(H)} = y

A
Since B+K is a K-inverse whenever B is, we can find a K-inverse B
with IIBf<y+e/2. Setting 6=e/2 we then get the desired F-inverse B"
An operator AE L(H) satisfying (3.12) is called a Fredholm
operator. An inverse B of A mod F is called a Fredholm inverse of
A . For more details cf.X,l The unique Fredholm inverse satis-
.
fying (3.11) with the orthogonal projections onto ker A and ker A
will be called the distinguished Fredholm inverse.
Next it will be our intention to show that the distingui-
shed Fredholm inverse has the properties of a "Green inverse",
intimately related to the generalized Green's kernel of the reali-
zation Z .
Proposition 3.5. Every Cs-differential expression of order <N over

Q , whether compact or not, is H-compatible in the sense of V,def.
6.4.
The proof is evident, since indeed the successive commuta-
tors L0=L , L1=[H,L] . ..... Lj+1=[H,Lj] all are compactly suppor-
ted and of order <N+j .
For an H-compatible expression of order N we now may apply

V,thm.6.8 and its corollaries.
ASAA_S
Proposition 3.6.For ADL0AN of prop.3.1 we have E L(H), and
ASAA-s
(3.18) - A E K(H) .
In other words, for every < N-th order differential expression L

with Cm-coefficients on a compact manifold 0 the operator ADLOAN
is of order 0 and is K-invariant under H-conjugation.
The proof is immediate, in view of the above mentionned thm.
since all terms at right of V,(6.11) are compact: Their differen-
tial expressions are compactly supported and there is an excess of
A-powers.
Theorem 3.7. For an elliptic differential expression L on a com-
pact manifold 0 the realization Z of V,(6.2) allows a 'distingui-
shed Green inverse' G , defined as an operator GEO(-N) (in the
sense of 1,6) such that
(3.19) ZG = l-Q Z G = 1-P ,
with the H0-orthogonal projections P and Q onto ker Z and ker Z*

respectively. We have ker Z , ker Z C H , and P, Q are conti-
nuous operators H_.-)-H. of finite rank.
Proof. Using thm.3.4, if L is elliptic, then the distinguished
Fredholm inverse B of ADL0AN is well defined. This implies that
the operator A=A0 is Fredholm. Moreover, from prop.3.6 above we
know that AEO(0) and that A is K-invariant under H-conjugation.
Thus I,thm.6.6 may be applied and we get ker A C Hm, ker A C H_.
Accordingly the distinguished Fredholm inverse B gives a distin-
guished Green inverse of A . Let G=ANB then we get LG=AB=1-Q, .
1-ANPA-N
as stated. Also GL=ANBAA-N = , with P of thm.3.4. One

ANPA-N
then confirms that P' = is the H 0 -orthogonal projection
onto ker L = ker(A AN). In particular (A AN)0 is Fredholm and
K-invariant under H-conjugation, so that ker L*CH. again. Q.E.D.
4. The Sobolev estimates for elliptic expressions on compact 0 .
In this section we again consider a compact manifold under

any H and dp whatsoever. We have condition (s) and therefore the
HS-chain of Sobolev spaces as in sec.3 above. First we note:
Lemma 4.1. For a compact manifold we have H. = C_(c2) and we get ,
Lu = Zu for all uE H_ and every expression L and the realization

VI.4. Sobolev estimates 182
Z of V, def.6.1.
This follows at once from the well known Sobolev imbedd*
lemma: For s>n/2+k the space Hs is (compactly) imbedded in C k (S2)
(cf. IX, prop.l.5). Here we shall discuss a somewhat weaker result
also implying lemma 4.1. (For a Sobolev imbedding of LP into Lq
cf. IV,prop.4.1.)
Lemma 4.2. For s>Ln/2]+k+l the space Hs has a natural imbedding

into the space Ck(12) defined as follows: The class
, of
squared integrable functions (with any two coinciding outside a
set of measure 0) contains precisely one function in Ck(S2) .
Moreover, there exists a constant cs k independent of u such that
(4.1) p up k < cs'k II uU s , for all u E HS , sj n/21 +k+l .

C (c)
Proof. We use Taylor's formula with integral remainder: A function
q(t) of one variable t, infinite differentiable on I , and with
derivatives of all orders (including 0) vanishing at 0 satisfies
t
It
(4.2) _ (t-T)N-1/(N-1)! (N)(T)dT
0
with the N-th derivative 4 since all terms of the Taylor se- E ,
ries at 0 vanish and only the remainder in integral form remains.

We apply (4.2) to the function 4(t) = u(x0+(l-t)z0)X((l-t)z0) ,
where u EC_ (In) is arbitrary and X E C,(]kn), X>0, denotes a cut-off

function at 0; we have X=l near 0, X=0 for Ixl>1/2. Also, x0,z0 E
2
1n are arbitrary, but IzOI=IzO =1 . N>0 denotes an integer to be
determined. With a change of integration variable 1-T-Tt we get

1
(4.3) u(x0) = dt tN-1/(N-1)!dN/dtN(X(tz0)u(x0+tz0))
0
Now we estimate, writing

1
< c dttN-1l,(tz0) Iu(6)(x0+tz0)I
(4.4) Iu(x0)I
0 I0I<N
where denotes another cut-off functions with the same properties
as X and IX(0)I<V
for all x and Iai<N, and with a constant c.

Then we integrate (4.4) over the unit sphere of In , with respect
to the variable z0 , and the surface measure dS . Using that
dx = rn-1drdS , with r=lx-x01 , we get
(4.5)
cJdxlx-x0IN-n,p(x-x0) Iu(9)(x)I
Iu(x0)I <
10l <N
Now we set N=[n/2]+1 and use Schwarz' inequality, noting that

Ix-x012(N-n)
dx = c0<- , with this choice of N. Thus,
11x-x 0I<1
(4.6) Iu(x0) I < c {J I2dx}1/2, x0E,,n

I0I<N
where c is independent of x0 .
We now may implant (4.6) onto a compact manifold P. Choosing

a finite atlas such that each chart S2' is contained in a chart S2"
in such a way that dist(S2',a52")>2 we find that
(4.7) Iu(x0) I < c INLj,x uN , X0 E 52 , u E C°°(R) .

0
where the sum is finite and the differential operators L. of
JSX0
order <N with Cm-coefficients satisfy
(4.8) IIL. ANN < c , c independent of x0 and j

0
(Also their number is independent of x0 ). The verification of the

latter fact requires some calculation involving decomposition of
Ljxo into sums of products of folpdes, similar as in (3.2), left
to the reader.
Combining (4.7) and (4.8) we get
(4.9) (lull <c IIuNIN , for all u E C°°(c) .

L°°(S2)
For an arbitrary u E HN , N=[n/2]+1 , there exists a sequence
u E C00(c) such that Ilu-uk0 ->. 0 , k+co . We then conclude uniform
k
convergence of uk (i.e., convergence in C(3) ), applying (4.9) to
the difference u.-uk E C°° . Hence there exists a limit of uk in
C(S2). It is clear that the limits in C(S2) and in H must coincide
whenever they both exists. Thus indeed uEC(1) , i.e., HNCC(S2)
Similarly, for N=[n/2]+l+k we may apply (4.7) to wu(a)
IaI<N , where u(a) is a local derivative of u and w cuts off in
a neighbourhood of the complement of the local chart used. Clearly
the expressions Lj,x w3X are of order <N+IaI < N+k . Hence (4.8)
0
may be replaced by
(4.10) IIL. w3XAN+kN < c , c independent of x0 and j

0
This implies the result for general k. Q.E.D.
Corollary 4.3. If 0 is not necessarily compact, but the triple

{c,dp,H} satisfies cdn.(s), then we still have the imbedding

Hs + Ck(S2) , as s>[n/2]+k+l , although the estimate (4.1) holds
only locally, with gu II and c s,k replaced by gu B , and
Ck(S2) Ck(S2' )
es,k,Sl, , respectively, for each S2'c 0 with compact closure.
All above proofs may be carried over to prove cor.4.3.
A simple consequence of thm.3.7 is the result below.
Theorem 4.3 (The Sobolev estimate). Let the expression L on S be

elliptic. For every s E R there exists a constant c s such that
(4.11) flullN+s< cs(fLull s+ Bugs) , for all uEH.
Proof. Using thm.3.7 we get
(4.12) 1 (1-P)vll = BBAv1I < 013 11 BAvll s, vEHm.

s s s
Here we may introduce v=A-Nu , u E C0 . It follows that
(4.13) BuNN+s = RA-Null s< BBBsHZuHs+ IIPA-Nulls.
Finally we use the fact that P E 0(-°) , so that BPA-Nlis<...

Therefore (4.11) follows from lemma 4.1, q.e.d.
Remark 4.5. Note that the term csllulls at right of (4.11) may be
replaced by cs tfu't , for just any tEl. Indeed, the expression
PA_N of (4.13)'still belongs to 0(-oo) , so that fPA-NB5,t<m for
any pair s,tE 1
Remark 4.6. It is possible to prove the following stronger result

For every a>O there exists a constant c(e) such that for an ellip-
tic expression L of order N we have
(4.14) Bugs+N < BQA lg )HZugs + c(e)Bu1s uEH

L-(Q)+6
To prove this we apply thm.3.7 to the self-adjoint elliptic
expression L L It is found that the bounded self-adjoint opera-
.
A0A0 (in H0) is Fredholm and admits compact conjugation. Similarly

all the operators ADA6-A = AN (L L-AHN),AN , as long as X<m2 =
L
(inf{laAI:(x,i;)EW})2. We conclude from there that A A 0 has only
0
discrete spectrum below mL . Moreover, all eigen functions to
eigenvalues less than mL are in H. . From this fact one concludes
the existence of a bounded self-adjoint B0 with norm BBlo<mL'+e
such that
(4.15) A0AB = BA0AO = 1-R ,
with a finite dimensional orthogonal projection R onto a subspace

of H. , i.e., the span of all eigen spaces to eigenvalues
< mL/(l+emL) . In particular B is positive self-adjoint. Now rela-
tion (4.15) may be used instead of the earlier BA=1-P to derive
the improved estimate for s=O . Similarly for general s, using the
operator As( s) As , with the Hilbert space adjoint A(s) of A in
the space Hs .
CHAPTER 7. THE SECONDARY SYMBOL SPACE.
In the present chapter we turn our attention to the set

it\W of the symbol space it of a general comparison algebra, consi-
sting of all points not contained in the wave front space. As
pointed out before, the wave front space W (i.e. the symbol space
of the minimal comparison algebra) is found naturally imbedded
as an open subset of the symbol space M for every comparison
,
algebra C meeting our general assumptions. This now will be dis-

cussed in detail in sec.l, below. If the manifold Q is compact,
so that the minimal comparison algebra is the only comparison
algebra, then we have it=W. In that case every differential expres-
sion on 0 is within comparison reach of the algebra J6 , as was
shown in VI,3. For an elliptic expression L then the realization
Z-X=(A-AAN)A-N
Z of V,def.6.2 has the property that is Fredholm

for every XET. In other words, there is no essential spectrum, in
the sense of [CHe1]). This fact is in direct relation to the fact
that the set it\W is void. Essentially it follows that the latter
set, or, rather its interior its = it\(Wclos) is the origin of the
,
essential spectrum of elliptic operators within reach of a compa-

rison algebra C, (assuming that E =K(H)) .
Accordingly we now engage in a detailed discussion of the

set its , which is called the secondary symbol space of the
algebra C . In this respect it may first be noticed that so far,
as long as only the minimal comparison algebra was under discussion,
we did not require any geometrical assumptions on the space 0 un-
der its Riemannian metric induced by H whatsoever, except for the
structure. On the other hand, for the investigation
of the secondary symbol we will have to impose assumptions at
infinity. We choose to work with a set of analytically formulated
conditions (cf. sec's 2,3,4 below) but will make little effort to
translate these conditions into geometrical properties.
Notice that an extensive study of the space its for a variety
VII.O. Introduction 187
of comparison algebras over n was done in [CHe1] and [C1],III,IV.

It appears natural from these results to seek the space ffis as a
subset of the boundary of a suitable compactification of the
cotangent space T 0 of S2
To be more specific, we first note that the multiplications
a(x), for aEA# , alone generate a Csubalgebra C # of C , isome-
trically isomorphic to the function subalgebra of CB(S2) generated

by A# . The maximal ideal space M# of C# will be a compactifica-
tion of the manifold Q . Actually, this is the smallest compacti-

fication of S2 onto which every function of A# allows a continuous
extension, as well known (cf.[Ke1]).
If a E EOCA# then Xa E EnJO=K(H), for X E C0(2), hence a=0
Thus as a C*-subalgebra of the commutative C
we also find C
A
algebra C/E Then the associate dual map of the injection
.
CA# + CIE defines a continuous surjective map is ffi } CA#
already used in ch.6 for the minimal comparison algebra JO

Now one may assign a formal algebra symbol TA to each of
the generators of C , by setting Ta (x,E) = a(x) , for a E A#
+q(x)) 1/2, for D=-ib08
T
DA
(x,C) _ (Wb0()E.+p(x))/(hOk(x)
jk xj +p
E D# . Clearly the formal symbols are bounded complex-valued
Cm-functions, defined over the cotangent bundle T*S2 of 0 . Thus
again, a smallest compactification & S2 of TtO exists onto which
all formal algebra symbols can be continuously extended. Again
* * *
1 0 is the maximal ideal space of the C -subalgebra of CB(T Q)
generated by the formal algebra symbols. Again one obtains a con-
tinuous surjective map 7:1 S2 - M # as associate dual map of the
,
A
*
injection CA# - C(& O) , since CA# also is a subalgebra of the
formal symbol algebra.

We are interested in the boundary 31 S1 = P*S2 \ T0 of the
compactification 152 . Notice that i-1 (0) is the bundle of cosphe-
res of infinite radius, recognized as homeomorphic to the wave
front space W in VI,2 and VII,1 .
In case of the algebras over In mentioned it proved that

the above homeomorphism extends to a homeomorphism between all
of ffi and all of 8PTh On the other hand, in case of certain
.
algebras over the cylinder Slxg one obtains no secondary symbol

VII.O. Introduction 188
space at all, although the space a&*SZ\W is nonvoid. In each of

these two cases, on the other hand, I may be found (naturally
identified with) a compact subset of aF S2 For a third example,
.
however, - i.e., the half space algebra of [C1],V - this no longer

is true: certain one-dimensional "filaments" no longer are subsets
of aP 0
In [CS],VI a result, identifying the entire M with a compact
subset of 9&*0 , wa's discussed under a number of special assump-
tions, and for a compact commutator algebra. The proof was simpli-
fied in [CM2] . McOwen [MO] , [Mi] extended this result to a
larger class of Laplace comparison triples and algebras.
Here we adopt some principles of McOwen's discussion by
imposing a set of conditions (m,) , naturally satisfied in
McOwen's case.
In essence, we request that the pointwise 'inner product'
(0.1) {D,F} = hjkVjck + pr/q , xE52,
defined for folpde's D=b3ax.+p , F=cJaxi +r E D# , also is well-
defined over aMA# = MA#\S2 . In other words, the function {D,F)(x)
of (0.1) must be in C # In addition, certain separation condi-

.
A
tions are required. One finds that such a condition has the effect
that, over each point x0EM # , the set R-1(x0) retains the vector
A
space structure of the cotangent space, or that of a lower dimen-
sional space. This, in essence, makes the indicated result
possible: IN is found as a compact subset of the space 31*0
In sec.2 we specify cdn's (mj) , and first show a local
result, identifying t-l(x0) , for x0EM , with a compact subset
#
of a Tk . In sec.3 we add more conditions and then find 1-1(x0)

as a compact subset of a hemisphere. This also allows a proof of
*
IN + a& S2 (sec.3) . On the other hand, in sec.4 it is found that
the compactification aE*S2 is not unique over aM # , insofar, as
A
the expression (0.1) may be replaced by the corresponding one
for any subextending triple {O,du",H"} and then the formal sym-
,
bol by the corresponding expression involving the coefficients of

H" . Then one obtains an imbedding Bi + (a&*Q)" if the conditions
are right.
VII.l. General symbol space 189
All of this is discussed for a variety of examples, espe-

cially those of V,4 and V,5(cf. sec.2 and sec.4) . In particular
some examples of Sohrab [S1] , [S2] , [S3] , are discussed and
partly generalized.
In general we do not assume that the algebra C has compact
commutators.
1. The symbol space of a general comparison algebra.
In this section we will discuss a general comparison algebra

C = C(A#,D#) of a triple {G,du,H} with generating classes A#, D#
satisfying cdn.'s (a0),(a1), and (d0) of V,1. Again no essential
self-adjointness of H0 will be required. In general we cannot
expect the commutators of C to be compact. For example let 0 = fin,
dy=dx , H=1-A , with the Euclidean Laplace operator A Let us .
choose A# as the class of a=b+c E C°°(ln) with c E CO°°(jn) _

{a E C"(In): a(a)(x)=o(1) for all a} , and with b an n-fold
periodic function with periods ej=(d ) . Let D# be the
jl 1=1,...,n
class of folpde's bJ3 j+p , with coefficients b], p E A#. Then
x
it is easily seen (along the proof of V,prop.5.5, discussing the
1-dimensional case) that not all commutators of operators in C are
compact.
Generally we denote by E the smallest norm closed 2-sided
ideal of C containing all commutators of elements in C .
Lemma 1.1. We have [A,B] E K(H) whenever A E C , B E JO .
Proof. It is sufficient to give the proof for the case that A and
B are generators of their corresponding algebras. For a E Am
D E D# we get [a,DA] E K by V,thm 3.1.For a E A# D E D# let X E,
m
CD be such that XD = D . Then aDA = aXDA =_ DAaX =_ XDAa (mod K)
so that again [a,DA] E K (we used V, thm.3.1 twice here). For
DED# , F E D# we get DAFAv = ADFAv + Cv = AFDAv +C'v = AFXDAv
m
+ C'v = FFXDAv + C"v = FAXDAv + C"'v = MDAv + C"'v , for all
vEAC', where we used V,lemma 3.6, V,lemma 3.7, V,lemma 3.6, and
V,thm.3.1, in that order. Thus we also get [DAFFA] compact. Note
that not even condition (w) is required, for each commutation,
in view of VI,(1.2), since we always were commuting expressions or
functions with compact support. This completes the proof.
Lemma 1.2. Let XjE C_(G), 0<Xj<1, and Xj=1 in DJ U...UGj, for j =1,
2,..., referring to a locally finite atlas {sL } with properties as

in VI,2 again. Then an operator A E C is in the minimal algebra J0
if and only if there exists a sequence of operators C1E K with
A = lim1;.(X1AX1+C1)
in operator norm convergence.

Proof. If A E JO then its symbol vA (with respect to JO/K) is a
function in CO(W) , so that we must have liml..(1-Xl)6A = 0 in
uniform convergence over W . This directly implies (1.1).
Vice versa, for any A E C we have X1AX1 E J0. Indeed, it is
sufficient to show this for the generators of C , in view of the
relation X1 = and the commutator property of
lemma 1.1. But for a E A we trivially get aX12E JO . Also for A
DA , D E D# we get X1DAX1 E JO as well, q.e.d.
Theorem 1.3. We have E 0 JO = K(H) .

Proof. Note that T = E n J0 is a closed 2-sided ideal of the alge-
bra JO , and we observed before that T contains nontrivial compact
operators, while JO was seen to be irreducible in V,1. It follows
that T J K, by V,lemma 1.1, so that we only must show that TCK.
Let A E T. There exists a sequence A.-*A such that A. is a finite
sum of terms X=B[C,D]E , where B,C,D,E E C Also, for a sequence
.
of functions X1 with the properties of lemma 1.2. we must have

lim1.(X1AX1+C1)= A in norm convergence, by lemma 1.2.
Combining these two facts, we get
1 Ni
(1.2) uA-C1-J7_iXrXljXrf < 11A-C1-XrAXrIi+gXr(A-l=1X1.)Xrp < e
as r and j are sufficiently large. This means that we only must

show that terms of the form XXX = XB(C,D]EX are compact, where
B,C,D,E E C , X E CO , due to closedness of K in norm convergence.
Applying lemma 1.1 we get XXX =_ BX[C,D]XE (mod K). Hence we
only must show that X[P,Q]X E K for every P,Q E C In turn then
.
this may be reduced to the same statement where now P and Q are
generators of C , again using lemma 1.2. Let again w E CO , wX=X
For a E A, D,F E D we have X[ a,DA] X = X [ wa,DA] X E K , again by
lemma 1.2. Finally, X[ DA,FA] X = X[ wDA,FAw] X , where FAw = FwA + C,
by V , (3.12) again. Accordingly, X[ DA , FA] X =_ X[ wDA , FWA] X E K, which
completes the proof of thm.1.3.
The significance of thm.1.3 is realized when we now start
discussing the symbol space BI of the algebra C . We noted before
that CIE = C" , as a commutative CC-algebra with unit, is isome-

trically isomorphic to an algebra C(BI) of continuous complex-
valued functions over a compact Hausdorff space Bt, by the Gel'fand
Naimark theorem. The space BI is explicitly given as the collection
of all maximal ideals of C" Its topology is the weak topology
.
of the function class induced by the elements of C" The space .
M was called the symbol space of the comparison algebra C and ,
the continuous function corresponding to A E C the symbol aA of A.

We will recognize the wave front space 01 of the corresponding
minimal algebra as an open subset of BI .
Consider the injection J0+ C . The commutator ideals of J0

and C are K(H) = K and E , respectively, by V, lemma 1.1.
We have K C E , by thm.1.3 , hence a map JO~+ C" is induced
by A" = A+K + A+E ,(i.e. we have the inclusion A+K C A+E, so that
the map is independent of A(=-A' , hence well defined). Now by
thm.1.3 this map is an injection. For if A,B E J0 and A+K , B+K
both are subsets of A+E , then B-A E J0fE = K so that A~ = By ,
This means that inclusion of cosets defines a natural injec-

tion J0 -. C' , which is a *-isomorphism, and will be considered
as an identification of J0" with a subalgebra of C" .
Proposition 1.4. J0 is a closed two-sided ideal of C . Correspon-

dingly J0with above identification, is a closed two-sided ideal
of C° .
Proof. We have A E C in J0 if and only if A = liml (X1AX1+C1)

in norm convergence, with some C1E K , and with the sequence X1 of
lemma 1.2 For B E C we get B(X1AX1+C1) = X1BAX1+Cl'
. Cl'E K ,
by lemma 1.1. Therefore BA = liml(X1BAX1+C1*) E J0 , q.e.d.

For a closed ideal 1 of a commutative C -algebra A one defi-
nes the hull h(I) as the collection of all maximal ideals contai-
ning I. The hull is a subset of the maximal ideal space M of A
and a closed subset, since we clearly have
h(I) (1{ {m E M:4 (m)=0)

x
xEI
with the symbol homomorphism 4:A + C(.M) , so that h(1) appears as
an intersection of closed subsets of M
The complement R = M\h(1) is an open (locally compact) sub-
set of M and it follows that the restriction 1I of the symbol
,
homomorphism m to I constitutes an isometric isomorphism I + C(M),

separating the points of R, but vanishing outside of R . One thus
may identify the functions x with their restrictions to R, arri-

ving at an isometric isomorphism from I onto CO(R) which must be
the symbol homomorphism of the commutative C*-algebra I .
In other words, the space R may be regarded as the maximal
ideal space of I and the symbol homomorphism of I appears as
,
the restriction of the symbol homomorphism of A to R and I .
In our case of I = JO° C C" the maximal ideal space of JC

was the wave front space W , so that we have a natural identi-
fication of l with an open subset of the symbol space I of C
On the other hand we again may consider the C -function al-
gebra C # obtained as norm closure of the algebra generated by A#
A
(either in Lw-norm of functions over 0 or in operator norm of
L(H) , which give the same, (cf.[C 1 ], III,(8.14)). In the case
of J0 (i.e., A#=Am) we had C #=CO(O). For general A# the class
A
C # will be a subalgebra of CB(O) the algebra of all bounded
,
A
continuous functions over 0 The maximal ideal space M
A# of C A#
.
will be a certain compactification of 0 (For more details about

.
such compactifications cf.[C1],IV,lemma 1.5) As in VI,l we get

.
an injection CA#- C. Using that CA# n K = {0), we also get a cor-
responding injection C CIE = C(I) That injection again defi-

.
A #+
nes a corresponding associate dual map t:N -' M # , and i again
A
is surjective, similarly as in VI,1 in the special case of J0
Now we have the following result.
Theorem 1.5. The symbol space El of a comparison algebra C contains

the wave front space W of the corresponding minimal algebra J0 as
an open subset. Moreover with the map t:1+M # defined above we get
A
(1.3) W = i-1(a) .
Proof. The only thing left to prove is (1.3) . Note that we have
m E w if and only if there exists an operator A E J0 with aA(m)#0.
Using lemma 1.2 once more we get A = liml(X1AX1+C1) , hence
(1.4) aA(m) = liml (Xl(1(m))2aA(m)) = a
But then we must have i(m) E 0 , since the compactly supported

VII.2. Secondary symbol space 193
functions xl vanish at each point of3M # . Hence we get Gict-1(SZ).

A
Vice versa, for m E 1-1(S2) let A E C be an operator with aA(m)#0
and let the function x E be 1 at x0= i(m) E 0 . Then lemma
1.2 implies that XAX E J0 while we get axAx(m)=aA(m)x(t(m))
aA(m)# 0 , hence m E W . This completes the proof.
Note that the wave front space was completely characterized
as the cosphere bundle S*0 of the manifold 0 (cf. VI, thm.2.2)
We now denote the closure of 91 in M by ffip , and call ffip the
principal symbol space of the comparison algebra C . The comple-
ment Ms = ffi\ffip will be called the secondary symbol space of C .
Our next effort will be aimed at the structure of the secon-

dary symbol space. Sec. 2, below, will give some answers, in
that respect.
2. The space M\W , and some examples.
In this section we will look at the secondary symbol space

ffis. We will obtain some information about ffis , but under certain
restrictions on 0 and the triple at infinity, and not as complete
as for the wave front space. Again no assumption on essential
self-adjointness of HO will be made.
In addition to cdns.(a0),(a1),(d0) of V,1 the classes
A# and D# now will be assumed to satisfy the following.
Condition (ml): For each pair D=bJ3xi+p , F=c j3xj+r of folpdets
in D# the bounded C-(0)-function {D,F} defined by
(2.1) {D,F} = hjkFjck + Fr/q , x E S2 ,
is a function in C #, the norm closed subalgebra generated by A#.

A
Condition (m2): A# and D# are complex vector spaces, and D# is a

left module of A#.(i.e., the pointwise product aD belongs to D#,
for all a E A# and D E D#) .
A condition like (m1) was first used by McOwen [M1] in his
thesis written with the author. Also the approach, below, to use
(2.1) as a local inner product is due to McOwen.
For any given D,FED# the function {D,F}:O admits a conti-
nuous extension to the space M # which is a certain compactifica-

A
tion of 0 . Then let us focus on a given fixed point x E M #, fin-
0
A
ite or infinite. The value Q(D,F)={D,F}(x0) at x0 of the function
{D,F) as D,F range over D# , defines a positive semi-definite
,
sesquilinear form on D# , given explicitly by (2.1) at finite

points, but only by continuity at points x0E M #\D . The set
A
(2.2) D# = {DED# Q(D,D)=0}

x
0
defines a linear subspace of D# Indeed, one obtains Schwarz'

.
12<
inequality IQ(D,F) Q(D,D)Q(F,F) for the semi-definite form Q,
which implies that Q(D,F)=O whenever either Q(D,D)=0 or Q(F,F)=0.
Then we get Q(cD+c'F,cD+c'F) = 0 whenever Q(D,D)=Q(F,F)=0 .
It then is easily seen that Q induces an inner product in

the quotient space D#/D# = S x Clearly, for finite x, the space
.
Sx is just the local Mn+l spanned by (b1(x),...,bn(x),p(x)) , with

inner product given by (2.1).
Proposition 2.1. The space Sx is finite dimensional for x E M #,

A
and we have 6x=dim S < n+l (n= dimension of Q). Moreover, 6x
is a lower semi-continuous function of x, defined on 2MA#=MA#\0
(That is, lim inf x}x 6

x > 6x .)
0 0
Proof. If D1,...,D6 E D# is (a set of representatives for) an

orthonormal base of S then {D.,Dl}(x0) = 6j1
x
j,1=1,...,6=6x0 ,
But the 6x6-matrix V=(({Dj,D1}(x))) has continuous entries over

the space M # , hence will stay invertible in some neighbourhood
A
of x0 In fact, a nonsingular 6x6-matrix W(x)=((w )) may be
.
ji
found such that W*VW = ((6jl)) . Then Vj=Jwj1D1 defines an ortho-
normal set of Sx , so that dim Sx=6x? dim Sx = 6x . Thus we get

0 0
lim infx3x 6x > 6x , or 6x is lower semi-continuous. In particu-

0 0
lar, due to the basic properties of A# and D# of V,1 it is clear

that 6x = n+l for all x E 0 , since A# and D# contain all C--func-
tions and folpde's.
We now arrive at the following characterization of the

secondary symbol space (rather, of the space ffi\W of all points of
M over infinite x). Assume given an orthonormal base D1,...,D6 of
Sx at x0E M #, and a set of Sx folpde's DjE D# with {Dj,D1}=6j1.
0 A 0
For a maximal ideal m E 1-1(x0) let us define
(2.3) n(m) _ (aD A(m),...,aD A(m)) = (nl(m),...,na(m)) .
1 S
Theorem 2.2. The map n: 1-1(x0) i Md , 6=6X is continuous and

0
injective from 1-1(x0) C i\W onto a compact subset of fs .

Moreover, for M E 1-1(x0) we have the symbol of the generators
defined by
aa(m)=a(x0) , aDJA(m)=nj(m) , a(DJA)*(m) = Fj(m) ,
(2.4)
aDA(m) = *(m) = 0 for all D E D#
Cr .
(DA) 0
Proof. The continuity of n.(m)=aDA (m) is evident, since every
J J
aA:ffi->M, for fixed AEC, is continuous over ffi, hence con-

tinuous. The injectivity of n is proven just like Herman's lemma
in (C11,IV,2 Suppose for two maximal ideals m,m- E 1-1(x0)
.
we have n(m)=n(m-) This means that aa(m)=aa(m-) for all a E A#

.
(due to aa(m) = a(i(m)) = a(x0) = a(i(m")) = aa(m") ) , and

also aDJA(m) = nj(m) = nj(m-) = ODJA(m"') by (2.3) , . Now let us
first assume the following.
(2.5) aDA(m) = 0 for all D E D# and m E 1-1(x0)

x
0
Assuming prop.2.3 it follows that aDA(m)=aDA(n ) = 0 for all

DEDxu. Hence we conclude that the symbols of all the generators
of C coincide at m and m"' (including the symbols of the adjoints
(DA) , which are just the complex conjugates of aDA
In particular we find formula (2.4) confirmed. Also, the
two homomorphisms CIE - T defined by A-aA(m) and A-aA(m"') ,
respectively, for A E C , must coincide, since they coincide for

the generators. But these homomorphisms determine their correspon-
ding maximal ideals uniquely, so that we must have m = m- , pro-
ving the theorem.
Proof of prop.2.3. For D E D# we have {D,D}(x0) = 0 . Thus there

0
exists a neighbourhood Ne of x0 (in M #) such that I{D,D}(x)I

A
e2/4, x ENe, for every e>0, by continuity of {D,D} on M # . There
A
exists a function a E C(MA#) with 0<a<l , a(x0)=1 , a=0 outside of
Ne/2. Let M=(sup{1,I{D,D}(x)I : x ES2})1/2, Let b be in the algebra,

finitely generated by A# with Ib-al < e/2M . Then we get
(2.6) {bD,bD}1/2<{(b-a)D,(b-a)D}1/2 + {aD,aD}1/2 ,
at each x E Q The first term at right is bounded by e/2, at each

.
x, as follows from (2.1). But the second term also is bounded by

e/2 , since a(x)=0 , hence {aD,aD}(x)=0 outside Ne , while
{aD,aD}(x)I < Ilall 2{D,D}(x) < e2/4 there.
Accordingly we conclude that
(2.7) sup {I{bD,bD}(x)l : x E S2} < e
Then, using (2.6) we conclude that also
(2.8) IbDAI < e
Finally we observe that
(2.9) sup {IabDA(m)l : mGM) = inf{IIbDA+EII E E E) < IbDAI < e

using that the Gel'fand isomorphism of a C -algebra is an isome-
try. Also, ab(m)= b(x0) , by construction, and Ib(x0)-lI =
(b(x0)-a(x0)I< Ib-all < e/2M , which implies ab(m) > 1/2 ,for
sufficiently small e . Thus (2.9) implies e> Iab(m)IIODA(m)I

< 2e
> IaDA(m)1/2 , or, IaDA(m)I Since e>0 was arbitrary, we
.
conclude that aDA(m) = 0 , q.e.d.

With thm.2.2 we now have the
ID
following model of the symbol space
M
M , from the structure of the inver- MsI I
s
se images 1-1(x0), x0EM 0(fig.2.1). w

A
M is a certain compactification of
A# am aM #
SZ . All functions in A# extend conti- A A
Fig. 2.1.
nuously to M Q is dense in M # , we write am # = M #\ S2
A A A A
Then t: H - M # is a continuous surjective map .
A
The inverse image of the manifold 0 , i.e., the set
W=t-1S2 , called the wave front space, is (homeomorphic to) the
bundle of unit spheres in the cotangent space (or better the
spheres of radius - ). In particular the sets t-1(x0), for x0EQ
are n-l-spheres. On the other hand, over any x0E am # the sets
A
I -1(x0) are compact subsets of M6 , with 6=6(x0) the dimension of
the local space Sx
0
Since the topology of H is the weak topology, determined

by the set CM, an explicit knowledge of the sets n(1-1(x0)),
for all x0EaM # , and of the functions n , will characterize the
A
topology of I as well. In particular we have the following.
Corollary 2.4. For a point x0E am assume the dimension Ix

#
A 0
locally constant (i.e., 6x=6x for all , with some
0 A#
nNx 0
neighbourhood Nx of x0 ). There exists a compact neighbourhood

0
Ox E Nx naM
#
of x0 in am # such that the map t-1(Ox0 )-0x 0XM6
0 0 A A
6=6x , defined by
(2.10) m + (t(m),n(m)) ,
with the map (2.3) , defines a continuous injection of t-1(0x

0
xM6
onto a compact subset of 0x Here the map n(m) is defined
0
with an orthonormal base of Sx which still is a base at xEOx

0
for x E 0x , although not necessarily an orthonormal one.

0
Proof. The continuity again is evident. For the injectivity we

observe that t(m)=t(m-) implies that m,m-Ei-1(x) , where x=t(m)
Then if also n(m)=n(m-) we may repeat the conclusion of thm.2.2,
to show that m=m" . In particular, D# is independent of the base
{Dv} , and the orthogonality of the base is not essential. Q.E.D.
Remark 2.5. We observe that 6 =n+l whenever xESt. Therefore the

x
assumption that, for some x0EaM # the dimension dx is constant in

A
a neighbourhood of x0 relative to M , not only to 8M # , implies
#
A A
that dx=n+l , in such a neighbourhood. In this case an analogue
of cor.2.4 for a neighbourhood 0x relative to M # is valid.
0 A
However, in this case we can make a stronger statement
under slightly stronger conditions, as will be discussed in sec.3.
Remark 2.5. If we have Sx=d=const. in some subdomain BCBM

A# '
then every point in B has a compact neighbourhood N such that

t-1(N) is homeomorphic to a compact subset of NxM6 In many con- .
crete examples this property may be translated into a manifold-

like structure of the subset t-1(B) C !\W .
Let us now examine some explicit examples.
Example (A) of V,4. We obtain the algebra s0 of [CHe1].

with the classes As, DS, as follows. A# is the function algebra
finitely generated by a(x)=(x)-1
, sj(x)=xja(x) , j=l.... n
D#
s
is the class of all folpde's with coefficients in A# . Clearly
s
A# and Ds satisfy (m1) and (m2) . We get
(2.11) {D,D} S3b3 bJa

J 1
+ pp , D = nj +p bJ,pE A#s
xj
In
In [C1],IV we have seen that MA # is the 'directional compac-
s
tification' of In , corresponding to the closure in In of the unit

ball 5(1n) , with the map s : x->x/(x) taking In onto { x <l } C in
Accordingly, aMA #= 3En is an n-l-sphere.
Clearly an orthonormal base of Sx is globally given by
(2.12) Do = 1 , Dv=-ia , v=l,...,n , for all x E Bn

v
x
Thus the dimension 6x=n+l is,constant, we get
(2.13) A0=D0A = A ,
Aj=Sj=DjA
, nj(m)=oA (m) , j=0,...,n .
The Aj are self-adjoint operators, hence nj are real-valued. Also

A0 is a positive operator, hence n0>0. A calculation shows that
(2.14) IV=0 Aj *A 1
(1'Yv=1In.12)112
so that 0< no . Thus we have the following.
Proposition 2.7. The space M\W is homeomorphic to a compact

subset of the set
fn={,EBn+l. Iv_olnvl2=1 , no>o)
(2.15) DEnxEn , where
Bn is topologically equivalent to the ball

It is clear that
En (of infinite radius). In [CHe1] is was seen that
(2.16) MW = aBnxBn = aBnxBn .
From V,4 we know that the algebra s0has compact commutators,

as a subalgebra of the algebra obtained in example (A) there.
Similarly, for the algebra N of V,4, example (A) we get
(2.17) ffi\ID = aMA # X En .
Example (B) of V,4. Let us use the function algebra A# and
folpde-class D# introduced in V,4,(B). Write xEO=M3k=gjxTk in

the form x=(x',x") , with x'EB3 , x"ETk . A function a(x',x")EA#
must also have its x"-derivarives decaying, as x- (i.e., as Ix'I
- ). This implies that functions in A# must be nearly constant
on the k-tori x'=const., Ix'I large. In VIII,2 we will consider a
larger class A# which, however, will involve non-compact commuta-
tors for the corresponding algebra C. The compactification MA
#
c
pjk
will look similar to the Stone-Cech compactification of M3k
One may use exactly the same basis Dv , v=0,...,n as defined in
(2.12) (with D. , v>j, now acting on a periodic function). Again
- . Again we get (2.14).
the Aj=D.A are self-adjoint, and A0>0
13
Accordingly it follows again that MW C aMA #xIn . However, in the
c
present case, we do not have equality. Rather, the secondary sym-
bol space MW)int is homeomorphic to B3xZk i.e., to the charac- ,
ter group of the group Mj,k (cf. thm.3.1 of [CS],ch.5 ) .
We postpone the discussion of other examples of V,4. Here

let us first discuss an example of different type.
Bn
Example (H): Let 0 be an open subdomain of , with smooth
smooth boundary, compact or not. Assume the boundary ao to be a
smooth n-l-dimensional submanifold of En. We will focus on two
special case only: either QUaO is compact, or 0 _ l+={xeRn:xl>0},

referred to as "cases (a) and (b)" respectively.
Consider the Euclidean Laplace-triple,in either case, with
the classes A# and V# of restrictions to Q of the functions
(folpde's) of As and Vs of example (A), above. Let C be the cor-
esponding comparison algebra.
Again we may use the orthonormal base Dv of (2.12),for
every xEaMA# . (We have MA#=cU2c2 in case (a) and MA#=B+=(g+)clos
with the closure of In in En .) The dimension is 6=n+l everywhere.

However, an investigation shows that the operators Aj no
longer are self-adjoint, although we still have (2.14). In par-
ticular, focusing on case (b), the operator Al , involving the
derivative normal to the boundary, is not even self-adjoint modulo
E . Accordingly the set n(t-1(x0)) , for a point x0 with x0=0
only is a subset of the complex unit sphere 1v=O1nv12=1, nEWn+1.
For a precise description of the space M cf.[C11,V,thm.10.3.

The description there is in a slightly different representation,
to be translated into the present form. Over points of the finite
boundary of 0 one only gets the (still well-defined) cosphere
bundle, and, in addition certain 1-dimensional 'filaments'. Over
points lxl=00, xl=0, one obtains the ball JEJ<co (i.e.n real, lnI=1,
c0>0 , and again some 1-dimensional filaments). For other points
Ixl=° we of course get the same as for D=Yn , A# , VS again.
In [C1],V we also investigate the commutator ideal E of case
(b) in detail, and thus obtain a complete Fredholm theory for the
algebra C. We have not analyzed the space ffi\T in case (a) , but
expect a very similar result.
Note that the last example pertains to the case of a genuine
boundary problem, where cdn.(w) is never true. In [C1] and [CE] we
also show that the elliptic boundary problem with 'Lopatinsky-
Shapiro'-condition is solvable with our algebra C
3. Stronger conditions and more detail on ffi\W .
It should be realized that our assumptions on A# and V# as

well as on the triple {O,du,H} are minimal, so far. In sec.2 we
have studied some examples which allowed a more complete descrip-
tion of the space ffi\W . We now first will look at conditions to
VII.3. Stronger conditions 201
allow a description of the sets i-1(x0) C I\W as compact subsets

of a real upper hemisphere ffl 1, just as in the examples of sec.2,
although only locally.
For the Rn-related examples of sec. 2 the following observa-
tion may be useful.
Remark 3.1. Suppose near a point x0E3M # the manifold Q 'looks

A
like Rn,, in the following sense: x0 has a neighbourhood Nx
0
relative to M 4 , such that = U is a chart. In particular,

A QnNxO
assume that, if U is considered a subset of In , then either x0 is
a point of a smooth boundary piece of U or JxoJ=- , and ,xl
for all xENx \R . Suppose then that, in the coordinates of U ,
0
the coefficients hjk(x) , q(x) extend continuously onto Nx , and

0
still define a positive definite (n+l)x(n+l)-matrix, at all points

of N x0 .Suppose also that D# contains folpdes D v such that
(3.1) D0 = 1 , Dv = -i3 v , v=l,...,n , for xEU .

x
Then we have dx=n+l in Nx , and it is possible to use the folp-
0
de's (3.1) as a base of Sx, although not necessarily orthonormal.

Accordingly cor.2.4 applies in the entire Nx ,with n defined
0
by (3.1). For a generator A =a(x), or =DA or =(DA)* of C we then

have the value of aA(m) at some mEt-1(x0) , x0EaM # , given by
(3.2) oa(m) = a(x0) , aDA(m) = b3(x0)nj b3(xO)nj.

' a(DA)*(m) =
for D=InbvDb3Dj+p, summation convention used from 1 to n,only
for non-greek indices.

In thm.3.2, below, we will show that, in this case, we get
(3.3) hJk(x0)njnk + q(x0)1no12 = 1 , n0>0, for nj = nj(m) .
Therefore we may write (3.2) in the form
(3.4) aa(m)=a(x0), aDA(m)=(b3(x0)Cj+p(x0))/(hjk(x0)Z* jCk+q(x0))1/2

.
using the coordinates j=nj/n0 , with a E E ]En
Here the vector _ is not always finite. It is
assumed to be a point in the directional compactification In
of Cn , defined similarly as n above. If, in addition, it can
,
be shown that all nj are real, then we have E,n (but we should
recall example (H) of sec.2, of a comparison algebra on a mani-
fold with boundary, where we do not get all nj real).
In this representation, the space N x$n appears as a part
x
0
of the compactification of the cotangent space T*c2 of 0 , genera-

ted by the bounded continuous functions (3.4) over 0 . Moreover,
the representations of t-1(x0) C M of thm.2.2 are obtained by a
change to 'projective coordinates'.
In the following we will seek to recover these features.
While we return to the general case, assuming only (a0),(a1),(d0),
and (m1),(m2), (and not (w)) we impose (some or all of) the follo-
wing additional conditions.
Condition (m3): For each x0EM # and open neighbourhood N of x0
A
there exists a function a E A# with 0<a<l and a=l near x0, a=0
outside of N and such that
, Da=alj3 E D# .
xj
Condition (m4): There exists an c > 0 such that
A-1-e[a,A]
(3.5) E C , A-'[D,A] E C , for all a E A# , D E D#
where the last expression should be interpreted as A-'(DA-(AD)**),

as a product of unbounded operators, to be continuously extended.
Condition (m5): We have A E C .
Condition (m6 )x The dimension 6x of S#=D#IV# is maximal (i.e.,

we have 6 = n+l).
x
We define
(m6) = U{(m6)x :x E 3M #} = U{(m6)x :x E M #},
A A
i.e., cdn.(m6) (without subscript x) means that (m6) holds for
all x E M # (it is always true for xEO) .
A
Note that (m3) is a separation condition comparable to the
condition describing completely regular algebras (cf.[C1],AII,5).
On the other hand, (3.5) may be compared to the statements of V,
lemmas 3.3 and 3.6. Using the techniques of V,3 one may reduce
(m4) to explicit conditions on a E A# or the coefficients of DED#,
but we will not make a systematic discussion here. Note that
(m5) holds if 1 E D#. For a weaker condition cf. IX,5 .
Theorem 3.2. Assume condition (mj) , j=1,2,3,4,5,6, and let Dj,

j=l,...,6, of thm.2.2 be formally self-adjoint. Then the image
set of the map n:i-1(x0) - M6 is contained in the sphere
(3.6) {n E ]Rd :
1v_lInll2 = 11 C In
Proof. Since the dimension 6(x) is lower semi-continuous, it has

to be constant near x0 if it is maximal at x0 Thus we conclude .
that dx= n+l near x0 .
We have Parseval's relation in S for the orthonormal set

x0
D1,...,D6 . For every D E D# we get
(3.7) {D,D} _ 1d_ll{Dv,D}I2 , at x=x0
Due to continuity of {.,.} as function of x this implies that
(3.8) I{D,D}(x) - Ea=1I{Dv,D}(x)I2I < E{D,D}(x) , as x E Nx

0
for a suitable neighbourhood Nx of x0 . Indeed, we already used

0
in prop.2.l that D-v(x)=Xwvu(x)Dm , for suitable W=((wvu)) close
to C(& )), is orthonormal at x, for every x E N . Since 6=n+1,

X
0
we will have (3.7) at x E Nx for D-v instead of Dv . This im-
plies (3.8) , by a simple estimate, as Nx is chosen sufficiently

0
small. (For more details cf. proof of IX, thm.2.1) .
Proposition 3.3. Let a E A# equal 1 and 0 , as x near x0 ,and x

outside Nx =Nx E, respectively, and let 0<a<l. Let AV = DvA ,
0 0'
v=l,...,6 Then for all u E AC0'(c7) we have the estimate
(3.9) IJ,,dv(a2(hJk(Au) (Au) k+gIAul2)- fIaAvul2)I < EQu02

x x v
This proposition follows from (3.8) by a calculation.
Actually, (3.8) implies a corresponding local estimate, and then
we may use that Bul12 = (u,Hu) _ flA-lufl (cf. V, (1.4) ) .
Using (m3) we now may bring the factor a2 of the first term
in (3.9) under one of the derivatives(Au)lj In detail we get .
J = J,,dpa2(hJk(-)Ij(Au)1k+gluI2) into the form
(3.10) 2J = (a2Au,A-lu)+(A-lu,a2Au)-2((Aa(DaA)+(DaA)*aA)u,u)
with the inner product (.,.). Apply the first relation (3. 5) for
(3.11) a2A = A(a2+aC+Ca+C2) , C = A-1[a,A] E C .
It follows that the first two terms can be written as
(3.12) (a2Au,A-lu)+(A-lu,a2Au) = 2(u,(a2+C')u) , C'=aC+Ca+C2
Finally (3.9) with (3.10), (3.12) may be polarized for
(3.13)
2
a av
1((a +C'-(Aa(D A)+(D A) aA)- JAa2Av)u,v)j<cE(Ilull2+Avg2 ),
V
valid for all u,v E ACS . All operators in (3.13) are continuous,
hence we may extend to all of H Denote the self-adjoint operator
.
at right by Z and introduce Zu=v , for
(3.14) IIZuf2 <E(fluu2+11Zu112) , u E H ,
which results in
(3.15) a - J(aAv)*(aAv)-C'-(Aa(DaA)*+(DaA)*aA)=Z , fZR<E/(1-E) .
v
Accordingly we also get loZ(m)I < E/(1-c) . However, for a
maximal ideal m E 1-(x0) we get aa(m) = 1 , and aaA (m) = aA (m)
V v
= n v(m). Also the symbol of all other terms of Z vanishes at m
Indeed, we get Da E D# , by (m3), while Da=O near x0 , by (m3)
Hence one may find a function X E A# equal to 1 near x0 and 0 on
supp Va , and set w=l-X E A# , wDa = Da . It follows that aD A(m)
a
w(x0)aD A(m) = 0, which makes the symbols of the last two terms
a
vanish. The same may be used for C' = A-1[a,A] . With a resolvent
integral we get
(3.16) C' = 1r-l J-A-1(H+A)-1[H,a](H+A)-1X-1/2da

0
*
Using V,(2.24) we get [H,a] = Da , where we again use wDa
= Da . (Note that (H+A)-1 , A-1(H+a)-1 , A-2(H+A)-2 E C , so
that the integrand is in C and has symbol vanishing at m . The

integral exists, by the standard estimates. Thus ac,(m) = 0.

As a consequence of (3.15) we therefore get
(3.17) I1 - Iv=11nv(m)I21 < e
This estimate holds for all e>0 , so that we get
(3.18) 1v=l1nv(m)I2 = 1 .
Note that self-adjointness of the Dj so far has not been

used. This condition will be essential if we try to show that
n(m) is real. Clearly the complex conjugate of nj(m)=aA (m) is
J
aA *(m) . Also A (DjA)* (ADj)** , using that DD . By the
second relation (3.5) we get
(3.19) (DjA)* = DjA + ACC , C E C
If the maximal ideal m considered has the property that aA(m)=0

then (3.19) implies at once that nj(m) is real, since the symbol
of the last term vanishes. On the other hand, we get
(3.20) [D.A,A] _ ((DjA - (ADj)**)A E E ,
**
so that Y = DjA - (AD satisfies aA(m)oy(m)=0 , which again
gives aY(m) = 0 (,i.e., n(m) real ) whenever OA(m) 9 0 Thus .
we indeed have proven our theorem.
Remark 3.4. Note that we have shown above that relation (3.5) 2
and (m5) imply relation (3.21) below. Similarly, (m4) and (m5)
1
imply (3.21)2 :
(3.21) [D,A1 E E , A-1[a,A] E E , for all D E D# , a E A#
Indeed we showed that the symbol of [D,A] vanishes for all m .
Similarly, A (A-1[ a,A]) = [a,A] E E . so that ac , with C=A-1[ a,A],

vanishes whenever aA90. But C=ACC', C'E C, which implies aC=O
also if aA=(a C) 1/e=0, or, again aC=O for all m it, It, i.e., C E E.
Corollary 3.5. Under the general assumptions of thm.3.2, if we

only require (m.), j=1,2,3,4,5, but not (m6) then we still have
(3.22) n(i-1(x0)) C B6 = {xER6 :

IxI<l} .
Moreover, if also the assumption of selfadjointness of the expres-

sions Dj is dropped, then we still have
(3.23) n(t-1(x0)) C {xEad : Ix1<1} .
Proof. With the present assumptions we can substitute (3.8), (3.9)
by the weaker estimates and
(3.24) (1+E)10du(a2(hJk(Au)Ixj(Au)Ixk+gIAu12) > Xa=1IIaAvul2
Here the left hand side may be treated exactly as above. One finds
that it is of the form (1+E)(u,Vu) where VEC has symbol aV(m)=1 .
Write Z=16=1Aa2Av , and conclude from (3.24) that (1+E)V-Z>0
hence (1+E)V-Z=T2 with a positive self-adjoint T E C From this .
we conclude that (1+E)aV-aZ=((y T)2>0 , which yields In(m)I2=az(m)

< (l+e)aV(m)=l+e . Since this holds for every a we get the sta-
tement (3.22) . Now (3.23) is trivial, q.e.d.
Let us now come back to a representation of the symbol space
IN as a compact subset of a certain compactification of the cotan-
gent space T O of 0 as discussed for examples earlier in this
,
section. For our generators a, DA we define formal symbols similar

to (3.2) by setting
(3.25) ta(x,E)=a(x),
for all (x,E)ET*D , where D = -ib32 j+p E D# We require (m4)

x
and (m5), hence can be assured that (AD)**=DA+E , EEE , by remark
3.4, so that (AD) need not be taken as generators, mod E
Note that the formal symbols are bounded complex-valued
CW(n)-functions over Q. Hence the class of functions (3.25), for
all aEA#, DEV# determines a smallest compactification of D, called
*
R , with the property that each function (3.25) admits a
*
continuous extension to Q (cf. Kelley [Ke1]) . This is
just the maximal ideal space of the algebra of bounded continuous
T*O
functions over generated by (3.25). The injection of the
algebra generated by A# into that algebra induces a surjective
associate dual map n: &0 - M # again, defined just as the map i
A
Let aIE*sl = E*R\T*D .
For the theorem, below, we will require a somewhat stronger

form of cdn.(m1), as well as cdn.(m7), below.
Condition (m1') : We have cdn.(ml) , and, in addition, p2/q E A#,

for all D=b33 .+p E D#
X3
Condition (m7) At each x E aM # there exists a folpde D E D#,

:
A
such that {D,D}0(x)#0 but {D,D}i(x)=0, with {D,D}j of (3.26).
,
Under cdn.(m1') we do not only have {D,D}(x) well defined

for points xE3M # , but even the decomposition
A
(3.26) {D,D} _ {D,D}0+ {D,D}1 , {D,D)O=IPI2/q , {D,D}l=hjkFibk
extends to 3MA# Then (m7) requests a nontrivial DESx which
'emphasizes the 0-order component'.

Clearly it makes sense to request (m1') or (m7) only at
a specific point x E 3M # Then we will speak of (ml') x or
.
0
A
(m7)x again.
0
Theorem 3.6. Assume that (m1'), and (mj) , j=2,3,4,5,7, hold.

Then the maximal ideal space fi of the comparison algebra C is
homeomorphic to a compact subset of the space aF*Q. Under this
homeomorphism the wave front space W is mapped onto the bundle
7r-1(S2) of co-spheres of infinite radius. Moreover, if we regard It
as a compact subset of aF S2, by virtue of this homeomorphism, then
for each mEM, and every generator a E A#, DA, DED#, the symbol is
the restriction of the formal symbol . That is, aa(m) and aDA(m)
equal the values of the continuous extension of a and TDA ('of
(3.25) ) to ff*0 at m , respectively.

Thm.3.6 was first proven in [CS],ch.6. (cf. also [CM2])
understronger assumptions (cdn.(14),of IX,3). The idea of the
present proof is due to McOwen [M1], where the result
was established for the Laplace comparison triple of a complete
Riemannian manifold, and the sets A# and D# of functions (folpdes)
with covariant derivatives (of the coefficients) vanishing at - .
McOwen assumes the additional condition that the dimension 6x0 is
constant at each end of 0, and that E = K(H) .

Proof. We only must provide an injection ffi-.21 *SZ with the proper-
ties of the theorem. This will define a continuous map since the
topology of M as well as that of 2B*S2 is determined as the weak
topology of the same function class (3.25).
One finds that, for x0EO,the set7r-1(x0)rig B*O is a sphere -
there is one boundary point of 2F 0 in each direction T-*-,
IEI=1. By VI,thm.2.2 we may define the injection over x0 as the

*
map E-E from the co-unit-sphere to Tr-1(x0)r12B 0 Therefore we .
only must focus on points x0EaM #. First assume cdn.(m6)x . Then,

A 0
for a formally self-adjoint orthonormal base D as in thm.3.2, we ,

v
SS-1
may define the injection map n t-1(x0)+ of that thm., with
:
S6-1
the unit sphere of (3.6) . Under the present assumptions we
may choose the base Dv such that DS=Dn+1 is a folpde with

1 , just by using a multiple of the assumed D#0 with {D,D}1=0
Then a conclusion like that in cor.3.5 may be used to show that

nn+l(m) = oD A(m)>0 (where, of course (m3) is essential). Accor-
d
H6-1 =
dingly, the map n now takes t-1(x0) into the hemisphere
{HES6-1:n6>0} .
Now we claim, that also n-1(x0) is homeomorphic to (the

HS-1
entire) . For pGTr-1(x0) let
(3.27) with (P)=TD A(p)
V
Observe that TD A is real-valued , by self-adjointness of Dv
and that 1v=1Cv2(p) = 1 , again follows from Parseval's relation.
Thus we get C:7T -1(x0)-S6-1 again, and the injectivity of C may be

proven as in thm.2.2 again. In fact, we also have r,>0 , so that
HS-1
C again maps into
To show the surjectivity, let us introduce 'projective coor-
dinates', for (x,E)ET S2 , with x near the point x0 , by setting
j=nj/no , n0>0 . We get
(3.28) (bjnj+pno)/(hjknjnk+gn02)1/2
Here we introduce n0=r(x)/q(x) , nj=hjk(x)ck(x) , with the (real)
coefficients of some folpde F=-icla j+r, into the right hand side.
x
1/
Then it will assume the form {F,D}/{F,F}2 .
For j=nj(x)/Jr(x)J
we thus have
(3.29) ±{F,D}/{F,F}1/2
In (3.29) set D=Dv . Also assume Dv orthonormal in some neighbour-

hood of x0 , and let {F,F}(x)=l , near x0 , without loss of gene-
rality. It follows that
(3.30) TD A(x,C) _ ±{F,Dv} ,
where the same sign holds for all v , since the sign is determined
as the sign of the coefficient r(x) If F runs through all unit .
vectors of the 6-dimensional space, at some x , then the 6-tuple

({F,D
v v=l,...,d
runs through the entire unit sphere. Thus, for
any point XES6-1 we have either X or -X assumed by (TD A(x,&)) .
The same property then must hold for the values gy(p) , as p runs
H6-1
through it-1(x But we already know that
) . (p) E Thus it .
0
(H6-1)int
follows that all points of must be s assumed. Since in
Ha-1
is closed, we then conclude that im = i.e., is sur- ,
jective.
It then follows that the map is well-defined. Let
p = v(m) , for mE1-1(x0) . We get
(3.31) Ta(p)=a(x0)=aa(m) , TD A(p)=Cv(p)=nv(m)=oD A(m)

U
so that indeed formal symbol and symbol coincide.

On the other hand, if (m6)x is false, i.e., 6x <n+l, then
0 0
BS-1={EEI6:I&I<1}
we have seen that n maps into the ball (cor.3.5)
Again we may assume that D6 satisfies at x0, and then con-

clude that nd>0 so that the map n really goes into the half-ball
,
B+d-1={nEBS-1:nd>0} .
But if 6x <n+l, then we get 6x= dim Sx = n+l>dim Sx = ax ,
0 0 0
for all x-'Q, and every neighbourhood of x0 contains such points.

Repeating the above conclusion we find that the vector (TD
V
CEBa-1
now assumes either C or -C, for every . The same again is
true for the vectors c(p), as p runs through w-1(x0) Thus, in .
B+6-1,
this case, the map also goes onto and we again may define
the map v(m) . Thus we have the same statement, and thm.3.6 is
VII.4. Structure of ffi\W; Examples 210
established.
4. More structure of M. , and more on examples.
Remark 4.1. Before we attempt application of the results of sec.3

let us point out that thm.3.2, cor.3.5, and thm.3.6 all have
'local' generalizations, in the following sense:
Under cdn's (m3)x and (m6)x only, for some given point x0
0 0
E 3M # , we still have the statement of thm.3.2 true for that

A
point, regardless of all other points, assuming the other (mj),
of course. Under (m3)x alone we still get cor.3.5, at x0 .
Again, under (ml')x ' (m3)x ,

with the other (mj) required for
0 0
thm.3.6 we get the homeomorphism of thm.3.6 at least between

C1(x0) and r-1(x0) , although perhaps not between ffi and 3F SZ
Here, by (m3)x we mean that the function a of (m3) exists
0
for that x0, and its neighbourhoods. By cdn.(mi')x we mean that

0
(ml') holds 'near x0', i.e. {D,D} and each coincide with some
function in A# near x0 . Similarly (mi)x , which may be substi-
0
tuted for (ml) , above.

We shall refer to these local generalizations as thm.3.2x ,
etc. Note that there also is a thm.2.2

x0
Let us point out again, that, speaking in terms of thm.3.6,

examples have been given which show that the set i-1(t ) , for
some x0E3M # , may range from the minimum (the set Wclo n -1(x0) Tr
A
to the maximum (all of ff-1(x0)) . In particular, example (B) of
V,4 shows that cases between these two extremes occur: For the
Laplace comparison algebras of the polycylinders G3k we may get
more than Wclo:h n-1(x0) but not all of n-1(x0) , rather, only
some sub-surfaces of the hemisphere Ha-1 are contained in I .
As a partial answer to the question for the set i-l(x0) we

prove the result, below.
Theorem 4.2. Under the assumptions of thm.3.6x , if there exists

0
VII.4. Structure of 81\W; Examples 211
a function aEA# with a(x0)90 , and aD.0AE E , where D0 denotes the
folpde assumed in thm.3.6, with {D0,D0}={DO,DO}0#0 , at x0 , then
we have
(4.1) t-1(x0) = t-1(x0)f'$Tclos
Proof. Let mEHts = ffi\Wclos = ffi\,s , and let 2(m)=x0. Since aDDAEE
we conclude that 0 = aaD A(m) = a(x0)OD A(m), or aD A(m)=0, since
0 d A
by assumption. Therefore the statement is an immediate
consequence of cor.4.3, below, q.e.d.
Corollary 4.3. Under the assumptions of thm.3.6x , with the

0
operator D=D0 of thm.3.6, we have
(4.2) t-1(x0)nIDclos={mE1-1(x0):aD A(m)=0}={mE1-1(x0):TD A(m)=0}

.0 0
Proof. First notice that TD A(m)=0 for points of Wclos, Indeed,
0
we know that, for an (x,F)EW we have so that the formal

symbol assumes the form
(4.3) T (x,E) = b0E?/(hOkE0E0)1/2 ,

for 01=1
DA 1 3 k 1
With the technique of the proof of thm.3.6 we express this as
(4.4) TDA(x,C) _ ({D,F}1/({F,F}1)1/2)(x) ,
where F is a suitable folpde. (Note that the sign of the 0-order

term no longer is essential. ) Since we have {DO,D0}1(x0)=0 , we
conclude from (4.4), that TD A(p) = 0 , for all p E Tr(x0)
whenever p E Wclos, which proves the above.

Now, again in the terminology of thm.3.6, we must show
that the restriction of the map p to the set {mEt-1(x0):r) 6(m)=0}
maps onto the sphere (or ball) ns=0 However, for xE12 near x0
.
the values of the vector (aD A(m)) , mE1-(W), are exactly the
v
values of (TD (({Dv,F}/({F,F}1)1/2)(x)) , taken over all
folpdes F, defined near x. From this one concludes that also in

the limit x-x0 every vector of the sphere or ball is assumed.
Remark 4.4. Let us emphasize again, that thm.3.6, and cor.4.3

VII.4. Structure of ffi\W ; Examples 212
require conditions (m4) and (m5), which imply that, for a basis
(mod A#) of self-adjoint folpdes DIV the corresponding generators
DvA are self-adjoint mod E . If these conditions are violated,
as in the case of a comparison algebra associated to a boundary
problem, (cf. example (H) of section 2), then N no longer must
be a subset of the space 3P 0 . In [C1] and [CC1] we discussed
such an example, where it turns out that certain additional
'filaments' of N occur, which cannot be found from the values
of the formal symbol.
Remark 4.5. As another feature, illuminating the role of the

formal symbol and the compactification 3&*sl let us observe that,
,
in all of the results of sec's 2, 3, 4, involving the sesqui-lin-

ear form {D,F} , D,FED# , we may use the form {D,F}^(x) of any
triple {O^,dp^,H^} (c> {O,dp,H} , defined as
(4.5) {D,D}^ = h^jkFibk+IPI2/q^ ,
and by polarization for {D,F}^ . Actually, one may use just any
similar sesqui-linear form, positive definite over each space
Mn+1 of n+l-tuples (Vu(x),u(x)), at every x E 0 , as long as the
estimate V,(1.6) remains satisfied with {D,D}^ instead of {D,D}.
It is clear then that we get corresponding conditions
(mj=1,6,7, as well as (ml')^, and the corresponding local
conditions (mj)^x etc. (Note that cdn's (mj), j=2,3,4,5, do not
depend on the form {D,F} ). Here it is essential that (ml
implies that {D,D}^<co for all DEV#. With the modified conditions
substituted for the ordinary ones we then again get thm.2.2,
thm.3.2, cor.3.5, thm.3.6, thm.4.2, cor.4.3 true again, with
exactly the same proofs, and including all the local versions.
However, the formal symbol changes with the form {D,F} . We
now must take
T^DA(x,E)=(b3Ej+p)/(h^]kEjEk+q^)1/2
(4.6)
*
instead of (3.25). Accordingly, the boundary 31 0 of the compac-
tification F 0 of T 0 also changes. In fact, one easily verifies,
* *
that there exists a surjective continuous map 8&^ 0 - 2F 0 , which
*
also is injective in WQ F^ 0 and acts as a homeomorphism between
the two images of W . Over the complement of W, on the other hand,
points may split into several points. In other words, the repre-
sentation of IN as a subset of 3r 0 changes with the form {D,F}
VII.4. Structure of M\W; Examples 213
Now let us return to the discussion of explicit examples

again. First we look at the comparison algebra of V, thm.4.1. In
details, we have 0 given as a subdomain with smooth compact boun-
dary M of a complete Riemannian manifold 2^ without boundary with
metric tensor h^3k and measure dp^ aQ is an n-l-dimensional sub-
manifold of Q^ . We have dp=duÎQ and the tensor hIk is the re-
,
striction of h^jk to N , but q is not necessarily a restriction of

q^. We have q > q^ and {D^,du^,H^} (c> {S1,dp,H} in the sense of
,
V,3. For formal reasons it is convenient, however, to impose

the following.
Condition (ql) : We have q(x) = q^(x) in N^\N , with some neigh-

bourhood N of 30 .
To satisfy V,(3.13) and (w) we again require (as in V,4) that
(4.7) q(x) > Y1(dist(x,3c))-2 , and V(log q^) = o(1) (in R^)
Now it turns out that the classes A' , Do of V,4 in general

will not satisfy conditions (m.), j=1,2. To discuss an example
only let us work with the form {D,F}^ of the above h^jk and q^
and the following more restricted classes called A# and Do
AC is the class of all bounded functions a E C'(c) such that
(4.8) Va = o,^(qÊ/2) , for some e, 0<E<1 .
(The difference to the class A# is that the functions a E A# must

have their derivative bounded near aQ , while Va for a E Ac may ,
not remain bounded, but only must be o(qC/2), O<e<1, near a2

DC is the class of all D=bJa j+p with bounded {D,D}^ and
x
{D ,D }^ (bounded only over N) satisfying the following stronger
version of V, (3.16): There exists an e , 0<e<1 , such that
(b'lk)=o0^(qê/2)
, b3(log q)lj=oN^(1) ,
(4.9)
V((bJ01j/(c)/q^))=0S1^(qÊ/2)
Again the requirements are distinctly different near - of V,

and near 3Q Near - the covariant derivative balk stays bounded,
:
near - it must be o(qe/2) Near 3N the last two conditions just

.
amount to
(4.10) Vp=0(1) , V((b0C)lj/)=0(1)

VII.4. Structure of ii\w; Examples 214
Clearly we have A# c Ac , D# c D# . A calculation, not dis-

cussed in detail will confirm the following.
Proposition 4.6. Let cdn.(gl) hold for a triple as described above

satisfying (4.7). Then the classes A# and D# satisfy (m1)^, (mj),
j=2,4,5,( as well as (miMoreover, A even is an algebra,
and DC satisfies the condition of thm.3.6^, for every xE3MA#, with
the expression D=l, at every point.
In particular (m4)^ is an immediate consequence of V, lemma
3.3 , and V, lemma 3.6, since K c C .
We now consider the comparison algebra C = C(A#,D#) . This

algebra clearly has compact commutators, and contains K(H) . More-
over,it also satisfies the conditions of thm.2.2, so that we get
the sets 1-1(x0) imbedded in M6 , for each x0 E 3M # .
A
Note that the functions of A# must have continuous exten-
sions to 52U812, since they have their first derivative bounded near
8S2 , by (4.6). Also A# contains all restrictions to 0 of functions
in C_(52Uac) . This shows that 352c8M # , In fact, we find that 3M #

A A#
is a disjoint union of 3Q and a set 8MA# 'at , where both '
sets are separated, i.e., have disjoint closures. Again, it is

easily verified that (m3)x holds for all points x0E352 , since 3S2
0
is smooth. The corresponding property (m3)x , for points in 3M #,

0 A
will be verifiable only under additional conditions on the Rieman-
nian space (2 , as well as on the potential q . For a more detailed
investigation of M\W in case of the Laplace comparison algebra of
a complete locally symmetric Riemannian manifold cf. McOwen [M1].
For a complete characterization of 1 in case of a Schroedinger
expression on In, for certain (spherically symmetric) potentials
q , cf. Sohrab [ S1] , [ S2] ,[S31.
Theorem 4.7. Under the above assumptions, with the decomposition
3MA# = 352 U 3MA# , we have
(4.11) 1-1(352) n m s = 0
That is, any possible points of the secondary symbol space His
_ fl\(Wclos) are over infinity, not over 30 .
Proof. We just apply III,thm.3.7 to conclude that aA E K(H), for

VII.4. Structure of M\W; Examples 215
every aEC_(OUdO) . For a point x0E3O one may find such a function
with a(x0)#0 . Since E=K(H),and (m3)x holds, we have the con-
0
ditions of thm.4.2 satisfied for {.,.}^ . It follows that there

indeed is no point of ffis over x0 , q.e.d.
The above discussion translates literally to the language
used in V,4(E) . Details are left to the reader. On the other
hand, just to illustrate the variety of problems encountered, let
us shortly look at example V,4(D). For the classes A# and D# used
there one finds that the form {D,D} is .well defined, but{D,D}^
is not generally defined for all DED# . Moreover one finds that
D# is not an Accordingly, none of the results of sec's
2, 3, 4 applies for these sets Ac , D# , although the algebra
C(A#,D#) has compact commutators and a well-defined symbol.
For example, one now might well expect nontrivial parts of the
secondary symbol space over 3O We have not tried to apply the
.
techniques described, although this should be easy and only a

formal task.
Similarly for the examples V,4(C) and V,4(F), which will not
be discussed in detail.
Finally let us discuss some 1-dimensional examples, exten-
ding an investigation by Sohrab [S2], where a complete knowledge
of the symbol space will be obtained. These are examples of
a Sturm-Liouville-triple V,(5.1).
Applying the Sturm-Liouville transform (V,prop.5.1(a)) we
may assume V,(5.1) in the form V,(5.3). In that form we assume the
interval I- to be which amounts to the condition that
s
(4.12) (K/p)1/2dx does not exist on both ends a,s
a
We assume the triple in the normal form {1,dx,q-3 X} , where
q(x)>1 on I . In order to reach the generality of [S2] all con-
ditions imposed below must be translated back to the general case.
In order to have only one point of M # at each end of the
interval we select A#=algebra finitely generated by C_(I) and
s(x)=x/(x) , and D#=span {DO=/q, D1=-i3x}(mod A#). We want E=K(H),

hence we require V,(3.13) and V,(3.15), V,(3.16), with q=q^. A
calculation shows that this requires V,(3.13) , i.e.,
(4.13) q' = o(q)

VII.4. Structure of M\Q1; Examples 216
as only additional condition. (Actually, we could allow the lar-

ger function class A#:A# of V,4, and still would get E=K under
(4.13) only, but this would give us many points in aM # over ±o'.)
A
Note that (4.13) will not allow'potentials of exponential growth,
while potentials of arbitrary polynomial growth often are accep-
table. In particular, the potentials q(x)=l+x2 and q(x)=1+x2 +x4
of the harmonic and the anharmonic oscillator in quantum mechanics
satisfy (4.13).
All conditions (m.), (ml') hold, as well as the assumption
of thm.3.6 (with We will show that
(4.14) IN = 3P*(R) ,
where the form {D,D} of hll=1 and q=q^ is used. We trivially have
W= Ix{-oo} u Ix{-} c1. For the secondary symbol space Sohrab looks
at the operator T = observing that the range of p2(m),

in the set 1-1x0 , for either one of the infinite points x0=±.
coincides with the essential spectrum of with a cut-off

function X=l near x0 , = 0 near -x0 .
Formally we get T = L-1, with a self-adjoint realization L

associated to the expression (in R = (--,o) )
32q-1/2+1 3xq-1 ax +1-q-1/2(q-1/2)

(4.15) L = -
Transforming L onto Sturm-Liouville normal form again we get

x _
(4.16) L` _ -a2 + l+q"/(4q2) -5q'2/(16g3)
y , -cc<y<o., Y= q
0
Let us assume that the potential in (4.16) has the limits l+B+ at
Then we have limit point case and the self-adjoint realization
L'"-1/2
is unique. The essential spectrum of is equal to the inter-
val [ 0, (l+B)-1/2] , where B=Min{B+} (cf.[ Wel] , [ Ne2] , [ DS3] ). More-
over, even the essential spectrum for a suitable realization over
one of the half-lines R± of an expression modified near 0 is
[0,(1+B+)-1/2]. Here we refer to L±-1/2 where L+=L"'+c0w/x2 for ,
example, with a cut-off function w=l near 0, wECp(I ), and a suf-

ficiently large constant c0 .
VII.4. Structure of ffi\j; Examples 217
It follows easily that the essential spectra of X,/T and of

L±
(with the back-transform of L±) coincide.
On the other hand $+=0 follows if S+ exists. Indeed, (4.13)
implies q'2/q3 + 0 and lim(q-1)" _ -46+
, If S+#0 this implies
.
1q-1Lyx2 , with y > 0 for every

, large (small) x Or, q=O(x-2 .
at that end. This contradicts q>1.

This argument shows that n2(m) has range [0,1], over each
end ±- of 1. Accordingly 1nl(m)1=(1-n2(m))1/2 also has range [0,1]
This may be used to show that, at each end ±- , the half-circle
(4.17) H1 = {g12+g22=1 , n2>0}
is in in n , so that we get (4.14). Indeed, if n2(m)+inl(m) = z ,
where z is any value of the circle JzJ=1 , Re z >0 , then the ope-
rator Tz= X(D0+iD1)A -z = X(Vq+ax)A -z is not Fredholm, since its

essential spectrum contains zero. Here X denotes any real-valued
cut-off function: XEC'(E, 0<X<1, X=0 in (-o,x0), X=1 in (x1,o).
This means that there exists a sequence ujEH , IIujII=l , with
L
]ITzu.II+O , u.E (ker Tz) . But then the conjugate-complex sequence
uj will satisfy the same, with respect to z , since D0+iD1 =ax+/-q
and A both are real operators, left invariant by complex conjuga-

tion. It follows that TZ is not Fredholm as well, so that
(-nl(m),g2(m)) also must be assumed by the map q , as another
maximal ideal of course. We have proven the following.
Theorem 4.8. Assume that q>l, that (4.13) holds, and that
exist. Then the comparison algebra C (with compact
commutator) has symbol space It given by (4.14) Topologically
.
this symbol space is a circle.

CHAPTER 8. COMPARISON ALGEBRAS WITH NONCOMPACT COMMUTATORS.
In this chapter we consider some examples of comparison

algebras with noncompact commutators. The most important example,
perhaps, has been discussed in detail in [C1],ch.5., in connection
with the half space R++1 = {(y,x):xE1k n, y,-R, y>0} . If the mani-
fold 0 has a regular boundary, onto which the expression H, and
measure du , as well as A# and V# may be extended, meeting similar
conditions as for interior points, then commutators are no longer
compact. However, both quotients, CIE as well as E/K(H) , turn
out to be function algebras. The maximal ideal space 1 of C/E
looks similar to the spaces encountered for complete manifolds
like
n except that it has certain one dimensional 'filaments'
,
over the boundary of 9 .
On the other hand we get
(0.1) E/K(H) __ C(N,K(h)) ,
with the Hilbert space h=L2(R+), and the symbol space ID of a well
investigated Laplace comparison algebra over In If the E-symbol
.
of an operator AEL(H) does not vanish, then A may be inverted mod

E , which corresponds to the first step of the common approach of
solving a boundary problem: That of using a fundamental solution

(or parametrix) of the differential equation to solve the differen
tial equation, though not yet the boundary condition.
This then must be followed in common theory by solving
another system of equations- for example a singular integral
equation over the boundary, in case one uses the well known
boundary layer approach. The latter corresponds to the inversion
mod K(H) of a (system) 1+E , with E E E .
Presently we do not study such manifolds with boundary, but

rather turn to a pair of different examples. First (sec.l) we con-
sider an algebra over I , where some of the generators are perio-
dic functions. (It was seen in V,5 that a commutator [ e1X,DA] is
VIII.O. Introduction 219
not compact) . It is found that again a 2-link chain results. Now

we get
(0.2) E/K(H) E C(N,K(t2)) , i = S1 U S1
with the Hilbert space t2= L2(Z) , and two disjoint copies S1+ of
the circle S1
In direct relation to the above special importance of L2(Z)
this algebra is invariant under translations by 2kir . Again there
is a tensor decomposition of H involved, together with a certain
unitary map of H. Results are related to those discussed in
Gohberg-Krupnik [GK), ch.l1, (cf. also Sarason [Sail , and [CMel1
where we expand on the present discussion).
Next, in sec.2, we look at a product manifold O=1n'>B, where
B is a given compact manifold. The simplest example would be the
circular cylinder XxSl , with the circle S1 We studied a compa- .
XxSI
rison algebra on in example (B) of sec.4, using an algebra
A# of bounded Cm-functions having all derivatives tending to 0,
at ±- including the derivatives in the 'S1-variables'
, That .
algebra had a compact commutator. In sec.2 we admit more general

generators, and obtain noncompact commutators. Again a 2-link
ideal chain is found. Again the second quotient is of a similar
form:
(0.3) E/K(H) __ C(ffi,K(h)) ,
where I is the symbol space of a certain algebra of singular inte-

gral operators over in In case of n=l we get I = R U I ,
.
a disjoint union of 2 copies of I In each case we have h= L2(B).

.
Again a tensor decomposition of H is involved, after a unitary

transformation. The ideal E is identified as a certain algebra of
singular integral operators with compact operator valued symbol.
In section 4 the corresponding is done for a manifold
with finitely many cylindrical ends. To make this discussion
possible we introduce the technique of 'algebra surgery' in sec.3.
This simply means that we cut out the portion of a comparison
algebra over some manifold 01 , corresponding to an open subset
U C 0 , and compare it with the corresponding portion of another
algebra C2 belonging to another manifolds 02 , but with the same
subset U . Of course algebra surgery also may be performed if com-
mutators are compact.
It should be noticed that sec.3 only discusses the basic
VIII.O. Introduction 220
method of surgery, at the example of a 'compact cut', - i.e. the

boundary of the cut-out portion is assumed to be compact (not the
cut-out portion itself). It is not hard to remove this assumption,
at the expense of additional conditions near the infinite parts
of the cut. We find that these additional conditions are too com-
plicated to state, except for very specific geometrical models.
For a discussion of surgery on a manifold with polycylindrical
ends we refer to a[ CDgl] , sec.5.
In all three examples considered here the symbol space of
C (i.e. the maximal ideal space of CIE) is a compactification of
the wave front space. In other words, the secondary symbol space
is void, fon each of these examples.
The theory of sec.4. has an application to a problem on In,
studied first by Nirenberg-Walker[ NW1J, then by Cantor [Ctl],
Lockhart [Lk] and McOwen [ M1] (cf. also [ LM] ). This involves dif-
ferential operators with coefficients constant at infinity and
homogeneous symbol in certain weighted Sobolev spaces, where
derivatives of different orders have different weights. Details
of this application will be discussed elsewhere.
We note that there are numerous other approaches to singu-
lar boundary problems on manifolds with cylindrical or conical
ends (cf. Agmon-Nirenberg [AN1] , Bruening-Seeley [BS1,2] '
Melrose-Mendoza [ MM] , Schulze [ Schu1] ). Also let us point to

a direct extension of our present theory discussed in [CFb] and
[ CMe1] It turns out that the E-symbol can be extended to the
.
entire algebra C One thus obtains Fredholm criteria without

.
involving a chain of two inversions.
1. An algebra invariant under a discrete translation group.
Let us consider the Laplace comparison algebra C on I of

example (G) of V,5 with generators V,(5.27). We have discussed the
commutator ideal E in V,prop.5.5. Now we want to obtain more
details about the ideal chain C D E D K(H) and the quotients CIE
and E/K(H). First we study the symbol of the algebra C .
Proposition 1.1. The compactification M of the function class

#
A
A# described by V,(5.27), as a point set, is given by collapsing
each circle {(x0,y):yESl}, for a fixed x0EI in the compact ,
infinite cylinder [-o,o]xSl , into a single point. The points of

VIII.l. Periodic coefficients 221
the two circular 'caps' 1(-co,y):y ES1} and {(co,y):y 61} correspond
to the maximal ideals {a(x): limk}-.a(2sy+2kir)=0} (and with the
limit k}+oo, respectively, for integers k) of the algebra C
A#
The topology is the weak topology induced by CA#. In particular,
the relative topology induced on the two caps is equivalent to

the Euclidean topology of the circle S1 . Every neighbourhood of
one of the points 6 of the circle caps contains an open neighbour-
hood of the point 0+2km, for all sufficiently large (sufficiently
small) integers k corresponding to the cap at
, respectively.
The proof is left as an exercise.
Theorem 1.2. The symbol space ffi of the algebra C generated by A#

andD# of (5.27) is homeomorphic to a compact subset of
,
M #x[-co,+wl. Using the homeomorphism as an identification we have

A
(1.1) ffi = M
A
That is, ffi is the closure of the wave front space in the above
product. The symbols of the generators are given by
(1.2) QA(x,E)=a(E) , cDA(x,E)=s(E) , ca(x,E)=a(x),
where the functions A , s and a must be continuously extended.

,
Proof. We use the argument of Herman's Lemma again (cf.[C1],IV,2):

The algebra C has the two commutative subalgebras C0 = C # and
A
C# = algebra span of the operators S=s(D) and A=a(D). These alge-
bras have the maximal ideal spaces M #, and (the interval)
A
respectively. Clearly both of them together generate the
algebra C.
Let r0:M-M # and iT #:M be the duals of the injections C0
A
iC and C# + C . Then u:ffi + MA#x[-co,+co] , defined as u=TrOxTr#
defines a continuous map onto a compact subset of the space at

right. This map must be 1-1, hence a homeomorphism, because
7T 0(m) _ 70(m') and n#(m) = 7#(m') , for m,m'E ffi implies that the
homomorphisms hm:C+M and hm,:C-M corresponding to m and m' coin-
cide on C0 and C#, hence on C . Then hm=hm, implies m=m'.
Now N must contain the set $x{-o,+oo} , identified as the
wave front space, by VII,thm.l.5. Indeed, the map n0 is identical

with i of the proof of VI,thm.2.2 over interior points of O.
The cosphere at each x0E I contains exactly two points which must
agree with the two points of 70-1(x0) Since N is closed, it must
.
contain the closure of W , i.e., the space (1.1). On the other

hand, N cannot contain any (x0,E0) in the product set having 0
finite. Indeed, let 0ECO(I) be such that Then f(D)EE
and also E C Using the associate dual map one finds that the
symbol of A=O(D) assumes the value at m=(x0,E0) On .
the other hand aA=O since AEE (by V,prop.5.5), a contradiction.

This shows that N coincides with the set (1.1). Then formula (1.2)
is a consequence of the fact that aa(m) = a(s0(m)), aECO
and af(D)(m) = (x#(m)) , O(D)EC , by virtue of the properties
of the associate dual map (cf.[C1],AII,5). Q.E.D.
Next we examine the quotient E/K(H). In fact it will be pos-
sible to obtain a much more precise control of the ideal E. First,
in that respect, it is useful to look at the C -subalgebra F of E
with generators
(1.3) E j=-Naj(D)eiTx E E , ajECO(l) , N=0,1,25... .
Note that the equation (1+E)u=f , for such an E, is equiva-

lent to the linear 2N-th order finite differences equation (for
the inverse Fourier transforms u' and f')
(1.4) u' (x) + j=-N

a.(-x)u'(x+j) = f' (W ) , x E I
J
Here (1.2) relates only the values of the vectors
(1.5) u'=(u )=(u'(x-j)) ...fA=(fj)=(f'(x-j))j=0,±1,...

i i=O,±l I
for each fixed x , O<x<l . In fact, for each such x we get an

infinite system of equations
(1.6) uj+Lk.. = fj , j=O,±l,... , with ajk=aj-k(j-x)

Najkuk
Note that the infinite matrix A=((ak))J.
J,k=0,tl.... ' where we
set a J. =0, as lj-kl>N , defines a compact operator on the Hilbert
space Z2= L2 (Z) . (Such a matrix, with the property that all
entries outside a finite number of diagonals vanish, is called
a Toeplitz matrix.) The present Toeplitz matrix has the additional
property that the elements in each diagonal have limit zero, as
JjJ-° , since aieCOW . This implies compactness of A . (To prove

this, observe that the matrix obtained from A by replacing all
entries ajk , with JjJ,JkJ<M , by 0 has norm <e , as M is large,
,
by Schurs Lemma [C1],I,l. This means that A differs from an opera-

tor with finite rank by an arbitrarily small amount,in norm. Thus
it must be compact.)
Notice that the vectors u° , f° of (1.3) vary with x , as
x E I , and that we have the 'periodicity condition'
(1.7) u°(x+l) = Y-1u°(x) , f°(x+l) = Y-1f°(x) ,
with the shift operator Y-1:C2i.22 defined by Y_1(uj) _ (uj_1)

Correspondingly the matrix A is a function of x and we have
(1.8) A(x+l) = Y-1A(x)Y1 , x E I .
In order to convert (1.7) and (1.8) into true periodicity one may
introduce a unitary deformation Yt 0<t<1 , of the shift matrix
,
Y1 into the identity operator. Note that Y1 is a unitary operator

2
of . For example one may define
(1.9) Y t ((u(j-k+t))). , tEl ,

t ,k=0,±1,...
with the entire analytic function u(T)=e17rTsin7rT/1rT, T90, V(O)=l.
It follows easily that Ymu° _ (uj+m) , so that indeed Ym shifts
the components of u°=(uj) by m units. Moreover we get Yt+T=YtYT,
as t,TEI (i.e.,Yt is a 1-parameter group). Also Yt is norm contin-
uous and even C0(l,L(t2)) and unitary. All this follows from the
fact that Yt, in essence, may be obtained as the conjugation by
the Fourier transform of Z of the family of multiplications (of
d function in L2([0,1])) by e2 rtit0 The Fourier transform of Z
.
is defined by
(1.10) u° _ (uj) -* u-(8) _ Zj=__uje27ii8 , 8 E [0,1]

Introducing u° (x)=Y-xv° (x), f° (x)=Y-xg° (x), A ° (x)=YxA(x)Y-x,
we find that (1.7) and (1.8) go into
(1.11) v° (x+l)=v° (x) , g° (x+l)=g° (x) , A(x+l) = A° (x)

In other words, v(x) and g°(x) now may be regarded as functions
on S1 , the one-dimensional circle obtained by identifying the
endpoints of the unit interval [0,1) . In fact we get v°, g° E
L2(S1,.22) , and, moreover, using that Yx:t2+22 is unitary,
(1.12) = J 1dtlly°112 .
uuu2=JRlu(x)I2dx S
Thus we have proven:
Proposition 1.3. The map u u° + v° defines a unitary map Z:H+H°,

with H=L2(it), H°=L2(S1,L2) . Moreover, for any EEE of the form
(1.3) the operator E°=ZEZ-1 E L(H°) is a 'multiplication operator'
(1.13) (E°v°)(t) = A°(t)v°(t) , v°(t) E H° ,
with a continuous function A°(t) E C(S1,K(t2)) .
Proposition 1.4. Conjugation with the operator Z-1:H°+H takes

the algebra FCE onto the class C(S1,K(t2)) of all multiplications
of the form (1.13) .
Proof. This is an immediate consequence of the Stone-Weierstrass

theorem for algebras of matrix functions: (i) At each point tES1
the algebra of all 'values' of functions A°(t) obtained from E of
2
the form (1.3) is dense in K(Z ), since by a special choice of the
functions a.(x) it always can be arranged that only the coeffi-
cient a°jk , for a given fixed index pair j,k, is different from
zero, while all other coefficients of the matrix A°(t) , for the
given t, vanish. (Note that conjugation by Yt does not change this
since Yt is unitary.) Also, (ii) F is invariant under multiplica-
tion by b(D) , with bEC"(R) , b periodic with period 1. Such
operation translates into a multiplication of A°(t) with the
function bEC(S1) . This proves the statement.
Proposition 1.5. A multiplication u°(t) + A(t)u(t) by an A° (t)

E C(S1,K(t2)) is in K(H°) if and only if A° (t)EO .
Proof. Let A(t0)#0 . There exists wEl2 with uwp=l and fA°(t)wp
> p>O in It-t0I<E , for some p>O, E>O Then the class of all
.
u°= w E C0'((t0-E,t0+E)) defines an infinite dimensional sub-

space of H° on which we get II A° u° ll p u° II . Hence A° (t) cannot be a
compact operator, q.e.d.
In order to obtain full control on the ideal E it now appears
natural to investigate operators of the form B°= Za(x)Z-1 , where
aECS(1g). These operators will only appear in the form B° E° or E° B° ,
with an E°EZFZ-1 . Getting an explicit form for B° implies that
we can carry out all investigations on E by looking at E°=ZEZ-1
only.
Note that a(x)E E K(H) for all a E CO(s). Also we may get
VIII.1. Periodic coefficients 225
restricted to the special case a(x)=s(x)=x/(x) 8S(l). Then we have

a(D)=(s'*) equal to the convolution operator with kernel s'=F-1s
The properties of s' are well known: Using [C1],II, for example,
one finds that s'ESad . The singularity at 0 has a homogeneous
expansion mod D with leading term c1/x We have s'EC-(1\{0}),
.
and it equals a function in S outside lxl<l . More directly,

s' may be expressed by modified Hankel functions, showing the
same feature, and even that it decays exponentially with all deri-
vatives, as lxl+oo Using a cutoff function XEC'(E), X=l near 0,
.
X=0 for Ixl>1/4 we write s' + w', where 1=Xs' , wv=(1-X)Q'

It follows that w'ES , hence wESCCO . Accordingly s(D)=p^(D)+w(D).
where one may neglect the second term for reason mentionned above.
In fact one may split again 1(x) = i-(x), where p-ÊCO,
using the asymptotic expansion of s' .
This discussion shows that it suffices to assume
(1.14) Tu'(x) = a(D)uv(x) = ip(z) = X(z)/z
with a cut-off function X as described above.
Proposition 1.6. For the operator T of (1.14) and T'=(ZF)p(ZF)-1

we have
(1.15) T'v°(x) = J*(x-y)Yx-yv'(y)dy v =(vj(x)) E

C'(S1,L2),
where Yt:t2+L2 denotes the operator of (1.9), and where we inter-

pret v° and T'v° as periodic functions (1) defined on I .
Proof. Given v '(x) _ (vj(x)) , where the v. are assumed defined

over I and periodic (1) . Let u = Z-lv . We then have u '(x)
(u'(x-j)) = Y-xv (x) . Or, using the periodicity of vk
(1.16) u'(x-j) _ Ik u(j-k-x)vk(x) _ Ik u(-(x-j)-k)vk(x-j)
We replace x-j by x and then apply the operator ' of (1.14), for
(1.17) f 1(x) _ 'u'(x) = Jdylk (x-yWu(-y-k)vk(y).
Finally let g° = (gj) = Zf . We get g° (x) = Yx(f'(x-j)). Or,
(1.18)
Note that near each x,y only one term of the infinite series 11
can be 90, since supp V C {IxI:1/4} , by construction of 1 , hence

interchange of sum and integral in (1.18) is justified. Using once

more the periodicity of vk we shift the integration variable, for
(1.19) gj(x) = Jdy*(x-y) lklu(j-l+x)u(1-k-y)1Pk(y)
Note that I1u(j-l+x)u(1-k-y) is the (j,l)-matrix element of the
product YXY-y = YX-y , using the properties of Yt of (1.9). Thus,
(1.20) gj(x) = Idy lp(x-y)yku(j-k+x-y)vk(y) ,
where the integral extends over I but may be extended from

,
y=x-1/2 to y=x+1/2 only, due to the properties of supp Q.E.D. .
It is convenient to introduce the periodic matrix kernel
(1.21) 'Y °(z) = rV(z)Yz , as JzI<1/2 , periodic (1) elsewhere.
Then (1.15) assumes the form
(1.22) Y'°v° (X) = J 1 `Y° (x-y)v° (y)dy
Moreover, using that the group Yt is not only continuous but even
CW(&) , in norm convergence, and that Y0=1, we conclude that
(1.23) `Y° (z) - Trcotirz = (z) E C(S1,L(k2)).
Proposition 1.7. For every E° E F° we have
(1.24) 'Y° E° - TrH° EA E K (H°)
where
(1.25) H°u(x) = J 1cotTr(x-y)u(y)dy
denotes the Hilbert transform of S1

Indeed, an operator E° is a multiplication by a compact
operator valued function E(x)E C(S1,K(t2)) , by prop.1.4. Thus
the right hand side of (1.24) is an integral operator with norm
continuous kernel taking values in K(.62) Such an operator must .
be in K(H°) , q.e.d.
At this point it will be useful to recall the topological
tensor structure of the Hilbert space H° = L2 (S1,22) . Defining
h = L2(S1) we find that
(1.26) H° = hg.C2
with the topological tensor product of the two Hilbert spaces,

as defined in [C1],V,8 (or [BC2]). On Sl we introduce the C*-alge-
bra SL of singular integral operators, i.e. just the unique
Laplace comparison algebra of the compact space S1 . Clearly SL
has compact commutators and its symbol space ffiSL coincides with
wave front space, since S1 is compact. Note also that the space
ffiSL consists of two disjoint copies of S1 since the unit sphere
,
in I consists of two points only.
Theorem 1.8. The C*_subalgebra E° = ZEZ-1 of L(H°) coincides with

the algebra SL. = AK(t2) . Moreover, the quotient E°/K(H°), and
hence also the quotient E/K(H) is isometrically isomorphic to the
function algebra C(ISL,K(t2)) This isometry is such that an ope-
.
rator E° = (multipliation by) x°(x) = ((ejk(W))) has the symbol

((ejk(x)) on both copies of Sl For an operator E of the form
.
(1.3) the symbol on ffiSL = S+ U S1 is given as
(1.27) IE(x) = E°(x) , x E S1+ , E° = ZEZ-1
Moreover, for an operator A = a(x)E, EE=-F , we get
(1.28) YAW = a(±')E°(x) , xE S1+
We shall offer only a skeleton proof. For details cf.[CMe1].

It may be shown that the generators V,(5.33) transform into
terms of the form
(1.29) X+E°T + X-E° + K , X+ = 1/2(1±iH°) ,
with the Z-1-conjugates E°+ of the two terms E+ E F occurring in

V,(5.33). Here it must be verified by a calculation that the Z_
conjugates of E+(x) equal the terms X+ , modulo a term W+ such
that W+E°+ E K(H) This calculation was prepared by prop's 1.6
.
and 1.7, but its details are left to the reader.

t2
Let e3 be the unit vector of with j-th component =1. From
V,lemma 1.1 we get K(HA) C E° Using prop.l.4 we then get all
.
multiplication operators a(x)M(ej)(e3) contained in E°, where jEZ

and aEC-(S1) are arbitrary. Then also H°a(x)M(ei)(ej) (=- E° , by
prop.l.7 . Now it follows that all operators AN(e3)(e3), for A E
SL are contained in E° . Repeating the same conclusion with ej
replaced by e3+e1 we find that also AM(e3)(el) , for AESL, j,lEZ,

is in E° Thus, letting KN denote the set of all infinite matri-
.
ces CEK(k2) with components zero outside IjI,Ikl:N , we find that

AOKEE° for all AESL , KEKN . In other words, SLIKKN C E° . In [C1],

V,8 (or [BC2]) we have shown that SLR = SLIRK(Q2) is the
closure of SMUKN) . Since E° is closed we get EADSLO, and equa-
lity follows since the opposite inclusion is trivial.
Finally, the statements about the quotients are immediate
consequences of an investigation for abstract algebras with sym-
bol in [BC2], sec.6. Q.E.D.
2. A C -algebra on a poly-cylinder.
Next we look at a Laplace comparison algebra with noncom-

r
pact commutator on a poly-cylinder Q=Rn xB. Here B denotes a com-

pact Riemannian space of dimension n" with metric dp2=gjkdxjdxk.
Accordingly, for the metric and the Laplace operator of 1 we get
(2.1) ds2= dt2+ dp2 , A = At + Ax ,

Ax=(Vg)-la
jgjka k
xi x
where Ax is the Laplace operator on B In (2.1) we are using the
'Euclidean
.
metric dt2=dt1 +...+dtn of Bn ={t=(tl,...,tn ):t.EX},
and the Euclidean Laplace operator At=j8 j2 . The summation con-

t
vention often will be used from 1 to n", as will be clear from
the context. We set n=n'+n", so that Q is n-dimensional.
Let H be the Hilbert space L2(0) = L2(c2,dS) , with the sur-
face measure of the metric (2.1). Let C C L(H)

be the smallest C*-subalgebra containing the (5 types of)operators
A=(l-A)-1/2,
(2.2)aEAB, sj(t)=t/(t), a jA, DxA, DXED#, j=l,..,n'.
t
Here (t)=(l+t2)1/2; we write AB=COO(S) while DB denotes the col-
lection of all Cam-vector fields on B . Also A is the unique posi-
tive inverse square root of the (unique) self-adjoint realization
(1-A) of the Laplace differential expression A There is a .
unique such self-adjoint realization because 0 is a complete Rie-

mannian space (cf.IV, cor.1.7). The functions aEA# and sj(t) in
in (2.2) represent the corresponding multiplication operators on
functions u(t,x) E H. Correspondingly, we denote by A# and D# the
algebra finitely generated by C_ and the first two kinds of gene-
rators, and the left module of A # spanned by {1, 8 ,
D x},
tj
VIII.2. Polycylinder algebra 229
respectively. The operations a Au and DXAu are well defined for

t
uE A-1C0(S)), a dense subspace of H (cf.V,l) and have continuous
extensions to H (cf. also (2.6), below).
We notice that H = h Rh is the topological tensor product
x
of the Hilbert spaces ht= L2(in ), and h = L2(B) (cf.[C1],V,8).
x
Let us write Lx = L(hx), kx = K(hx), Lt = L(ht), kt = K(ht).
It^may be observed that the topological tensor products Lt'Lx=Ltx'
, k 9L = Kt are all well-defined C -sub-
K(H) ,
tN x
Lt9
xx
algebras of L(H) , where K(H) C Kx , Kt C Lxt C L(H) all are pro-
per inclusions. In fact, Kx and Kt are proper closed two-sided
ideals of Ltx , and, of course, K(H) is a proper closed ideal of
all the others. Evidently, Ltx (but not Kx or Kt) contains the
identity operator I .
Note that we may write H = L2(In ,hx) as the space of all

functions over In ' with values in h = L2(B) such that
x
(2.3) 11u02 n,dtUu(t,.)N2 <

)
R
by Fubini's theorem. Correspondingly, if CB(In',kx) and CB(Rnl,Lx)
Ins
denote the classes of bounded norm continuous functions over
taking values in kx and Lx, respectively, then a function P E
CB( ln ,Lx) has a natural interpretation as an operator in L(H),
defined by u(t,.)-; Vt)u(t) . Moreover, this operator is in Kx
whenever 4D E CO(I,kx) (cf.[C1],V,B). (We indicate a proof of this
fact in cor.2.4.) This establishes an isometric *-isomorphism of
CB(l;Lx) (as a Banach algebra with norm
(2.4) sup{NA(T)N : T E 11 ,
is the norm in Lx) into L(H). In the following we

where IIA(T)II
will use this interpretation of functions as operators in L(H)
In order to find the commutator ideal E of the unital C-
algebra C with generators (2.2) we conjugate the generators with
the Fourier transform in the t-direction. In detail we have
n,e-itTu(t)dt T E Ins
(2.5) Ftu(T) _ (2n)-n'121 , ,
which defines a unitary operator of ht . By conjugation with Ft

we, of course, mean conjugation with Ft0Ix, where Ix denotes the
identity operator. (However,we will write this as Ft-AFt
l ,for
A E L(H). ) First consider the Ft-conjugations of A, D .A, D A :
t
X
Ax)-1/2
(2.6) A-(T) _ ((T)2- , TjA-(T) , DxA-(T) ,
with Dt.=-iat. and above interpretation of a function as operator.
Since B is compact we have A"(t) E K(hx), for all T (cf.III,

cor.3.9). Moreover, we even get A"'(T) E CO(lnt,kx), where CO CCB
denotes the class of functions with limit 0 at infinity. In fact,
A -(T) even is analytic, in norm topology of L(hx) , and we have
IIA- (t)II < (T)as well as IIA-(0)-lA-(T)II<l , so that T.A-(T)
J
DxA-(T)=(DxAx)(A-1 A-(T)) are functions in CBU nt ,Lx)CL(H) , con-
firming the boundedness of the last two types (2.2). Also, writing
T(T)=(l+T2Ax)-1/2
Ax=A-(0), the generators (2.2) correspond to

,
(2.7) a(x), p.* , AxT(T) , (iTjAx)T(T) , (DxAx)T(T)
with Vj=Ft1s. and ut*= convolution in ht

ITI1-CA
Proposition 2.1. The operator functions AXT(T) and xT(T)
for each fixed s , O<E<l , are in COM,kx) .
Proof. Just note that /C=AxT(T)ECO(]k,kx), while T(T) and ITIAxT(T)
belong to CB . All operators are positive self-adjoint, so that
AXT(T) = AET(T)1-E E CO, (TI1-EAXT(T)
= AE(ITIAxT(T))1-6(CO , as
products of a bounded function and a function with limit 0, q.e.d.

Now we first will describe the algebra generated by the
commutators of the Ft-conjugated generators (2.7) .
Proposition 2.2. All commutators of the Ft-conjugated generators

and their adjoints are contained in the algebra
(2.8) G^ = CO(]Rnt,K(hx)) + K(H) C K x
Here a function C(T) E CO(&nr,K(hx)) must be interpreted as an

operator in Kx in the manner described above.
Proof. We will use the resolvent integral technique used in V,3.
Let the operators ( 2 . 7 ) be denoted by G1,...,G5, in the order lis-
ted (we write t=t3, at=atj, T=Tj, etc). Clearly [ Gi,G2] _[ G3,G4] =0.
We get (DxA-)*-DxA-=[A ,D*]ECO(k n',kx), by (2.14) below, and
adjoint invariance of DB). All other generator(classe)s are self-

adjoint. Hence the adjoint generators need no special attention.
For the commutators [G.,Gl],j#2, we use the well known res-
olvent integral representation of A-=G3: Let R(s)=(s+(T)2-Ax)-1
Then we have
G3=A-(T)=AT(T)=((T)2-A
(2.9)
0
(cf. VI,(1.6)), with a norm convergent improper Riemann

integral, in the algebra L(hx) . We get
(2.10) [G3,G5]=[A-,Dx]A-, [G4,G5]=[A-,Dx](TA°), where A-,TA-ECB,
so that it suffices to show that [A-,D x] E CO(R,kx)

Similarly,
[Gi,G4l=(TA»)(A"'-1[ax,A"']),
[Gi,G3]=[ax,A`],
(2.11)
px=[ax,Dx]EC'(B).
[Gi,G5]=pxA-+(DxA -)(A--1[ax,A-]), with
Accordingly we also must show that A"-i[ax,A-] E CO(R,ftx)

Both these facts are consequences of prop.l.3, below. Before we
discuss it we turn to the commutators [G2,GI]. There we find it
practical to work without the Fourier transform (2.5), writing
(2.12) A = 1/Tr S(r) _ (r+1-At-Ax1

0
Instead of 'diagonalizing the t-variable by using the Fourier

transform' we will consider the x-variable diagonalized later on.
Let Aj =FtGi Ft-1 be the generators (2.2). Then we have
(2.13) [A2,A4]=ptA+(8tA)V, [A2,A51=(DxA)V, V=A-1[A 2,A31, PtEC
Note that ptECO(RnT) hence p+AEK(H) by a well known result

,
(cf. III,thm.3.7). We claim that all 3 commutators [A2,A11,

1=3,4,5, are in K(H). Again this is a trivial consequence of prop.
1.3 below, so that all of prop.l.2 has been reduced to prop.1.3.
Proposition 2.3. For c with 0<c<1 we have
(2.14) A"-1-E[ax,A-] E COOL n',hx) , A--E[Dx,A-] E CO(Rn',hx)
(2.15) A-1-E[sj(t),A] E K(H) , si(t)=tj/(t) .
Remark. As a consequence of prop.2.3 the algebra C satisfies con-

dition (m4) of VII,3 as required later on (cf.thm.2.8).
For the proof of prop.2.3 we use (2.9) for
(2.16) A --1 E[ Gi,G3] =1/Tr Li=[ a(x),Ax] .

0
The integrand in (2.16) is a norm-continuous (even analytic) func-

tion F(s,T), with values in kx Indeed, one may write .
(A--1-ER(s))(LlAx)(A-1R(s)//-s)
(2.17) F(s,T) _
and use the estimates
0<n<2
IIA_"R(o)I<(1+s+T2)p/2-1, IIA`-nR(s)1<(1+s)n/2-1,
(2.18)
x
easily derived from the spectral decomposition of the self-adjoint
operator -Ax>0 of hx. Note that L1 is a folpde on B independent of
s,T, so that the second factor in (2.17) is a constant in L(hx).
Analyticity of the first and third factor is a consequence of ana-
0((l+s)(E-1)/2)
lyticity of the resolvent R(s). These factors are
0(s-1/2(1+s)-1/2),
and respectively. Thus F(s,T)=0((1+s)-1+E/2/yrs-)
uniformly for all TEIR . Also both factors are in kx since B is
compact, insuring the compactness of the resolvent R(s) of the
Laplace operator Ax (III,thm.3.1). This implies existence of
the improper Riemann integral (2.16) in norm convergence of L(hx)
and uniformly so , for TEX Thus the integral is in
.
For more detail in such a proof cf. V,3).Moreover, since 0<E<1

is arbitrary, one may use this for E+6<1 , with some 6>0 This .
gives a factor A-6ECO, whence the first (2.14) is CO, not only CB.
Similarly we may use (2.9) for
(2.19) A--'[ Dx A-] =1/7J(A`

0
where the integral exists for analogous reason. (Now the second
factor contains the second order operator [Dx,Ax] over B so that
again the factor is a constant in L(hx).)
For (2.15) we use the other resolvent integral (2.12), for
P
(2.20) A-1-E[ s(t),A] =1/rr M1=[ s(t),At]
0
Now, with respect to an orthonormal base of the Hilbert
space hx consisting of eigen-functions of the self-adjoint opera-
tor Ox, the operator S(r) is diagonalized, with respect to hx in
the tensor product decomposition of H. Its diagonal components are
(2.21) S.(r) _ (r+(A.)2-At)-1
with the eigenvalues X of the positive self-adjoint operator Ax

At=(1-pt)-1/2
on B. In analogy to (2.18) we conclude that, with
(2.22)
IIAtnS.(r)II<(l+r+A])fl/2-1,
BAt-nS.(r)II<(l+r)'1/2-1
, Aj=(1+X2+Ot)-1/2
for j=1,2,..., and At=(l-At)'1/2

Notice that the entire relation (2.20) is 'x-diagonalized',
i.e., decomposes into a set of countably many relations, invol-
ving At and S. instead of A , j=1,2..... . Again one may write
the integrand Fi(r) as a product of three factors:
(At-l-esj(r))(M1At)(A-1Sj(r)/V-r)
(2.23) F.(r) _
where the second term is a constant in L(H) . For the first and
0((l+r)(e-1)/2)
third term we get estimates and
0((aj)-d(1+r)(b-1)/2r-1/2),
for any 0<6<l. Thus
(2.24)
0((a.)-ar-1/2(l+r)(e+S)/2-1)
IIF.(s)II =
3 3
The right hand side is integrable, as long as e+6<1 . The integral
fFj(r)dr = Cj is an operator in K(ht), and we get
(2.25) HCj II = 0(( aj) - d) .
Accordingly the operator (2.20) corresponds to a diagonal matrix

((Cj6kl))k,1=1,2,,,, . with diagonal components converging to 0
in norm. This indeed implies that (2.20) belongs to K(H). (It is
limit in L(H) of the sequence of diagonal matrices Tk obtained
by setting all Cj , j>k, equal to zero, while the matrices Tk are
in K(ht)HF(hx) C K(ht)ZK(hx) C K(H) , with the class F(hx) of
bounded operators of finite rank over hx .) This completes the
proof of prop.l.3.
Corollary 2.4. The C*-algebra G^ and its Ftl-conjugate G are

subalgebras of Kx = Ltt kx
Proof. It is sufficient to show that C Kx . For any
orthonormal basis TOi:j=1,2.... }, and the orthogonal projection
Pn onto the span of TO,.... ,O } we get uniform convergence
PnC(T)Pn+C(T), TEl ,for every C(T) E CO(ln ,K(hx)). But the opera-
tor PnC(T)Pn is a finite sum of operators cjl(T)
E CO(s) . Thus PnC(T)Pn E kx and the limit C(T) as well, q.e.d.
Proposition.2.5. The commutator ideal GO of the C -subalgebra of

L(H) generated by the operators of the form G1,G3,G5 contains the
algebra CO(1nt,hx).
Proof. For a moment consider only the C -algebra B generated by
G1,G3, and G5 By virtue of the isometric isomorphism mentionned
.
initially in this section, the generators of B belong to the

function algebra CO(Int,L ). Hence the algebra B may be interprete
x
as a subalgebra of CO(I' ,Lx). For a fixed T=T0E gnu the values
BT ={A(T0):A(T)EB} form a *-subalgebra of Lx . It is clear that
0
BT is a *-subalgebra of the Cat-subalgebra CT of Lx generated by

0 0
the values of the functions Gj, j=1,3,5 at TO , i.e., by the
operators
(2.26) ax ,
A-(T0)=(-Ax+(T0)2)-1"2
,
DxA-(T0) ,
where ax and Dx run through all the functions (folpdes) over B

Moreover, C evidently is the closure of B . Also C is just
TO TO TO
the minimal comparison algebra in the sense of V,1 generated

by the triple {B,-Ax+(T0)2,dS} , on the compact manifold B By .
V,lemma 1.1, it follows that CT and even its commutator

0
ideal contain all of hx . But commutators in CT are compact,
since B is compact. Therefore the commutator ideal of CT equals

0
h
x
. Since that commutator ideal is the closure of the commutator
0
ideal ET of the finitely generated algebra, we must have ET
TO TO
dense in hx . On the other hand E0 clearly is contained in the

0
commutator ideal of the algebra BT , and even in the localization

0
GOT at TO of the commutator ideal GO . Thus we conclude that the

0
algebra GOT of 'values' of GO at T is dense in k.x, for all TEXn'.
We also find that the algebra B contains A_(T) , hence also

contains every f(T,Ax), for a general fECO(]Rntx(0,11), by the
spectral theorem and the Stone-Weierstrass theorem. Hence B con-
tains all functions *(T)Ex, with lECO(Xn') and the projection ope-
rators EX of the spectral family of Ax-1 . Note that EX are of
finite rank, and that EN3l , strongly, as N+-. Since GO is an
ideal of B , it follows that GO contains
(2.27) GON = {1P(T)ENA(T)EN : A(T)E GO, ip(T)ECO(Xn)} .

But GON is a self-adjoint algebra of (finite) jNxjNmatrix-

valued functions, separating points in the following sense. For
every e>0,and jNxjN matrix P there exists K(T)EGON such
that K(T2)=0 and IK(T1)-Pl<e (This follows from the above). By the
matrix version of the Stone Weierstrass theorem this implies
JN
that GON=CO(In' L(T )). Since this holds for all N we find that
GO contains all these matrix algebras. But for a general
iN
A(T) E CO(In' kx) we get AN(T)=ENA(T)EN E CO(En' L(C )). Also
AN(T)-A(T) ; 0, in C0(lntkx) (with the norm (2.4)), since A(T) is
compact and EN converges strongly to 1 Since GO is closed we
.
conclude that A(T)EGO , so that indeed CO(l,kx) C GO , q.e.d.

Returning to our task of describing the commutator ideal E
of the cylinder algebra C we conclude from prop.2.5 and V,lemma
1.1 that E contains the C*-algebras GO' = FtGOFt1 and K(H).
It is convenient again to work with the Ft-conjugated ideal E^
Ft1EFt , containing the sum GO+K(H)=SO (which in turn contains
all the commutators of the generators G., j=1,...,5)
Notice that SO is invariant under left and right multipli-
cation with the (functions in CB(I,kx)) G1,G3,G4,G5, but not
under multiplication with G2 Accordingly E- must be properly
.
larger than SO.

Specifically G2=sj(Dt) is a singular convolution operator
with Cauchy-type singular integral and kernel p.=sj', so that
a product K(T)G2 for K(T)ECO(ln ,kx), appears as an infinite
,
matrix of singular integral operators on Ins, if we introduce

some orthonormal base of hx.
Theorem 2.6. Let SQ be the C -algebra of singular integral opera-

tors over ht generated by the multiplications in CO(In ), and the
operabrs a(M)s .(D), aECO(e'), j=l,...,n'. Then the ideal E^ coin-
cides with the J topological tensor product SQ. = SQNkx .
Proof. Notice that SQ coincides with the minimal comparison alge-
bra of the triple {in',dt,l-ot} (cf. ch.Vl). Thus it contains the
compact ideal kt of L(ht), and SQ. contains ktNkx=K(H). Therefore
it is trivial from the above that SQ. contains all the commutators
EGj,G11 j,1=1,...,5 and that it is a closed *-ideal of C
, .
Hence we have E'CSQ.. To show equality we introduce a fixed ortho-

normal base l'2' . of the space hx and first consider K(T)E GO
such that K(T) takes span into itself and its ortho-
gonal complement to 0, for all T. Thus the infinite matrix vani-
shes outside its first N rows and columns. Let CON denote the
subalgebra of CO(I ,kx) of all such finite matrices, for a given
fixed N. Clearly CON is isometrically isomorphic to CO(ln ,L(CN)).
Now we observe that K(T) and L(T)s(Dt) , with K,LECON ,
belong to E^ , and generate the algebra SQN = SQNL(&N) for each ,
N=1,2.... . Also we again find that S9, is the norm closure of

USQN Therefore indeed q.e.d.
We now come to our main task: The description of the symbol
chain of the cylinder algebra C First let us look at the ideal
.
quotient E/K(H) In that respect we observe that the algebra

.
SQ is a subalgebra of the algebra SI of singular integral opera-

tors on In' generated by Sj=sj(Dt)=(pj*), j=1,.... n', and the mul-
tiplications with functions in C(Bn ) , with the 'ball compactifi-
Ins Ins
cation' of , having one infinite point in each direction
ItOl=l (cf.[C1],IV,l,problems). The special comparison alge-
bra SI C L(ht) has symbol space
(2.28) MI = {(t,T)EC(1n'xEn') :
ltI+ITI=0}
(cf. ex.(A) of V,4, or [CHe1]). The subalgebra SQ of SI con-

sists precisely of all operators in SI with symbol vanishing at
ITI , as follows from the Stone-Weierstrass theorem, looking at
the symhols of the generators. (Note that the generators are
written as multiplications by functions of the variable T and
convolutions by functions of T as well, since we consider the
Ft-conjugated ideal E^ . Accordingly we have t and T reversed,
compared to the normal notation for space and momentum coordinates.)
It follows that SQ/kt is isometrically isomorphic to the function
algebra CO(f) with the locally compact space
compact space
(2.29) 1 = NQ = {(t,T)E NI :
Itl=o' ,
ITI-} = 2Bn'xgn' .
In case of n'=l this space is a disjoint union of the two

sets {oo}x1K=1- and {_co}x$=1+ Both E± are copies of I , with the
variable T running over 1. In the general case n'>l the space I is
connected, and is a product of the infinite sphere 3Bn'=%n'\ln'
with in
Ins
Clearly this also is just the wave front space , but
with t and T interchanged. We have proven the following result:
Theorem 2.7. The quotient algebra E/K(H) is isometrically isomor-

phic to the function algebra CO(L,kx),sothat E is a C*-algebra
with (compact operator valued) symbol, with symbol space F In .
the special case n'=1 we have
(2.30) CO(E,kx) = CO(]E-,k x) (D COW,kx) .
The symbols of the generators of E are given as compact ope-

rator valued functions of (t,T), for tE3En' (i.e., t0Mnl,
ltol=l) and TEIn , as follows:
Let A1, j=1,...,5, be the generators (2.2) of C, in the order
listed (so that Gj are their Fourier transforms). Then [A1,A2]=0=
IA3,A4]. The symbols Y(A3'As]' YIA4,A5] are independent of t, as
1E311 ,the value given by the terms of (2.10), respectively, where
A-(T)=AXT(T) , while the commutator IA"',Dx] is obtained from the
resolvent integral from (2.19). Similarly 1[A1,A7], j=3,4,5, are
independent of t, as ltl=°°, and, for TEIn1, their values are

given by (2.11) and the resolvent integral (2.16).
Also, for the symbol of a product Aj(Ak,Al] or [Ak,A1]AI
is obtained by multiplying with the corresponding func-
Y[Ak Al]
tion G. of (2.6) , from the left or right, respectively. Also,
for j=2, the symbols of these products equal the product of
](T) with the value of the function s(t) (extended conti-
Y[Ak' A
11
nuously to 81n1) at t. More generally, for every function bEC(Zn1

the operator b(t)IAk,Al] E E has the symbol
(2.31) Y = b(t)y(A ,A 1](T) , k,192 .
k
Next we turn to the discussion of the quotient C/E , i.e.,
of the symbol and symbol space of C
*
Theorem 2.8. The C -algebra CIE is isometrically isomorphic to the
algebra C(ffi) of continuous complex-valued functions over a compact
space ffi, called the symbol space of C. Here M is (homeomorphic to)
the bundle of cospheres with infinite radius of the compactified
poly-cylinder InfxB considered as a compact C--manifold with boun-
dary (i.e., the product Bn1xB of the unit ball Bn1={tEgn1:ltl=l}
In1
in with the compact manifold B ).

Let A1,...,A5 be the generators (2.2) again. Then the C-sym-
bols aA =0A i.e., the functions in C(ffi) , associated to

J J
Aj by the above isomorphism, are given as explicit functions of

(t,x,T,E) as follows:
a
Al = a(x) , a
A2
= t/( t) ' a
A3
= 0
(2.32)
aA4 iT/(T2+I X2)1/2 , aA5 = (b3Ei)/(T2+IEI2)1/2 f
with
(2.33) ICI = Dx = b3(x)3 j+p(x)

x
in local coordinates of B , where tE,n , xEB , while (T,E)ESt'x,
the co-sphere at (t,x) with infinite radius. (Actually the last
two symbols are the limits of the full symbol quotients
(2.34) T/(1+T2+IM I2)1/2 ,
as (T,E) is replaced by and p-).

Proof. We may just apply the general results of ch.VII, first
verifying their assumptions. Let us shortly summarize these facts.
First of all every symbol space of a comparison algebra contains
the wave font space W , normally identified with the bundle of
unit spheres in the cotangent space T*c of the underlying manifold
St (cf.VI,thm.1.5). The space W is an open subset of 1 . It
precisely coincides with the set of points (x,t,T,g), with ItI- .
Essentially this follows whenever C can be shown to contain all
Cs-functions and all operators DA for a general first order ,
differential expression with C--cofficients and compact support.

In order to study the points at ItI=0 we require the compac-
tification P*Q of T*0 induced by the formal symbol quotients
(2.34) together with the functions a(x) aEAB and t/(t) , and , ,
(1+T2+II2)-1/2
(in other words, by the formal symbols of the
5 types of generators (2.2)). (That is, P 5 is defined as the
smallest compactification of T*(1 onto which all above functions
can be continuously extended.) It is readily verified that P*St
*
is given as the compactification of T S2 obtained by adding the
*
infinite sphere {oo(T,E) : ITI2+IEI2=1} to each fiber T x of the
Ste=PnxB
cotangent bundle T*cc , of the compactification of S2

*
In other words, P a is the disjoint union of all balls {(t,x)}x1n
of infinite radius, as (t,x)ESt c
Moreover, S2c coincides with the compactification M of 52

A#
defined by the functions a(x), t/(t)EC"(c), introduced in ch.VI,
and the algebra C satisfies conditions (ml'), and (mj), j=2,3,4,
5,7. (We noted before that (m4) is implied by prop.2.3. Cdn.(m5)
is trivial: We have AEC. Cdn.(ml'), first used by McOwen, requires
that the functions
lp(x)I2 and gjk(x)Fj(x)bk(x)

(2.35) ,
are in A#, for every Dx=bJ(x)8 j+p(x)ED#. This again is trivially

x
true, since A# contains all of C'(B) . Furthermore the conditions
(mj) involve some separation conditions which can be satisfied by
enlarging A# and D# in such a way that the generated algebra C
remains the same. Details are left to the reader.
As a consequence we may apply VII, thm.3.6. The conclu-
sion is that the symbol space N of C is a compact subset of the
boundary 31 0 = ff SAT *S2 of our compactification & X52 , containing
the wave front space W , i.e., the bundle of cospheres of infinite
radius over Q.
Since N is compact, it must contain all points of the
infinite cosphere bundle over3nc as well. Thus it remains to be
shown that no other point of 31 S2 is contained in ffi In parti-
.
cular none of the points Itl=co, T2+IEI2<_ can be contained in N

This, on the other hand, is a consequence of VII,thm.4.2.
To indicate at least the idea, we find that AEE , for our
(1+T2+I02)-1/2,
present algebra, while the formal symbol of A is
i.e., is tO for the latter type of points. Hence such points
can not be in IN , since the symbol of AEE must vanish, while VII,
thm.3.6 implies that symbol and formal symbol coincide for the
t
points of 3P 12 which are in N
3. Algebra surgery.
Let us assume cdn.(w) (at least) , in this section.

We already started discussing the relation between the mini-
mal comparison algebras of different triples, although so far we
only allowed a change of H and dp over a given compact set K C 2
In this section we will look at the comparison algebras of
VIII.3. Algebra surgery 240
triples over two different manifolds 01 and 2 . Suppose two

2
triples {c21,dp1,H1} , and {12,dp2,H2} are given, and assume that
an open subdomain U1 of Q1 is diffeomorphic to a subdomain U C 02,
2
in such a way that the two triples correspond to each other over
. In order to have simple notations we will regard the two
U1-.U2
sets Uj identified by the diffeomorphism.
This will bring a configuration
as sketched in fig.3.1. We are given
a manifold U and two different exten-
,
sions 01 and 02 of U. The two triples

coincide over U . That is, we have
du1IU = dp2IU H1JU = H2IU,
, or
Also we will assume that the boundary 1
3U of U is compact, although U itself

U (noncompact)
may be noncompact.
The last assumption may be remo-
ved, at the expense of additional 02
conditions at the infinite ends of 3U. Fig.3.1

For a discussion of manifolds with polycylindrical ends involving
such noncompact cuts cf.[CDg],sec.5.
We shall discover a 'common part' of the two comparison
algebras C=C(0,D!), induced by classes A , D! over fj , j=1525
whenever the classes of restrictions of functions (folpde's) to U
coincide, i.e.,
(3.1) AU = A#JU = A2IU , DU = V U = D2JU
We introduce the Hilbert spaces H1=L2(c1,du1), H2=L2(22,d112),

and the space H=L2(U,dpj) . Any CD(U) function will also be con-
sidered a C0-(Q.)-function (with support in U ), without change in
notation. Similarly the functions u E H1 with support in U will
also be considered functions in H or in H2, and vice versa. Also,
for A E L(H1) and Q E C0-(U) we may regard A as an operator in
L(H2,H1) and similar.
Clearly we will get corresponding maximal ideal spaces M
Al
and MA2# , and corresponding associate dual maps il:Il- MA # and
12:M2; M # . Note that the restriction to U of an a E A# induces

A2
an algebra homomorphism C(MAl#) i C(MAU#) which identifies the
closure of U U 3U in MA # with the space MAU# , by virtue of its

i
(injective) associated dual map. Similarly for M # , so that

A2
MA # is found as a closed subspace of both spaces M # and M #
U Al A2
i.e., as the closure of U U 3U
Note that the common restriction
(3.2) {U,dp,H} , dp = dujJU , H = HjjU , j=1,2 ,
represents a comparison triple of its own. As we know, this tri-

ple cannot satisfy (w) , unless 3U is void. However, as long as we
use sets A# , D# of functions and folpde's vanishing near 3U ,
we may use the techniques of VI,l and V,2 to achieve (w) or more
by modifying the potential near 3S2 .
Now, for the remainder of sec.3, we assume the following

separation condition.
Condition (m3)U : For every aEA# and DED# with a=O near 3U and D=O
near 3U we have XUa E A# , XUD E D# , with the characteristic
function XU of UCc.
Notice that (m3) U is always true if U U 3U is compact, and
that (m3)U follows trivially from (m3)
Accordingly we now define classes A# and D# on U by setting
(3.3) A#,O= {aEAU : a=0 near 3U} ,

DU,O= {DE=-DU : D=O near 3U}
and then consider the comparison algebra C = C(AU 0'VU 0) C L(H)

We notice that AUTO in general does not contain the constant
functions. Therefore this is one of the exceptions to the general
rules of V,l In general C will not have a unit (except only in
.
case where 3U=0 ). However, as a consequence of (m3)U it follows

that AU 0 satisfies (al') of V,l. Also, we conclude that, for
a,E
every AU D E DU the trivial (i.e.=0) extension to S.
0' 0
is contained in Aj , , respectively.
or Dj
We claim that C is the 'common part' of C1 and C2 , in the
following sense.
Theorem 3.1. Let V. C C. denote the closed two-sided ideal of all

A E C. with symbol]aA having support in the set tj-1 (M A #). Then
U
we have
(3.4) V. = C + E. , C = XUVjX3
with the commutator ideal E. of C. , and the characteristic func-

tion XU of the subdomain U C Q.. Here in the first relation (3.4)
an operator AEC must be interpreted as AXUEL(Hwhile in the
second relation (3.4) we mean the restriction to its invariant
subspace H of XUBXU for B E V.
,
Proof. First we notice that a statement like VI,prop.2.3 is valid,

with exactly the same proof:
Proposition 3.2. The ideal V. C C. is generated by E. and all

a, DA. with a E A,,0 , D E DU,O (extended zero outside of U ).
After this prop. we must use prop.3.3, below. From (3.6) it
follows that COXU C V. , for the finitely generated algebra CO
corresponding to C . Taking closure we even get CXUCV., using
that
(3.5)
1AXUfL(H.) = BAIIL(H)
, for all AEL(H)
Accordingly, C + Ej C Vj Vice versa, for AEVj let AkEC? be an

.
approximating sequence, RA - AklIL(H )}0 , where Ak is a finite
sum of finite products of the generators of prop.3.2. Using

prop.3.3 again we substitute each DA. in the products of Ak by
J .
DA', and obtain an operator AkEC=CXU with Ak=A0+ Kk , Kk E Ej
Therefore IIA - AkXU KkII ; 0 , k-. Applying XU from both sides
we conclude that Ak + XUKkXU 3 XUAXU in L(H) where AO = XUAXU

,
and Kk = XUKXU now stand for the restrictions of these operators

to their invariant subspace H . We claim that Kk E E . Indeed,
let yECp(sL ), Y=l near 3U. From VII,thm.l.5 we know that YKk, Kky
E K(H.) Thus, with 6=1-y, Kk=SKk6+Ck, CkEK(Hj)
. Now KkEE fol-
.
lows from thm.3.4, below, and from the fact that XUK(H)XUCK(H),
since XU acts as a projection, hence a contractive map Hj+ H In .
the limit k- we get AO = XUAXU E C C Cj , and the second (3.4)

follows.
Finally consider A-AO=KECj. For w E A with supp w E 2j\BU
we get wKw = wAw - (wXU)A(wXU) . Using that AEVj has symbol with
support in UUDU we conclude that awKw (W01 )2aK = 0. Since this
holds for all such w we find that aK(m)=0 for all mE t(M # 3U).
A
However, the set t-1(3U) C Wj is nowhere dense in ffi , since this
is just the co-sphere bundle over Q . Hence the continuous

function aK:ffi-C, zero in i l(M #), must vanish identically. This
A
proves that A-A0 E Ej , and completes the proof of thm.3.1.
The existence of the potential q' , in prop.3.3, below, is
a consequence of IV,thm.5.1.
Proposition 3.3. Let H'=H+q' satisfy (w), where q'>0,q'=0 outside

a given neighbourhood N' of 3U , and let A'=H'-1/2 Then we have .
(3.6) DAj - DA'XU E K(H C Ej, j = 1,2,
whenever supp D C USN' .
The proof will again use resolvent integrals. Let R(A)

=(Hj+a)-1, R'() = (H'+A)-1. For u E CD(Qj) write
(3.7) Iu = DAju - DA'XUu = D(Aj-A')Xu + Cu , C E K(Hj)
where V,lemma 3.3 was used, as well as the fact that K(H)XUCK(H.)
Also we wrote D=DX with X E C-(U), 0<X ql X=1 in supp D, =0 in
N D N,clos Then
.
f-X-112dADX(R(A)-R'(A))Xu
(3.8) (I-C)u = W-1
0
where we write
(3.9) X(R-R')X = [X,R]X - (XR'-RX)X .
For v E im(H0+a) we get
(3.10) XR'v - RXv = R(H+A)XR'v-RX(H'+X)R'v = R[H,X]Rv .
In view of (w) for H' we may extend continuously to H . Similarly

for [X,H] , using (w) for Hi . Accordingly we get
(3.11) X(R-R')X = R[H,X](R-R')X ,
which may be used in (3.8) for
I = C + i-1 f'A-112
(3.12) dXDR(a)[H,X](R(X)-Rt(A))X
0
The integral in (3.12) has a compact integrand, since the folpde

[H,X] has compact support. Also we use an estimate like V,(3.9)
again, to show that we have a norm convergent improper Riemann
integral. This shows that I E K(H) q.e.d. ,
Next let us compare the commutator ideals.

Theorem 3.4.The commutator ideal E of C is given as algebraic sum
(3.13) E = XEjX + K(H) ,
with any cut-off function XEC0(U) , 0<X<1 , X=l outside some com-
pact subset of U U 3U . Here again we consider X extended zero to
Q.\U , and, for EEE, interpret XEX as an operator in L(H)
Proof. Note that the following operators form a set of generators
(mod K(H)) for Ej (with the finitely generated algebras C?)
(3.14) A[B,C] : A , B , C E CY
Similarly for E , where now A , B , C E CO , and where the A' of

prop.3.3 may be used for CO, instead of A Now such a generator .
of E is contained in Ej as a consequence of prop.3.3, so that

,
ECEj , all by virtue of the interpretation E = EXU E L(Hfor

EEE, of course. Also we get K(H)CK(Hj )CE , so that Ei DXEX + K(H).
For the converse we first note that trivially XKX E K(H) , for
every KEK(K.) . Also, for a generator E. of the form (3.14), con-
sider XEiX = XA[B,C]X Here we use the fact that there exist
.
'wider cut-off functions' X'E C-(c), 0<X'<1 , X'=O near an, with
XX'N=X , for every N=1,2.... . Note that X' commutes mod K(H
with every generator a , DAB , (ADD) of Ci Indeed, the gra- .
dient VX' has compact support, so that V,thm.3.1 applies to a

commutator while trivially [X',a] = 0 It follows that .
we can write
(3.15) XEjX = X(X'(AO[BO,CO] + K)XI)X ,
where A0 has been constructed from A by replacing in the finite

sum of finite products of generators a, DAB, every a by aO=X'a
and every DAB by D0A' , with DO=X'D and A' of prop.3.3. Simi-
,
larly for B and C , going into B0 and CO Then we have X'KX' E

.
K(H) , and X'A0[BO,CO]X'E EO , proving that XEjX C C. Q.E.D.
Corollary 3.5. We have
(3.16) EXU = Eon (CXU)
That is, an operator AEC has AXU E E3 . if and only if AEE

Proof. We know that EXUEEI . , by (3.13). Suppose AXU E Ej for some
AEC . With the decomposition 1=y+6 used in the proof of thm.3.1
we get A = 6A6 + K , where KEK(H) C E . Also 6A6=(6XU)A(6XU) E E,
by (3.13), since AEEj , q.e.d.
Theorem 3.6. The quotient algebras Vj/Ej and C/E are isome-
trically *-isomorphic under the canonical map A+E -, A+Ej , AEC
The symbol space of C/E is homeomorphic to the set i.. (MA #) C ffij,
U
with the map m m-, for m E C/E , m"'E Vj/Ej given by m = m- n C
or m= m U E. . For AEC the C-symbol equals the restriction QAlM
of the Cj-symbol to ffl i. (MA
U#
The proof of thm.3.6 is evident, after the above discussion.
In particular, every *-isomorphism is an isometry ((C1],AII,thm.
7.17, or [R1], IDxl] ).
Remark 3.7. Let us give a special consideration to the case where

U U M is compact. In that case, clearly, we have AU 0 = C0(U)
and DU is the class of all folpde's over U with compact support.
Hence, in that case, the algebra C = CU coincides with the mini-
mal comparison algebra JD of the restricted triple {2,dVJU,HJU}
Its commutator ideal is K(H), and its symbol space is the restric-
tion of the cc-sphere bundle to U The homeomorphism of thm.3.6
.
is just the injection of that restriction into the co-sphere

bundle of 0.
Remark 3.8. Now consider the case where U U 3U is non-compact.

Then either 3MA1# = aMAIU# , or 3M A,# = aMAIU# U 3MA,V# , where
V = c\(UU3U) is the interior of the noncompact complement of U

A similar alternative holds for j=2 . Now the homeomorphism of
thm.3.6 will put the cosphere bundle parts ti l(U) of ffij, j=1,2,
into correspondence. But it also will establish a homeomorphism
between the spaces tll(aMAI #) and 1-1(amA2 #) , under which the
U U
symbols of generators remain invariant.
Remark 3.9. We also get relation (3.17), below, as an equi-

valent of cor.3.5 for E K(H) and K(H.)
,
(3.17) K(H)XU = EXUn K(H
(The proof is left to the reader in fact (3.17) remains true if

;
we substitute L(H) for E ). This will be useful in the nextfollo-

wing section, where the algebra C/E turns out to be a function
algebra again.
VIII.4. Cylindrical ends 246
4. Complete Riemannian spaces with cylindrical ends.
In the present section we consider a noncompact (infinitely

differentiable) Riemannian manifold 0 of dimension n , with a dis-
joint decomposition into a finite number of subdomains (fig.4.1 )
0 = S20 U S21 U ... U S2N U 3 3 1 U . .. U a S2N ,
where the 12j are mutually disjoint with compact boundaries aS2j
(4.2) aS20 = 3S21U ...Ud0N , aS2j n aS21 = 0 , j,l > 0 ,
each aS2j ,j=1,...N being a (connected) n-l-dimensional submanifold

of 0 . Moreover, SZ0U3O0 is compact and,
, R+ [0,oo)
(4.3) 0 U 30. = B.xg+ _ , with aS2j = Bjx{0} .
In other words, using the terminology of [KG 1J , the manifold 0

has N ends, and is a half cylinder Bjxg+ at each end , where Bj
is a compact n-l-dimensional Riemannian space, and where B.x1+
carries the product metric
(4.4) ds2=dp2+dt2, dp 2=metric of B,1
with Euclidean metric on R On such

a manifold 0 we now just consider
the Laplace comparison algebra C
for generating sets A# and D#
satisfying (a0), (a1), (d0). Fig. 4.1
In general we cannot expect compact commutators. For the in-
vestigation of C=C(A#,D#), and its commutator ideal E=E(A#,P#) it
appears natural to use algebra surgery of sec.3 to link the alge-
bra C with corresponding C., j=1,...,N, where C. live on the
full cylinders RxB.= 00
Notice that for this type of surgery we indeed get 'compact
cuts', i.e., the cylindrical end Q. is a common subdomain U. of 0
3 3
and with compact boundary aS2j . For the corresponding task
involving a poly-cylinder of sec.2 , having n'>l, (i.e., some
noncompact part of 0 is identified with a subdomain of a poly-
cylinder) we get a noncompact boundary, however. Such type of
algebra surgery is attempted in [CDgll.
First of all it is clear that the wave front space T of C
is just the cosphere bundle of 0 , and the restrictions of the
symbols of generators to W are the functions
(4.5) a(x) , bJ(W )j , aEA# , D=bJ3 j+p E D#

x
for a multiplication operator a , and for DA , respectively.
The important question to be answered is about the
secondary symbol space M\01 i.e., we ask for points of BI
,
over infinite points of the compactification M # . At each end

A
the open subdomain 0, of the manifold S2 is identified with an
open subdomain Uj = Bjxl I =(0,oo) of the cylinder 0.
, ,
Thus we have the configuration of thm.3.1 , assuming that

the generating sets A# D are matched with corresponding sets
,
A# , D# on S2 . In that respect let us choose A# and D# in accor-

3 J 3 # # J 3
dance with (2.2) and the proposed sets A D following (2.2), ,
chosing n'=l and B=BJ . , j=1,.... N This and (3.1) describe the .
sets A D# uniquely, since in any part of 0 away from its ends

,
we just get all Cm-functions (folpde's).

Next we look at the commutator ideals E. = of the
Laplace comparison algebras C. = j=l,...,N These
were studied in sec.2. In fact, thm.2.6 gives a precise descip-
tion of the Ft-conjugated ideal E.^ of E. , where Ft denotes the
Fourier transform (2.3) in the t-direction of 0 . .
J
Remark 4.1. The ideal E. contains the operator A. = (l-A.)-1/2

3 1
0
where 4j = D2 + Ax3j denotes, the Laplace operator on Stj , with
the Laplace operator A on B . Indeed, we get
x,j J
(4.6) Aj^ = F.t1AjFt = (1+T2-Ax,j)_1/2 (T)-1 -+ 0 lTI

Also the operator valued function S(T) _ (1+T2_A x,j ) takes values
in K(hx'j) hx'j = L2(B.)
, because the manifold B. is compact.
Thus we indeed get AjÊ CO(l,hx'j) C Ej^ or, A.E E. ,
The above remark shows that VII,thm.4.7 may be applied to

each of the cylinder algebras Cj , showing that the secondary
symbol space is empty. (In particular we have MA #
I
= Bas
easily seen. One thus confirms easily that all assumptions of VII,
thm.4.7 hold.) In view of remark 4.8 we have the following result.
Theorem 4.2. The secondary symbol space of the Laplace compari-

son algebra C(A#,D#) is void. The symbol space Iffi of this algebra
coincides with the compactification of the wave front space 1
under the symbols (4.5) of the generators

Next we look at the commutator ideal E(A#,D#) Here we may .
apply thm.3.4, which will link E with E. . Again 0 and have

the common subdomain Uj=2j=BjXE+ Using thm.2.4 we get a com-
.
plete description of the commutator ideal E. Then we may gene-

.
rate the comparison algebra - here called C. - of the subdomain

U.C. SZwith the classes (3.3) and call its commutator ideal E.-.
J J J
Note that Cj" and Ej- also live on the manifold 0 , where
they again are generated by the common subdomain Uj=cj of 0
from restrictions of the classes A# and D# defined on 0 .
For each j=1,...,N choose a cut-off function 1(T), depending

on t only, and meeting the requirements of thm.3.4. Then, we get
E C'(N) For any EEE we have 0E , E K(H) , by
VII,thm.1.3, and [Pj,E]EKr(), by V,thm.3.1 and ECC. Accordingly
(4.7) E = y* EP + jj=1pjEipj (mod K(H))
Using thm.3.4 this implies the proposition, below.
Proposition 4.3. The quotient algebra E/K(H) allows the direct

decomposition
(4.8) E/K(H) = E1-/(H1 ® ...® EN"/(HN') , HL2(O.)

On the other hand, we now look at E as a subalgebra of Ej,
again using thm.3.4 and remark 3.9:
Theorem 4.4. We have, with !z . K(hj) , hj=L2(Bj)
(4.9) E3.°/K(H.-) = C(Ekx'j) Ii = E , j=1,...,N.
Proof. We apply thm.3.4 and conclude that, for EEE ., we have

J
(4.10) E + K(Hj) = K + K(Hj) *jEj*j + K(Hi) ,
with some EEEj, and H=L2(0). Since K(Hj-)CK(Hj) with the imbed-
ding of thm.3.4, we get a map u:Ej-/K(Hj")} Ej/K(Hj) , which is
an isomorphism, by virtue of (3.17). On the other hand, thm.2.7
implies that E./K(Hj) = C(E+,kx'j)®C(E-,kx'j). Also, from (2.31)
we conclude that the symbols of operators of the form 4) jEj
vanish identically on E . Again, checking on the symbols of all
the generators, as they are listed in thm.2.7, one finds that

there are enough symbols of this form to generate the entire
C(tk ), by the Stone Weierstrass theorem. We still introduce
x,j
the notation Ij = E + _ I for each j=1,...,N, and find (4.9)
,
established, and our proof is complete.

We now summarize the results about the commutator ideal E
below. Remaining proofs are standard and will be left to the
reader.
Theorem 4.5. For the quotient algebra E/K(H) we have
(4.11) EIK(H) = C(&l,fzxej) a) C(E2,hx,j) ® ... ® C(EN,kx,N) I
where each I. is homeomorphic to I , and where fzx'j are the com-

pact ideals of the spaces L2(B.)=h3.. Moreover, E (mod K(H)) is
generated by
(4.12) $jax,jlj*j, *jDx,jAj,pj : Dx'jED#j, j=1,...,N,

ax'jE
with the classes and of of (2.2), for B=B. . In (4.12)
the support of the cut-off functions may be chosen arbitra-
rily far out - i.e. within t>R , for arbitrarily large R.
The generators (4.12) have symbol
(4.13) a
x,jAx,] T.(T),
J D
x,jA2x,j T?(T),
J
TEE =1 ,=0 on E
J
,
k'
k#j
with
)-1/2
(4.14) A
x,j = (1-A x,j )-1/2 T.(T) = (1+T2A2
x,j
i
Thm.4.5 has an application to a singular elliptic boundary

problem on In considered by Nirenberg and Walker [NW], which we
shall not discuss in detail.
CHAPTER 9. Hs-ALGEBRAS; HIGHER ORDER OPERATORS WITHIN REACH.
In the present chapter we attend to two different, but

related problems. First, we propose to also study comparison alge-
bras in L(Hs) , where Hs is an L2-Sobolev space on 0 of order s.
Here the spaces Hs always are understood as spaces of the HS-chain
induced by the (self-adjoint) Friedrichs extension H of HO5 for a
a given triple {O,dp,H} (cf.I,6). In that respect we must assume
cdn.(s) in order to be in agreement with the customary definition
of Sobolev spaces.
Second, we will focus on more general N-th order expressions
within reach of a given comparison algebra C (cf. V. def.6.2).
Every compactly supported expression already is within reach of
the minimal comparison algebra, as we know from V,6. However, fm a
general algebra C on a noncompact 0 the general expression L no
longer is within reach. Thus we will ask for criteria to decide
whether a given L is within reach.
The relation between these two tasks is discussed in sec.l
below. There we also discuss the organization of the present
chapter. In particular we point out that some of the theorems
have to be 'recycled', in the following sense. The first applica-
tion only applies to the minimal comparison algebra J in L(H)
0
(or in L(Hs) ). That brings certain higher order differential
expressions 'within reach' of JO , hence of larger algebras, so
that, in turn, the same theorem now may be applied again, to get
larger classes within reach, etc.
Unfortunately this makes it necessary to allow as assum-
ptions, in many theorems, a variety of special cases, so that
the theorems look complicated.
IX.l. Sobolev comparison algebras 251
1. Higher order Sobolev spaces, and Hs-comparison algebras.
In this section, and in the following we assume cdn. (s):

All operators Hm are essentially self-adjoint, for m=1,2,... .
Then it becomes possible to work with general Sobolev spaces Hs

as already introduced in V,rem.6.3 (and rem.6.3'). We recall the
definition:
For s > 0 we define Hs = dom A-s , and
(1.1) (u,v)s = (A-Su,A-S V) , (lulls = (u,u)1/2 , u,v E Hs
For s < 0 we define (lulls and (u,v) s by (1.1) for all u,v
A_s
E H = dom and then Hs as the completion of H under ilulls .
The spaces H. and H_. are defined as the intersection and

the union of all Sobolev spaces:
(1.2) H. = n{H : sell , H__ = V[Hs : s E 11 .
In particular H. is considered a Frechet space under the

topology of the class {Il.lls s E 1)
: The space H_. is provided
.
with the inductive limit topology. For details cf. [C1],III,(l.23)

(but this will not be required in the sequel).
Notice that the collection (H:Isl<co) defines an HS-chain
in the sense of 1,6, generated by the unbounded self-adjoint ope-
rator H -i.e., the Friedrichs extension of the minimal operator
H0 of our expression H .
By definition Hs consists of special L2-functions, as s>0
For s=l and uEC0 the Sobolev norm Ilull1 coincides with a (weighted)
L2-norm of the (n+l)-vector-valued function (u;Vu), by V,(1.3).
Correspondingly, every norm Ilullk , k=2,3,..., may be expressed as
a weighted L2-norm of the 'jet' (u(a):Ial<k) of local derivatives
of order < k . We shall discuss this in sec.6, where covariant
derivatives will be used, for a global representation, under some
assumptions (cf. also [C1],III,(2.7), (2.8)).
In VI,cor.4.3 we already discussed a weakened version of the
Sobolev imbedding lemma, stating that H C Ck(S2) for s>[n/2]+k+l
s
Accordingly Hs consists of smoother functions as s grows larger,
and H.C C().
For s<0 the general uEHs no longer will be an L2-function,
but still may be interpreted as a distribution. Indeed, for s<0
and u E H , v E H-s = don A s , we get
IX.1. Sobolev comparison algebras 252
(1.3) I(u,v)I = I(u,v)ol = I(A-su,Asv)I < 'lullsllvll_s .
Accordingly, the sesqui-linear functional (u,v) , extends conti-

nuously to a sesquilinear functional over HsxH_s , which again
will be denoted by (u,v) , for the moment. For every u E Hs the
map v - (u,v)=lu(v) defines a continuous linear functional over
Hs , and we get Ilulls , by (1.3)
lllull and "_" for v=A-2suEHs
=
This shows, that Hs coincides with the set of linear func-

tionals over H-s , under the induced pairing HsxH_s 3 M In other .
words, u E Hs , s<0, may be characterized by the values of the

linear functional (u,v) v E H_ , S.
Note that the above pairing is identical to the pairing

I,(6.13) introduced for abstract HS-chains. By lemma 1.1, below,
the functional v+(u,v) is characterized by its values over CD(Q).
Moreover, the restriction to C_ of the functional is a distribu-
tion over N , as an immediate consequence of VI, cor.4.3.
In this sense it is true that all the Sobolev spaces Hs
Isl < are subsets of V'(1) , the space of all distributions
over 0 . Note also that we have Hs C Ht , as s > t (including ,
In fact the real-valued functions s } (lull for a fixed

s
u E V'(N) are defined and nondecreasing on the half-line (-=,s0)
if u E H s This is a trivial consequence of the definition.
.
Unweighted L2-Sobolev spaces over In have been introduced in

[C1],ch.III, using distributions and the Fourier transform. The
same approach is used in [C3] as well. The present definition
is easily related to the earlier one if we observe that 1-A =
F-1(l+lx12)F in case of the triple {,n,dx,l-A} , with the
,
Fourier transform F, and its inverse F-1 Note that this triple.
satisfies cdn.(s) and that (s) is quite essential in proving

,
equivalence of the above definition with that of [C1],V,(2.1)

The lemma below is an easy consequence of I,prop.6.1 and the
remarks on density of A_NCD(c2) in H following V,def.6.1.
Lemma 1.1. The space C_(c) is dense in Hs for --<s--

After the preliminary introduction into Sobolev spaces of
1,6 (from an abstract view point) and VI,4 (mainly directed to
compact manifolds) we will have to make a more systematical inve-
stigation for noncompact spaces now. In particular we will show
independence of Hs of the choice of the comparison triple in a
sense similar to that derived for comparison algebras: One may
change du and H on a compact set without changing Hs . Also, if

the manifold is changed outside a region U C St with compact boun-
dary then Hs does not change over U , in a manner to be discussed
(thm.1.2 and thm.1.3 below, to be proven in sec.5). However, our
most important task will be the discussion of comparison algebras
Cs C L(Hs) . Such algebras will have a similar application as
the algebras CC L(H) discussed in the previous chapters. We will
use them to characterize the Fredholm properties of realizations
of differential expressions within their reach.
For the concept of algebra to be studied next
we again require classes A D satisfying (a0), (a1) (or (al')),
and (d0) of V,1. In addition to (s), and (m2),(m3) we now must
impose stronger conditions on A# and D# . We choose to work with
a subselection of the conditions (lj), j=1,2,3,4,5, below. (Note
that (11) implies (m2) and (13) implies (m3). )
,
Condition (11): A# is an algebra under the pointwise product of

functions, and D# is a Lie-algebra under the com-
mutator product [.,.] of folpde's. Also, D# is a
two-sided A#-module, as under (m2), and A#, inter-
preted as a set of zero order differential expres-
sions, is a subset of D#. Also, for aEA#, DED# we
have [a,D]EA# .
Condition (12): D# contains the formal adjoints of its folpde's.
Condition (13): Condition (m3) holds, and, in addition, for any

a E A#, satisfying ago at a closed set Q C M#
A
there exists a function b E A# with ab=1 near Q
Condition (14): The expression H can be written as a finite sum of

terms of the form aDF , with a E A# , D,F E D# .
Condition (15): For every aEA# and DED# we have [a,H] E D# , and
[D,H] is a sum of finitely many terms bEF, bEA#,
E,F E D
In addition (with one exception) we will either impose (m5),
or even the following formally stronger condition.
Condition (m5') We have 1 E D#

: .
Clearly (m5') implies (m5), but (m5') means that A even be-
longs to the generators of C , hence is in the finitely generated

algebra CO This will be formally useful for the Hs-comparison
.
algebras, below.
The exception, mentioned above, is the minimal algebra J.
It is clear that its generating classes A# Dm satisfy (1.),
,
j=1,2,3,5, while, on the other hand, (m5) will not generally hold.
The case A# = A# , D# = Dm , nevertheless, must be admitted in
all results of sec's 3f. Actually, we know that J0 is a compact
commutator algebra, and it will be seen that (m5) will not be
required if the algebra C has compact commutators.
Note that the other exceptional case, not satisfying (a1),
and giving a non-unital C -i.e. the classes AU 0' DU 0 of VIII,3 -
needs no special consideration here. For, in order to achieve cdn.
(s) in spite of the truncation boundary 3U , one will have to work
with a potential q tending to infinity at 3U. This implies (m5),
(cf. prop.2.3).
For a pair A# , D# satisfying (m5), and (1j), j=1,2,5,
(in addition to (a0), (d0) and either (a1) or (al')), we
introduce the new notation A!=A#, and revised family DI=D#+{al},
where now A! D! generate exactly the same algebra C, but (m5')
,
is valid for A! and D!. Clearly A!, D! also satisfy (1 .), j=1,2,
J
since the constants commute with all of D! and A!. However, (al')
might be violated for A! , D! , although valid for A# D# (In , .
case where the minimal algebra J0 is under consideration, we will

use thm's 3.8, 4.2, and 4.3 with A! = Am , DI = Dm, since
then the adjunction of the constants to D# might give a larger
algebra.)
It is clear that (m5') follows from (a1) and (11). We will
show in sec.2 that (a1), and Cl) j: j=1,2, imply (m4), and we
will see that (m5) even follows from (al') and (11) alone, if 'q
grows large where D# gets scarce', as to be specified (prop.2.3).
It follows by a calculation that {(11):j=1,2,4} implies (15)
while {(lj):j=1,2,5} does not necessarily imply (14) .
In thm.2.1 we will show that (14) is a consequence of (m6)

assuming some of the other conditions (m.) and (1.) . Actually, we
3 3
will show that (14) is equivalent to a 'global Parseval relation'
of the form
(1.4) {D,F} = LN,p=lave{D,Dv}{DU,F} for all D,F E D#
with a symmetric positive matrix ((avu)) of functions avuE A#

and a system D1,...,DN E D# . The same relation (1.4) also

follows locally from (m6)x for an orthonormal base Dj,j=1,...
0
..,n+l, at x0, and then can be pieced together globally, using the
compactness of M # .
A
It should be emphasized that condition (14) may amount to a
restriction of the Riemannian space at infinity. It implies
that the metric tensor hIk can be globally written as a finite
sum javbvcv , with bounded functions av and tensors bv, cj still
satisfying a set of conditions, as principal parts of D,F E D!

(cf.[CS1,VI,L.2.5, [CM11, [CM21, and sec.3, below ).
To say this again in different words: We presently depend on
(15) which is difficult to achieve for classes A! , D! , nontri-
vial near infinity, except through (14). But (14) often cannot be
satisfied. However (15) holds trivially for the generators Am
Dm of the minimal algebra JO , regardless of any other properties
of the triple or the space The above difficulty will in part
.
be removed later on, as follows: After proving our results for

the algebra JO (and its Hs-equivalents) we will be able to substi-
tute (15) with a more general condition, but will have to go seve-
ral times through the same chain of arguments, under slightly
varying conditions. This forces us to make split assumptions.
Also, all the results, below, concerning our above Sobolev
spaces, will only require the higher order theory for the special
case of the minimal algebra JO .
Theorem 1.2. Let two triples {c,dp,H} , {c,dv,K} on the same

manifold be given, where H=K dp = dv holds outside some compact
,
set M C Q . Then we have cdn.(s) valid for {Q,dp,H} if and only if

it holds for {c,dv,K}. For each sE [-co,o1 the Sobolev spaces Hs of
the two triples coincide as subsets of DI(Q) The corresponding
.
Sobolev norms are equivalent (although the inner products and Hs-
adjoints in general are different).
Theorem 1.3. Let two triples {Qj,dvj,Hj} j=1,2, be given, where

and 02 both are extensions of a manifold U such that U appears
1
as an open subdomain with compact boundary 3U of 2 1 and 22 ,each.
Let condition (s) be satisfied for both triples, and assume that
the restrictions to U of the two triples coincide:
(1.5) dp = dp1IU = dp2IU , H = H1IU = H2IU .
Then, if the Sobolev spaces of the triples are denoted by Hs

then for every open set V C U with Vclos C U we have
HSIV = H2 IV , for all s E I
Hs IV = {uIV : u E Hs}
Moreover, if 2U is a smooth n-l-dimensional (compact) sub-

manifold of Qj , then the statement also holds for V=U .
(Recall that the restriction uIV of a distribution u E V'(2) to
the open subset V is defined as the restriction of the linear
functional to the testing functions with support in V.)
The proofs of thm.1.2 and thm.1.3 depend on techniques
involving higher order commutators to be prepared in sec.3.
We will discuss these proofs in sec.5.
For a given triple on a manifold 1 consider a chart U as a
common subset with the manifold In . Assume that U is a subset
with compact closure of a larger chart V Then all assumptions
.
of thm.1.3 hold, and it follows that the subset of u E Hs with

support in U coincides with the subset of Hs(Rn) with support in
U . Here we first must use a measure du and expression H different
In
from the surface measure and Laplace operator of , on some
compact subset containing U but we then may use thm.1.2 to
,
remove this and obtain the Laplace comparison triple of In

Thus we have the following.
Corollary 1.4. At each point x0Ec there exists a neighbourhood

N of x0 such that the set HsIN of restrictions uIN of u E Hs to N
if related to In by the local coordinate map, precisely consists
of all restrictions of u E H(In) to N .
Indeed it is a direct consequence of thm.1.3, that HsIN

H`sIN , where H-s is formed with the manifold In and a triple
equal to the Laplace triple outside a compact set, containing N
but coinciding on N with the given triple on Q. Then one uses thm.
1.2 to show that H-s is just the ordinary Sobolev space Hs(Rn)
This corollary connects our present Sobolev spaces with
those of In , discussed in detail in [C1],III. In particular we
obtain the Sobolev imbedding, generalizing VI, cor.4.3:
Proposition 1.5. We have Hs C C(O) , as s>n/2 , and H C Ck(Q),

s
as s>n/2 + k . In particular, for any triple over D2, we have
C' (D) C H C C0(Q)
Note that we are not attempting an investigation of the

spaces Hs at infinity. For example, for S!=Xn it is known that the
functions of H. have all derivatives vanishing at infinity. More
generally one could ask how Hs will be influenced by a change of
hjk
the metric tensor or the potential q. Discussions of this
type will not be made here.
Assume now that we are given a function algebra A! and a
Lie-algebra D! satisfying (m5'), and (1.), j=1,2,5, as well as
(a0),(al)(or (al') for D# only, D!=D#+{al}), (d0), (or also, we
admit the case of the minimal classes).
Under these assumptions we will show that the operators
(1.9) a , DA , AD , for a A! , DED!

belong to L(Hs) (cf. thm.4.2). Accordingly, for s E B , the
Banach algebra Cs = C5(A!,D!) obtained as norm closure in L(H5) of
the finitely generated algebra C0 of the operators (1.9) will be
called the Hs-comparison algebra of the triple {S2,du,H} , and the
generating sets A! and D! Similarly we introduce the algebra C.
.
as closure of C0 in 0(0) For a precise definition of the opera-

.
tcrs (1.9) one may think of D=bJB j+p in terms of distribution

x
derivatives. Or else, one also may think of the closures (in Hs)
of the unbounded operators ajHs, D0,(AIHs), (AIHs)D0, respectively.
For technical reasons we introduce a second type of finitely
generated algebra, called CO , obtained from the generators
(1.10) a , DAj , AJD , for a E A! , D E D! , j=1,2,....
Clearly we get C0 C while C = Co whenever (m5') holds,

since then we have A E CO, hence DA3= (DA)Aj-'E CO, and similarly
A2D E CO. If (m5') does not hold, then (1.9) and (1.10) still
generate the same Banach-subalgebra of L(H), (and of L(Hs) or 0(1)
as well, as will be seen later on). Indeed, from (al') we get
existence of a function X E A! such that XD=DX=D, for any given
D E D! , and with OX of compact support. Then we get DA=(DA)(AX)
2
+D[x,A2], where the last term is in KO), by V,lemma 3.2. By (11)

we have x E D! so that the first term is contained in CO
, Also, .
C D K(H), by V,lemma 1,1, thus it follows that DA2 E C. A similar

conclusion implies DAB E C, j=3,4,... , and AIDE C, j=2,3,... .
In thm.4.2 we also will show that, under above conditions,

the algebra Cs is a C -subalgebra of L(Hs), while the ideal chains
C and Cs are in agreement, and the symbol spaces remain the same.
In fact, under suitable additional conditions, such as (14), even
the symbols of the generators remain independent of s.
2. Closer analysis of some of the conditions (1.) and (m.).
In this section we assume (a0),(d0),(al)(or (al')), and more

cdn's as stated. In thm.2.1 we show that (m6) implies (14), assu-
ming some other (mj) and (lj) . Actually, we show that (14) and
(**) D=b03x.+p E D# implies p E D#, and D0=b33 . D#
imply a 'global Parseval relation' of the form
(2.1) {D,F} = JN,=lava{D,Dv}{D,,F} for all D,F E D#
with a symmetric positive ((a,X)), avXElk, and a system D1,..,DNED#

while (2.1), even with avxEA , and without (**), but with inver-
tible ((avx)) implies (14). Thus (m6)x gives a local (2.1),i.e.,
0
DvED#
(2.2) H = 16,X-1DvaVADX , , near x0,
from which a global (14) may be pieced together, if (m6) holds.
Theorem 2.1. Assume cdn's (m1),(lj),j=1,2,3. Then (m6)x implies

0
(2.2), while (m6) implies (14), and even (with certain DEED#)
(2.3) H = 6=6(x)=n+l
with a partition {xj:j=l,...,N} of MA#, xjE A#, au,=a,u=ai.E A#.
Vice versa, (14) and (**) imply (2.1) with certain DvED#, avxE I
Proof. Assume (m6)x for x0E 3M #. With some orthonormal base {D.}
0 A
of Sx we get ((a.l(W))) invertible for xE N=Nx , where
0 0
(2.4) ajl(x) = {Dj,D1}(x)E A# , aj1(x0) = Sjl

IX.2. The conditions (1) 259
Using (13) on det((a J.l)) we get X((ava(x)))-1=X((avX(x))), with

ava
functions E A# , with a cut-off function X E A#, in accordance
with (m3), for x0 and N . We introduce ((bvx)) _ ((av,A))1/2, and
find that Dv=lXbvXDX , v=l,...,5, is an orthonormal base at every
xE N (Perhaps D )"'E D#, but still D v"E S ). Use Parseval's relation

x
for the DJ J. Define the local matrix H"= ((hv,X-))v
a=0 n by
hvA =hvX for v, a>1, =q-1 for v= A=0, =0 for v=0, a>l, and v>1, u=0,
for a moment. Also let D,-= Jbva j+b0 , and then define the matrix
x
B=((b)) v=0,...,n,a=1,. .6 , with'v=row-index, u=column-index.
X
The above is valid in local coordinates, near points of Q. dense
in M Observe that then Parseval's relation can be written as
A#
(2.5) b*H-b = b(H-BB*H"')b , for all b E !):d , xE NO

H"'-1=BB4
This relation may be polarized, and then yields . Or,
(2.6)
which is valid for all x E N , u E Md . If we substitute u0=u(x),

uj= u .(x), for some u E C' (O) , and multiply by X(x) , we get
IX
(2.7) J0dpX2(hJku u k+gIul2) =

1b_1IIXDv""uII2
, u E C .
IxJ Ix
Now we get Dv = XcvXDX- , with the matrix

((cva))=((bvX))-
((avX))1/2 .
As a consequence the right hand side of (2.7) may be
written in the form
(2.8) Lsa=iJ0dpX2avXDvuDXu
Finally, assuming (m6) U(m6)x, use (13) to construct a par-

tition of unity {X1,...,XN} on the compact space M # , where we
A
require that XjE A# , j=1,...,N , while JX? = 1 on M # , and such

A
that (2.7) with (2.8) hold for every Xj with suitable
Dv = Dvj . Taking a sum over j we then get
(2.9) (u,u)1 = LN=l1 v,a=1(XjavADvju,DXju) for all u E CQ(O)

Here we finally may polarize and change notation for (2.3),

i.e., the first half of the theorem.
Vice versa, let (14) and (**) hold. Accordingly we can write
(2.10) H = Iv=lavDvFu , av E A# , D') , Fv E V#
where all coefficients may be assumed real-valued, since H has

real coefficients. Using (11),(12),(m1) one may rewrite this as
IN
(2.11) H = v=1D Fv
with changed, but still real-valued DV,FV Let Gay , X=1,...,R, .
be a basis of span {D,,, D,,, F,,, F :v=1,...,N}, with the principal

parts DV , FV of Dv Fv, We may express Dv and Fv as linear com-
.
binations of the G. and (2.11) takes the form

,
(2.12) H = iR,X=lavXGv*GX ' a,,=

a,X= a.X E R
Now use that H is self-adjoint, so that H = 1/2(H+Hh) . We get
IR
(2.13) H = X=lavXGv*GX
where again the avX have been changed. Now the matrix ((avX)) is
real and symmetric.
Returning to our old notation we have proven
Proposition 2.2. If condition (14) is valid, then we have
(2.14) H = IR,X=lavADv*DX , with avX=aX=aXv E A# , Dv E V#
where all functions and folpde's are real.

Now we write
(2.15) Dv = bj3xj+b , B =
n,v=1,...,N
where B will be considered an (n+l)xN-matrix , i.e., j indicates
the rows and V the columns. It may be assumed that DV, v=1,...,M,
have bR=0 , while D v=M+1,...,N, are of order zero. Now let
denote the sum (2.14) taken only over v,X with at least one >M
While g' formally is of first order, it really is of order zero,
0 0
due to bXDv 0 0* _ (DVV
+ (bXDv) 0+D 0* )b v
0 0 0
+ [bX,Dv], where the right hand
side is sum of products of two zero-order terms in V#. We thus may
repeat the above proceedure once more on 1', to arrive at a new
matrix ((aV), )) with a.VX=O as v<M , X>M or V>M , X<M . Comparing

coefficients in (2.14) we then find that, with H- of (2.5) we have
(2.16) H--1 = BAB*
Then (2.16) implies H_=(H-B)A(H-B)* which translates into
(2.17) {D,F} = IN X=lava{D,DV}{DX,F} , for all D,F E D#

s

Next we will look at condition (m5) . This condition is a
consequence of (11), if (a1) holds. On the other hand, under (al')
only, prop.2.3, below, shows that (m5) follows for the algebra C
of the 'truncated classes' AU,O' DU,0 of VIII,3. (We, of course,
then must work with a potential q going to - at 2U as was used ,
in VIII,prop.3.3.)
Proposition 2.3. Suppose D# contains a sequence of functions

(i.e. zero order expressions) X. j=1,2,..., such that ,
0 <Xj <1 , Xj = 1 in Qj , where UQj = 0 , while

(2.18) limj,(inf{q(x): x E S2\Q}) = - .
Then condition (m5) holds.
Proof. Let Aj = XjA E C . From V,(1.7) it follows that
(2.19) 0A1 - A'1 = 0(1-Xj)AII < 0 , as j--

Hence A E C , q.e.d.
Proposition 2.4. Conditions (15),(m5) and (w) imply cdn. (m4) .
Proof. First, in the special case of the minimal classes, we have

a compact commutator algebra, for which (m4) was proven in V,3
(since a and D have compact support, V,lemma 3.3 and V,lemma 3.6
apply ). Similarly, in the general case, we again use a resolvent
integral of the form (and with the contour of) V,(3.6). Since we
(A2-X)-1
assume (m5) we have A2 E C , hence the resolvent R(X) =
is in C for all X E r\{0), with the path r used in (4.6). (If C is
non-unital, we at least get AER(X)EC, which is enough). We get
(2.20) Ia,R(X)]=AR(X)[H,a]R(X)A2
2 , ID,R(X)I=A 2 R(X)IH,DIR(X)A 2
[a,R]A-1-e , A-1-e[a,R] E C.
Using (15) we get [H,a] E D!, so that
[D,R]A-e
Similarly, A-£[D,R] E C , due to (15) again. Moreover,
,
the integrands are norm continuous (even analytic). Using V,(3.9)

we get estimates like those used in V,lemma 3.3 or 3.6, so that
the integrals converge in norm. This completes the proof.
3. Higher order expressions within reach of C or Cs.
In this section we first assume cdn's (s),(1.) j=1,2,5,

and either (m5'), or else the minimal classes Am, Vm. Later on we
shall replace (15) by a weaker condition. Let C C(A!,D!) C L(H)
and C0 be the comparison algebra and the finitely generated alge-
bra introduced in sec.1, repectively. Our first use of (15) will
be to identify larger classes of expressions L within reach of C
or C. , as specified in V,def.6.1.
By (11) the class D! is a Lie-algebra, under the commutator
product, and a two-sided A!-module. Let L' denote the correspon-
ding universal enveloping algebra, defined as the algebra of all
differential expressions of arbitrary finite order, of the form
j
(3.1) L = X. 1a.
II1=1 Djl , a,'= A! , Dj1E D!
Clearly this defines an algebra under A! since D! is a two-sided ,
A!-module. We may set a =1 except for Ni = 0 , .

j
By Lm we denote the class of all operators in L of the form
(3.1), with NJ < m , for all j=1,...,M . Since 1 E D! , by defi-
nition (or else D! = Dm ) , we may assume exactly Nj=m terms in
each product of (3.1) , except for m=0 where L0 = A! Also we .
get L1 = D! , while, for m=0,1,2....., Lm is a linear space of

m-th order differential expressions over 0, containing its formal
adjoints, and all m-th order Co-expressions with compact support.
Note that (14) amounts to the condition that H E L2. Also
(15) means that [a,H] E L1, [D,H] E L2 , for all a E A!, D E D!
Let A!0 and D!0 denote the ideals of A! and D! generated
by the commutators [a,D] E A! , and [D,F] E D! , respectively.
Let L,,o C L,, LN,o C LN denote the subsets of all L of the form
(3.1), where at least one factor of each product is in D!0. Here
it is no loss of generality to assume that always the first
(always the last) factor is in D!0.
To be more specific, we set L0 = A!0 . A set of generators
0,0
of L1
1,0
= D!0 is given by {a[D,F] : aEA!, D,F E D!), or also by
IX.3. Higher order expressions 263
{[D,F]a : aEA!, D,FED!} , or by {a[D,F]b : a,bE A! , D,FE D!}

For m=2,3....., a set of generators is given by
D.,D,FED!}, or else {[D,F]Dl...DN-1: D.,D,FED!}.
.DN-1[D,F]:
{D1"
In particular, L. 0 again is an algebra invariant under multi-
plication by A! and adjoint invariant. For the minimal classes
,
it follows easily that A!0=A!=Am , D!0=D!=Dm Lm O=Lm . There- ,
fore, since a general pair always contains the minimal classes,

we find that also a general Lm o contains all Cam-expressions of
s
order m with compact support.

The following modification of condition (15) will be useful:
Condition (15'): We have CH,a]E L1,0 , CH,D]E L2 for all
2,0
a E A! , D E D! .
Note that (14) not only implies (15), but even (15'). Also,
the minimal classes satisfy (la') not only (15) ,Some of the .
results below need the stronger condition (l5') instead of (15).
Proposition 3.1. The class L.O is a two-sided ideal of L_ .

e
Moreover, we have LL', L'L E Lm+m, for L E Lm LIE= Lm, , and ,

0 s 0
(3.2) [L,L'] E Lm+m' 1 0 for L E Lm , WE Lm? .
The proof is a calculation, and is left to the reader.

For every differential expression L E LN consider the two
linear operators
)e.*
(3.3) (L.0 AN)4 , (ANL 0
formed as closures of products of unbounded operators over H ,
with the minimal operator LO of L The closures are well defined

.
since AN is bounded, and L has a well defined formal adjoint.
Proposition 3.2. The operators (3.3) are in L(H), and, moreover,

they even are in the comparison algebra C(A!,D!) In fact, they .
even are in the finitely generated algebra C9 of sec.l. In parti-

cular, using the terminology of V,def.6.1, every expressions LELN
is within reach of C and C , as an expression of order N
Remark: Please observe that C and C, have the generators (1.10),

each algebra with its own way of generation. In particular, the
ideal of compact operators appears on its own, and is not needed
among the generators.
Proof. Note that the set of generators (1.10) contains its
adjoints, and that also LN is adjoint invariant. Thus we have

(A NLO) ** =(L"A N) * E C if we can show that (LOA N ) E C. and it ,
is sufficient to consider the first kind of operator (LOAN) **

Accordingly we essentially may repeat the proof of VI,prop.3.1,
noting that cdn.(l5) just supplies the essential property.
It is clear that the assertion holds for N<l , since
O**
a=(aA) , 1** just are the generators of C
and DA=(D,0A) . For
a product DF E L2 we get, as in the proof of V,lemma 3.2,
(3.4) DFA2u = DA2Fu + DA2[H,F]A2u , for all u E im H0 .
Using (l5) we may substitute [H,F] by a finite sum of products

D'F' , for which the second term (3.3) assumes the form
*
DA2D'F'A2u = (DAM' A) (F'A2)u All these terms involve opera-
.
tors of C0 only. Since in HO is dense, we thus get ((DF)OA2)**

E CO , proving the assertion for N=2. (The first term, of course,
gives (DA)(D*A)* E C..)
Assume now the statement proven for LN ,N>2 . Consider a
A-N-1CO
product LM , LEL2 , MELN-1 C LN . For uE we get
(3.5) LMAN+lu = LA 2MA N-lu + LA 2[ H,M] A N+lu
since AN±lu E C_ . Using (15) and prop.3.1 we get [H,M]ELN,0 C
LN+l 0 , so that the second term represents an operator in C..

(It may be written as (L0A2)**([H,M]OAN+l)**u. ) The space
A-N-1C0
is dense in H in view of cdn.(s). Accordingly (3.5)
implies that LMAN+1 E C_ Since the general expression of LN+l
.
is a sum of expressions LM of the above type, the prop. follows.

In order to simplify notation we will, for the present sec-
tion, use the notation L for the realization Z of V,def.6.2, cor-
responding to an expression LELN In other words, the expression
.
L and the unbounded (differential-) operator L:HN+ HO=H will be
denoted by the same symbol L . (We recall that HN = dom A-N =

im A N. For u e HN we write = ANv , and the define
(3.6) Lu = LANv = (LOAN)**v
which is well defined, by prop.3.2.) Note that this notation

already was in use for DES! , where we used to write DA , inter-
preting dom D = H1 = im A . Also we observe that L C L0** The .
notation is ambiguous, due to the fact, that an expression L of

order N also has every order R>N . However, all these various
operators, denoted by L , have the same closure L0
N N L)**A
Proposition 3.3. For L E LN 0 we have ALA E E, (A .
Proof. Again it is sufficient to focus on the first kind of opera-

tor, using that (ANL)**A = (AL*AN) Looking first at Am=A!, Dm
.
D! , we find that then the operators listed are compact. Indeed,
from prop.3.2 we conclude that LAN E CO is a finite sum of pro-

ducts of terms (1.10), with compactly supported a and D. But then
Aa, ADA3, Aj+1D all are compact, by III,thm.3.7, so that ALAN
is compact as well. Since K(H) c E , the statement follows.
In the general case we always have (m5). For N=0 we must
look at A = A[a,D]. Note that [DA,a] _ (DA)(A-[A,a])+[D,a]A E E
By prop.2.4 we have (m4) valid as well. Thus VII,rem.3.4 implies
that the first term at right is in E , hence [D,a]A E E
*
and A[D,a] _ (A[a,D]) E E as well. Note that we have (m4
again by prop.2.4. To discuss the case of N=1, let L=a[D,F]
aEA!, D,FED! Now we get, with proper interpretation of [D,A],
.
(3.7) [DA,FA] _ [D,A]FA + A[D,F]A + [A,F]DA E E .
We already noted that [D,A] E E, [A,F]EE, by VII,rem.3.4, since we

have (m4) and (m5). Therefore (3.7) implies A[D,F]AEE. But we get
(3.8) Aa[D,F]A = ([A,a]A-1)(A[D,F]A) + aA[D,F]A ,
where all terms at right are in E . Therefore the statement holds

for N=l .
For N=2 the typical term of L is of the form F[D,E]. We get
(3.9) AID,E]A2u = (AF)(A[D,E]A)u + (AF)[A,[D,E]]Au , uE H2 ,
where the right hand side operator is in E , by the above.

Finally,for general N>2 the typical term in LN+1,0 may be
,
written as LM , with LFEL2 2,0 , M E LN-1 Then we may write (3.5)

.
again, multiplied by A from the left. Again we get ALA2MA N-]EE

ALA2[ H,M]AN+l E E , completing the proof.
In the following we introduce more general algebras of diffe-
rential expressions of arbitrary order, within reach of an algebra
C or Co. The purpose is to replace cdn's (15), (15') by weaker
conditions.
Let P = N=OPN be a graded algebra of differential expres-

sions, invariant under formal adjoints, and a 2-sided L.(A!,D!)-
module. In details, we assume the expressions in PN of order <N,
that PN C PM as N<M (Cm-coefficients, of course), and that
,
(3.10) P P
N M C PN+M' PNLM C PN+M' LNPM C PM+N' N,M=0,1,2....
.
In addition let the following conditions (pj), j=1,2, hold:

Condition (pl): [H,L] E PN+l for all L E PN , N=0,1,2,...,
Condition (p2): Every L E PN is within reach of L(H) .
Notice that (pl),(p2) imply that all LEPN are H-compatible,

in the sense of V,def.6.4.
We shall say that an algebra P. with above properties is
within reach of an algebra A C L(H) if every P E PN as an ,
expression of order N is within reach of A

, Also, if P. is
.
within reach of C_(A!,D!) for some classes A!, D!, with C.CO(0)
of sec.1, then we say that P is within general reach of A! , D!
For P0, within reach of C5(A!,D!) , if we even have ALA NE Es(A!,D!),
for all LEPN, N=0,1,... we shall say that P. is of principal
,
symbol type (in C ), - the latter only for s E I .

s
For PW with above properties define the classes QN
= PN+LN'
N=0,1,... Then Q. =
. N=0QN clearly again is a graded algebra
of differential expressions and P. is an ideal of Q.
We will ask whether Q,, again is within reach of C(A!,D!),
given that P,, is within reach of C(A!,D!) . Cor.3.4 below gives
an affirmative answer, using the following amended condition:
Condition (15)(P,,): We have [H,a] E Q1, [H,D]E Q2, for aEA!, DED!.
In case of an algebra Pte, satisfying (pj), j=1,2, and the
above, and of principal symbol type, we also formulate a relative
cdn.(l5') ,
as follows (setting QN,0 = PN+LN,0' Q- = U QN ):
Condition (15')(P):Wehme [H,a]EQ1,0, [H,D]EQ2,0, for aEA!, D(=-D!.
Notice that (15)(P.) implies that Q,, = UQN satisfies (p.),
j=1,2, by the corollary, below. In fact, this even is true without
the condition that P. is within reach of some comparison algebra,
by the same arguments.
In the corollary, below, the special case of the minimal
classes Am, Dm is excluded. (Since Am , Dm satisfy (15) the ,
corollary would be useless.)
Corollary 3.4. Under the above assumptions on P,,, if cdn.(l5)(P,,)

holds (but not necessarily cdn.(l5)), then the graded algebra QW
is within reach of C(A!,D!). Moreover, if in addition P0, is of prin-

cipal symbol type (in C0= C) , then the algebra Q, 0
= P+L_ 0
also is of principal symbol type (in C).
Proof. Since P E PN already is within reach of C we only must
show this for 1. E LN again. For the latter one may repeat lite-
rally all arguments of the proofs of prop.3.2 and 3.3, using the
ideal property of Pte. In (3.4), for example, the commutator now is
a sum P2+L2 , where L2 is treated as before, while P2EP2 gives the
term (DA2)P2A2EC. Similarly for (3.5), also in view of prop.3.3.
4. Symbol calculus in Hs
In this section we keep all assumptions of sec's 1 and 3

valid. Particularly we assume cdn's (s), (1j), j=1,2, and (15)
(or (15')), and either (m5) for a triple {O,dp,H}, and classes A!
A#, D! = D#+{al), or we use the minimal classes A!=Am, D!=Dm
Or else, we assume (15)(P0,) (or (15')(P0,)), instead of (15), (or
(15')) , relative to some algebra P_ within comparison reach of
C(A!,D!) , satisfying (pj),j=l,2, and (3.10) (and of principal
symbol type (in C)). In the latter case we exclude the minimal
classes, and always assume (m5') .
In the interest of a unified notation, let us always at

least formally work with an algebra of expressions P. , where we
set PN = {0} in the case where no preconceived such class is
given. Clearly this defines the trivial algebra P_ , satisfying
all assumptions. It is within general reach of every A!. D!, is
of principal symbol type, gives an algebra P0,, and an L0,-module,
satisfies cdn's (pj), j=1,2, and (15)(P.) , (15')(P0,) , for the
trivial algebra P0,, coincide with the ordinary conditions (15) and
(1b'). We appoint that for the minimal classes Am , Dm only the
trivial algebra P, is to be used. We then always define QN=PN+LN
in case of the tri-
QN,O=PN+LN,O ) so that QN LN ' QN,O = LN,O 5
vial algebra P. .
Let us recall the order concept introduced in 1,6 for gene-

ral HS-chains. Here we use the chain {H:sE[--,+-]} of Sobolev
spaces, of course. We shall adopt an attitude similar as in theory
of pseudodifferential operators to not distinguish in notation
between the various extensions in L(HsHs-m) of a given operator
A E 0(m). If AjC is given by a differential expression L then we
IX.4. Symbols over HS-spaces 268
even use the same notation L for all above operators as well as
the realization Z of V,6 This will bring us away from the
.
unbounded operator notions of ch.l, and will be practical in view

of the fact, that there is a unique such extension, for every s,
and a unique closed realization, all in view of cdn.(s).
Focusing on the algebra Q_ = P + L. of expressions, we find
that every L E QN is within reach of C and even is H-compatible,
in the sense of V,6, by cor.3.4 and the remarks preceeding it.
Thus we find that L defines an operator in 0(N), by V,cor.6.10.
Proposition 4.1. For A=LAN, B=(ANL)* , with LEQN , we get
(4.1) AsAA-s-A = CA=AC'=E , AsBA-s-B =C"A=AC I"=E' , sEl
where E, E',C, C', C", C"'E C may depend on s . Moreover, if we

require (15')(P-) instead of (15)(P-) , we even get E,E'E E .
Proof. For a given s we apply V,thm.6.8, with M > Isi . Also we

use that Lj=(ad H)2L EQN+j (or EQN+j,O, in case of (15')), so that
(4.2) LjAN+3E C , and LjAN+j+1=ALjAN+j+ [LAN+j,A] E E,
by cor.3.4, correspondingly, depending on the assumption made.

(For the minimal classes we must use XA for A , in the commutator
with suitable X.) Similarly the term ALN+lA11+1 in V,(6.12) is
in E or of the form CA, CEC, respectively, hence also the
,
reminder gives a term of the proper form for V,(6.13). (For the
minimal classes Am, Dm the resolvent R(?) only is in C + {al}
but C is an ideal of C+{al} so that the integrand of V,(6.12)
still is in C or E resp. This gives the first (4.1) while the
second follows by taking adjoints, q.e.d.
For thm.4.2 below we recall the natural isometry Ar:Hs+Hs+r
(Actually, Ar stands for the extension (Ar)s of Ar to Hs, but
the subscript 's" can safely be dropped). Note that
(4.3) L(Hs+r) = Az'L(Hs)A-' , K(Hs+r) = ArK(Hr)A-r ,
For the remainder of this section we assume, in addition to

the conditions mentioned earlier, that P- is within general reach
of the classes A!, D!, and of principal symbol type in Cs, for sEl.
Theorem 4.2. The generators a, DAD. ADD, aEA!, DED! of (1.10)

are in L(Hs), for every sEl, (and in 0(0)), so that all comparison
algebras Cs (including C,, ) are well defined norm closed subalge-

bras of L(Hs) (or 0(0)). We have LAN, ANL E Cs , -°°<s<- , for all
LEAN . Every Cs sEl, contains its HS-adjoints, i.e., Cs is a
,
C -subalgebra of L(Hs). For and every A in the algebra

finitely generated by C,,O and {LAN, ANL: LEAN, N=1,..} we have,
(with the commutator ideal Et of Ct)
(4.4) A-SAAB - A E Ct, and 'E Et' in case of cdn. (15')(P-),
Also,
AsCrA-s
(4.5) Cs+r = , Es+r = ASErA-s , for all s,r E 1
In particular, the class Q. is within general reach of A!, D!

and Q is of principal symbol type in every C (A!,D!)
Proof. Prop.4.1 implies that a DA , AD LAN , ANL E 0(0), hence
,
EL(Hs), justifying the definition of Cs, also for s=-. To show

LAN, A N LEC5 one repeats the argument of prop.3.2, as. in cor.3.4,
using that P is within reach of CS, --<s<°°. Inspecting (3.4) and
(3.5) one indeed confirms AECs Similarly, in case of (151)(P,,),
.
one confirms that Q_ 'O is of principal symbol type (in Cs ), sEX.

A2sA*A-2s,
For an operator AEO(0) the H-adjoint
S is given by
with the H-adjoint A*, since (u,Av)s= (A2sA*A-2su,v), for u,vEH,.
We get A2sA*A-2s E Cs , by (4.4) with s,t replaced by -2s and s.
Thus the adjoint invariance of Cs follows from (4.4).
Next we note that (4.4) for t=O implies
(4.6) A-sCQ,OAs C C , for all s E I
However, if A. converges to A in L(H) , then A-sA.As E L(Hs)

converges to A-5AAsE L(Hs), and vice versa. Therefore we may
take closures in (4.6) and get
(4.7) Cs C AsC0A-s
This conclusion may be repeated with CO=C replaced by Cs, and

A-SC SAs
As by A-s, using (4.4) for t=s.Then we get C=C0C , or,
equality in (4.7). Thus we have the first relation (4.5) , for
r=0, and the second follows since, of course, the commutator ideal
is invariant under a conjugation with an isometry. Similarly for
general r.
Accordingly, (4.5) and the adjoint invariance both depend
on proving (only the first part of) (4.4).
However, (4.4), for t=0 and s E Id , is a consequence of

(4.1) as far as the generators are concerned, and it then
,
follows trivially for all A E C-0. Similarly for A finitely gene-

rated by ANL, L E 2N. For general t we cannot use prop.4.l,
LAN,
but must go back to V,thm.6.8 instead, showing that all terms at

right of V,(6.11) are in Ct. Indeed, we get LjEQN+j, hence LjAN+2j
E Ct , as already shown above, using that P. is within reach of
C. . Also the integrand of S m,s is in C t again. In fact, we have
ALM+1AM+jE Ct. In case of (m5') the functions S1, SM also
s
take values in Ct since AECt , and S3 are functions of A

Therefore the integrand is in Ct . On the other hand, in case of
the minimal classes, the resolvent is in Ct+fal} , hence the S1
are in that algebra. But Ct is an ideal of Ct+{al}, so that again
the integrand is in Ct. Thus we only must prove convergence of
the integral in L(Ht) .
For that we may look at the L(H)-convergence of the integral

V,(6.12) with LM+IAM+N+l replaced by At(LM+IAM+M+l)A-t= Vt, since
At and A-t commute with SJ(1) But we get V E L(H) , by another
.
t
application of the expansion V,(6.11). Accordingly the integral
V,(6.12) with Vt indeed converges in L(H), by V,thm.6.8 and we
get SM'-s E Ct as stated in (4.4).
,
Finally, for the second part of (4.4), assume that

(l5')(P.) holds. We may assume (4.5) true now, which only depends
on the first (4.4). Thus, in order to show that the right hand
side of V,(6.11) ,
t
(for -s), is in E , we must prove that Wt_
A-t LjAN+2jAt E E, as j>l, and A-tS At E E , where we know that
M,-s
LjE 2N+j,0 From cor.3.4 we get LjAN+2j E E. Again we use

.
V,(6.11) on Lj , and for -t instead of s, to express W. by
Lj+kAN+2j+2k
and another remainder. The first terms are in E as ,
posted. The remainder is an L(H)-convergent integral with inte-

grand in E . Thus WtE E , and L.AM+2!E Et . The A-powers may be
taken into the remainder S and will give an integral V,(6.12)
M+N+l M,-s,
with ALM+1A replaced by V. above. We just showed that VtE E
(we have Vt=Wt, for j=M+1). Thus also Sm s E Et , and we get
(4.4) for A = LAM . Similarly for the other generators B = ANL.
The same follows trivially for AEC.0, or for A finitely generated
N
by LAN, A L , with LEAN and the proof is complete.
,
In thm.4.3, below, we now summarize the consequences of thm.

4.2. These facts will be pursued in closer detail in X,4, where

it will be seen that the Fredholm theory of an Hs-comparison
algebra is similar, or even the same, as that in the corresponding
H-comparison algebra. The assumptions are the same as for thm.4.2.
Theorem 4.3. The quotient algebras Cs/E5 , s e I , all are isome-

trically isomorphic, and also the quotient algebras Es/K(H5), SEEM,
all are isometrically isomorphic. In each case a *-isomorphism
Cs/Es } Ct/Et and a *-isomorphism Es/K(H5) - Et/K(Ht) both are
induced by the natural isometry A - ArAA-r , r=t-s .
Accordingly, the maximal ideal spaces Ms and ffit of Cs/Es and

Ct/Et are homeomorphic under the associate dual map of the above
isometric isomorphism, and, if E/K(H) is a function algebra
CU,K(h)), as in the examples of VII,1,2, and 4, then also
(4.8) Es/K(Hs) = C(Es3K(h)) ,
where all spaces Is are homeomorphic to E = E0 .
Moreover, under condition (l5')(P.), if the homeomorphisms

used to identify all the spaces ffiwith M _ MO then the opera- ,
tors in the algebra finitely generated by CO and LAN, ANL , for

LEQN, N=0,1,..., have their Cs-symbol independent of s
The proof of theorem 4.3 is evident. We will continue this
discussion in X,4. Note that a dependence of the Cs-symbol of
a (finitely generated) operator A on s is possible, if only
(15)(P00) , not (15') holds. We shall not discuss such problems
here.
Finally let us shortly look at the (Frechet-)algebra C_
already mentioned. Note that the algebra C.0 C 0(0) also has a
closure C, in the natural topology of 0(0) i.e., the locally ,
convex topology of all operator norms
(4.9) IlAlls = sup{llAulls: Hull s<1) , s E IR .
By the Calderon interpolation theorem this is a Frechet topology;

the class {IIAllj:jEZ} defines the same topology (cf.I, thm.6.3) .
If we assume (15')(P,) then a C.-symbol a:C.+C(ffi), with the common

space M=ffis of thm.4.3 still may be defined as the restriction to
C. of the Cs-symbol (which proves independent of s) (cf.X,5).
In the special case of a certain comparison algebra for
Laplace-fin
it was shown by E.Schrohe and the author [CSch], that
the symbol of this Frechet-algebra still maps onto C(M). Moreover,

the same is true for the (unique) Frechet comparison algebra of
a compact manifold Q (cf. Schrohe [SchI]). (In both cases we have
Es=K(Hs), so that the symbol of the second kind is trivial.)
For a more detailed discussion of the abstract elements of
the Frechet comparison algebra Cam, and the Hs-comparison algebras
Cs cf. X,4.
5. Local properties of the Sobolev spaces Hs.
In sec's 5 and 6 we will look somewhat more closely at the

spaces Hs of sec.l, which, so far, were little more than the
domains of the powers of the self-adjoint operator A=H-1/2. Here
we will discuss the proofs of VI,thm.1.6, and thm. 1.2, thm.1.3,
concerning local properties which were postponed so far. In sec.6
we will focus on equivalence of Ilufk for an integer k>O, with a
,
weighted L2-norm of the 'k-jet' of u

First let the assumptions of thm.1.2 hold: We are given
(5.1) (O,dp,H) , {O,dv,K)
on the same manifold 0, coinciding outside some compact set. Only

the minimal sets A# and D# are needed in this discussion. We assu-
m m
me condition (s) for the first triple. The sets Am , Dm , with the
first triple, then qualify for all results of III,3,V,3, V,6 and
will give us a minimal algebra of expressions- explicitly known
as the set of all C'(O)-differential expressions of compact sup-
port, with LN consisting of all such expressions of order <N .
Although the spaces 0(m), for the first triple (5.1), are
established, we will not use them, or the notation of sec.4,
until thm.1.2 is established, for reason of clarity. However,
once we have thm.1.2, we will have a unique H. and Hs , and
it follows that all classes 0(m) for H and K coincide. Thus the
notation of sec.4 then will be clear and convenient again.
Note that y = {dp/dv}1/2 defines a non-vanishing C'(O)-func-
tion, with y=l outside a compact set. We have dp=y2dv , and a
change of dependent variable u=yv will take the second triple
onto the form
(5.2) {O,dp,K'} ,
K' = y-1Ky ,
which has the same measure and inner product as the first (5.1).
IX.5. Local properties 273
Proposition 5.1. The second triple also satisfies condition (s).

Proof. Clearly (s) holds for {S2,dv,K) if and only if it holds for
{ft,du,K'} . Thus it is sufficient to prove (s) for K` , i.e. one
may assume that dp=dv
Note that
1 + LJA2J E J
KPH-3
(5.3) Pj = _ (H] + (K3-H3))H 3 =
A=H-1/2
with , where Lj=K!-Hi has compact support, hence L3EL2j
This operator is Fredholm, since we have
(5.4) aP = ((k01(x)EjEl)/h31(x)EiEl))3
J
by VI,thm.2.2, since C=J has E=K(H) , while its symbol space ffi is
the 1-point compactification of the cosphere bundle. Moreover,
since aP is always positive, the Fredholm index must be zero,
since the symbol can be connected by a continuous non-vanishing

family of symbols to the identity function (cf. X,thm.1.9(A)) .
(Notice the symbol is =1 for all x ouside a compact set.) Also,
(5.5) (Hiu,P.u) =
follows first for u E HjC , then also for all u E H , by conti-

nuous extension. Thus Pju=O implies u=0. Since PI is Fredholm and
has index zero, it must be invertible.
Now, to show that the minimal operator Ki is essentially
self-adjoint, assume that fEH, (f,KJu)=O for all uEC (Q). We may
write this as 0 = (f,P.H3u) _ (Pi f,Hju) , u E CO . But since A
0
is essentially self-adjoint, we have in H3 dense in H , so that
P. f = 0 . Since P. is invertible in L(H), we conclude that f=0.
Therefore in K3 is dense in H Since K3 > 1 we conclude that K3
.
is essentially self-adjoint, by I, thm.3.l, q.e.d.

Denoting the Sobolev norms of the triples (5.1) by 11.0
ss H
and I1.11 s K , for a moment, we note that
IIy-1K s/2ufl
(5.6) Iluus H = IIH-s/2uII , Iluns = , as uEC'(c)
K
where 11.1 is the norm of L2(S2,dp) . For a proof of thm.1.2 it is

sufficient to show that both norms (5.6) are equivalent, for
every s E I .
Proposition 5.2. For a given sEI the two norms (5.6) are equiva-
Ks/2)'3d
lent if and only if the operator (H-s/2 and its inverse are
bounded (in L(H)).
Indeed, C. is dense in both Sobolev spaces, by lemma 1.1,
(since we have (s) for H and K), we get equivalence if and only if
(5.7) ls H
cl'ull < IIu11
s ,K
< c2 11u11for
s ,H
all u E C
Setting H s/2u=v , and using that H-s/2C0 is dense in H we con-
clude that (5.7) holds if and only if
(5.8) c111v11 < Ily-1(H s/2Hs/2)* vll < c211v11 , for all v E H ,
which means that (H s/2 K s/2 ) *' is bounded and has a bounded
inverse since y is bounded and bounded below, and by symmetry.
Next we apply I,lemma 5.1 :The operators H and K+ are self-
adjoint with respect to the same inner product, and we have
K`1H-! E J C L(H), while the inverse HjK-j exists in J C L(H) as
Hs/2K.-s/2
well. Hence we get K`S/2Hs/2 bounded for s>0, and
bounded as well. Taking the adjoint of these operators as well,
1 s/2 -s/2
we conclude that (W = Y (K YH ) E L(H),
and that these operators have bounded inverses.
Finally we apply V, thm.3.8 for A=y AS=K-s/2, noting that ,
CK-S/2
the remainder is of the form , with C E K(H) Since Y-1E .
L(H), we get y-1Ks/2y = Ks/2 + CKs/2 , C E K(H) = E . Or,
(1+C)Ks/2H-s/2=Y1Ks/2YH-s/2EL(H)5 Hs/2K-s/2(1+C')
(5.9) E L(H),
with compact operators C,C' . Clearly l+C , 1+C' are Fredholm,

PSKs/2H-s/2 Hs/2K-s/2Qs
hence and , with orthogonal projections
PS, QS onto spaces of finite codimension are bounded. Set JS =
Hs/2-s/2
K .
This operator and its inverse are preclosed. If
is not bounded, there exists a sequence u.EH with u.-O, IIJsujll=1.
We must have PSJsuj-0 , hence vj=(l-PS)Jsu. is bounded in a fini-
te dimensional space. It may be assumed convergent,by compactness
of closed balls in im(l-PS) . Thus uj+0 Jsu.->v , Ilvll=l . Since
Js is preclosed, this gives a contradiction. Similarly for JSl
We also have made progress in the general proof of VI,thm.
1.6, insofar as the boundedness of the operators Vt Vt-1 was ,
shown for all tel. To confirm the statements about the cosets Ut
Vt+K(H), we must extend VI,prop.l.8. As mentioned above, the nota-
tion of sec.4 now is in order: all operators in 0(m) are consi-

dered as operators H_-H_ , but also are identified with their
continuous extensions to any Hs, as maps Hs;Hs-m. Specifically we
have H S, Ks E 0(2s) .
Proposition 5.3. For 0 < s,t,a,T E I we have
(5.10) Ho[Hs,K-t]HT E K(H) , whenever o+T < s+t+1/2
Proof. With R(X)=(H+a)-1, S(p)=(K+11)-1 as in the proof of VI,prop.

1.8 we set up a formula like VI,(1.17) again, but now we will use
identity VI, (6.8) to express H-s and K-t by integrals over
higher powers of the resolvents. Instead of the commutator formula
VI,(1.18) we need formula (5.11), below.
Proposition 5.4. Under cdn.(s) we have
(5.11) [Rm(A),Sl(P)] = Rm(a)S1(U)L(X,P)S1(P)Rm(a)
for all X,p > 0 , and all l,m=1,2,..., where L(A,p)=L1 m(A,p)
denotes the polynomial (with rPq=2(m+l-p-q)+3 )
=lp=llq=lLpgaP-l11 q-1,
(5.12) L(X,p) =[(H+X)m,(K+11)1] LPgELr ,0
Pq
where we again use the classes A!=Am , D! = Dm
The proof is a calculation, noting that R(a) S(p) E 0(-2). ,
One has Lpq = cpq[Hm-p+lKl-q+l] , using the binomial theorem, and
the fact that [HP,Kq]=0, as p=0 or q=0. Also, a straight-forward

extension of VI,prop.1.9 is required. We indeed get LpgE Lr ,0.
P9
Now we prove prop.5.3. Using VI,(1.6) for H-s and K t write
(5.13) Ho[H-s,K-t]HT = lP=1Fq=l JdA JdPIpq(a,u)

0 0
with
Ipq=cpgm+p-s-211 1+q-t-2 HoRm(a)S1(11)LpgSl(P)Rm(A)HT,

(5.14) cpoER.
For given s,t,a,T >0 , if 1 and m are chosen sufficiently large,

0((Xp)-1-e) ,
then the integrand will be norm continuous and e>0,
so that the improper Riemann integrals exist in L(H) . Also we get
Ipq(X,P) E K(H) so that (5.13) is a compact operator.
,
Indeed, the existence of the integral at a=0 or P=0 just

requires s<m , t<l , since p,q>l. Write
AaRmHoK6-aS1PY)(Ko_6LpgKT-6)(UYS1K6-THTRmaa)
Ipq=cpq(Au)-l-6(
(5.15)
=cpgJpq , Jpq=A.B.C ,
with the three parentheses called A, B, C , where
(5.16) a=(m+p-s-l+e)/2 , y=(l+q-t-l+e)/2
while $ must be chosen such that B is bounded. From m>s , 1>t

we also get 0<a<m , 0<6<1 , for small s An inequality like
.
V,(3.9) (cf. also VIII,(2.22)) implies

NAufl = O(fHa+a-mKS-aSluYull) = O(IKa+6-mS11Yull) = 0(IIKa+S+Y-m-lull),
so that fAll=O(1) whenever O aa+$+y-m-1=6-(m+l-p-q_+2-2e+s+t)/2, or,
(5.17) 2S < m+l-p-q+2-2e+s+t = rpq/2 + 1/2 - 2e +s+t ,
with the order rpq of the expression Lpq . For the operator C we
obtain exactly the same estimate, (i.e., CE L(H) if (5.17) holds.
On the other hand we get BEK(H) if 2s-a-T>rpq/2. Accordingly,
6 satisfying both estimates can be found if rpq/2+a+T<rpq/2 -2e
+s+t +1/2 In other words, we must require (5.10). Then it also
.
follows that the integrand is norm continuous, q.e.d.

The proof of VI,thm.l.6 now only requires the use of prop.
5.3, to establish the group property and adjoint invariance of the
cosets Ut , since the boundedness of Vt was shown. In VI,prop.l.7
we have demonstrated how to do this if t is small. We now may
use prop.5.3 instead of VI,prop.l.8 to do this for general t.
Details are left to the reader.
Next, in this section we focus on the proof of thm.1.3.
We will use all the notations introduced in VIII,3. Specifically
Aj=Hj-1/2
let . Since this is a matter of Sobolev spaces which
require cdn.(s), with our definition, we will generally assume
that both triples {Oj,duj,H} satisfy cdn.(s). On the other hand,
the prop's below clearly all are local. Thus cdn.(s) may be avoi-
ded, using a method similar to the one in early VI,l.
We need a few preparations.
Proposition 5.6. Under the assumptions of thm.1.3 let ¢EC'(U) ,
4=0 near 3U , =l outside some neighbourhood N of 3U . Then
(5.18) es for all u E CE(O) .

Proof. It is sufficient to prove (5.18). Indeed we may write

Aj-s(W = Aj-s+2m(lHm.u) = AjtPv, with H = H1IU = H21U,
t=-s+2m , where j satisfies the same assumptions as , and also
V¢ _ , and where the integer m=1,2,..., may be choosen such that
t>0 Then (5.18) for all v, i and t instead of u,
. and -s will
imply our statement for the general case.
Let XEC_(U) , 0<X<l , X=0 near 8U , Xc= , and let w=l-X
We get
(5.19) HwAlt$ull = J1+J2 ,
where we apply V,prop.6.7 to the term J2 . (Here we work with the

operator A = SAO involving the H-compatible expression of order
zero.) The terms J. in (5.19) will be treated separately.
Proposition.5.7. With , w, satisfying the above assumptions,

there exists a constant y = y such that
r,t,4,w
(5.20) u E CD(S21)
where rEl is arbitrary.

Proof. It is enough to prove (5.20) valid for every sufficiently
large integer r=2N. Thus let N>t>0 . Again, we may replace ¢u at
right of (5.20) by u For if that is true for all such ,w then
.
we use it for 4- instead of with - = 1 and replace u by ¢u in

the inequality achieved. Also we may replace u by H2Nu for the ,
equivalent relation (we used cdn.(s))
(5.21) yuull2 , u E C'(S22)
Again in (5.21) we may replace u by u with some function i with

V = 1 , and then get (5.22), below, equivalent to (5.20).
(5.22) yllull2 , u E C1(S22)
But this is an immediate consequence of V,(6.9) and (6.10).

For we may assume i
= 0, and also, ((ad H1)1w) = 0 There- .
fore, if the expansion V,(6.9) is applied to A=w, with N=O, s=t,

and the adjoint is right-multiplied by , then the remainder is
the only non-vanishing term of the expansion. By choosing M large
the reminder term can be made L2-bounded, even compensating the
H1N yQ,Pu112 q.e.d.
power . Hence we get IIwA1H1N1ull1 < ,
(Note that V,(6.9)-(6.10) are valid since w is H-compatible.)

IX.S. Local properties 278
Proposition 5.8. For every p,r,t>O we have
XA2X)Ak-r
(5.23) E K(H) , ],k=1,2
Proof. This is a matter of the following identity, where we write

Ry(a)=(Ai2-a)-1. We have, with Mi=(ad Hk)jX , for N,N'=0,1,...,
XA12R1(a)X - XA22R2(A)X =
(5.24)
(-l)NXN+N'AI. 2NRiN(a)(MNAl2Rl()MNt- MNA22R2()MN,)A12N'R1Nt(M)
For N=1, N'=0 relation (5.24) coincides with VII,(3.11). For more
general N,N' it follows by induction, using that
(1-AHD)-1 = Tj , so that AMN+l = MN(1-XH1) -
MNT1-1 - TJ-1MN , i.e.,
2
(5.25) AAi 2Ri(a)MN+lA12R1(A) = Ai R (a)MN - MNA12R1(a).
i
Relation (5.25) serves for an induction proof of (5.24).

On the other hand, (5.23) follows from (5.24) with the same
resolvent integral procedure used over and over, q.e.d.
Finally, to complete the proof of prop.5.5, we write
(5.26) XA24u + (XAt1 x - XA2t
where the first term is and the second term as well,

due to (5.23), q.e.d.
We now may approach the proof of thm.1.3. The second state-
ment is easily reduced to the first. If DU is smooth, then the
diffeomorphism identifying the two subdomains U. of Qj to con- ,
stitute the open subdomain U may be extended beyond 3U such that

01 and Q2 contain the commom subdomain W , with compact boundary,
while UL U C W . Perhaps then the triples are different over W\U.
However, we then may use thm.l.2 to remedy this, at least in a
subdomain of W , still satisfying the conditions. Then (1.6)
holds for U as a subset of W.
,
To prove (1.6), let V C U be an open subset, with VclosC U.

Since 3U is compact, there is a compact neighbourhood N of 3U ,
relative to UUBU such that U\NnV=O Then we may select a func-

.
tion meeting the conditions of prop.5.7, with q=l outside N ,

so particularly =l on a neighbourhood of Vclos .

By V, prop.6.7,
applied to the 0-order operator E C-(St with [H,q,] E Dm , we
find that is a bounded operator of L(H1.) , hence qE L(Hs).
Now let u E HS. There exists a sequence ukE C0(S21) such that
uk-r u in Hs .We have u E Hs as well, and ctuk4 u in HS. However,
we get uIV = since u may be interpreted as the product
of the distribution u and the C"-function , and since =l near
V . Also, the distribution u has support in U since supp C U
Accordingly, we may interpret $u as a distribution on S22 as well,
and also we get cu k- u in D' (02) , i.e., in weak convergence of
distributions over S22 Also, by (5.18), the sequence uk is
a Cauchy sequence in H2 , hence converges in H2 to some vEH22
Since this also implies distribution convergence (on St 2), we find
that v=4u E H2 This shows that every restriction ujV of some
.
s
u E H1 may also be obtained as a restriction to V of v=mu E H2
Since the assumptions are symmetric, this implies the statement
of thm.1.3.
6. Sobolev norms of integral order.
In this section, we look at Sobolev norms Ilullk for the case

of an integer k>0 Clearly, with the inner product (.,.) of H ,
.
Iluuk2 = (A-ku,A-ku) = (u,Hku) , u E C0'(S2)
expressing Ilullk in terms of derivatives of u. In tin, where H has

constant coefficients, one may integrate by parts for
,
(6.2) Ilull 2 = llu(a)02
k a
I
with positive (multinomial coefficients) ya (cf.[C1],III,(2.11)).

In order to derive a similar formula under more general
conditions, let us assume a triple {St,dc,q-0} in Sturm-Liouville
normal form, i.e., du = dS = Fdx , h= (det((hjk)))-1 . (From

III,prop.6.1 we know that this is no loss of generality, but it
should be kept in mind that cdn (q3), below, refers to this form.)
We assume {S2^,dp^,H^}={P,du,H}, and q>l on 0 . Moreover, we will
use the Riemannian co-variant derivatives induced by the metric
tensor hjk (cf.app.B.).
Under these general assumptions we impose the following.
Condition (q2): The Riemannian curvature tensor R (cf.(B.10))

IX.6. Integral order 280
is bounded on n , and its co-variant derivatives of

all orders also are bounded over 0 .
Condition (q3) : The potential q satisfies the estimates 4kq =

k
O(ql+k/2) , k=1,2..... x E S2, where V q denotes ,
the k-th order co-variant derivative (tensor) of q.

Our problem formally suggests introduction of the norms
{IqN/2u12+Iq(N-1)/2Vu
(u)N =
(6.3)
N = I92( u) 2du = IIgN/2u112 + .. . + IIVNuII2 ,
with the formal covariant tensor 4k of the k-th covariant deri-

vative, and with Itl IIt" for t=(t., ), defined by
,
11...ik
(6.4) ItI2= h11]1...h1kOk t.l IItu2=

l" ItI2du
1 ..1 3
Theorem 6.1. Under cdn.s (q2), (q3) the Sobolev norm 11-11N , on
the dense space C' C HN, is equivalent to the norm <.>N. That is,
(6.5) clN < IIuIIN < c2N , for all u E C0_(S2)
where c1 , c2 are positive constants independent of u .
Proof. The second estimate (6.5) essentially is formal: For N=2m,

an even number, one writes (6.1) in the form Hull N2= IIHmu112.
In order to simplify the formalism introduce, for a moment, the
(n+l)x(n+l)-matrix ((gvu)) 9jk=hJk , g00=q ,
n
goj=gj0=0
and its inverse
, 0 .((gvu)) Use summation conven- .
tion from 0 to n (over greek indices only, one up, one down). Also
define a0k=-hjk, j,k90, a00=q, a03=a00=0, j90 , and, with formal
tensor notation, (greek indices from 0 to n), [u]k the 'tensor'
with components uv = VV ...V u, with 70u=u V u=ulv ,
., v V '
1 k 1 k
as V # 0 .
Recall that the metric tensor hjk acts like a constant; it
has all its covariant derivatives zero. Suppose, for a moment,
that also the co-variant derivatives of q vanish (i.e.,q=const.).
Then we may write
u1vl umvm
( 6 . 6 ) H ...a u
u1v1...umvm
This is just the contraction of the tensors (a)x(a)x...x(a)

with m factors (a)= (avu) , and [u) N . Thus it is a matter of

Schwarz' inequality that
(6.7) IHmul2<(guvgoTauoavT
)m(givi...guNvNuu
u uv
.. V
1.. uN 1 N
The first term at right is a constant. Integrating over 0 we get

the second estimate (6.5). The second term at right of (6.7) may
be written explicitly as
(1)gION-lul2+
IVNul2+ ... + (I)gJIVN-7uI2 +...+ gNIuI2
(6.8)
< (u) N
However, in general the derivatives of q will not vanish.

Thus (6.7) must be replaced by a more general representation,
containing additional terms reflecting the interchange of q with
co-variant differentiation. The point then is, that the estimates
of cdn.(g3) give
v ., vN
(6.9) Hmu = a i u vl...vN
(avi...vN)
where the 'tensor' satisfies the same estimates as
(a)x(a)x...x(a), above, so that (6.7) still follows. (If a
derivative 'lands' on q instead of u then we will get one more ,
power , but one less derivative on u, precisely as required).

This proves the second relation (6.5) for an even N The , .
proof for an odd N is similar. One will set N=2m+l and work with
Hmu as before. Then, however, IIuIIN2 = (Hmu,Hmu)1 so that one ,
more derivative must be worked under the coefficients of Hm

There is no new point, however. Details are left to the reader.
For the first inequlity (6.5) we require the prop.,below.
Proposition 6.2. Under cdn's (q2) , (q3) we have
(6.10) H(uIi
.i ) _ (Hu)Ii1 i
+ O((u)N
1 N N
Proof. This is a matter of pulling out the covariant derivatives

from the expression H . Clearly the interchange of A = hjkV.Vk
with V il.. 0. = 0"' will give additional terms involving the curva-
iN
ture tensor R , and its derivatives up to order N-1 . We get
j j
IN p
(6.11) D4' - 0"A _ p-1 pti 1.. iN 1 u Ijl ...j p
with bounded tensors pt in view of (q2)

, .
On the other hand, Leibniz' formula gives
(6.12) 4-(qu) = q0-u + jy(V'q)(4"u) ,
where V', V" are covariant differentiations for complementary sub-

sets (i 31. ..i ), (ik ..ik ) of (i1...iN) with jl...jr and kl..
3r 1 $
..ks in increasing order each, and where the sum is taken over
all such splittings of (i1...iN) , with certain individual con-
stants y for each term. Again (q3) must be used; it will give
qV-u = V-(qu) + O((u)N-1) From (6.12) we get the corresponding
.
for Au , where it must be used that q>l Q.E.D. .
For the first inequality (6.5) we introduce <u,v>N, the

form obtained by polarizing the hermitian form (6.8), integrated
over 0 (i.e., N is equivalent to (<u,u>N)1/2 ; only the posi-
tive binomial coefficients distinguish (6.3) and (6.8)).
An integration by parts gives
<u,v> = gulvl...guNVN u
ul...11N
((v ul...uN
(6.13) N+1
= <u,Hv>N + 0(N<v>N) = <u,Hv> + O(IIuONIIvIIN)
by induction. Iterating (6.13) , we get
<u,u>N+l = <Hu,Hu>N-1+ O(IIuUN2 + pull N-lIIHuIIN-1)

(6.14)
= <Hu,Hu>N-l+ O(IIulIN+l2) .
Continuing on we either get <u,u>N+l = IIHmUII2 + O(OutIN+l2) if ,
N+1=2m is even. Or else, the first term at right takes the form
(Hmu,Hmu)1 if N+l = 2m+l is odd. In either case (6.5)
,
follows, q.e.d.
Note that thm.6.1 is applicable to both of our typical
cases, that of a regular boundary, with a singular q , and the
case of a complete manifold with q=1 or q slowly increasing.
,
For example, the expression H = x-2-d2/dx2 in the interval ,
(0,°) has a q satisfying (q3) , as easily checked.

IX.7. Examples 283
7. Examples for higher order theory. The secondary symbol.
In this section we will consider examples of classes A!

D! satisfying the general conditions imposed. Clearly we must
insist on cdn.(s), for each example considered. As we know, the
conditions Cl.) in particular give us the possibility of looking
at operators A=LAN as members of a comparison algebra. It is
easily checked that many of the classes A# , D# singled out so
far do not give an algebra or Lie-algebra. For example, regarding
A#, Dc AC,
, DC of V,4 and VII,4 one finds that Ac, AC are alge-
bras, but DC , DC are not in general Lie-algebras.
On the other hand, some of the classes used for In and Jk
have this property. For example, let us look at IC1
and define A! = A , with the class A defined there. (In detail, A
= A! is the class of all bounded C(I)-functions having their
partial derivatives of all orders vanishing at infinity. ) Let D!
denote the class of all folpde's with coefficients in A , with
the standard coordinates of In, of course. Then it is easily seen
that this pair A!, D! has all the properties (lj) In particular
.
D! is a Lie-algebra. Also, the universal algebra L_ is completely

known, in this case. LN is just the algebra of all differential
expressions of order < N, with coefficients in A , and L_=ULN
On a general manifold Q we always have complete control of
the minimal classes Am , Dm , and their enveloping algebra, which
consists of all expressions with compact support. The generated
algebra will be the minimal algebra JO or, in L(Hs)
, the cor- ,
responding minimal algebra Js C L(Ns) Now, thm.6.1 will imply

.
a simple generalization of V,(1.8) to higher order expressions,

allowing us a better control of operators A=LAN in our algebras.
Lemma 7.1 Assume H in SL-normal form, with (s),(q 2)1 (q,). Let an
<N-th order expression L over 0 be written in the form
(7.1) L = NaVN + N-laVN-1 + ... + 1aV + 0a ,
with (locally Lw-) symmetric contravariant tensor fields
ja = (jai1 J) defined over 0 , where jaVJ denotes the j-fold
contraction with the formal tensor V : V. ...V.

J J 11 1j
For such expressions let a norm 'ILIIH be defined by
IX.7. Examples 284
IILIIH = JN Ohq-(N-0)/2j.ajiI
(7.2)
j= 3 L '(Q)
with the local norm I.aI(x) of the tensor a. There exists a con-
3
stant c, independent of u and the choice of tensors ja, such that
(7.3) IILANuUI < cUILIIHPull , for all u E A-NC (O) C H
Proof. Clearly (7.3) is equivalent to
(7.4) IILvh < cIILIIHIIvON , for all v E C0'(c2)
On the other hand (7.4) simply is a matter of Schwarz' inequality

in view of (6.5), q.e.d.
Theorem 7.2. Assume H in SL-normal form with (s),(g2),(g3). Then

the minimal comparison algebra J0 contains every operator A=LAN,
with L of the form (7.1), with continuous tensors ja, satisfying
(N-J')/2
(7.5) 1 j al = o (q s j=0 ,1,...,N
0
Moreover, let R. = URN, where RN denotes the class of all expres-

sions of the form (7.1) with C-(O)-coefficients satisfying
, j=0,1,...,N , k=0,1,...
o2((/)N-J)
(7.6) Vk(ja) = .
Then the collection R.= N=ORN is an adjoint invariant graded alge-

bra satisfying (pl),(p2), and (3.10) for the minimal algebra J0
Also, R. is within reach and of principal symbol type, for J0.
Furthermore we have A=LANE O(0),and AAE K(H5), for LERN, sEi.
Proof. We know that every A corresponding to Co-tensors a with
compact supports is in JO (cf. VI,prop.3.1), since such L has com-
pact support. In fact, such A even is contained in the finitely
generated algebra C0 corresponding to the minimal classes, hence
AEJ5 sEI
, . For a general continuous a satisfying (7.5) there
exists an approximation under the norm IILIIH
] by compactly supported
tensors, and the corresponding sequence of operators converges to
A in L(H), so that A E J0. Next a calculation confirms that R. is
invariant under formal adjoints. In view of V,cor.6.10 it is clear
that the remaining statements of thm.7.2 depend on the verifi-
cation of cdn.(p1) for the class R. defined. (Note that (p2) is
a consequence of the H-boundedness of A, already shown.)
In other words, we must verify that [H,L] E RN+l for LERN.
This is shown by a calculation which we shortly indicate: Let
IX.7. Examples 285
L. = a.V , with a j-tensor a. Using all rules of covariant dif-

ferentiation we get
(7.7) IH,LjI = IH,al.V3 + a.IH,oi1 ,
where the second commutator already was considered in prop.6.2.

Using (qj) , j=2,3 , one finds that the same arguments give
a.IH,VII E RN+l, using (7.6) for a=ja On the other hand,
.
i ...i
(7.8) [H,a] = IA,a1 = 2(a l 0)Ik Vk + (Aa) ,
with the contravariant derivative 'Ik, (i.e., covariant, raised

with the metric tensor). Again it is evident, looking at the esti-
mates (7.6), that IH,a1.V E
It is clear that R_ = URN of thm.7.2 is within reach of J0
For J0 the classes LN and LN 0
coincide. Every A=LAN, for LERN was
seen to be a limit in norm of a sequence L AN, with compactly sup-
k
ported Lk, i.e., LkE LN Thus ALAN= limk+-6ALkANE K(H)=E, q.e.d.
O.
It is clear that R_ also is within reach (and of principal
symbol type) with respect to any arbitrary comparison algebra
C(A!,D!) with the property
(n4) A!RN , RNA! C RN e D!RN e RND! C RN 5
(which implies (3.10)), since J0 CC and K(H) C E are always true.

This fact proves useful for an A! D! for which (15)(R_) or even
,
(15') can be confirmed, as in ex.7.3, below.

On the other hand, while thm.7.2 did not state that R. is
within general reach (or princial symbol type) for the minimal
classes, it did assert LANEO(O) , ALAN E K(Hs), for LERN .
Therefore, while thm.4.2 and thm.4.3 do not apply in their present

form, they will apply if the operators LAN , ANL , for all LERN
are added to the generators (1.10) of Cs, the other conditions
granted, as the reader is invited to check.
Example 7.3. For a triple in SL-normal form satisfying (qj),
j=2,3, consider the following classes A! , D! .
A! is the class of all bounded C°°(O)-functions with all

(Riemannian) co-variant derivatives o0(1) .
D! is the class of all C-(c2)-folpde's D=b4 + p with b=0(1)
and p= and
(7.9) Okb = o0(l) , 4kp =
A calculation confirms that, for these classes, we get (°0*)

IX.7. Examples 286
above, as well as cdn's (11), (12), and also
(7.10) [H,a] E 1 , [H,D] E R2 , for all aEA! , D(=-D!
In other words, A! and D! also satisfy (151)(R_) . Accordingly,

cor.3.4 applies : All operators in Q.=R.+L. are within reach of
C(A!,D!) C L(H) , We also obtain (slightly enlarged) comparison
algebras in L(Hs) and in 0(0), containing LAN and AN, for LEAN'
,
Also all expressions L E QN RN + LN give operators

0 = 0
A=LAN in K(Hs), sER.
We observe that the class D! used here is less restricted
in the first derivatives of the coefficients, than the class D#
of V,4, since E<l is required in V,(3.16). However, we have no
curvature restriction in sec.5, instead. Also the first derivati-
ves of functions in A! are stronger restricted as those of A#
Rn
Following the example of let us look at the above pair
A!, D!, in the special case of the Laplace triple of a complete
Riemannian manifold 0. The corresponding comparison algebra was
discussed by the author in [CS],IV,11, under cdn.(14) (which ,
poses a rather severe restriction), and, almost in the present

generality, by McOwen in his thesis [M0]. In fact, our discussion
here uses some ideas of McOwen in [M0], in a generalized form.
Example 7.4. We are given the Laplace triple {l,ds,l-A} of a

complete Riemannian manifold I , with A!, D! as in example 7.3.
This triple is in SL-normal form, with q=l satisfying (q3).
cdn.(s) follows from IV,thm.l.8. We have the additional require-
ment (q2), i.e., the Riemann curvature tensor and all its deriva-
tives are bounded. A! and D! consist of all bounded functions
(folpdes) with covariant derivatives of coefficients 0 at - .
Notice that A! C AO , D! C D# , with the classes of VII,4.

Recall that the algebra C =C(A!,D!) has compact commutators, hence
gives the simplest ideal chain. Also, cdn's (11), (12) hold
trivially, for A!, D! (we assume Q^ = R , of course).
We still observe, that although the algebras C and Cs of the
present example are well defined on a general space Q with bounded
curvature tensor (having bounded covariant derivatives), we get
(7.11) limx,.,,,b1Riljk = limx+.(b1Ijk-biIkj) 0,
for every tensor (bi) occuring as principal part of some DC-D!

as a consequence of app.B. This shows that the curvature is
IX.7. Examples 287
restricted at -, if folpdes DED! exist which are not in R1 For

.
example, on a manifold 0 having a chart {xE In lxl>l} near °°

:
if there exist n vectorfields with linearly independent limits at

w , occurring as principal parts of folpde's in D!, then we must
have limx.,R = 0 , R being the curvature tensor.
Note that even on In (or Mjk) the comparison algebra
C(A!,D!) is inconveniently large, for some discussions. Its
symbol space HI exhibits some features of the Stone-Cech compacti-
fication. This is why we often prefer to work with the smaller
algebra S of [CHe1], corresponding to the functions and folpde's
having directional limits at infinity. Similarly, working on a
noncompact Riemannian space one might tend to restrict the geo-
metrical structure of the space at infinity, and in addition the
generating classes.
From example 7.4 we conclude that compact commutator
algebras which are 'free at infinity', in the sense that their
symbol space contains unrestricted cotangent spaces also over
only can be achieved on spaces Q which are (asymptotically) flat
at infinity. On the other hand the few examples of non-compact
commutator algebras with a 2-link ideal chain we have discussed
in VIII,l, 2, 4, and in [C1],V, may be offered for examination,
in view of the following points.
(a) The longer ideal chain may reflect a more complicated
structure of the compactification M # at °°, as in the first exam-
A
ple of VIII,l. Or, (b) it may result from a nonvanishing curvature
tensor at °°, as in VIII,2. Or, (c) it may reflect the necessity of
boundary conditions (at °°) as in [C1],V .
In every case examined, existence of a finite set of genera-

tors of D! (mod A!) gives a simpler theory.
In case of example 7.4 - Laplace comparison algebra of the
bounded functions and folpdes with covariant derivatives vani-
shing at °°, on a complete Riemannian manifold, we now want to
supply a criterion, allowing the construction of the secondary
A = LAN , for LELN
Above we have discussed validity of (s),(lI),(l2),(15)(R.)
It is found that (m1'), (m4), (m5), (m7) hold trivially, but we
now require the separation condition (m3). Then VII,thm.3.6 is
applicable and gives the secondary symbol of DA, for DEED!. On the
other hand, for L E LN , the operator A = LAN first must be writ-
IX.7. Examples 288
as a sum of products of such DA , (mod E), before its secondary

symbol can be obtained. For the principal symbol we know, of
course, that only the principal part of L matters, and that o A1W
equals the principal part, restricted to S S2
Theorem 7.5. Let all above assumptions be satisfied.
(a) The operator A = LAN , for a product
(7.12) L = D1D2..... DN e DjE D! , j=1,...,N
is contained in C , and its symbol is given as the restric-

tion to I of the product
(7.13) TD
ATD2 A " " TD A
1 N
(b) Let x0E 2MA! have a neighbourhood N in M with the
A!
following property: There exists a cut-off function X with the
properties of cdn.(m3)x such that, in local coordinates any-
where at NnD we have

11...1. 11...1.
(7.14) L = IN Oa 0V1 ...41 j E C 'M
j=0 ja
1 0
where (with rb , k depending on j, of course)
(7.15) for all j=O,...,N,
with finite sums, and with bounded tensors (kb) of one variable
index, defined over all S2 such that ckE A! ,r(rb13 1) E D!
x
Then we have A=XLANE C. Moreover, the secondary symbol
of A over x0 is given explicitly as the restriction to the (pos-
sibly void) set t-1(x0) c &*S2 of the continuous extension of the
function (7.16) (which is contained in CB(T 0))
(7.16) ...C1 )(hkm(x)ekEm+1)N/2

1 0
(c) For any LELN,O, and, more generally, for LE RN , we get

A=LANE C., and the secondary symbol of A is zero. Moreover, we
have A E K(Hs), for all sE 1, whenever L E RN is of order <N , so
that its entire symbol vanishes.
Proof. Assertion (a) is trivial for N=l and follows by induction:
Write
(7.17) D1.... DNDAN+1=(Di.... DNAN)((DA) + (A-N(DA)AN-(DA))) ,

IX.7. Examples 289
where the first term's symbol can be written in the form (7.12),
by induction, while the last term is compact, by cor.3.4.
The key to (b) is found in (c): Any commutator [D,F] has
coefficients vanishing at infinity, by V,(2.13), where the deriva-
tives may be written as co-variant derivatives. Hence the formal
symbol vanishes there too. Accordingly, using (a), it is clear
that the secondary symbol is zero for any LAN , where LEAN, since
for such L there is at least one commutator in each of the N-fold
products. It thus follows that, for the construction of a local
decomposition of LAN into N-fold products the operators DA may
be treated as if they were commuting. Hence the polynomial decom-
position corresponding to (7.15) gives the proper symbol. Also,
since LAN , for LERN , has vanishing secondary symbol, it is suf-
ficient to calculate modulo RN . This completes the proof.
Remark 7.6. Notice that thm.7.5 will be valid for the Cs-symbol
as well, if the operators LAN ANL
, for LER_ , are added to the
,
generators (1.10). Also we observe that McOwen's result has (m3)

replaced by another condition - that the dimension 6x of VII,2
be constant on connected components of BMA, .
A local version of thm.7.5 is valid too, requiring only

(m3)x at some x0E BMA,
, Also, one of the other representa-
.
tions in VII,2,3,4 may be applicable, if (m3)x fails.

0
Example 7.7. Consider the comparison algebra C of VIII,4 again.

Here we have a Laplace triple on a complete Riemannian manifold
again, but the classes A! and D! of example 7.4 appear as too
restrictive. The algebra C of VIII,4 corresponds to a different
choice of A! and D! , which we discuss in X,5. These new classes
contain functions (folpde's) with covariant derivatives not
tending to zero at infinity (only the t-derivatives must vanish).
On the other hand, there exists a limit function (or folde), at
each end, which is not in general true for example 7.4.
Example 7.8. For the one-dimensional example of Sohrab, discussed

in VII,thm.4.8, we trivially have cdn.(q 2) satisfied, since R=0
and then must impose (q3) , which simplifies to
(7.18) q(k) = dkq/dxk = 0(q1+k/2) , k=1,2,...,
assuming SL-normal form, of course.

IX.7. Examples 290
Covariant derivatives are ordinary derivatives, in this

case. Accordingly, LN consists of all N-th order expressions
aj=O(q(N-j)/2), a,(k)=o(q(N-j)/2 k>0
(7.19) with
Control of the secondary symbol is easy, in this case. We get
(7.20)
for all LELN
Example 7.9. Consider again the triple of V,thm.4.1 (and VII,4),

now considered with cdn's (q2), (q3), and without a sub-extending
triple. We require cdn.(s), hence VII,(4.7) valid for every posi-
tive y , not for yl only, and then use the classes A! , D! of
example 7.3.
Here we only notice that the conditions imposed are consi-
stent. For example, consider the 1-dimensional example
(7.21) sl _ (0,o) , du = dx , H = -d2/dx2 + l+x-3 .
A calculation shows that (q3) holds as well as (s) (IV,thm.1.5.).

Accordingly we will get a special case of example 7.3. The dis-
cussion of classes of differential operators within reach of the
generated comparison algebra should follow that of example 7.3.
Also, of course, we will have to characterize the secondary symbol
space of this algebra. This will not be investigated here.
CHAPTER 10. FREDHOLM THEORY IN COMPARISON ALGEBRAS.
In this final chapter we are going to discuss the main

application of theory of comparison algebras to the derivation of
Fredholm properties, and essential spectrum properties for diffe-
rential operators within reach.
Note that the general concepts were developed in [C1], and
they were demonstrated there in relation to elliptic differential
operators on In , or on the half space R++l There is no change
at all when we apply these abstract concepts to general manifolds.
A more detailed introductory discussion is given in sec.l, below.
In particular we choose to discuss this abstract theory
after our full scale investigation of the symbol spaces of a
comparison algebra, above, to direct the focus of attention accor-
dingly. Also, many proofs will be omitted.
Note that we shortly touched the concepts of Fredholm opera-
tor, Fredholm inverse (and Green inverse) in VI,3, while results
on compact manifolds were discussed, requiring this terminology.
In sec.l we discuss the general facts around Fredholm opera-
tor, Fredholm index, and C -algebras with symbol. In sec.2 we dis-
cuss some facts concerning the Fredholm properties of differential
operators within reach of a comparison algebra, and about essen-
tial spectra. In sec.3 we discuss comparison algebras with compact
operator valued symbol, and operators acting on crosssection of
vector bundles over a noncompact manifold.
In section 4 we discuss abstract Fredholm theory for a C*-
algebra with a two-link ideal chain. In sec.5 we turn to the
Frechet algebra of operators of order 0, bounded over all Sobolev
spaces, reviewing some abstract general facts, discussed in [C1]
X.1. Fredholm operators in C -algebras 292
1. Fredholm theory in C C L(H).
In this section we will continue the abstract discussions of

ch.I, looking at the concept of Fredholm operator in general,
bounded or unbounded, and, especially, at the theory of Fredholm
operator and index in a comparison algebra CCL(H) , H=L 2 (Q), and
for differential operators within reach. This material has been
discussed by the author and his associates in [C 1 1, [BC 1 1, [BC 21,
[ CM1] , [CM 1, [ CHe1] , [ CS] . Some slightly different approaches

2
are given by Coburn [Cbi1, Douglas [Dgj], Barnes et al.[BMSW]
Accordingly we will omit some detailed proofs, referring to [C1],
where our views were presented in close detail.
In these notes we have focused on the Hilbert spaces Hs
s E 1, and the Frechet space H, only. Accordingly we restrict our
review to Hilbert spaces only, except in sec.4, where the Frechet
algebra C, and differential operators within its reach are studied
In Hilbert space many features become simpler, and discussions are
less costly. However, most features extend to bounded or unboun-
ded operators in Banach spaces as well (cf. [C1],App.AI) .
On the other hand, a general Fredholm theory in Frechet

spaces seems to have pathological features. For example, the
linear group of all invertible operators of L(X), for a Frechet
space X , no longer needs to be open in L(X) Therefore it per-
.
haps is even more interesting that the Fredholm theory of C. still

remains intact. It looks like that of a C-algebra (cf.sec.4).
In fact, C. is a 4-subalgebra of L(H), in the sense of
Gramsch [G11. As a subalgebra of L(H) it contains all its inver-
ses in the larger algebra, just as this is known for C -algebras
Moreover, in [CSch], [Schl] it was seen that, at least for some
important special cases, the symbol homomorphism of C. still is
a surjection onto C(ffi) (This property is known to be generally
false for comparison algebras in LP-spaces or other Banach spaces.)
Perhaps, in the interest of a more systematic approach, the
present discussion should have been in order at the begin of ch.5.
However, we wanted to establish the Banach algebra structure of
a comparison algebra as the focus of our attention. Thus we were
pushing aside the important motivation of this theory from theory
of regular and singular elliptic boundary problems in partial
differential equations. The latter will proceed now.
Returning to abstract theory of (unbounded closed) linear

operators in Hilbert space of I,1 we recall that a closed operator
A: dom A -> H , in an abstract (separable) Hilbert space H, is cal-
led a Fredholm operator if (i) in A is a closed subspace of H
and (ii) we have
(1.1) dim ker A < - , and codim im A = dim H/(im A) < - .
It is not hard to see that (ii) implies (i), but it is use-

ful to keep (i) in mind, as an important property of Fredholm ope-
rators. In particular, for operators with closed range (with im A
closed) the second condition (1.1) may be stated by requiring
A).1_-
that dim(im For a Fredholm operator one defines the
.
Fredholm index by setting
(1.2) ind A = dim ker A - dim(H/im A) = dim ker A - dim ker A".
The concept of Fredholm operator is nontrivial only for

infinite dimensional spaces. This may be the reason for the fact
that they were systematically studied only in the 1950-s, although
the concept is quite simple. Special classes of Fredholm opera-
tors(integral operators and differential operators) were studied
in the late 19-th and early 20-th century (cf. Fredholm [Fh1]
and F.Noether[Nt1], who showed that Fredholm operators with
nonvanishing index occur in the theory of the 2-dimensional obli-
que derivative boundary problem).
We summarize the principal properties of a Fredholm operator
in thm.1.1, below. We will express the statement that A is Fred-
holm by stating that 'A is ', or that ' A has property '.
Theorem 1.1.(a) Property (D and ind A are invariant under small

bounded and arbitrary compact perturbations; (b) Property is
t
adjoint invariant, and ind A = - ind A ; (c) The product AB of
closed Fredholm operators A and B is always a well defined closed
operator; we have (AB) = B A and ,
(1.3) ind(AB) = ind A + ind B ;
(d) A closed operator is if and only if either OE Rs(A) or 0 is

an isolated point of Sp(A) , and an eigenvalue of finite multipli-
city for A and A
Remark 1.2. Point (a) of thm.l.l may be expressed as follows.

(a): For a closed A with preperty , and B E L(H) , C E K(H)

we have A+C E as well as A+cB E ( , for small c , and
(1.4) ind A = ind (A+C) = ind (A+cB) .
The same statement holds for unbounded operators B and C which

are 'A-bounded' and 'A-compact', respectively, i.e.,
(1.5) BI(dom A) E L(dom A,H) , CI(dom A) E K(dom A,H) ,
where dom A is regarded as a Hilbert space of its own under the

graph norm I,(2.8) .
For a proof of thm.l.1 cf.[C1] (for Banach spaces) , [CLa]

for Hilbert spaces, with generalized perturbations, also [Nbi]
for generalized perturbations in Banach spaces. For a discussion
of bounded and unbounded Fredholm operators between different
Banach spaces cf.[C1l,AI .
We also will consider Fredholm operators between different

(separable) Hilbert spaces. However, since any two separable Hil-
bert spaces H1 , H2 are isometrically isomorphic (with an iso-
morphism U defined by mapping j -"Vi for any fixed pair {4i },
of orthonormal bases of H1 and H2) we simply define A E Q(H1,H2)
AU-1
to be Fredholm if is a closed D-operator in H2 , and define
= ind(AU-1) . This reduces our discussion to the case of 0-opera-
tors in Q(H) , H=H2.
The key to our present theory of Fredholm theory of compari-
son algebras is the following fact stated for bounded Fredholm
operators. Let O(H) denote the class of bounded c-operators.
Theorem 1.3. (a) An operator AEL(H) is Fredholm if and only if

there exists an operator B E L(H) such that
(1.6) AB = 1+E , BA = 1+F ,
where E and F are operators of finite rank (with finite dimensio-

nal image) .
(b) An operator A E L(H) is Fredholm if and only if its

coset A = A + K(H) mod K(H) is an invertible element of the
C*-algebra L(H)/K(H)
(c) Let [Al be the homotopy class of q in the group ,
GL(L(H)/K(H)) of invertible elements of L(H)/K(H) These homo- .
topy classes form a group Hd under operator multiplication. There

exists a group homomorphism u:HcD - Z such that
X.1. Fredholm operators in C-algebras 295
(1.7) ind A = u[A-] , for all A E VH) .
Remark 1.4. An operator B E L(H) satisfying (1.6) is called a

Fredholm inverse (or (D-inverse) of A . The homomorphism u of
(1.7) is called the index homomorphism (of the C -algebra L(H)).
Clearly a Fredholm operator will have many Fredholm inverses.
Among these we have focused on the special Fredholm inverse in
[C1],AI and, with respect to an HS-chain, the distinguished Fred-
holm inverse (VI,3).
For a proof of thm.1.3 cf.[C1],AI for Banach spaces.
,
9e
Now let us recall the following property of C -algebras, not
normally shared by other Banach algebras:
Proposition 1.6. Let A be a invariant sub-algebra of the C -

algebra 8 , where both A and B have (the same) unit. Then, if
A E A C 8 is invertible in 8 , its inverse A-1 is contained in A
For a proof cf. [C1] , All , or [ Rl] .
Combining thm.1.4 and prop.l.6 we now get a result useful
for Fredholm theory in comparison algebras.
Theorem 1.7. Let A be a COY-subalgebra of L(H) containing the ,
identity operator and the compact ideal K(H). Then an operator

A E A is Fredholm if and only if its coset Av = A+K(H) is an
invertible element of the quotient algebra A- = A/K(H)
Indeed, we then have A/K(H) a C -subalgebra of L(H)IK(H)
both with the unit 1+K(H) , so that prop.l.6 applies.
Remark 1.8. Under the assumptions of cor.1.7 we also may replace

the homomorphism u by a homomorphism from the group of homotopy
classes of GL(A/K(H)) to 1 (cf. [BC1] ).
Now, if A = C is a comparison algebra, as studied in

chapters V - IX, then the algebra C/K(H) has a very simple struc-
ture at least in the following two cases.
Case (A): The commutator ideal E is K(H) (Such comparison alge-

bras were investigated in V,4 and VII,2).
Case (B): The commutator ideal E is not K(H), but has the property
that E/K(H) is an algebra of compact-operator-valued functions
over a locally compact space (such as the two examples discussed
in VIII,l and VIII,2.)
Note that other cases are possible: Not every comparison is

either in case (A) or (B) . In particular there may be ideal
chains longer than the two-link C D E D K(H) , with other proper
ideals fitting between E an K(H).
First consider case (A) . Then C/K(H) = C(81) , with the
symbol space . A continuous complex-valued function aA(m) over
I
the compact Hausdorff space M is invertible if and only if it

never vanishes. Therefore we have the following result.
Theorem 1.9(A). Under cdn.s (a0),(a1),(d0) only let the compa-

rison algebra C have compact commutators. Let HI denote the symbol
space of C (cf.VII,l), and aA the symbol of AEC. Then an operator
AEC is Fredholm if and only if its symbol aA(m) , mEI , does never
vanish. There exists a group homomorphism u from the group f1(ffi)
of homotopy classes [a] of non-vanishing continuous complex-valued
functions a over M such that
(1.8) ind A = u([aAI) for all A E MOM .
The proof of thm.1.9(A) is an almost trivial consequence of

our above discussion. For a more detailed derivation cf. [BC1]
For a partial discussion of case (B) we refer to sec.4
below. Two inversions, modulo the two ideals E and K are needed,
in this case. We also refer to [ CPo] , [ CDgl] and [ CMel] , where
we treat a useful generalization for the special 2-link-algebras
of VIII,1,2,4 , and for more general such algebras. Instead of the
two inversions we there have only one symbol, assuming singular-
integral-operator-values at certain points of the symbol space.
2. Fredholm properties of operators within reach of C.
Note that thm.1.9(A) gives a very simple necessary and

sufficient criterion for the Fredholm property and index of opera-
tors in the algebra C provided that we know details about the
,
symbol space 1I of the algebra, and also know how to explicitly

obtain the symbol of an operator AEC That precisely was the
.
intent of our effort, in the preceeding sections.

The present thm.1.9(A) only represents the simplest non-tri-
vial case of a chain of similar results. Later we will discuss
a similar result for case (B) as well. Also some cases involving
longer ideal chains will not be discussed. Also we will have to
X.2. Differential operators within reach 297
deal with the case of a more general matrix A = ((A.k)) of

J
operators AjkEC, where even infinite dimensional matrices will be
admitted. Also, this will lead us into comparison algebras acting
on L2(O)-crossections of a vector bundle on c .
Let us use the present example to explain the principle

of comparison: Often one will not be interested in property
(or the index) for operators in C, but rather, for an unbounded
operator of the form
(2.1) L = AQ-1 , with A, Q E C .
Example 2.1. Let Q = A = H-1/2 , with the expression H of some

triple {O,dp,H} , and let A=DA , with some DED. Assuming cdn.(m5),
it t2en follows that A and A both are in C However, the operator .
A-1 = H1/2 is an unbounded closed Fredholm operator of H , with

index 0 since it is self-adjoint, positive definite, and >1
, ,
(thm.l.1(d)). Suppose now that A is Fredholm. For example, assume

that C has compact commutator, and that aA(m)#O for all mE ffi. Then
apply thm.1.9(A)). Then AA-1 is an unbounded Fredholm operator,
and ind(AA-1) = ind A , by thm.l.l(c),(d) , since ind A-1 = 0
However, we find that

AA-1u,
(2.2) Du = bJu .+pu = for all uE dom(AA-1)=dom A-1 = H15
1x
with the first Sobolev space H1. In other words, Z=AA-1 is a rea-
lization of the first order differential operator D, with domain
H1, in the sense of V,1 . Thus,if aA does never vanish on M ,
then we have found a realization Z of the differential operator D

which is Fredholm. In fact, since A-1:H1-H acts as an isometry,
(cf.V,l) one concludes at once that the realization D constructed
(with domain H1) is if and only if A is ' i.e., if and only ,
if aA(m)#O , mEffi .
The above conclusion is not very far reaching, since there

are much simpler methods available, to solve the single first
order partial differential equation Du=f, with or without boundary
conditions. (In fact, looking at the symbol aA = bJ(x)Cj over
the wave front space W C ¶ it is clear at once that the condi-
,
tion aA90 on W cannot be satisfied, unless n=l.)

This changes, however, if we either consider a matrix
A=((Ajk))j,k=l,...,m of operators Aj1E C or an operator linked
,
to a higher order expression L E LN as in VIII,3. We look at ,

matrices of operators in C in sec.3, below, although this is

the simpler problem. Here we next look at an LELN
Example 2.2. Consider the case Q=AN, for some integer N=1,2,...,
with a differential expression LELN. Here we must ask for cdns.
(s),(m5),(lj), j=1,2, (apart from (a0),(a1),(d0)), and (15)(P.)
for a suitable class P., within reach of A!, D!, so that IX,prop.
3.2 or cor.3.4 hold. (Note that (s) may be replaced by (sk), for
sufficiently large k, as easily seen.) Also, cdn.(m1) or a cdn.
(m1)^, for suitable q^, may be useful. Since compactness of commu-
tators is required, we assume a triple as in early V,4, with a
large constant y of V,(5.3) to get cdn.(s ), not only (w), and
cdn.(g1), with classes A!CA#, V!CD#. Or else, we assume the case
of IX,example 7.3, with cdn's (q2), (q3). In any such case we get
(2.3) A = LQ = LAN E C .
Thus the conclusion of example 2.2 may be repeated. We have

im AN = dom A_N = HN, with the Sobolev space of order N of IX,l.
Formula (2.4) defines a closed realization of the expression L
(It was already discussed in V, def.6.2 , and called Z there.
Here we find it more convenient to use the same symbol for the
expression L and its realization.)
(2.4) L = AA-N , dom L = HN = im AN
This realization L is a closed '-operator, if and only if
(2.5) aA(m)#0 for all mEffi .
We again are interested in checking on (2.5) explicitly.

In that respect note that formula VI,(3.9) indeed gives the symbol
on the wave front space. We get
(2.6) as IfI=h3kEjY1,
where L=a(x,D), in local coordinates. In other words, aAIW is just

the principal symbol of the expression L, at the unit co-sphere,
in the conventional sense, and the condition (2.5), restricted
to g!clos, amounts to the statement that
(2.7) IaN(x,E)I > Ti >0 for all E ICc? = W ,

i.e., that the expression L is uniformly elliptic on 0, with res-
pect to the metric tensor hjk .
It is important to observe that we are not yet in control of

the full symbol, except in a case where the secondary symbol space
is empty. However, we have the results of VII,2, 3, 4, and IX,7 to
investigate the secondary symbol. For example, under (ql) of
VII,4, and (m1'), (m3), (m7), or the corresponding "'-conditions,
we conclude from VII,thm.4.7 that there is no secondary symbol
space over aQ , although there may well be points of the secondary
symbol space over DMA#\aQ=DMA# , as our examples of V,4,
VII,2, VII,4, and IX,7 show.

Even if the space ffis is completely known, as in case of VII,
thm.4.8, for example, the secondary symbol of A=LAN, for a higher
order expression L is known only if a representation of A as
a sum of products (D1A)(D2A)...(DkA) is explicitly given. For the
factors DA of such a product we may invoke VII,thm.2.2, thm.2.3,
thm.2.6, etc. Again, under (q2), (q3), we may refer to IX,thm.7.5,
given its assumptions.
In some cases, however, a complete knowledge of the symbol
is not required, while significant statements still can be made.
Note that the differential operator L of example 2.2, regard-
less whether Fredholm or not, always is a restriction of the clo-
A-N
sure of the minimal operator L0 , since (A-NICO)clos =
due to cdn.(s). Assuming the expression L to be formally self-
adjoint, the operator L will be hermitian, and a restriction of
t'he closure L0 of L0 . In fact, we now have the following.
Theorem 2.3. Assume cdn's (a0), (al), (d0), (s), (11), (12),
(m5), (15)(P,) with respect to a suitable class P, within reach
of C , and of principal symbol type, and assume that C has compact
commutator. Also assume that a0 = a N # 0 on all of M Then .
n
for every self-adjoint uniformly elliptic expression LELN (i.e.,
L satisfies (2.7)) the realization L of (2.4) is a self-ad-
joint operator.
Remark 2.4. The condition' (or aA#0) on Ms ' follows from

c N#0
VII,cor.4.3, if q=l for large x of QUMM , since then D=D0=1 may be

taken in VII,thm.3.6.
N/Proof of thm.2.3. Let us define B = A2LA N/2 , where, for a
moment we use the convention of IX,4 (i.e., regard all operators

and expressions as operators on H.). It follows that (the H-clo-
sure of) B defines a bounded self-adjoint operator of H. Moreover,
fromIX,prop.4.1 we conclude that
(2.8) A - B = A - AN/2AA_N/2 = AC, with C E C ,
so that also BEC , since AEC . In fact, we get aA=aB on ID , since

aA=O on W .
Since B is a self-adjoint operator, its symbol must be real-
valued. Since we assumed L uniformly elliptic, it follows that aB
= aA is bounded away from zero on W. On the other hand we assumed
aA00 on Ms = M\(,clos) Since A also is self-adjoint, aA is
.
real-valued as well. Hence we conclude that
(2.9) (m) # 0 for all m E H1

(B+ieA N )
since aB90 in a neighbourhood N of pjclos , while ffi\N C ffis is com-
pact, so a N bounded away from 0 on ffi\N , and the two terms cannot
A
cancel each other, one being real, the other purely imaginary.
Conclusion: B+iCAN is Fredholm, hence L+ie= A-N/2(B+iEAN)A-N/2
also is Fredholm, whenever e90 as follows from thm.l.l(c). From

,
the same thm. it also follows that

-N/2 =A-N/2 (B-iCAN)A_M/2 = L-ie ,
(2.10) (L+ie =A (B+iCAN )A-N/2
in the precise sense of adjoint of unbounded operator, using that

both operators A_N/2 and (B+ieAN) are Fredholm. Hence it follows
that L L , or that L is self-adjoint, q.e.d.
For a general, not necessarily self-adjoint, but uniformly
elliptic expression we can derive knowledge on the essential
spectrum of the realization (2.4) from the knowledge of the
secondary symbol. Here the essential spectrum of an unbounded
closed operator AEQ(H) is defined as the collection of all AEtC
such that A-A is not a Fredholm operator. From thm.2.l it follows
that the essential spectrum consists precisely of all points A E
Sp(A) which are not isolated point-eigenvalues (of A or A*) of
finite multiplicity. (Note that other definitions of the concept
'essential spectrum' are in common use (cf. Kato [K2], for exam-
ple )
Theorem 2.5. Assume cdn's (a0), (a1), (d0), (s), (11), (12),
(m5), (15)(P0), as above, and that the algebra C has compact

commutators. Let oQ=oAN#0 on all of Ms . Let LELN be a uniformly
elliptic expression (satisfying (2.7)), and assume that the set
(2.11) {oA(m)/o N(m) : mEM 5}

A
leaves out at least one complex number. Then the realization of L
defined by (2.4) is a closed operator (and the closure of the
minimal operator L0). Moreover, the essential spectrum of this
realization is given by the set (2.11).
Proof. In essence we repeat the arguments of the proof of thm.
2.3 in the following simpler form. First, the realization L
again is a restriction of L0* Next, we write
.
(2.12) L-A = (A-AAN)A-N
where AA=A-aAN E C , and aA X =CA # 0 on Mp=wclos ,

due to oA=O
on 1% , and (2.7). There exists at least one XoEM, not contained

in the set (2.11). We conclude that oA =oA-Xoa N#0 on all of ffi
Ao A
so that AXo is Fredholm. Accordingly L-ao , hence L is closed.
Also L-Xo is Fredholm, by thm.1.1(c). The argument may be repea-
ted for any XEM not assumed by the set (2.11), showing that
L-A is Fredholm for every such A. Hence the essential spectrum is
a subset of the set (2.11). On the other hand, if A is contained
in the set (2.11), then we know that A-aAN is not Fredholm.
But the operator A-N:HN=dom AN + H is an isometry between the
Sobolev spaces HN and H as we know. Relation (2.12) may be
,
interpreted in two different ways, either the product is a product

of two unbounded operators, or a composition of and
A-N:HNH . Both interpretations describe exactly the same map,
(even though not topologically). With the second interpretation we
AAA-N
find that is not a bounded Fredholm operator HN-3.H , hence
either codim im = co or dim ker = - or both. But the bounded
,
operator HN+H and the unbounded operator of H with domain don. L

= HN have exactly the same image and kernel. Therefore L cannot
be Fredholm as well, q.e.d.
The above examples were designed as a demonstration, sho-
wing how a (singular) elliptic problem may be handled with the
'comparison techique'. We selected the simplest nontrivial case
for this demonstration. Results like thm.2.3 and thm.2.6 have
straight generalizations to the case of a (finite or infinite)

matrix of differential expressions, as discussed in sec.3, and
to expressions formally acting on the type of vector bundles of
sec.3 as well. We shall not discuss details.
Perhaps we should also point to a possible use of resolvent
integrals for realizations of the form (2.4). A criterion like
thm.2.5 gives very precise control on the Fredholm domains of
the analytic operator family LT = L-X They are just the connec-
:
ted components of the complement of the (closed) set of (2.11).

By a result of Gramsch [G2] it is sufficient to know existence
of the resolvent in only one point of a Fredholm domain, to con-
clude existence of the resolvent as a meromorphic function defi-
ned in the entire domain, with poles clustering at the boundary
only. Such knowledge then allows definition of functions of L as
resolvent integrals.
Let us finally note that one needs not necessarily choose
the operator Q of (2.3) as a power of A. Clearly, Q-1 should be
Fredholm, and perhaps should have other properties. It could be
self-adjoint (and positive definite), for example. Note that the
condition aQ#O on ffis, required in thm's 2.3 and 2.5, does not hold
for the problems of Sohrab (VII,thm.4.8).There we have aA = 0 on
the entire symbol space, including ffi .
s
Our technique for those thm's depended on the fact, that
the principal symbol space coincides with the set where aQ=O .
In general one could redesign the results, replacing the set ffip
by ker aO
Note also that we have been using the comparison technique
in [C1],V,14, with an operator Q (called T there) of a different
kind. Each T used there represents an elliptic boundary condition
(of Lopatinskij-Shapiro type), and we were solving a regular
(not boundary-value problem.
X.3. Systems and Vector bundles 303
3. Systems of operators, and operators acting on vector bundles.
We continue our review of Fredholm theory in comparison

algebras, and now discuss operators with compact operator-valued
symbol, as in IC1],V,8 and [BC2] ,Recall that such algebras
.
already were used in VIII,1,2,4. In particular it was found that

the commutator ideal E if not compact, often has a natural com-
pact-operator-valued symbol. Here, however, we work from the start
in a Hilbert space H- which is a topological tensor product
(3.1) H_ = hiH , H = L2(0) ,
where h is a separable Hilbert space. (We normally assume h=k2

= L2(Z+) , Z+ = {1,2,3.....}). In [BC2] the space H also is an
abstract Hilbert space. Here we consider L(H) as the home of our
comparison algebra C, acting on the complex-valued functions of
H = L2(M. We now want to consider the C' -algebra
(3.2) C' = kffiC , where h = K(h) .
Note that an operator KEk is an infinite matrix K = ((kjk)),

acting on a sequence of h=f2 by matrix multiplication. All matri-
ces with only finitely many non-zero entries are necessarily com-
pact, and every compact matrix is a limit in norm convergence of
a sequence of such matrices of finite rank. Correspondingly, it
can be shown that every operator of kNL(H) is an infinite matrix
(3.3) A = ((Ajk))j,k=1,2,... e where Ajk E L(H)
Vice versa, a given such matrix (3.3) is in kNL(H) if and

only if it is a limit, in norm convergence of L(H '), of a sequence
of matrices (3.3) each having only finitely many non-zero
entries Ajk We have K(H) = k9K(H) C koL(H) C L(H-)
. all with ,
proper inclusions, and K(Hf) is a two-sided ideal of kk(H)

We get A E C if and only if all entries AjkEC , j,k=1,25 .... ,
and, again, A is limit of a norm convergent sequence of such

matrices with only finitely many non-zero entries.
For such operator A E C we define a symbol a by setting
(3.4) Cr (m) = ((a (m))) . .

A Ajk ,k=1,2,...
Now it follows easily that aA:M - k is a continuous function
in norm topology of k = K(h) . We have the result, below, where
X.3. Systems and Vectorbundles 304
E = K(H) is assumed for the remainder of sec.3, although an exten-

sion to the 2-link-case would meet no obstacles.
Theorem 3.1. Let (a0),(a1),(d1) hold, and let CCL(H) be a com-

parison algebra with compact commutator
(a) Then an operator B = 1+A, with A E C is Fredholm if
and only if the function 1+aA(m) is invertible in L(h), for every
mEj . (b) There exists a homomorphism u from the group HWI) of
homotopy classes [a] of continuous maps a: GL(l+k), underffi -
operator multiplication as group operation, into Z, such that
(3.5) ind A = u[aA] , for all A E C' n V H') .
For detailed proofs we refer to [BC21 thm.7 and thm.l0. ,
There we also discuss the precise definition of the topological

tensor products ho-H and PkQ , for CC subalgebras P C L(h), and
Q C L(H) .
The most important applications of thm.3.1 aim at (finite

dimensional) matrices of operators with entries in a comparison
algebra C and, more generally, at operators acting on crossec-
,
tions of a (finite dimensional) vector bundle over B Note that, .
in case of a compact manifold n the homomorphism u of thm.3.1

,
essentially coincides with the relation between topological and

analytical index expressed by the Atiah-Singer-Bott index theorem.
Finding u explicitly, i.e., obtaining an explicit index
formula for a comparison algebra C' on a noncompact mainfold, will
be a task of algebraic topology. We are not equipped to discuss
this problem here, but, perhaps should make a few comments.
Remark 3.2. A homotopy A:[0,11xffi 3 GL(l+k) , in detail,
(3.6) A(t,m)=a(t,m) + K(t,m) , 0<t<l , mEffi ,
always can be deformed into a homotopy U:[ 0,11 -* U(l+k) , where

U(l+k) denotes the group of unitary operators in the algebra l+k
C L(h) This follows, because a polar decomposition A = BU is
.
possible within the algebra l+k , where B=(AA*)112 ,

U=A(AAC)-1/2.
Here U is unitary, while the invertible self-adjoint positive

definite operator B may be deformed into the identity, by the
deformation titB+(1-t) E GL(l+k) .
Remark 3.3. The problem of discussing the index homomorphism u

may be reduced to a more conventional form in homotopy theory by
inducing the Bott group U(°°) , defined as follows. For N=1,2,...,

let hN C h = 12 be the subspace of all sequences u=(ul,u2,...
with uN+l=uN+2= ... =0 , and let U(N) denote the subgroup of
U(l+h) consisting of all unitary matrices equal to 1 in h1 .
Regarding U(N) as a topological group with (Euclidean) topology

one defines U(°) as inductive limit of the sequence U(N) , with
inductive limit topology. Concretely, we thus have
(3.7) U(oo) = U.- U(N)

3 -1
A subset MW(-) is closed if and only if MrU(N) is closed, for all

N=1,2,... . A compact subset of U(oo) can have a non-void inter-
section with at most a finite number of U(N) .
Note that a continuous map A: X - U(l+k) for any compact ,
space X can be written uniquely in the form
(3.8) A(x) = a(x)(1+k(x)), xEX , with aEC(X,S1) , kEC(X,K) ,
with the unit circle S1 = {cEc:Ic1=1} in T . On the other hand,

every continuous map F:X+U(oo) is continuous from X to U(l+h), and
vice versa, a continuous map G:X + U(l+h) of the form (3.8), with
a(x) = 1 can be continuously deformed into a map F:X-U(oo)
, .
(Essentially, to construct F from G one uses the fact that U(oo)

is dense in the class of all unitary operators on It of the form
1+K , KEk, in norm topology of L(h) , and the polar decomposition
of rem.3.2.)
This reduces the discussion of the homotopy classes of thm.
3.1 to that of homotopy classes of maps 1 ; UN , with the unitary
group UN of large dimension N.
Note that the theory of NxN-matrices A=((A.7k)).,k=1,...,N
of operators AjkEC may be treated as a special case of thm.3.1.
Indeed, for such a matrix we define an operator A- E C' by
(3.9) A-=((A-jk))j,k=1,2,..' A-ik=Ajk, for j,k<N, =6jk otherwise.
Clearly A- will be a Fredholm operator of L(H ) if and only if

A is a Fredholm operator of CN = L(fNW , and A, A" then will
have the same index. The symbol of A will of course be defined as
(3.10) aA(m) = aA-(m)IhN
noting that aA(m) is reduced by the decomposition h=hN ® hN .

It is just as easy, of course, and, in fact, simpler, to
redo the above theory with h replaced by the finite dimensional

.
space hN = (EN
In the following let Cm = L(Cm')ffiC , where C denotes any
comparison algebra. Let us come back to the comparison principle
of (2.1), and again set Q = A = ((Adjk)) and A = DA , with ,
(3.11) D = ((D.
7k
)).
,k=l,...,m
,
k
D. E D
AA-1
Again we get a realization D = of the (mxm-matrix of)
expression(s) D That realization always is a restriction of the
.
closure of the minimal operator D0 , regardless of any condition

(w), or (sk) In the results below we skip looking at the Fred-
.
holm index. An abstract index formula always follows in a similar

way as in thm.1.9(A), thm.3.l,etc. Obtaining a concrete index
formula should be easiest, in the present cases, but will not be
a subject of our concern.
Theorem 3.4. Assume cdn's (a0), (a1), (d0), and that C has com-
pact commutator. The above realization D is a closed (unbounded)
Fredholm operator of L(Hm), with HRL-TmNH, if and only if
(3.12) the mxm-matrix aA(m)=((aD A(m))) is invertible for mEI

jl
Note that no cdn.(w) or (s) is required in thm.3.4. The

proof is an immediate consequence of thm.3.1.
Let us try to discuss the result. Again, we first consider
condition (3.12) restricted to the wave front space Wcffi There .
it just means that the first order mxm-system D=((Djk)) is uni-

formly elliptic, with respect to the tensor hjk . That is, if
(3.13) D = -iB1(x)3 . + P(x)

X3
locally, with mxm-matrix-valued functions Bj(x) , P(x) , then
the function D(x,O) = B3(x)E. (globally defined as a func-
tion in C_(T*O,L(Mm)) ) is invertible for every (x,E)40, and
satisfies
(3.14) HD-1(x,i;)II < X < - , as xEQ, CE]kn, h3k(x)EiEk=1
with the matrix inverse at x,C .
For a compact manifold 0 this is the only condition, since

then I = W . On the other hand, if 0 is non-compact then there
still may be comparison algebras with Ms = 0 . For example, this
X.3. Systems and vector bundles 307
is true under the assumptions of VII,thm.4.7, if 04 0 is compact.

On the other hand, our examples of IC11, and those of Sohrab
show that we may have It s# 0 Then we can apply thm.3.4 only if
.
IDs and aA"ffis are accessible.

Note that thm.3.4 at least gives a sufficient criterion for
the Fredholm property, if we have the assumptions of VII,thm.3.6
(or thm.3.6", for a suitable {St",dp",H"} (c> {S2,dp,H}). Then the
symbol is a restriction of the formal symbol of VII,(3.25). For
a compact commutator algebra C the operator D of (3.11) then is
Fredholm if (though not necessarily only if) the formal symbol
se
TDA is invertible at each point of 88 0
Finally we will use our above theory to also consider opera-
tors acting on crossections of a vector bundle over Q. We have
avoided vector bundles, so far, because our C -algebra approach
to symbol and symbol space is developed best for complex-valued
symbols, while its extension to vector bundles is a simple
abstract matter, to be discussed now.
Let V be a vector bundle over Q with fiber an inner product
space Vx of dimension m, with norm IuIx , u E Vx . We only con-
sider (C'-)vector bundles of the form V=VA#IS2 = 0-1(Q), where VA#
is a (continuous) vector bundle over the compactification M # of

A
0 induced by A# (cf.VII,l), with bundle projection 0 We assume .
the euclidean metric 1.1x of the fiber defined for xC M # , conti-
nuous in x , and C' over Q.

For the above we assume given a comparison triple on 2 , and
generating classes A# D# satisfying (a0),(a1),(d0) , and then
,
will introduce operators acting on HV = L2(Q,V,dp), the space of
crossections u of V satisfying Ilull2 = J0Iu(x)l2dp <
Let MA#=U1U...UUN be a finite atlas of the (compact) bundle
VA#, and 1=V)1+...+PN a subordinate partition of unity, V .IOE C'(0)
Under the above assumptions we will construct an isomorphic

imbedding of the Hilbert space HV into our above space HR=hRNHCH°°
with orthogonal projection PV:H-+HV in CR C C C = C(A#,D#). For

two such vector bundles V and W one then may consider operators
of the form A"=PWA`PV , where A^'E C" . We get A"E C'; the restric-
tion AÎHV = A is a map in L(HV,HW), and the set of such maps will
be denoted by CVW Fredholm results for AE CVW will be obtained.
.
Let e ( 6 1 ) 1 1=1,...,p, denote the standard base

.
,
p 9P
of MP. Define a linear map Uj from the space of local crossections
over Uj into the fNNTm-valued functions over U. by u(x) + eou(x),
in the bundle coordinates of the chart U.. Here ij(x)u(x) +
eNTiP.(x)u(x), where both sides may be extended 0 outside U. to ob-
tainJglobal crossections of V and the product bundle x(gNN(m)=Z
respectively. A (global) crossection u of V may be written in the
form u = and a map O:TV + TZ, between the global crossec-
tions of f V and Z is defined by Ou = (where each iju
is mapped in the coordinates of Uj ).
The above construction defines a bundle homomorphism V+Z

and an isomorphism: For an xE Uk , in the bundle coordinates of
U
k
, we get Ou(x) = JN ejOR. ( ju), with the transition-matrices
J=1 N Jk
Rjk. Thus Ou(x)=0 if and only if ipju=0, j=1,...,N , i.e., u=0
In particular, if ujl denotes the local crossection in U.
with ujl(x)=em in the coordinates of U., then ijuj1= vjl defines
a global crossection, and fk(x)=IN=1e3mvj1() , xED , defines a
global m-frame over M , C' over Q . Write V(x)=span{fl'" .,fm}.
We may identify the space MNNCm with the space hR , R =mN

of rem.3.3, writing eN0em = eN(0-1)+k . Then the map 0 identifies
the space HV with the closed subspace of all measurable functions
over 0 with value u(x)(= V(x), such that IIuIIV of (3.15) is finite:
(3.15) IIuIIV= JO lu(x)Ix2du , u E HV ,

Iu12={1mjk()ujuk}1/2,
with a positive definite matrix ((mjk(x))), mjkE C # . Moreover,

A
the global frame fk is C # as well.
A
It follows that the orthogonal projection PV(x):h+V(x) (with
the inner product of h) has matrix coefficients in C # and hence ,
A
is a (multiplication) operator in CRC C
For two vector bundles V,W, satisfying the above assumptions
consider the class CV W above. It is natural to introduce a symbol

GA (mod EV W=PWE°°IV, where E'=hnE) by setting
(3.16) OA(m) = PW(1(m))oA_(m)IV(1(m)) , m E M
Clearly aA may be regarded as a homomorphism Vffi } WN , with the

liftings V., Wffi of V and W onto M (by means of the map i of VII,l)
Again consider only the case of E=K(H)
Theorem 3.5. An operator AE CVW is Fredholm in L(HV,HW) if and
only if its symbol aA E hom(VM,WM) is an isomorphism onto WM
Proof. Let m,m' be the fiber dimensions of V,W. Consider the
case m<m' ; in the other case we may take adjoints. The above glo-
bal frame fl.... fm for V(x) (and gj for W(x)) allows us to define
a global isomorphism V(x)-W(x) by assigning fj,gj. Using the cor-
responding isomorphism HV}HW we may get restricted to the case
V(x)CW(x). For m=m' get V=W and note that A°=l-PV+APVEC, is iff
A is , reducing thm.3.5 to thm.3.1. For m<m' aA cannot be surjec-
tive, and A cannot be The projection P onto span{gm,} is non-
:
0
compact, but aP AP =0, hence POAPVEK(H-), a contradiction, if
0 V
A is . Q.E.D.
It is easy to derive an abstract index formula of the form
(3.5) again. Details are left to the reader.
Also, most of the results of this section can be generalized
to the case of a multi-link ideal chain, with no or very little
change in proof.
4. Discussion of algebras with a two-link ideal chain.
In this section we want to look at case (B) of sec.l.

That is, in detail, we assume that we have
(4.1) C D E 3 K(H) ,
where the commutator ideal E of the comparison algebra C properly

contains K(H) , and where
(4.2) E/K(H) = C(ME,K(h)) ,
with a locally compact space ffiE , called the symbol space of E

and the compact ideal K(h) of a (separable) Hilbert space h
We of course also get
X.4. 2-link ideal chains 310
(4.3) A/E = C(N) ,
with another (compact) space N _ ffiC called the symbol space of

,
C , since by design C/E is a commutative C'S-algebra with unit,

so that the Gel'fand-Naimark theorem applies.
We have discussed two examples of the present form in VIII,
1, 2, and a third more complicated example in VIII,4. Another
example, corresponding to the regular elliptic boundary-value
problem of a half-space, was discussed in [C1],V .
Again it should be mentioned, that we know examples with

longer ideal chains, i.e., E/K(H) is not a function algebra, but
(4.4) E D F D G D ... D K(H) ,
with each successive quotient of the form C(X,K(h)). (cf.Dynin

[Dyl]). For most examples the space It has a concrete meaning; it
occurs in some tensor product decomposition H=hNH- of H=L2(D,dp)
On the abstract side of this discussion, Kaplanski [Kpl]
has introduced a class of C -algebras, called GCR-algebras, having
a similar ideal chain, perhaps of infinite length (cf.Dixmier
[Dx1],p.87).(There the type of algebra is called 'algebre post-
liminaire'.) Sakai [Skl] has shown that a C*-algebra is 'GCR'
if and only if it is of type 1, in the sense that every represen-
tation is of type 1).
It is easy to formulate an abstract result extending thm.
4.1, below, to longer chains. However, we will not discuss this,
since all our examples are of type (A) or (B).
Presently, for a two-link-chain, we have the following
result.
Theorem 4.1.(a) An operator A E C is invertible modulo E if and

only if its (C-)symbol aA, a certain continuous complex-valued
function over ffi = ffiC does never vanish; (b) If an E-inverse P of
AEC exist, so that we have
(4.5) AP = 1+E , PA = 1+F , with E,F E E ,
then 1+E and 1+F will be Fredholmif and only if the operator-val-
ued functions 1+TE ,l+TF are invertible at all points of ME, and
the inverses are bounded on NE , where TE TF denote the E-symbols
,
of E and F, respectively; (c) If the E-inverse P exists, and G, H

are K(H)-inverses of 1+E and 1+F, then PG and HP are K(H)-inverses
of A, and A is Fredholm; (d) Vice versa, if AEC is Fredholm, then

there exists a K(H)-inverse P of A, which also is an E-inverse,
so that (4.5) holds with E,F E K(H) c E We then have oA g 0
.
for all mEM and TE = TF = 0 on ffic

, so that 1+TE = 1+TF = 1
,
are invertible.
The proof is a standard application of theory of commutative
C -algebras.
Let us practice once more the comparison technique, by
applying thm.4.l to the example of VIII,4, involving a complete
Riemannian manifold with cylindrical ends. First, in that respect,
recall that the secondary symbol space of that algebra C is void,
by VII,thm.4.2. Second, we note that we have cdn's (s),(ml1),(m6
(m.), (1j), j=1,2,3,4,5, and the assumption of VII,thm.3.6, all
true. Accordingly we have the result, below.
Theorem 4.2. Let C be the comparison algebra of VIII,4. Then we

have A = LAN E C , for every LELN . Moreover, the operator A=LAN
admits an E-inverse if and only if L is uniformly elliptic on S1
with respect to the tensor h;k .
Indeed, IX,prop.3.2 is valid, and the algebra C has no

secondary symbol space, so that the theorem is evident.
As a consequence of thm.4.2 we must look at the E-symbol
of E = 1-AP and F=1-PA , for a suitable E-inverse P of A, in order
to decide whether or not A is Fredholm. Clearly, again, the reali-
zation L=AA-N of (2.4) of the expression L will be a closed
Fredholm operator if and only if A is Fredholm.
First let us investigate the class LN (of IX,3) in the pre-
sent case. The description of the classes A# D#, is found in
,
VIII,4(i.e, they are determined by VIII,(2.2) , at each end of 0).

Note that the classes A# D# of VIII,4 were 'minimal', in
,
the sense that we were interested only in supplying enough gene-

rators for the algebra we had in mind, and to satisfy our basic
conditions. Presently we should enlarge these classes (without
changing the comparison algebra C ), in order to have 'large'
classes LN of differential expressions.
Clearly we may admit into A# any function aE
A
without enlarging the algebra. Since D# is a left A#-module, and
D# a Lie-algebra we must watch about properties of the derivatives
in order to keep our cdn's (mj),(lj) preserved. A more careful
investigation shows that we can reach the following classes LlpLN
of N-th order differential expressions.
Definition 4.3. The class LN consists precisely of all expres-

sions of order N with Co-coefficients, defined on 0 , such that
at each end, in the coordinates (x,t) of VIII, (2.1),L has the form
(4.6) L= limt;-Li(t)=L (eo), limt-.(a kLi)(t)=O k>O
Remark 4.4. Recall, that, at each end, we have a 'chart'
(4.7) U= {(x,t): xE BJ . , O<t<oo}
In local coordinates (x1,...,xn-i) of the compact manifold B.
formula (4.6) means that L=jLj(t)3t , where
L.(t) _ I a.,a(x,t)Dxa , limt aj,e(x,t)=aj, (x,oo),

(4.8) lal<.N i
limt Sta. a(x,t) = 0 , k > 0
0O Js
A quick calculation shows that the classes LN are generated
in the sense of IX,3 by the function algebra A!=L# and Lie-alge-
bra D! = L# , while the Stone-Weierstrass theorem may be used to
show that A# and D#A are dense in A! and D!A , respectively, in
norm convergence of L(H) so that C(A#,D#) = C(A!,D!)
, .
Let us go somewhat more into the detail of construction of

an E-inverse P , for an operator A=LAN , since it is evident, that
thm.4.1 is of little use for Fredholm theory, unless P is known,
since the symbols otherwise cannot be obtained.
a(l-AP) , (,_PA)
N Let L E LN be uniformly elliptic, so that aA #Oon
O HI , for
A=LA We focus on one of the cylindrical ends
. Std=B1.x1, which
will be called S2°= BxR Correspondingly we have H°
. C°, A!° ,
D!° ,A°, all defined as the corresponding concept on the straight

cylinder O°, with only 2 ends, which has the common domain U=Bxl+
with 0 , onto which we apply algebra surgery.
One may construct a uniformly elliptic expression L°EL°N
coinciding with L on U , and then observe that
(4.9) XLANX - .XL°A'NX E K(HU) , HU = L2(U) .
This shows that construction of the E-inverse P , and even

the later construction of a Fredholm inverse can be accomplished
separately, at each end. Once we construct a Fredholm inverse

P° = Pj for A° = L°A°N , at each end, a Fredholm inverse P is
given by
(4.10) P JjM M= number of ends,
with cut-off functions *j , j=1,...,M , and a suitable
xMp 2=1
making . Also, P0 is just any operator in JO such that
aAP=1 over a compact neighbourhood of supp 0 .

Working on the straight cylinder Bxl = S24 now, let us drop
the superscript"' again, keeping in mind, that L=L° only is
determined on the right end, t--, and is quite arbitrary for small
negative t. According to (4.6) the limits Lj(oo)=limt+,L.(t)
exist, and we may define the 't-translation invariant' expression
(4.11) L_ = IL (-)a3
Since L is arbitrary for negative t we may assume that L,, also is

the limit of L, as t-+-- . Then one concludes that L,ELN , and
(4.12) A-A,, = LAN - LMAN E JO .
The theorem, below, is self-explanatory.

Theorem 4.5. If P,, is an E-inverse of A_ , then we have
(4.13) AP. = J + 1 + E , P,,A = J'+ 1 + E' , J,J'EJO , E,E'E E
with the minimal comparison algebra JO Then, if we find K(H)- .
inverses 1+F, l+F', F,F'E E of 1+E and 1+E' , respectively, we

,
get
(4.14) A(P,,(1+F))=l+K+C, ((1+F')P,)A=1+K'+C', K,K'EJO, C,C'EK(H).
Indeed, we get (A-A,)P,, E JO , P,,(A-A,,) E JO , explaining

(4.13) , and then (4.14) is evident.
It is clear now that the right hand sides of the relations
(4.14) are contained in the compact commutator algebra J=1+0
'
so that existence of K(H)- inverses for 1+K and l+K' is a matter

of checking on the symbols. In addition, a proper choice of the
cut-off function Xj in (4.10) will allows us to use any opera-
tor G. G' E 1+J0 satisfying aGa(l+K)=1, aG,a(1+K,)=1 near t=-.
Thus we have proven the result, below.
Theorem 4.6. Assume L of thm.4.2 uniformly elliptic. At each end

let L° = L(co) be the translation invariant limit operator of L,
and let A° _ (1-A°)-1/2 be the (translation invariant) inverse
comparison operator square root. Then A= LAN is Fredholm if and
only if, at each end, the translation invariant limits AA=L°A°N
are Fredholm operators of L(H°)
One will be tempted, of course, to conjugate AA with the
t-Fourier transform Ft of VIII,2. Note that this takes A° onto
the operator
(4.15) AA _ ik=OAk(TAx)kTk(t), Ak=ikLk(-)AX
with
, (1+T2Ax2)-1/2
(4.16) Ax=(l-Ax)-1/2 T(T) _
Recall also that the commutator ideal here is represented

by the class CO(1,k)+K(H°) k=K(L2(B))
, cf.VIII,2. ,
A simple left multiplication of A^ by a Fredholm inverse P° of

the elliptic operator A0 on B will convert (4.15) into the form
(4.17) PA^ = 1+K(T) , K(T) E CB(R,k) .
This, however, is no inversion mod E . Still the problem is

reduced to an equation involving compact operators over the
compact manifold B (For more facts of. [CPo7.)
.
Remark 4.7. It is clear that the operator K(T) of (4.17) is not

only continuous but even analytic in the parameter T . Using a
result of Gramsch [G2] it therefore follows that the Fredholm ope-
tor-valued analytic operator function 1+K(T) has a meromorphic
inverse if only it is invertible at a single T=T0 . Existence of
such TO indeed can be established, using the Sobolev estimate of
VI, thm.4.3, by a simple estimate, due to Agmon and Nirenberg,
of. [AN1]. Then, of course, 1+K(T) must be invertible at all
T E T near the real axis, except at most countably many without
finite cluster. In fact, it follows that there are at most fini-
tely many such points in some strip IIm Xl<n , n>0 .
We will get a Fredholm operator if none of these exceptio-

nal points lies on the real axis. While in general this may or may
not occur for a general expression L it always will be true
,
eatLe-5L
for the expression for all but finitely many 161< n, 6E]k.
X.5. Sobolev spaces 315
5. Fredhoim theory and comparison technique in Sobolev spaces.
First let us point to a serious restriction of applicabi-

lity of the present discussion. We generally assume here cdn.(s),
which is never satisfied for a comparison triple with any nonvoid
regular boundary, - i.e., if any open nonvoid subset of 2M # is
A
an n-l-dimensional manifold 30 such that dp and H extend to OL0,
satisfying all the conditions they otherwise satisfy on o .
Thus, in particular, while much of our L2-theory applies to

the regular elliptic boundary problem, with Lopatinskij-Shapiro
type boundary conditions, this is not so for theory in Sobolev
spaces.
In fact, even our definition of Sobolev spaces Hs does not
supply the commonly used Sobolev spaces if (s) is violated. In
that case there is a choice of self-adjoint extensions of H0m
for some integers m. One will get the common Sobolev space Hm
only if Am is replaced by the inverse square root of the extension
giving the largest domain for the corresponding form (u,Hmv)=
(u,v)m , which then is the Sobolev space Hs This normally does
.
not correspond to the Friedrichs extension (which gives the smal-

lest domain for the form). In case of a regular boundary the Frie-
drichs extension corresponds to the Dirichlet boundary condition,
while the Sobolev space corresponds to the Neumann condition.
In particular also, the As to be substituted for As in IX,(l.l)
no longer will be the power of a fixed operator A .
Thus we assume cdn.(s), for the remainder of sec.5. Also we

require the same conditions as in IX,4, i.e., (1.), j=1,2, (m5),
and (15)(P.) , for a suitable P. within general reach of A# D# ,
of principal symbol type. (We always have (aj), j=0,1, and (d0),
of course. )
Under these assumptions IX,thm.4.3 is valid in its strong
form. A comparison algebra Cs is defined in Hs, sEa, from the
generators IX,(1.10). Any two such comparison algebras Cs, C. are
isometrically isomorphic, under the natural isomorphism Ar:Hs-Ht
r=t-s. For AECs we have A-rAAr E Ct , and vice versa. The commu-
tator ideals also map onto each other, under this conjugation.
The symbol spaces are homeomorphic under the associate dual
map of the isomorphism Cs/Fs - Ct/Et induced. Also, for A E CO
and even for A finitely generated by LAN , ANL , LC-QN=PN+LN '
we have A-rAAr-A E E , so that the symbol of A in Hs coincides

with the symbol of A in Ht, if the associate dual map is used to
identify the two symbol spaces. With that interpretation we obtain
a common symbol space 1 for all algebras Cs , and the symbol aA
is independent of s, whenever AEC, , or A is finitely generated
from LAN , ANL , for LE `PI
, and CO
Taking closure in the topological algebra 0(0) we obtain

the Frechet algebra C, Since convergence Aj-A in C. implies con-
.
vergence A.-A in every Cs , it follows that every AEC, has its

symbol independent of s . Accordingly we define a symbol for the
Frechet algebra C. by setting aA equal to the symbol of AECcCCS
in Cs , for any fixed s, this definition being independent of s.
We also get any two algebras Es/K(Hs) Et/K(Ht) , s,tEg, ,
isometrically isomorphic, under the same conjugation with

Ar , r=t-s Thus, if the algebra E=E0 has a symbol of the second
.
kind, i.e., if E/K(H) is isometrically isomorphic to C(E,K(h)),

as in the examples of VIII,l, 2, and 4, then the same holds true
for all the algebras Es/K(H5), sEE . We get Es/K(H5)=C(E,K(h)).
If y:E - K(H) denotes the symbol homomorphism of E , then the Cs-
symbol yE of an operator E Cs is given by
(5.1) YE = YE , Es = A-SEAS E C
S
where we know from IX,(4.4), here valid in its strong form, that
ESE E so that yE is defined.
,
It now is natural to focus on the commutator ideal EO of the

finitely generated algebra C2 Note that Es is the closure of EO
.
in L(Hs), s E E and in 0(0)

, for s=- . For an operator EEEO we
again have yE defined for all sEE , and the same is true for EEE_,
by continuous extension in 0(0) . While we found the Cs-symbol of
an AEC. independent of s , in all cases examined, the correspon-
ding property for the Ea-symbols would depend on the relation
(5.2) A_r EAr - E E K(H) for all EE E0, rE E .
A check on our examples in sec.VIII,1,2,4 , and [C1],V reveals

that (5.2) normally is not satisfied.
As a substitute for 'yE=const.' , using V,thm.6.8, it fol-
lows that Es is an entire analytic function of s , defined for all
real and complex s, and taking values in E for all s . All deri- ,
vatives of Es (for s) again are in E .

X.S. Sobolev spaces 317
Then, using that the symbol homomorphism y: E/K(H)-+C(E,K(h))

is an isometry, one concludes that also yE is an entire function
s
of s . Indeed, we get, with Fs=dEs/ds ,
(5.3) supPYE _YE -(s-s')YF II < IIEs-Es,-(s-s')Fs,u = o(Is-s'I).

s s' s'
showing that the complex derivative of yE for s exists and is
s
a continuous function over E , with values in K(h) again.

Now, for a general operator E E E , the Frechet closure of
E0 in 0(0) , we obtain E as limit of a sequence E E E0 , for which
i
the sequence of symbols yE converges uniformly in s on compact ,
is
subsets of I1: . Therefore yE is entire again. We have proven:
s
Proposition 5.1. Under the assumptions of sec.5 the symbol ys(e),

for EEE. eEE , is an entire function of s with values in K(h),
,
continuous in s and e , and CO(E) , for fixed s. Moreover, YE even

is entire as a function E-+CO(E,K(h)). For fixed eEE and EEE' the
(entire Fredholm-operator-valued) function 1+yE(e), either has a
nontrivial kernel for every (real and complex) s, or else has a
meromorphic inverse, singular only at discrete points of M
We have the following analogue of thm.4.l:
Theorem 5.2. Under the above assumptions, an operator A E Cs has

an Es-inverse if its (primary) symbol aA does never vanish on the
space ffi. In case of AEC. the primary symbol is independent of s
so that an E -inverse exists either for no s or for all s E E .
An 'Es-inverted' operator K=1+E , EEEs is Fredholm if and

only if its Es-symbol 1+ys is nonsingular for all eEE, and its
inverse is in 1+CO(E,K(h)).
A Fredholm operator of Cs must have an Es-inverse, hence
a non-vanishing primary symbol. Moreover, it must have an Es-
inverse B such that AB and BA have a non-singular Es-symbol on
all of E .
Among the examples offered one finds that the Hilbert space
h of IX,(4.8) always has a concrete meaning as a factor in a topo-
logical tensor decomposition
(5.4) H = H0= h»h' ,

where then the E-symbol yE of a generator EEE0 may be easily cal-
X.S. Sobolev spaces 318
culated from that decomposition. Generally the operators As do not

separate relative to (5.4).
For example, in the algebra of VIII,2 (with generators
VIII,(2.2)), setting n'=1, we have It = hx , h'= ht and ,
(5.5) As = (1 - a2r _ Ax)-s/2
with at Ax
, but not As acting on the factor spaces. After an
,
Ft-conjugation, as in VIII,2, we get the operator function
(1+'r2Ax)-1/2
(5.6) MS(T) = AxsTs(T) E K(h), with T(T) _
acting as multiplication operator on the factor h' Here .
it may be observed that Ms(T) provides an isomorphism It -+ hs
(although not an isometry), depending on T , for each fixed T

where, for a moment, hs denotes the L2-Sobolev space of order s
of the compact manifold B , with respect to the metric on hand.
Thus Hs , in this case, corresponds to the class of all
hs-valued functions u(T) over I with f- (IM-s(T)u(T)II2dT < - .
One confirms easily that, in this case the symbol 1E for ,
an EEE0 assumes values in the algebra CB

, the , finitely genera-
ted algebra within the unique Laplace comparison algebra CB of the
compact space B Thus 1+yE(e) , for fixed eEE ,
. is an operator
of 0(0) , relative to B , and is K-invariant under H-conjugation,
in the sense of I,def.6.5, in view of IX,(4.4), and (5.6). Accor-
dingly, I,thm.6.6 applies, and it follows that ker(l+ys) and
ker (1+yE) have dimension independent of s . Accordingly, in this
case, although yE is not independent of s , it still follows that
l+yE(e) is invertible for one s=s0 if and only if it is invertible
for all s .
To apply the last observation for construction of Fredholm

inverses of an AE C_ we would require an E--inverse B of A We .
will not attempt its construction here, and leave further inves-
tigations to the reader.
Instead we now get restricted to the case E=K(H) . We know
that then also Es = K(Hs), for sEl by IX,thm.4.2.
,
As in VI,3 two operators A, B E 0W are called Green-

inverses of each other, if we have
(5.7) AB - 1 E F. , BA - 1 E F. ,
where F_ denotes the ideal of operators of finite rank in 0(-°),

as introduced in I, problem 6.8. There we stated that the 0(0)-
closure FW coincides with the class of K_ of operators AE0(0) with
AEK(H ) for all sEl .
s
Starting with any comparison algebra C. = C.(A!,D!) , under
the assumptions of the present section (with C of compact commu-
tator), we define the classes (subspaces of 0(s))
(5.8) 'PCs = {AA s :

AEC_} , 'PC. = U{'PCs: sEg} .
It then follows that 'PC. is a graded algebra containing the

(differential operators of the) expressions in 9 within general
reach. In details, we have
(5.9) 'PCs.TCt C 'PCs+t s,t,El a QN C TCN , N=0,1,2,... .
Also TC0D K_ D Fw .
K=AA-s
For an operator E 'PCs the symbol aA of A is called
the symbol quotient of order s.
Theorem 5.3. An operator KE'PCs admits a Green inverse in 0(-s)

if and only if its symbol quotient of order s does never vanish
on the symbol space M . Moreover, any such Green inverse is con-
tained in TC-s .
The essential point of the proof is that operators in TCO=Cm

are K-invariant under H-conjugation, in the sense of I,def.6.5, so
that I,thm.6.6 applies for all Fredholm operators. This in essence
implies the statement for s=0, while the general case is easily
reduced to s=0. For details cf. [C1],IV,thm.4.1.
Theorem 5.4. Under the present assumptions, with a compact commu-

tator algebra C, the algebra C. is a i -subalgebra of L(H)
Proof. Suppose A G C_ admits an inverse B E L(H) Then it is .
Fredholm in C , hence aA90 on M By thm.5.2 it then admits a

.
Green inverse BO E C . In fact, the proof of I,thm.6.6 shows that

the distinguished Green inverse B0:(ker A)1 im A of VI,3, with the
orthogonal complement in H. is in C. . But that must be B, since
im A=H , ker A = 0 , A being invertible, q.e.d.
App.A. Functions on manifolds 320
Appendix A. Auxiliary results, concerning functions on manifolds.
Let the manifold 0 satisfy the general assumptions of III,1.

In particular we assume the existence of a countable locally fini-
0jclos
te atlas {Qj j=1,2,...} , where each
: is compactly contai-
ned in some U. , where {U.) is another locally finite atlas of 0
Let 0` be an open subdomain of 0 (where 0 _ 0- is permitted.)
Suppose f(x) and g(x) > 0 are functions over U- and 0 , respecti-
vely. We will use the Landau symbols in the following sense:
Write f=O(g) (in c') if f(x)/g(x) is bounded over c-; write
f=o0(g) (in D-) if f=O(g) (in 0-) and limx}.(f(x)/g(x))=0 (in U).
(That is, for e>0 there exists a compact set K C 0 such that
lf(x)/g(x)l < c for all x E 0-\K .) We shall write f=O(g) , and
f=o(g) , (without "(in 0`)" , etc.) if no confusion can arise.
Lemma A.1. Let f,g be as above, and let g be continuous over U

If f = o0(g) , then there exists a positive C-(U)-function ip
such that f = 0(*) (in R') , and V = o0(y) .
Proof. Let (x) = f(x)/g(x) , so that we have q(x) bounded over

U` and limx ¢(x) = 0 Consider. extended to 0 by setting fi(x)
= 0 outside 0, then the limit still is zero. With our partition
w. define nj = x E supp wj } Observe that nj>0, and

.
0 For there exists a compact set Ke C 0 such that

.
Ifl<e outside K6 for every e>O Only finitely many of the sets
.
0. = supp wj can have points in common with Ke Else there exists .
X. E U nK e with a limit point xE 0 Ke Due to localy finite cove-

.
3k ik
rage x0 has a neighbourhood N contained in only finitely many U.
so cannot be a limit point of the x. . Hence, for e>0 there
k
exists NE such that all supp wj , j>Ne are completely outside Ke.
Accordingly, nj = supp wj} < e .
Now we just define the function where
the sum is locally finite, hence represents a C'(0-function.

Since y is only continuous, the function $ thus defined is not yet
C_(U) but is at least continuous. It also satisfies the other
conditions: For any x E U we get (x)I = Jnjwj(x)
= p(x)/g(x) . This implies Jf(x)I = p(x) , i.e.,

App.A. Functions on manifolds 321
f = 0(p) (in SZ-) . Also we note that limx4Injwj(x) = 0 , so that
= o0(g) . Indeed, let KN = S21U....US2N Clearly KN is compact,
while x E S2\KN implies that 1j=N+lnjwj(x) <
sup{nj : j>N+l} - 0 , as N - . To fully establish the lemma we

now must make a C"-correction of 4 which does not disturb the
other conditions already established. This is accomplished in
lemma A.2, below.
Lemma A.2. Let f,g E C(Q) , g > 0 . Then there exist positive
C%1)-functions y , 6 such that
(A.1) f(x) < y(x) < f(x) + 6(x) , x E S2 , 6 = 0Q(g) .
Proof. Let ej = 2-JMin{g(x) x E .jclos} Clearly ej> 0 , since

:
.
UJclos are compact, and g(x) > 0

Using the coordinate transform .
QjclosC
of Uj we may regard S2jC Uj as subsets of in Let wj > 0 .
w. E C0(S2j) ' jj=lwj(x) = 1 in a , i.e., {wj} is a partition of
unity . Using regularizing techniques and the compactness of

clos
0 it is possible to find fjE C0'(S2 such that
(A.2) wj(x)f(x) < f.(x) < wj(x)f(x) + ej , x E S2j .
For each j let Xj(x) E CD(Uj) , 0<Xj<1 , Xj(x) = 1 , as x E S2.

Then (A.2) implies that
(A.3) f(x) < If.(x) < f(x) + I x(x) , x E S2 .
Here y(x) = Ifj(x) , and 6(x) _ I ejXj(x) are well defined
C"(c2)-functions, since the sums are locally finite. Moreover,
(A.4) 0 < 6(x)/g(x) < 2-j .

LxEsuPP(Xj)
For every N=1,2,..., a compact set KN may be found such that

supp XJ .
n KNmpl = 0 for all j < N due to local finiteness of
, ,
{UTherefore the right hand side of (A.4) is < JI+12-J=21-N
as x E Kcompl. We get lima->.,6(x)/g(x) = 0 , or, 6 = wS2(g), q.e.d.

App.B. Covarient derivatives 322
Appendix B. Covariant derivatives, and curvature.
By 'covariant derivatives' we mean the Riemannian covariant

derivatives, with respect to the connection defined by the
Christoffel symbols of the metric tensor hjk , where
(B.1) ((h 7.k )).,k=l,...,n _ ((hjk))-l .
Let bj and cj be covariant and contravariant tensors, (i.e.,

tangent or cotangent vector fields, respectively, depending
smoothly on xES2 . Then we have
(B.2) bj lk = bj
x
k - r.kbl , c] lk = Cl lxk + rmk cR1,
where the Christoffel symbols
(B.3) rjk = hilrjk,l ,

rjk,l =2(hjllxk + hlklxj - hjklxl)
are symmetric in j and k, and are not tensors. Both covariant

derivatives are tensors with two variable indices.
Covariant derivatives of tensors with multiple indices are
defined similarly. For example,
aijlk - aijlxk - aihrjk - ahjrik
aIk = a xIx +
ah hi
k + ahjr hk
i
In general,
r1r2 rI r2 rl a-1 a+l . a
as1...r jr ...
as1s2... i = aS152, +
l Ix i a ... rr1
(B.6)
ar1...
Ss1...ss-llss+l... ssi
r1
Higher covariant derivatives are defined recursively. For

example alljk = (a'lj)lk The covariant derivativative of the
:
tensor a'lj , of one covariant and one contravariant index. Etc.

For sake of completeness one defines alj = a That is, the . .
Ix
first covariant derivative of a (scalar) function a(x) is its
gradient.
Covariant derivatives do not necessarily commute, even if
the coefficients are smooth. We have
App.B. Covarient derivatives 323
(B.7)
aIik - alki '
for a scalar a . However , for a tensor ai and b1 we get
ailjk - ailkj _ a1 ijk
(B.9) biIjk - b1lki = -b11lljk '
with the Riemann curvature tensor Rijk and Ri1jk . Here we have
(B.10) Rlijk= r±klXj- rijlXk+ rijrmk' Rilkj =

hirRrlk..
Fikrmj-
Similarly
(B.11)
aijlkl - aijllk - aihRjk l + ahjRikl
etc. It is known that the metric tensor hjk as well as hjk ,
behave like constants under covariant differentiation, i.e.,
(B.12) hjkll = 0 , hjkll = 0 .
Also, covariant derivatives satisfy the usual differentiation

rules. For example,
(B.13) (ajkbJk)I1 = ajkllbjk + ajkbjk11 ,
etc. Finally we note the symmetry relations of the Riemann tensor:
(B.14) Rijkl
-Rjikl -Rijlk ' Rijkl+ Riklj+ Riljk= 0
Also the Ricci tensor will be used:
(B.15) R. . = Rijk
as well as the divergence formula
(B.16) j .
hjka,jk = x IXk
(Notice that, as for all tensors, a covariant derivative may be
'raised', giving a contravariant derivative:
(B.17) ajlk = hJlajll , etc.

App.C. Summary of cdns(x) 324
Appendix C. Summary of the conditions used.
In the following we give a list of the various conditions

used to describe features of comparison algebras and their genera-
ting sets, in alphabetical order, with the section listed, where
they were introduced.
(a0) (V,l): a E A# implies a E A# and A# D C'(c)

,
(a1) (V,1): A# contains the constant functions.

(a1') (V,1): For every D E D# there exists an a E A# such
that Va has compact support, and aD = Da = D .
(d0) (V,1): D=b33xi +pE D implies I5=F3axj+pE D#, and D con-
tains all Co-coefficient folpde's with compact support.

(11) (IX,1): A# is an algebra under the pointwise product
of functions, and D# is a Lie-algebra, under the commu-
tator product of folpde's. Also, D# is a two-sided A
module, as under (m2), and A#, interpreted as a set
of zero-order expressions, is a subset of D# , Also,
for a E A# , D E D# we have [a,D] E A#
(12) (IX,1): DED# implies that D*ED
(13) (IX,1): Cdn.(m3) holds, and, in addition, for every
a E A# with a#0 at a closed set Q C M# there exists
a function b E A# with ab=l near Q

(14) (IX,1): The expression H can be written as a finite
sum of terms aDF, with a E A# , D,F E D#
(15) (IX,1): For every a E A# D E D we have [a,H] E D
,
and [D,H] is a finite sum of terms bEF , bEA E,FED#

(15') (IX,3): We have [ H,a] E L1,0, [ H,D] E L22,0
for all a E A! , D E D!
(l5)(P.) (IX,3): Cdn.(15) with L, replaced by Qj=Lj+P]
for a graded algebra P_= UPN within reach.
(I5')(P_) (IX,3): Cdn.(15') with Lj'O replaced by Qj'O
P1.+L7,0 , with an algebra P.=UPN within reach and of
principal symbol type.
(m1) (VII,2): For each D E D# the expression
{D,D}(x) = h]k(x)5J(x)bk(x) + p(x)p(x)/q(x)

defines a function in A# .
(ml)^ (VII,4): Same as (m1), but with {D,D} replaced by

{D,D}" , formed with the coefficients of some triple

",dU",H"}
{Sj (C> {c,du,H}

(ml') (VII,3): In addition to {D,D} of (mI) also {D,D}0(x)
_ Ip(x)I2/q(x) defines a function in A# .
(VII,4): Same as (ml') , but with respect to a

triple {Y",du",H"} (cf.(ml)" ).
(mlt
(ml)"x (ml')x" (VII,4): The correspon-

(mlx , ) x , '
ding conditions, (ml),..., stated only for the given x.

(m2) (VII,2): A# and D# are complex vector spaces, and D#
is a left A#-module, i.e., aDED# for all aEA# DED# ,
(m3) (VII,3): For each x0E M and open neighbourhood N of

A#
x0 there exists a function aEA# with 0<a<l , and a=l
near x0, and a=0 outside N, and with Da=hJka 8 kED
xi x
(m3)x (VII,4): Condition (m3) imposed only for x=x0 , not
necessarily for any other xOEM # .
A
(m3) U (VIII,3): For every aE A# , DE D# with a=0 D=0 near ,
3U we have XUa E A# , XUD E D#, with the characteri-

stic function XU of U C 2.
(m4) (VII,3): There exists an e>0 such that A-1-e[a,A]E C,
and A-EID,A1= A-'(DA-(AD)**) E C, for all aE A#, DE D
(m5) (VII,3): We have A E C .
(m5') (IX,1): D# contains the 0-order folpde 1

(m6)x (VII,3): The dimension dx of S# = D#/D# , with
Dx = {DED# {D,D}(x)=0} is maximal (i.e., 6x = n+l)
:
(m6) (VII,3): Means (m6)x for all x E M

A#
(m6)" (VII,4): Cdn. (m6), but with respect to {D,D}^
formed with the coefficients of some triple
{P",du",H"} (c> {c,du,H)
6)x"
(m : Cf. (m6)" .
(mi)x (VII,3): There exists a DOE D#, such that {D,D}(x)90,

but {D,D}1(x)=0 .
(m7) (VII,3): Means cdn.(m7)x for all x E M

#
A
(m7)x" (VII,4): Means cdn. (m7)x ,but with {D,D} replaced
by {D,D}" , as n (m6)"
(m7)" (VII,4): See (mi)x" .
(p) (V,6): (Same as 'H-compatible, V,def.6.4).

(pl) ( I X , 3 ) : [H,L] E PN+l for all L E PN , N=0,1,... .
(p2) (IX,3): Every L E pN is within reach of L(H) .
(ql) (VII,4): We have q=q^ in some neighbourhood of 3o

(q2) (IX,6): The Riemann curvature tensor and all its co-
riant derivatives are bounded over 0 .
(q3) (IX,6): The potential q satisfies Vkq=O(ql+k/2), k=1,..

(s) (V,1): For all m=1,2,... the minimal operator (Hm)0
=
H0m of the m-th power HM is essentially self-adjoint.
(sj) (V,1): (where j=1,2,...,oo) (H3)0= H0J is essentially
self-adjoint.
(w) (V,1): The minimal operator H0 is essentially self-
adjoint.
List of symbols used
A<B, B>A 14 E 219,236,248

ACB, BDA 2 e(x,y) 62/63
A# 126,129 E(A) 15
AC 213 graph A 2
A# 144 mn 199
c
En
198 Hd 78
blj , blb 322f Hs, H. 30,157,160,251
B(x,q) 86 H°° , Hm 303
B 78 i , 10 129
<c), (c> 76 K(H) 24
cdn.(x3.) 324f L(H) 2
CB , CB(X), CB(X,X) 229 ffi 131

CO , CO(X), C0(X,X) 230 ffis , fp 189
C 129 M # 187
A
CO 129
C 257
0(g) , o0(g) 320
S <p), (p> 115
C 271
C0
P(X,Y) , P(X) 2
257
E*0 187,206
C A# 187
<q), (q> 115
D# 126,129 2(X,V) , Q(X) 2
D 213 Rs(A) 12
D# 144 Sp(A) 13
c
d(x,y) 86 Sx 194
dom A 2 T 0 , T 0x 161
D
x
194 w 161
dx 194 (x.) 324f
{D,F} 193 ( x) 146
E 126
References
[AN1] S.Agmon and L.Nirenberg, Properties of solutions of ordinary
differential equations in Banach space; Commun. Pure
Appl. Math. 16 (1963) 121-239.
[Al1] W.Allegretto, On the equivalence of two types of oscilla-
tions for elliptic operators; Pac.J.Math. 55(1974)
319-328.
[Al2] , Spectral estimates and oscillation of singular dif-
ferential operators; Proc.Amer.Math.Soc.73(1979)51.
[Al3] , Positive solutions and spectral properties of
second order elliptic operators;Pac.J.Math.92(1981)
15-25.
[APS] M.Atiyah, V.Patodi, and I.Singer, Spectral assymmetry and
Riemannian geometry; Math. Proc. Phil. Cambridge
Phil. Soc. 77 (1975) 43-69.
[AS] M.Atiyah and I.Singer, The index of elliptic operators; I:
Ann. of Math. 87 (1968) 484-530; III: Ann. of Math.
87 (1968) 546-604; IV: Ann. of Math. 92 (1970) 119-
138; V: Ann of Math. 92 (1970) 139-149,
[BMSW] B.Barnes, G.Murphy, R.Smyth, T.West, Riesz- and Fredholm
theory in Banach algebras; Pitman, Boston 1982.
[BDT] P.Baum, R.Douglas, M.Taylor, Cycles and relative cycles in
analytic K-homology; to appear.
[B1] R.Beals, A general calculus of pseudo-differential operators;
Duke Math.J.42 (1975) 1-42.
[B2] , Characterization of pseudodifferential operators and
applications; Duke Math.J. 44 (1977) ,45-57 ; cor-
rection, Duke Math.J.46 (1979),215 .
[B31 , On the boundedness of pseudodifferential.operators;

Comm.Partial Diff.Equ. 2(10) (1977) 1063-1070.
[Bz11 I.M.Berezanski, Expansions in eigenfunctions of self-adjoint
operators; AMS transl. of math. monographs, vol.17,
Providence 1968.
[BW1] B.Booss and K Wojciechowski, Desuspension of splitting
elliptic symbols I,II; to appear.
[BdM1] L.Boutet de Monvel, Boundary problems for pseudo-differen-
tial operators; Acta Math. 126 (1971) 11-51.
[BS1] J.Bruening and R.Seeley, Regular singular asymptotics; Adv.
Math. 58 (1985) 133-148.
[BS2] , The resolvent expansion for second order regular
References 329
singular operators; Preprint Augsburg 1985.

[BS3] , An index theorem for first order regular singu-
lar operators; Preprint Augsburg 1985.
[BC1] M.Breuer and H.O.Cordes,On Banach algebras with a-symbol;
J.Math.Mech. 13 (1964) 313-324 .
[BC2] part 2 ; J.Math.Mech.14 (1965) 299-314

,
[Ca] A.Calderon, Intermediate spaces and interpolation,the complex

method; Studia Math.24 (1964) 113-190.
[Ca1] T.Carleman, Sur la theorie mathematique de 1'equation de
Schroedinger; Arkiv f. Mat., Astr.,og Fys. 24B
Nil (1934).
[Ct1] M.Cantor, Some problems of global analysis on asymptotically
simple manifolds; Compos. Math. 38 (1979) 3-35.
[CV1] A.Calderon and R.Vaillancourt,On the boundedness of pseudo-
differential operators; J.Math.Soc.Japan 23
(1971) 374-378.
[CV2] , A class of bounded pseudodifferential operators;
Proc.Nat.Acad.Sci.USA 69 (1972) 1185-1187.
[Che1]J.Cheeger, Spectral geometry of spaces with cone-like sin-
gularities;Proc.Nat.Acad.Sci. USA (1979) 2103.
[CGT] , M.Gromov, and M.Taylor, Finite propagation speed, ker-
nel estimates for functions of the Laplace opera-
tor, and the geometry of complete Riemannian mani-
folds; J. Diff. Geom. 17 (1982) 15-53.
[Ch] P.Chernoff, Essential selfadjointness of powers of generators
of hyperbolic equations ; J.Functional Analysis
12 (1973) 402-414 .
[CBC] Y.Choquet-Bruhat,and D.Christadolou, Elliptic systems in

Ns 6-spaces on manifolds, elliptic at infinity;
Acta Math.146 (1981) 129-150.
[Cb1,2 1 L.Coburn, The C -algebra generated by an isometry, I:Bull.
Amer.Math.Soc.73 (1967) 722-726; II: Transactions
Amer.Math.Soc.137 (1969) 211-217.
[CL1] L.Coburn and A.Lebow, Algebraic theory of Fredfholm opera-
tors; J.Math.Mech.15,577-584.
[CM] R.Coifman and Y.Meyer, Au dela des operateurs pseudodifferen-
tielles; Asterisque 57 (1979) 1-184.
[CC1] P.Colella and H.O.Cordes,The C -algebra of the elliptic
boundary problem;Rocky Mtn.J.Math. 10 (1980)
217-238 .
References 330
[Cd1] E.A.Coddington,The spectral representation of ordinary self-

adjoint differential operators;Ann.of Math.60
(1954) 192-211.
[CdL1] and N.Levinson, Theory of ordinary differential
equations,McGraw Hill,New York 1955.
[C1] H.O.Cordes,Elliptic pseudo-differential operators,an abstract
theory.Springer Lecture Notes in Math.Vol.756,
Berlin,Heidelberg,New York 1979.
[C3] ,Techniques in pseudodifferential operators;mono-
graph,to appear.
[Cu] , On some C -algebras and Frechet-*-algebras of
pseudodifferential operators; Proc.Sympos.P.Appl.
Math.43 (1985) 79-104.
[CQ] , A matrix inequality, Proc.Amer.Math.Soc.ll (1960)
206-210.
[CC] , On compactness of commutators of multiplications
and convolutions,and boundedness of pseudodiffe-
rential operators; J.Functional Analysis,l8
(1975) 115-131.
[CD] , A pseudo-algebra of observables for the Dirac
equation; Manuscripta Math.45 (1983) 77-105.
[CE] , A version of Egorov's theorem for systems of
hyperbolic pseudo-differential equations;J.of
Functional Analysis 48 (1982) 285-300.
[CEd] , Ueber die Eigenwertdichte von Sturm-Liouville
Problemen am unteren Rande des Spectrums; Math.
Nachr.l9(1958), 64-71.
[CF] , A pseudodifferential Foldy-Wouthuysen transform;
Comm.in Partial Differential Equations, 8(13)
(1983) 1475-1485.
[CFo] , On maximal first order partial differential ope-
rators; Amer.J.Math.82 (1960) 63-91.
[CG] , On geometrical optics;lecture notes,Berkeley,1982
[CI] , The algebra of singular integral operators in in;
J.Math.Mech.14 (1965) 1007-1032.
[CL] , On pseudodifferential operators and smoothness of
special Lie-group representations;Manuscripta
math. 28 (1979) 51-69 .
[CP] , A global parametrix for pseudodifferential

operators over In ;Preprint SFB 72 (Bonn) (1976)
References 331
(available as preprint from the author).

[CS] , Banach algebras,singular integral operators and
partial differential equations;Lecture Notes
Lund,1971.(Available from the author).
[CTJ , An algebra of singular integral operators with
two symbol homomorphisms;Bulletin AMS ,75 (1969)
37-42 .
[CWS] , Selfadjointness of powers of elliptic operators

on noncompact manifolds;Math.Ann. 195 (1972)
257-272
[CPo] , On the two-fold symbol chain of a C -algebra of
singular integral operators on a polycylinder;
Revista Mat.Iberoamericana 1986, to appear.
[CDgl] , and S.H.Doong, The Laplace comparison algebra of
a space with conical and cylindrical ends; To
appear; Proceedings, Topics in pseudodifferential
operators; Oberwolfach 1986; Springer LN.
[CEr] _,and A.Erkip, The N-th order elliptic boundary problem
for non-compact boundaries;Rocky Mtn.J.of Math.
10 (1980) 7-24 .
[CHel] ,and E.Herman , Gelfand theory of pseudo-differential

operators; American J.of Math. 90 (1968)681-717.
[CHe2] Singular integral operators on a half-line;
Proceedings Nat. Acad. Sci. (1966) 1668-1673.
[CLa] , and J.Labrousse, The invariance of the index in the
metric space of closed operators;J.Math.Mech.
15 (1963) 693-720.
[CM1] and R.McOwen The C -algebra of a singular elliptic
,
problem on a non-compact Riemannian manifold;

Math.Zeitschrift 153 (1977) 101-116 .
[CM2J Remarks on singular elliptic theory for

complete Riemannian manifolds; Pacific J.Math.
70 (1977) 133-141 .
[CMe1] , and S.Melo, An algebra of singular integral opera-

tors with coefficients of bounded osecillation;
To appear.
[CSch] and E.Schrohe, On the symbol homomorphism of a
certain Frechet algebra of singular integral ope-
rators; Integral equ. and operator theory (1985)
120-133.
References 332
[CW] _,and D.Williams, An algebra of pseudo-differential ope-

rators with non-smooth symbol ;Pacific J.Math.
78 (1978) 279-290 .
[CH] R.Courant and D.Hilbert,Methods of mathematical physics,I:

Interscience Pub. ltd. New York 1955; II: Inter-
science Pub. New York 1966.
[Dd1] J.Dieudonne, Foundations of modern analysis,Acad.Press
New York 1964 .
[Do1] H.Donnelly, Stability theorems for the continuous spectrum

of a negatively curved manifold; Transactions
AMS 264 (1981) 431-450.
[Dgl] R.Douglas, Banach algebra techniques in operator theory;
Academic Press, New York 1972.
[Dg2] , Banach algebra techniques in the theory of
Toeplitz operators, Expository Lec. CBMS Regional
Conf. held at U. of Georgia, June 1972, AMS,
Providence, 1973.
[Dg3] , C -algebra extensions and K-homology; Princeton
Univ.Press and Tokyo Univ.Press, Princeton 1980.
[DG1] J.Duistermat and V.Guillemin, The spectrum of positive ellip
tic operators and periodic bicharacteristics;
Inventiones Math. 29 (1975) 39-79.
[DS1,2,3 1 N.Dunford and J.Schwarz, Linear operators, Vol.I,II,III,
Wiley Interscience,New York 1958,1963,1971.
[Dx1] Dixmier, Les C-algebres et leurs representations; Gauthier-
Villars, Paris 1964.
[Dx2] Sur une inegalite de E.Heinz, Math.Ann.l25 (1952)
75-78.
[Dv1] A.Devinatz, Essential self-adjointness of Schroedinger ope-
rators; J.Functional Analysis 25 (1977) 51-69.
[Du1] R.Duduchava, On integral equations of convolutions with dis-
continuous coefficients; Math. Nachr. 79 (1977) 75-98.
[Du2] , An application of singular integral equations to
some problems of elasticity; Integral Equ. and
Operator theory 5 (1982) 475-489.
(Du3] , On multi-dimensional singular integral operators,I:
The half-space case; J.Operator theory 11 (1984)
41-76 II: The case of a compact manifold; J.
;
Operator theory 11 (1984) 199-214.

[Du4] , Integral equations with fixed singularities; Teub-
References 333
ner Texte zur Math.,Teubner,Leizig,1979.

[D11 J.Dunau ,Fonctions d'un operateur elliptique sur une variete
compacte; J.Math.pures et appl. 56 (1977) 367-391
[Dyl] Dynin, Multivariable Wiener-Hopf operators; I: Integ.eq. and
operator theory 9 (1986), 937-556; II: To appear.
[Ei11 L.P.Eisenhart, Riemannian geometry; Princeton 1960.
[Er1] A.Erkip, The elliptic boundary problem on the half-space;
Comm. PDE 4(5) (1979) 537-554.
[Er2] , Normal solvability of boundary value problems in
half-space; to appear.
[Es1] G.Eskin, Boundary problems for elliptic pseudo-differen-
tial equations; AMS Transl. Math. Monogr. 52
Providence 1981 (Russian edition 1973).
[Fe I] C.Fefferman, The multiplier problem for the ball; Ann. Math.
94 (1971) 330-336.
[F11 J.Frehse,Essential selfadjointness of singular elliptic
operators;Bol.Soc.Bras.Mat. 8.2(1977),87-107
[Fr1] K.O.Friedrichs, Symmetric positive linear differential equa-
tions; Commun.P.Appl.Math.11 (1958) 333-418.
[Fr21 Spectraltheorie halbbeschraenkter Operatoren
Math.Ann.109 (1934) 465-487 ,685-713 and
110 (1935) 777-779 .
[FL] K.O.Friedrichs and P.D.Lax,Proc.Sypos.Pure Math.Chicago,1966.

[Fh11 I.Fredholm, Sur une classe d'equations fonctionelles; Acta
Math.27 (1903) 365-390.
[Ga1] L.Garding,Linear hyperbolic partial differential equations
with constant coefficients;Acta Math.85 (1-62) 1950
[Gf1] M.P.Gaffney, A special Stoke's theorem for complete Rieman-
nian manifolds; Annals of Math. 60 (1954) 140-145
[GT1] D.Gilbarg,and N.Trudinger, Elliptic partial differential
equations of second order; 2-nd edition; Springer
New York 1983.
[GS] I.Gelfand and G.E.Silov, Generalized functions; Vol.1,Acad.
Press, New York 1964.
[Gl1] I.M.Glazman, Direct methods of qualitative spectral analysis
of singular differential operators; translated by
Israel Prog. f. scientific transl. Davey,
New York 1965.
[Gb] I.Gohberg, On the theory of multi-dimensional singular inte-
References 334
gral operators; Soviet Math. 1 (1960) 960-963.

[GK] I.Gohberg and N.Krupnik, Einfuehrung in die Theorie der ein-
dimensionalen singulaeren Integraloperatoren;
Birkhaeuser, Basel 1979 (Russian ed. 1973).
[G1] B.Gramsch, Relative inversion in der Stoerungstheorie von
Operatoren and ip-Algebren; Math. Ann. 269 (1984)
27-71.
[G2J , Meromorphie in der Theorie der Fredholm Operatoren
mit Anwendungen auf elliptische Differentialopera-
toren; Math. Ann. 188 (1970) 97-112.
[Gr1] G.Grubb, Problemes aux limites pseudo-differentiels depen-
dent d'un parametre; C.R.Acad.Sci.Paris 292 (1981)
581-583.
[Guil] V.Guillemin, A new proof of Weyl's formula on the asympto-
tic distribution of eigenvalues; Adv. Math. 55
(1985) 131-160.
[GP1] V.Guillemin and A.Pollack, Differential topology; Prentice
Hall, 1974.
[GK] I.Gohberg and N.Krupnik,Einfuehrung in die Theorie der ein-
dimensionalen singulaeren Integrale; Birkhaeuser,
Basel,Boston,Stuttgart 1979.
[GLS] A.Grossmann,G.Loupias,and E.Stein, An algebra of pseudodif-
ferential operators and quantum mechanics in
phase space; Ann.Inst. Fourier,Grenoble 18 (1968)
343-368.
[Hgl] S.Helgason, Differential geometry, Lie groups, and symmetric
spaces; Acad. Press, New York 1978.
[Hz1] E.Heinz, Beitraege zur Stoerungstheorie der Spektralzerle-
gung; Math.Ann.123 (1951) 415-438.
[H1] E.Herman, The symbol of the algebra of singular integral
operators; J.Math.Mech. 15 (1966) 147-156 .
[H11] E.Hilb, Ueber Integraldarstellungen willkuerlicher Funktio-

nen; Math.Ann. 66 (1908) 1-66.
[HP] E.Hille,and R.S.Phillips, Functional analysis and semi-groups
Amer.Math.Soc.Coll.Publ. Providence, 1957.
[Hz1] F.Hirzebruch, Topological methods in algebraic geometry;
Springer, Grundlehren, Vol. 131, 1966, Berlin Hei-
delberg, New York.
[Ho1] L.Hoermander,Linear partial differential operators;Springer,
Berlin, Heidelberg, NewYork 1963.
References 335
[Ho] L.Hoermander, Pseudo-differential operators and hypoelliptic

equations; Proceedings Symposia in pure and appl.
Math. Vol 10 (1966) 138-183.
[Ho3] , The analysis of linear partial differential ope-
rators; Springer New York I: distribution theory
and Fourier analysis, 1983; II: diff. operators
with constant coefficients, 1983; III: pseudo-
differential operators, 1985; IV, Fourier inte-
operators, 1985.
[I1] R.Illner, On algebras of pseudodifferential operators in
LP (,n); Comm.Part.Diff.Equ. 2 (1977) 133-141.
[Kpl] I.Kaplanski, The structure of certain operator algebras;
Transactions Amer.Math.Soc.70 (1951) 219-255.
[K1] T.Kato, Perturbation theory for linear operators; Springer
New York 1966.
[K2] , Perturbation theory for nullity, deficiency and
other quantities of linear operators; Journal
d'Analyse Math. 6 (1958) 261-322.
[K3, , Notes on some inequalities for linear operators;
Math.Ann.125 (1952) 208-212.
[Ke1] J.L.Kelley, General topology; Van Nostrand, Princeton 1955.
[KG1] W.Klingenberg, D.Grommol, and W.Meyer, Riemannsche Geometrie
in Grossen; Springer Lecture Notes Math. Vol.
55 New York 1968.
[Ko1] K.Kodaira, On ordinary differential equations of any even
order and the corresponding eigenfunction expan-
sions; Amer.J.Math. 72 (1950) 502-544.
[Kr1] M.G.Krein,On hermitian operators with direction functionals;
Sbornik Praz. Inst. Mat. Akad. Nauk Ukr. SSR
10 (1948) 83-105.
[Ku1] H.Kumano-go, Pseudodifferential operators; MIT-Press 1982.
[Le1] N.Levinson, The expansion theorem for singular self-adjoint
linear differential operators; Ann.o.f Math.ser.2
59 (1954) 300-315 .
[Lk] R.Lockhart, Fredholm properties of a class of elliptic ope-

rators on noncompact manifolds; Duke Math.J. 48
(1981) 289-312.
[LMgl] J.Lions and E.Magenes, Non-homogeneous boundary value
problems and applications; I, II, Springer,
References 336
New York 1972.

[LM] R.Lockhart and R.McOwen, Elliptic differential operators
on noncompact manifolds; Acta Math.151 (1984)
123-234.
[Lw1] K.Loewner, Ueber monotone Matrixfunktionen; Math.Zeitschr.
38 (1934) 177-216.
[MaM] R.Mazzeo and R.Melrose, Meromorphic extension of the resol-
vent on complete spaces with asymptotically con-
stant negative curvature; preprint Dept.Math.MIT,
1986.
[M0] R.McOwen, Fredholm theory of partial differential equations
on complete Riemannian manifolds; Ph.D. Thesis,
Berkeley 1978.
[M1] R.McOwen Fredholm theory of partial differential equations
on complete Riemannian manifolds;Pacific J.Math.
87 (1980) 169-185 .
[M2] R.McOwen On elliptic operators in In ;Comm. Partial Diff.

equations, 5(9) (1980) 913-933.
[MS1] E.Meister and F.O.Speck, The Moore-Penrose inverse of Wiener
Hopf operators on the half-axis and the quarter
plane; J. Integral equations 9 (1985) 45-61.
[Me1] R.Melrose, Transformation of boundary problems, Acta Math.
147 (149-236).
[MM1] R.Melrose and G.Mendoza,Elliptic boundary problems on spaces
with conic points; Journees des equations diffe-
rentielles, St. Jean-de-Monts, 1981.
[MM2] , Elliptic operators of totally characteristic type;
MSRI-preprint, Berkeley, 1982.
[Mi1] J.Milnor, Topology from a different viewpoint; Univ.of Vir-
ginia press, 1965.
[Ms1] J.Moser, On Harnack's theorem for elliptic differential
equations; Comm.P.Appl.Math.l4 (1961) 577-591.
[MP1] W.Moss and J.Piepenbrinck, Positive solutions of elliptic
equations; Pac.J.Math.55(1978) 219-226.
[NSz1] B.v.Sz.Nagy, Spektraldarstellung linearer Transformationen
des Hilbertschen Raumes; Springer, New York 1967.
[Nb1,2 1 G.Neubauer, Index abgeschlossener Operatoren in Banach-
raeumen I: Math.Ann.160 (1965) 93-130; II:Math.
Ann. 162 (1965) 92-119.
[vN1] J.v.Neumann,Allgemeine Eigenwerttheorie Hermitescher Funk-
References 337
tionaloperatoren; Math.Ann.102 (1929) 49-131.

[Ne1] M.A.Neumark, On the defect index of linear differential ope-
rators; Doklady Akad.Nauk SSSR 82 (1952) 517-520.
[Ne2] , Lineare Differentialoperatoren; Akademie Verlag
Berlin 1960.
[Nt1] F.Noether, Ueber eine Klasse singulaerer Integralgleichun-
gen; Math. Ann. 82 (1921) 42-63.
[NW1] L.Nirenberg, and H.F.Walker, The null spaces of elliptic
partial differential operators in Rn J. Math.
;
analysis and Appl. 42 (1973) 271-301.

[Ppl] J.Piepenbrink, Nonoscillatory elliptic equations; J.Diff.Eq.
15(1974) 541-550.
[Pp2] , A conjecture of Glazman; J.Diff.Eq.24(1977) 173-176
[P1] S.C.Power, Fredholm theory of piecewise continuous Fourier
integral operators on Hilbert space; J. Operator
theory 7 (1982) 52-60 .
[PW1] M.Protter, and H.Weinberger, Maximum principles in diffe-

rential equations; Prentice Hall, Englewood
Cliffs 1967.
[RS] M.Reed and B.Simon, Methods of modern mathematical physics;
Academic Press, New York. I: Functional analysis,
1975; II: self-adjointness and Fourier analysis,
1975; III: Scattering theory, 1978; IV: Spectral
theory, 1978; V: Perturbation theory, 1979
[Re1] F.Rellich, Halbbeschraenkte Differentialoperatoren hoeherer
Ordnung; Proceedings International Kongr. Math.
Amsterdam 1954 , vol.3, 243-250.
[RS1] S.Rempel and B.W.Schulze, Parametrices and boundary symbo-
lic calculus for elliptic boundary problems
without the transmission property; Math.Nachr.
105 (1982) 45-149.
[R1] C.E.Rickart, General theory of Banach algebras; v.Nostrand
Princeton 1960.
[Ril] F.Riesz, Ueber die linearen Transformationen des komplexen
Hilbertschen Raumes;Acta Sci.Math.Szeged 5 (1930)
23-54.
[RN] F.Riesz and B.Sz.-Nagy, Functional analysis; Frederic Ungar,
New York 1955.
[Rol] D.Robert, Proprietes spectrales d'operateurs pseudo-diffe-
rentiels; Comm. PDE 3(9) (1978) 755-826.
References 338
[Sa1] D.Sarason, Toeplitz-operators with semi-almost periodic

symbols; Duke Math.J. 44 (1977) 357-364.
[Schu1] W.Schulze, Ellipticity and continuous conormal asympto-
tics on manifolds with conical singularities; to
appear; Math. Nachrichten.
[Schu2] , Mellin expansions of pseudo-differential opera-
tors and conormal asymptotics of solutions; Pro-
ceedings, Oberwolfach conference on Topics in
pseudodifferential operators 1986; to appear.
[Sk1] S.Sakai, C -algebras and W -algebras; Springer, Berlin
New York 1971.
[Ro1l W.Roelcke, Ueber Laplaceoperatoren auf Riemannschen Mannig-
faltigkeiten mit diskontintuierlichen Gruppen;
Math.Nachr.21 (1960) 132-149 .
[Schr1] E.Schrohe, The symbol of an algebra of pseudo-differential

operators; to appear.
[Schr2] , Potenzen elliptischer Pseudodifferentialoperato-
ren; Thesis, Mainz 1986.
[Se1] R.T.Seeley, Topics in pseudo-differential operators, CIME
Conference at Stresa 1968.
[Se2] Integro-differential operators on vector bundles;
Transactions AMS 117 (1965) 167-204 .
Ca
[S1] H.Sohrab, The -algebra of the n-dimensional harmonic
oscillator; Manuscripta.Math.34 (1981),45-70 .
[S21 C -algebras of singular integral operators on the

line related to singular Sturm-Liouville pro-
blem; Manuscripta Math. 41 (1983) 109-138.
[S3] , Pseudodifferential C -algebras related to Schroe-
dinger operators with radially symmetric poten-
tial; Quarterly J.Math.Oxford 37 (1986) 105-115.
[So] A.Sommerfeld, Atombau and Spektrallinien; Vieweg,Braunschweig
1957
[Sp1] F.O.Speck, General Wiener-Hopf-factorization methods; Pitman
London 1985.
[St1] M.H.Stone, Linear transformations in Hilbert space; Amer.
Math. Soc. Coll. Publ. New York 1932 .
[T1] M.Taylor , Pseudo-differential operators; Springer Lecture

Notes in Math. Vol. 416, New York 1974 .
[T2] , Pseudodifferential operators;Princeton Univ.Press,

Princeton 1981 .
References 339
[Ti1] E.C.Titchmarsh, Eigenfunction excpansions associated with

second order differential equations;part I,2nded.
Clarendon Oxford 1062; part II, ibid. 1958.
[Tr1] F.Treves, Introduction to pseudodifferential operators and
Fourier integral operators; Plenum Press,New York
and London 1980
[Up1] H.Upmeier, Toeplitz,C -algebras on bounded symmetric
domains; Ann. of Math.(2) 119 (1984) 549-576.
[W1] A.Weinstein, A symbol class for some Schroedinger equations
In
on ; Amer.J.Math.107 (1985) 1-21.
[We1] H.Weyl, Ueber gewoehnliche Differentialgleichungen mit Sin-
gularitaeten, and die zugehoerigen Entwicklungen
willkuerlicher Funktionen; Math.Ann.68 (1910)
220-269.
[We2] , The method of orthogonal projection in potential
theory; Duke Math.J. 7(1940) 411-444.
[Wd1] H.Widom, Asymptotic expansions for pseudodifferential ope-
rators on bounded domains; Springer Lecture Notes
No 1152, 1985.
[Wi1] E.Wienholtz, Halbbeschraenkte partielle Differentialopera-
toren vom elliptischen Typus; Math.Ann.135 (1958)
50-80.
[Z1] S.Zelditch, Reconstruction of singularities for solutions of
Schroedinger equations;Ph.D.Thesis,Univ.of Calif.
Berkeley (1981).
Index
condition (xj),x=a,d,... 324

A-bounded 294 covariant derivatives 134f, 322
A-compact 294 curvature (Riemannian) 279f, 323
adjoint operator 3
adjoint relation 5 defect index 7
algebra surgery 239 defect spaces 6,38
definite case 41
boundary, regular 67 Dirichlet condition 53,68
boundary condition 41,68,315 operator 60,68
,Dirichlet 53,68 problem 87
,Neumann 315 realization 59
boundary space 44 domain (of an
bounded operator 1 unbounded operator) 2
Carleman alternatives 41 E-symbol 218f

change of dependent for periodic
variable 93f coefficients 227f
Christoffel symbols 134,322 for polycylinder
closed operator 2 algebra 235f
closing 39 elliptic differential
closure of an operator 2 expression 37
co-sphere bundle 170 equivalent triples 93
covariant derivative 134,322 essentially self-adjoint 5
commutator of differential essential self-adjointness 123
expressions 133f of H0 103, 111
, compactness 137f of H0m 105f,lll
commutator ideal E 126,131
commuting unbounded operators 17 expresssion (differential) 36
compactness criteria 71f within reach 127,155,262
of commutators 144 extension (of an operator) 2
comparison operator 61,127f
comparison algebra 127f,157 4 (property) 293
, minimal 129,161f folpde 126,129
comparison triple 128 formal adjoint 36
complete spaces, Fredholm domain 302
with cylindrical ends 246 index 33,293
341
operator 33,180,293 maximum principle 91

inverse 180,295 minimal operator 37
minimal comparison algebra 129
distinguished 34,180 multiplication operator 134
, special 295
Frehse's theorem 118f noncompact commutators,
Friedrichs extension 11,71f algebras with 218f
fundamental solution 42,78 normal forms 93,94
Sturm-Liouville 95,151
graph (of an unbounded
operator) 2 order classes 30,267,318
Green inverse 161,181,318 ordinary diff.expression 37
, distinguished 181
Green's function 78f periodic coefficients 154f,221
poly-cylinder algebra 228f
H-compatible expresssion 157,160 positive operator 14
Hs-comparison positivity of the Green's
algebra 257,267f,315f function 100,130
Harnack inequality 87f preclosed operator 3
Heinz, lemma of 24 principal part 36
hermitian operator 5 symbol type 266
HS-chain 29 symbol space 193
hull 191 projection 14
hypo-elliptic expression 37 positive square root 12
isometry, betw.Hs-spaces 31,268 reach, expression

indefinite case 41 within 127,155f,186,262f
irreducible C -algebra 130 algebra within 266
general 266
K-invariant under real operator 10
H-conjugation 33,181 realization,of a differential
expression 35,37
Landau symbols 320 regular endpoint 45
Laplace comparison operator 93 regularity, boundary 67f
limit circle case 41,53 Rellich's criterion 75,77,179
limit point case 41,53 resolvent 12
Loewner, lemma of 24 compactness 71f
expansion of
maximal null space 47 of H-compat. expr.158
maximal operator 37 formula for As 139,163
342
set 12 square root of an operator 12

Riesz-v.Neumann extens.thm.8 strongly hypo-elliptic 38
Sturm-Liouville problem 51,53
second order express.on 0 132 subextending triple 76
secondary symbol space 193 summation convention 60
rel.to formal symbol 207 surgery, algebra 239f
local charac.195,211,214 symbol of an operator 191,137
for Schroedinger opera- of an algebra 191
tors on In 215f function 170
self-adjoint operator 4 formal algebra 187,206
semi-bounded operator 10 symbol space ffi 131,189,191,145
separated boundary cdn's 48 , principal 193
singular endpoint 45,49 , secondary 193
Sobolev space 61,157,159,251f systems of differential
first 128 expressions 303f
on compact Q 157
on noncompact 0 251f tensor 132,322
of integral order 279f tensor products,
local 272 topological 226,229,303
Sobolev estimate 181f,184 two-link ideal chains 218f,309
imbedding 182,217,251
, comparison unbounded operator 1
algebra in 253,257,315 Vector bundles, compar-
spectral density matrix 52 ison algebras on 303f
spectral family 15
spectral theorem 12f wave front space 161f,169f
spectrum 13 Weyl's lemma 39,42,60
essential 265

Document - 2019-03-07T080258.777SPECTRALTEORIJA PDF

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Document - 2019-03-07T080258.777SPECTRALTEORIJA PDF

Hochgeladen von

Copyright:

Verfügbare Formate

LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES'

Managing Editor: Professor J.W.S. Cassels, Department of Pure Mathematics

4 Algebraic topology, J.F.ADAMS

CAMBRIDGE UNIVERSITY PRESS

Cambridge University Press

© Cambridge University Press 1987

This publication is in copyright. Subject to statutory exception

First published 1987

Library of Congress Cataloguing in Publication data

Cordes, H. O. (Heinz Otto), 1925-

(London Mathematical Society lecture note series ; 76)

ISBN 978-0-521-28443-1 paperback

The main purpose of this volume is to introduce the reader

son algebra of Rn and the half-space were presented in [C1] .

In the course of laying out theory of comparison algebras

recall the basic facts of theory of Fredholm operators, partly

associates. Accordingly it is almost impossible to recall in

Berkeley, September 1986 Heinz 0. Cordes

Chapter 1. Abstract spectral theory in Hilbert spaces 1

1.3. On the spectral theorem for self-adjoint

4.2. Essential self-adjointness of H0 109

9.3. Higher order differential expressions

In this chapter we give a short introduction into spectral

1. Unbounded linear operators on Banach and Hilbert spaces.

The term "(unbounded) linear operator" (between Banach spaces

from a dense linear subspace dom A of X to Y. The space dom A C X

The term "unbounded linear operator" will be used with the

(1.1) sup {UAuU/UuI : 0 9 u E dom A} < - ,

is necessasily continuous, hence admits a unique extension to X

Since unbounded linear operators are not linear maps from X

product needs the following special interpretation: For A,BEP(X,Y)

(1.2) dom (A+B) = dom A n dom B , (A+B)u = Au+Bu for u E dom(A+B),

(1.3) dom AC = {uEdom C : CuEdom Al , (AC)u=A(Cu) for u E dom AC

where it is assumed that dom(A+B) and dom AC are dense in X (or

A linear operator A E P(X,V) is uniquely characterized by

(1.4) graph A = {(u;Au) : u e dom Al

of the cartesian product XxY = {(u;v) : uEX , v(=-Y} , where (u;v)

(1.5) dom A = {uEX:(u;v)ET for some vEV} , Au=v , for u E dom A ,

and then we have graph A = T .

For two linear operators A,B E P(X,Y) we shall say that A

Notice that the cartesian product XxY of two Banach spaces is

a closed subspace of Xxy . The class of all closed operators in

In the following we will be mainly interested in unbounded

(1.6) ((u;w),(v;z)) _ (u,v) + (w,z) , I1(u;w)II = (I1 ull2+I1w11

The following facts, regarding adjoint and closure all work

we will get restricted to the case X=Y=H in all of the follo- ,

denotes the orthogonal complement in the graph space (with respect

Notice that A , if it exists, is closed, since all orthogonal

Moreover the element v thus defined for each u E dom A is uniouely

exists a sequence wk E dom A with (wk;Awk) - (0;z) Substitute

w = wk in (1.8), and conclude that (u,z) = 0 for all u E dom A

is not dense in H then there exists 0 9 z E H with ((0;z),(u;v))=

dom A is dense. This shows that then indeed A is a well defined

All continuous linear maps in L(H) are closed,hence have an

rator allows a spectral decomposition as a direct generaliza-

tion of the principal axis transformation of a symmetric matrix.

(1,9) (u,Av) _ (Au,v) for all u,v E dom A ,

(where dom A = H ).A general unbounded operator A satisfying (1.9)

More generally, two operators A, B E P(H) will be said to be

(1.10) (u,Av) = (Bu,v), for all u E dom B , v E dom A .

The above conclusion, showing that hermitian operators have

2. Self-adjoint extensions of Hermitian operators.

In this section we discuss the v. Neumann-Riesz theory of