Beruflich Dokumente
Kultur Dokumente
The books in the series listed below are available from booksellers, or, in
case of difficulty, from Cambridge University Press.
Spectral Theory of
Linear Differential Operators and
Comparison Algebras
H. 0. CORDES
University of California, Berkeley
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521284431
A catalogue record for this publication is available from the British Library
and
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
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" - ,
(1.7) graph A TA
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.
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
.
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
)
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
( 2 . 1 ) A C B = B C A
The two spaces D+=D+(A) are called the defect spaces of the closed
1.2. Riesz-v.Neumann extension 7
(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 :
=
(2.7) n(Ati)ull = (IIAuN2+llull2)1/2 hl(u;Au)H , u E dom A .
where W0 is just any isometry W0, W0 = W+nD+C V+. Then we have the
1.2. Riesz-v.Neumann extension 9
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
But we have
is, in fact
which shows that the map 4)-W04)-4), taking W0 onto Z0 ,
lows. Also it is clear that D+ and D+- have the same dimension.
We have proven:
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
1.2. Riesz-v.Neumann extension 12
tion A>c (or A<c) then A is essentially self-adjoint if(f) for some
1.3. Spectral theorem, general discussion 14
(3.7) H = im P $ ker P .
tion onto the complement (im Q)n(im P)1 . In particular this holds
if and only if
(3.8) PQ = QP = 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
and
(3.12) HA C Hu , as A < p .
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 ,
in view of the orthogonality. One may use this fact to define the
(3.20) 11
A A A
1.3. Spectral theorem, general discussion 17
(3.21) 111u112 =
(3.22) A = J-XdE(X) ,
and
(3.24) Au = J-AdE(X)u ,
U _ (T2)1/2 = 1--0(1/2)(TA2
(4.2) - 1)0
J j
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.
clo
(4.3) H = ker TA $ Um TA)s.
1.4. Proof of the spectral theorem 19
(4.5) H = Hx ® KX ,
= 2R(A+i)(A-A)R(A-i)u , uEH ,
implying 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
implying (4.14).
Next let A',AER , a'<A . For uEH., write
The first term, at right, is EHA, and the second term wEHXtnKX
using the commutativity (4.14) . Thus the second term satisfies
(4.19) HE
Aj
ull2
-< (E Aj u,AE Aj u) -< aJ IIE Ajull2 , uEH
(A-A0)-1
_ Sp(AO) .
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
, .
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 .
B1/2A3/2B = (A1/2B1/2)'`AB
that C = E L(H) The operator C is .
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
z z
we get A B v = Ait(AsBs)(Bity) well defined, since the operators
Alt and Bit are unitary, of course. We get
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) .
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
.
6. On HS-chains.
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)
,
up, as p<A , and that Ilull2 = JX2Re(s)dlluX112 <. . The same represen-
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. ,
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
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
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) ,
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
e
a(As)=dim ker As is nonincreasing. Similarly, ker(A*)-sDker(A ) -t,
1.6. HS-chains 34
In
1. Linear differential operators on a subdomain of
a a
Da = D1a1D2a2...D n , 8a = 81a1...an n
a
xa = x1a1...xn n , for x = (x1,...,xn) , etc.
(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
with the inner product and norm of the Hilbert space H = L2(C)
(1.5) Lt = Daaa
laT<N
Using Leibnitz'formula
6
*
one may write L in the form (1.1) again, with new coefficients
(1.7) aa(x)Ea # 0
ja =N
Particularly an ordinary differential expression
**
(1.13) dom LO = dom LO ®D+ ® D_
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.
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 <
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,
11.2. Boundary problems , ODE 44
where the functions X and may be kept the same. In the limit
e -> 0 , c > 0 , one obtains the equation
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 ,
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.
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
(2.16) ua"=
(u(a),...,u(N-1)(a))
, us"'= (u(g),...,u(N-1)(6))
,
(2.18)
g2N=W+
orthogonal direct decomposition $ W_ We get positive.
respectively,
II.3. Singular Sturm-Liouville problems 51
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)
1/2
(3.7) 10
X14 -< (a 4jj *'All) i/2(a
4jj
+ a
411
)
(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,
(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)),
, .
=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-
.
dary condition at 0, i.e., the condition (4.3) with a=l, bzO . Or,
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
and
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
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
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
This proves that the above isometry maps onto HP , since C0(1+)
11.4. 2-nd order spectral theory 58
is dense in H
P
Next we prove the inversion formula (4.8). From (4.25)
we get
t+d t+d
(4.31) f(x)dx = j $Q)( ) u(x;X)dx)dp(A)
t - t
We assume E C U +) , and then may interchange the integrals at
0
right, then divide by S and let d -> 0. In the limit we get (4.8).
JI
412da for a pair of left-continuous nondecreasing real-valued
-K-1a
(0.1) H = Khjk3 k + q ,
xj x
with function K,q>O , and a positive definite tensor ((hjk))
It then is self-adjoint with respect to the inner product (u,v)
we denote the Hilbert space L2(O,dp), with inner product and norm
In ch.5 we will assume that (1.4) holds, and then denote the
Friedrichs extension by H again. This will be a self-adjoint ope-
rator,satisfying H > 1 We then shall speak of a comparison ope-
.
yk=0(-1)k(lk2)(1-H-1)k
S2
h0k(u Ixj - um Ixj )(u Ixk - um Ixk )du -> 0 , m
III.1. PDE on manifolds 62
e(x,y) _ as n > 2 ,
(1.9)
= log p(x,x-y) , as n=2
III.l. PDE on manifolds 63
=(hjk(x)zjzk)1-n/2
(1.15) f(x,y,z) ,
y
where fj again has all the properties of (1.14) , with the same
homogeneity degree as f . This follows, because a function a(x-y)
III.1. PDE on manifolds 64
(1.18) v(a)(x) = I
a<B,BAa
where all ea,$ are of the form (1.14) (degree 2-n) In particular .
(1.20)
x .(y)dy
with h0 , h0 again of the form (1.14) (degree 2-n) .
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 ,
(1.24) u(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
III.1. PDE on manifolds 66
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
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
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- .
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
Note that the extended H maps even functions into even functions
and odd functions into odd functions. Since (the extended) f is
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-
.
g = 0 so that we have
,
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.
(3.1) 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
product, to free the term (uv) from its gradient. For more detail,
cf. the derivation of formula (6.6), below.
If we assume that
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.
Proof. We get
J82
JaA2h3k(Bu) (Bu) kdu = Uy(AA6u)112+ K(Khjka klul2dp.
Ix7 IxJ
Ix Ix
Here we estimate the last term at right by
Ipvl2 = j=llaxjv12 ),
with the Riemmannian metric of In , (i.e.,
with the sums taken over all multi-indices a with aj > 1 , and
JQIVv12
(3.14) dx > C JQIv12 dx .
L-1/2w
Proof of theorem 3.1. Setting u = in (3.15) , (3.16) we get
IIL-1/2ITBI <
B-1/2 , TB = (L-1/2{zl,...,zM})1
(3.20)
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. ,
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
xj x
over In , where
Xhjk+(1-X)Sjk
(4.2) K'=XK+(l-X) , q'=Xq+(1-X) , h'Jk =
with
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.
,
mes the form u-Ou=Gv. Then the iterates Gk=OkG , k=l,...,n , and
0n+1,
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)
111.4. Greens function 80
exists for all x,y including x=y . It is easily seen that the
,
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,
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
111.4. Greens function 81
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
ged with the inner product, since the family {cx:xEIn} is C(In,H).
It follows that
Accordingly,
Theorem 4.4. For any triple {c2,du,H} satisfying (1.4) the Frie-
drichs extension has the form of an integral operator
0((x-y)-n) for every n>0 for k=n-1 Finally, for any x0EH , and
.
(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.
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
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},
,
(4.29) IYI2-n + Y(Y) = .2-n
where we have
0(ly15/2-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,
a0(t)=sup{Tn-5/2IY(TYO)l, Tn-3/21yoVy(Ty0)l/(n-2):
(4.32) T<t}
t5/2-na
(t) < (t,y0) < t5/2-nX+(t) ,
(4.33)
(n-2)t3/2-na+(t) (n-2)t3/2-nX-(t)
< .
(1/2)1/(n-2)n (3/2)1/(n-2)n
(4.34) < t < .
.
(4.35) 2/9(n-2)nl-n
<
6(n-2)nl-n
(4.37) In = JW (g(x0,y)-Sn2-n)Hu(y)duy
TI
n IYj
111.4. Greens function 86
(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.
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' .
(5.9)
Eajla.I2/1lajl2
J(u) = > A0 = J(p0)
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
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
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
,
.
222.5. Harnack inequality 93
hjk
the tensor on 2 suggests introduction of the Riemannian metric
((hjk))-l
(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)
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.
-
- JK-1(yKhJky )
kuv du . Accordingly, by a calculation,
Jxj j x
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.
{K2/h}-1/4
(6.9) q = Hy/y , Y =
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
111.6. Normal forms; positivity of g 96
necessarily q > 1 .
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
(6.18)
< C inf {vj(x) : x E 01} < C sup {vj(x):x E 01) = C
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
It follows that
JSt Xhkly
(6.21) J52 2
hkly
lxk vI xldp < 3 Ixkvl xl dp
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.
Theorem 6.8. Assume that condition (1.4) holds. then the Greens
function g(x,y) of the Friedrichs extension H of the minimal,
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
so that Hvll1<211uhl=c
Next we note that, for uEHOC" , we have Hvj= HG.-u-HGu = 0
(6.25) 0 , as
(This follows first for all 4EC0 and then for H1, since Ilvjlll
,
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 .
0 y0 y0
may assume that N x0rN = 0 . For any SEC-(N
0 x0
) >0 , we conclude
y0
(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
,
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 ;
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 .
y/p2>p-2Vp2/p2
(1.6) q+c0> VK near - of NO , and q+c0> near 3C ,
limx;3 (q(x)(dist(x,aC))2 = + ,
( 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-
,
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
(2.9) Vol (< ena < CenP < C(A/c)n/2 = c1An/2, as x E Ep, P > P1.
c4/X-ran/2
(2.8), for A < .
Or
c5A2-n
(2.10) A' > , as p > pl
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) < °° ,
erlr
(3.3)
j < C
+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
IV.3. Proof of theorem 1.1 113
_ 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
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
(3.6)
X2 j'2-uj'2 1
+11=0 j+1+lj-1-1
2(X.+ PI +.-1+1X1
uj)
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 ,
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
(3.14) X?/(2X )
J J -'f
<
J J J 1
a.J+1+1a.J_1_1
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
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
Taking the cube root we thus get (3.22) , also using prop.3.4.
The proof is complete.
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
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
because we have
(3.30) -(l/KT)' =
TK'/KT+1 > c- ,
t = 1/Y - 1 > 0 , c_>O .
IV.4. Frehse's theorem 118
(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.
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
n)
with the Lp(.n)-norms ( IVuI=(flu1J)1/2 , integrals over
with
2n-1/2(n-l)/(n-2)(A+/a-)1/2
(4.11) C1 =
(4.12)
&X2/Pupgn/(n-2)
< (PC2)1/PIIIVX,2/Pulip , for all 2<p<-.
J+1 J
2/pJ l
(4.18) HullL_(B ) = lim pX uh
R/2 J R/2 J J pi.
=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
(tR2)(2-p)/2pllVCll2
(4.23) 11CI2p/(2-p) < , 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)
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
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 .
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
(4.33)
Jdx1Iu(x)ln/(n-1)
<
{Jdxllulll 1Jdxldxjlul. }1/(n-1)
fl
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
,
(0.1) a , DA , aE A# , DE
(0.2) C D E D K(H) ,
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.
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
.
H-1/2 = 1k_0(_1)k(1k2)(1-H-1)k
adjoint positive square root A =
(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
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.
operators in C .
A2n
hence (A4.)2, A2, (A'. )2n in strong operator convergence.
-
,
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 .
(2.3) A = DA , B = AMA .
(2.7)cl= sup{(hijhklmikmjl+h]kFibk/Iql+hjkc3ck/IqI+IPI2/IgI2)1/2
which yields
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
, ,
(2.12) D = b33.
x + p , F = claxj + r
(2.14) rji k =
hil(h .1
k + hkl j - hjk 1
Ix Ix Ix
write
Similarly ,
hjkbi hik
= [(h'jbk Ix3 + - b3)u v du + (u,D v)
Ix3 Ixj Ix1 Ixk c
h]Jbk +
hjkbi - hik bj
(2.21) lxJ lxJ 1X3
Accordingly,
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
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)
AC2A2-E
(3.3) [a,A2] = A2-eC1A +
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
Lemma 3.3. Let 0 < s < 2, and let the constants E,r,p,t,T satisfy
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)
, 0 < s < 2
(3.9) AsR(A) = 0(lal-l+s/2) , A E r
fC3H=O(lal-1+(l+s-e-r-P)/2) IIC411=0(lal-1+(l+s-e-t-T)/2).
(3.10) ,
b01k + bk13 =
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
Lemma 3.5. Let D satisfy (3.16) for some 0<e<1 Then the .
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).
V.3. Compactness of Commutators 142
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 +
.
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 = ,
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.
[DA,As] = i/2Tr )
Xs/2R()[A2,D]AR(a) dA
r
Lemma 3.7. For some 0<E<1, let the two differential expressions
D,F , as in (2.12), satisfy the conditions
4 1q" ehikPlx-Plxk
also are bounded over 0 (or at least R\K ), and have limit zero
as x- (in 12") .
(4.8) Ix'(t)I
EV(PP
(4.9) < JJ( <11
V.4. Compact commutator algebras 147
x) 2a + X(n+2B"-X)(x)2B -
Let us first consider the case B>a+l Then the second term .
(4.12) e- naiVal =
0((x)B--n(B E+n/2)) = 0(1)
Notice that the right hand side of (4.13) is <1 , for all B"">0 ,
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
For a concrete example we now set q"=1, and then, with A#,
D# as above, introduce the classes A#=A# , and
k
b3
Ixk +b , b3q , p Ixj/,q , (bl l) Ixj/V-q , all, = 0(1),
Ixj Ixj/q I x
with any given x0 E I then the triple (5.1) assumes the form
with
dp-
= K-dy , K_(Y) =
p"'(y) _ q-(y) =
(5.12) q(x) > nk(x - a)-2 near a, or q(x) > nk(x - S)-2, near B ,
holds for q with (5.12), for nl = 12 , and (s) _ (s,) holds for
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, .
(B-X) 2X
(5.24) q0(x) = X(B-x) X-1 - = A(B-x) A-1 x.B
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)-
X X-()lj=-Na-.(D)e13X + K, where
(5.33)
A-NCO
= HkC' = im H dense in H , since H is > 1 and essentially
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
.
+ XMA2R(a)((ad H)M+1L)A2M+N+2RN+1(a) , x E r ,
Proof. We get
Note that (6.6) amounts to (6.5) for N=O. Suppose (6.5) holds
V.6. Expressions within reach 158
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 ,
(6.9) AsA =
(-1)3(s/2+j-1) ((ad H)OL)As+2j+N + RM
0
where
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
L(H) ), defined in the strip -2M-2 < Re s < M+1 and that even ,
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
.
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).
=
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
,
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.
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 ,
(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
.
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
(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 .
H-s S-1
(1.6) = m-lsin rs /(M-1 J'(R(A))mam-s-lda, s>0 , m=1,2,...
0
VI.l. Local invariance 164
,
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
with
(1.8) C(A)=A-1/2(A-1R(A))(AMA)(A-lA°)(A°-1S(A))=O(A-1/2(l+A)-1),
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
,
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
.
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 .
Jr(U1-a)-1X_1/2dA
(1.9) U-1/2 = i/2,
(1-a)-1)X-1/2dX
(1.10) U_1/2-1 = i/2n Jr ((J+l-x)-1 -
(J+1-a)-1-(1-a)-1
(1.11) = -J(1-a)-1(J+l-a)-1 E J0(K)/K, X E r
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)
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
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).
.
(1.19) VD = [S(p),D] _
(K1/2S(ir))(K-1/2[ DFK] K-1/2)(K1/2S(u)) ,
R(X)S(p)DEFS(p)R(X)v = R(X)DS(p)ES(p)FR(X)v
(1.21)
X-sp-SHaR(X)S(p)DEFS(p)R(X)HT
(1.22)
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
.
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 .
Ia(1-Xl)A(m)I=(1-Xl(t(m)))IaA(m)I<sup{IoA(m)I:1(m)EOj\Kl}.0, 1--,
(2.5) {(x,E) E 0i X In , 1 } ,
x
with a C0--function X , equal to 1 near t(m)= x , hence aA (m)
3
(u,A(-a k)Av) =
x x
- (u,A(D 3 + q¢3)Av) .
$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
(2.9) M. _ {(x,E) : x E 0i , E E n 1 } .
x]
To verify that u is 1-1 we recall that the maximal ideal mEffi.
is characterized by its corresponding homomorphism J -' M , given
VI.2. The wave front space 173
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
, ,
We now must find out how the operator A changes under conju-
ation by 0 Clearly the differential expression H goes into
.
(2.13) xi x
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
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
1
Accordingly, a Ar e(m) = m r 1a (A-s
lA ) (m) = nra (A-,
lA ) (m) . Thus
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
.
( 3 . 3 ) L = ...... , Dj
D E Dm .
(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
(3.8) L = a(x)DX
a , x E 0'
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
(3.13) 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
VI.3. Reach of the minimal algebra 179
(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
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 .
ASAA-s
(3.18) - A E K(H) .
1-ANPA-N
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.)
with the N-th derivative 4 since all terms of the Taylor se- E ,
(4.5)
cJdxlx-x0IN-n,p(x-x0) Iu(9)(x)I
Iu(x0)I <
10l <N
VI.4. Sobolev estimates 183
may be replaced by
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
A
*
injection CA# - C(& O) , since CA# also is a subalgebra of the
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
#
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,
VII.l. General symbol space 190
A = lim1;.(X1AX1+C1)
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
VII.l. General symbol space 191
A
continuous functions over 0 The maximal ideal space M
A# of C A#
.
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.
(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
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 .
1 S
(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
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
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)) ,
, 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
(2.13) A0=D0A = A ,
Aj=Sj=DjA
, nj(m)=oA (m) , j=0,...,n .
(2.14) IV=0 Aj *A 1
VII.2. Secondary symbol space 199
(1'Yv=1In.12)112
so that 0< no . Thus we have the following.
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
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#
(3.6) {n E ]Rd :
1v_lInll2 = 11 C In
0 0'
v=l,...,6 Then for all u E AC0'(c7) we have the estimate
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 .
VII.3. Stronger conditions 204
(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
(3.13)
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
.
which results in
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
(3.18) 1v=l1nv(m)I2 = 1 .
**
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 .
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 :
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 .
(3.25) ta(x,E)=a(x),
A
such that {D,D}0(x)#0 but {D,D}i(x)=0, with {D,D}j of (3.26).
,
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-*-,
S6-1
the unit sphere of (3.6) . Under the present assumptions we
(3.28) (bjnj+pno)/(hjknjnk+gn02)1/2
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
VII.3. Stronger conditions 209
we thus have
(3.29) ±{F,D}/{F,F}1/2
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 .
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
B+d-1={nEBS-1:nd>0} .
0 0 0
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.
(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
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 .
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.
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
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.
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
(4.7) q(x) > Y1(dist(x,3c))-2 , and V(log q^) = o(1) (in R^)
(4.9)
V((bJ01j/(c)/q^))=0S1^(q^E/2)
amount to
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
(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 .
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
(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
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 ,
where z is any value of the circle JzJ=1 , Re z >0 , then the ope-
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
.
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
.
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 .
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,
(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
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
,
is defined by
(1.12) = J 1dtlly°112 .
uuu2=JRlu(x)I2dx S
Thus we have proven:
We replace x-j by x and then apply the operator ' of (1.14), for
(1.18)
Note that near each x,y only one term of the infinite series 11
Moreover, using that the group Yt is not only continuous but even
CW(&) , in norm convergence, and that Y0=1, we conclude that
where
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
VIII.l. Periodic coefficients 227
2. A C -algebra on a poly-cylinder.
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 .
(2.4) sup{NA(T)N : T E 11 ,
t
X
VIII.2. Polycylinder algebra 230
Ax)-1/2
(2.6) A-(T) _ ((T)2- , TjA-(T) , DxA-(T) ,
G3=A-(T)=AT(T)=((T)2-A
(2.9)
0
[Gi,G4l=(TA»)(A"'-1[ax,A"']),
[Gi,G3]=[ax,A`],
(2.11)
px=[ax,Dx]EC'(B).
(A--1-ER(s))(LlAx)(A-1R(s)//-s)
(2.17) F(s,T) _
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
.
gives a factor A-6ECO, whence the first (2.14) is CO, not only CB.
Similarly we may use (2.9) for
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
(2.22)
IIAtnS.(r)II<(l+r+A])fl/2-1,
BAt-nS.(r)II<(l+r)'1/2-1
, Aj=(1+X2+Ot)-1/2
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),
(2.24)
0((a.)-ar-1/2(l+r)(e+S)/2-1)
IIF.(s)II =
3 3
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
.
(2.26) ax ,
A-(T0)=(-Ax+(T0)2)-1"2
,
DxA-(T0) ,
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
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
.
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 ,
(2.28) MI = {(t,T)EC(1n'xEn') :
ltI+ITI=0}
(2.29) 1 = NQ = {(t,T)E NI :
Itl=o' ,
ITI-} = 2Bn'xgn' .
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
a
Al = a(x) , a
A2
= t/( t) ' a
A3
= 0
(2.32)
with
(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
3. Algebra surgery.
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
(3.4) V. = C + E. , C = XUVjX3
VIII.3. Algebra surgery 242
(3.5)
1AXUfL(H.) = BAIIL(H)
, for all AEL(H)
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 .
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
VIII.3. Algebra surgery 243
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
I = C + i-1 f'A-112
(3.12) dXDR(a)[H,X](R(X)-Rt(A))X
0
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
'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- .
we can write
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] ).
where the 12j are mutually disjoint with compact boundaries aS2j
chosing n'=l and B=BJ . , j=1,.... N This and (3.1) describe the .
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
VIII.4. Cylindrical ends 249
(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
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 .
Fourier transform F, and its inverse F-1 Note that this triple.
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.
Clearly (m5') implies (m5), but (m5') means that A even be-
IX.l. Sobolev comparison algebras 254
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 , .
..,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
Sobolev norms are equivalent (although the inner products and Hs-
adjoints in general are different).
Hs IV = {uIV : u E Hs}
DvED#
(2.2) H = 16,X-1DvaVADX , , near x0,
(2.2), while (m6) implies (14), and even (with certain DEED#)
(2.3) H = 6=6(x)=n+l
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
(2.6)
((avX))1/2 .
As a consequence the right hand side of (2.7) may be
written in the form
(2.8) Lsa=iJ0dpX2avXDvuDXu
IN
(2.11) H = v=1D Fv
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
(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,
denote the sum (2.14) taken only over v,X with at least one >M
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
IX.2. The conditions (1) 261
in VIII,prop.3.3.)
(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
IX.2. The conditions (1) 262
[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,
,
Note that (14) not only implies (15), but even (15'). Also,
the minimal classes satisfy (la') not only (15) ,Some of the .
N N L)**A
Proposition 3.3. For L E LN 0 we have ALA E E, (A .
(3.10) P P
N M C PN+M' PNLM C PN+M' LNPM C PM+N' N,M=0,1,2....
.
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
,
4. Symbol calculus in Hs
vial algebra P. .
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
.
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
Also,
AsCrA-s
(4.7) Cs C AsC0A-s
Lj+kAN+2j+2k
and another remainder. The first terms are in E as ,
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
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,
(5.5) (Hiu,P.u) =
IIy-1K s/2ufl
(5.6) Iluus H = IIH-s/2uII , Iluns = , as uEC'(c)
K
Proposition 5.2. For a given sEI the two norms (5.6) are equiva-
IX.5. Local properties 274
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 .
(1+C)Ks/2H-s/2=Y1Ks/2YH-s/2EL(H)5 Hs/2K-s/2(1+C')
(5.9) E L(H),
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-
IX.5. Local properties 275
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). ,
with
AaRmHoK6-aS1PY)(Ko_6LpgKT-6)(UYS1K6-THTRmaa)
Ipq=cpq(Au)-l-6(
(5.15)
=cpgJpq , Jpq=A.B.C ,
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
.
(5.20) u E CD(S21)
XA2X)Ak-r
(5.23) E K(H) , ],k=1,2
XA12R1(a)X - XA22R2(A)X =
(5.24)
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) -
2
(5.25) AAi 2Ri(a)MN+lA12R1(A) = Ai R (a)MN - MNA12R1(a).
i
,
(6.2) Ilull 2 = llu(a)02
k a
I
{IqN/2u12+Iq(N-1)/2Vu
(u)N =
(6.3)
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,
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
(6.7) IHmul2<(guvgoTauoavT
)m(givi...guNvNuu
u uv
.. V
1.. uN 1 N
< (u) N
(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 ,
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 ,
(6.10) H(uIi
.i ) _ (Hu)Ii1 i
+ O((u)N
1 N N
..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
.
<u,v> = gulvl...guNVN u
ul...11N
((v ul...uN
(6.13) N+1
= <Hu,Hu>N-l+ O(IIulIN+l2) .
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.
,
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
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
(N-J')/2
(7.5) 1 j al = o (q s j=0 ,1,...,N
0
, j=0,1,...,N , k=0,1,...
o2((/)N-J)
(7.6) Vk(ja) = .
i ...i
(7.8) [H,a] = IA,a1 = 2(a l 0)Ik Vk + (Aa) ,
and p= and
(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-
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))
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
,
(7.20)
(1.2) ind A = dim ker A - dim(H/im A) = dim ker A - dim ker A".
9e
Now let us recall the following property of C -algebras, not
normally shared by other Banach algebras:
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.)
X.1. Fredholm operators in C -algebras 296
if aA(m)#O , mEffi .
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 .
(2.6) as IfI=h3kEjY1,
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.
Theorem 2.5. Assume cdn's (a0), (a1), (d0), (s), (11), (12),
X.2. Differential operators within reach 301
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
(3.11) 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
.
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
tion A^IHV = 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.
.
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
is a (multiplication) operator in CRC C
For two vector bundles V,W, satisfying the above assumptions
X.3. Systems and vector bundles 309
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.
(4.1) C D E 3 K(H) ,
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
,
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.
xMp 2=1
making . Also, P0 is just any operator in JO such that
(4.11) L_ = IL (-)a3
get
with
, (1+T2Ax2)-1/2
(4.16) Ax=(l-Ax)-1/2 T(T) _
eatLe-5L
for the expression for all but finitely many 161< n, 6E]k.
X.5. Sobolev spaces 315
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 '
X.5. Sobolev spaces 316
(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.
,
is
subsets of I1: . Therefore yE is entire again. We have proven:
s
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
with at Ax
, but not As acting on the factor spaces. After an
,
(1+'r2Ax)-1/2
(5.6) MS(T) = AxsTs(T) E K(h), with T(T) _
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.
,
(5.7) AB - 1 E F. , BA - 1 E F. ,
X.5. Sobolev spaces 319
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.
Ifl<e outside K6 for every e>O Only finitely many of the sets
.
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 .
Lemma A.2. Let f,g E C(Q) , g > 0 . Then there exist positive
C%1)-functions y , 6 such that
QjclosC
of Uj we may regard S2jC Uj as subsets of in Let wj > 0 .
(B.2) bj lk = bj
x
k - r.kbl , c] lk = Cl lxk + rmk cR1,
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
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 '
with the Riemann curvature tensor Rijk and Ri1jk . Here we have
Similarly
(B.11)
aijlkl - aijllk - aihRjk l + ahjRikl
(B.14) Rijkl
-Rjikl -Rijlk ' Rijkl+ Riklj+ Riljk= 0
(B.15) R. . = Rijk
(B.16) j .
hjka,jk = x IXk
(Notice that, as for all tensors, a covariant derivative may be
'raised', giving a contravariant derivative:
A
(m3) U (VIII,3): For every aE A# , DE D# with a=0 D=0 near ,
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
Ca
[S1] H.Sohrab, The -algebra of the n-dimensional harmonic
oscillator; Manuscripta.Math.34 (1981),45-70 .