SIMPLE NASH-MOSER IMPLICIT FUNCTION

THEOREM

Autor(en):

Raymond, Xavier Saint

Objekttyp:

Article

Zeitschrift:

L'Enseignement Mathmatique

Band (Jahr): 35 (1989)

Heft 1-2:

L'ENSEIGNEMENT MATHMATIQUE

PDF erstellt am:

10.12.2016

Persistenter Link: http://doi.org/10.5169/seals-57374

Nutzungsbedingungen
Die ETH-Bibliothek ist Anbieterin der digitalisierten Zeitschriften. Sie besitzt keine Urheberrechte an
den Inhalten der Zeitschriften. Die Rechte liegen in der Regel bei den Herausgebern.
Die auf der Plattform e-periodica verffentlichten Dokumente stehen fr nicht-kommerzielle Zwecke in
Lehre und Forschung sowie fr die private Nutzung frei zur Verfgung. Einzelne Dateien oder
Ausdrucke aus diesem Angebot knnen zusammen mit diesen Nutzungsbedingungen und den
korrekten Herkunftsbezeichnungen weitergegeben werden.
Das Verffentlichen von Bildern in Print- und Online-Publikationen ist nur mit vorheriger Genehmigung
der Rechteinhaber erlaubt. Die systematische Speicherung von Teilen des elektronischen Angebots
auf anderen Servern bedarf ebenfalls des schriftlichen Einverstndnisses der Rechteinhaber.
Haftungsausschluss
Alle Angaben erfolgen ohne Gewhr fr Vollstndigkeit oder Richtigkeit. Es wird keine Haftung
bernommen fr Schden durch die Verwendung von Informationen aus diesem Online-Angebot oder
durch das Fehlen von Informationen. Dies gilt auch fr Inhalte Dritter, die ber dieses Angebot
zugnglich sind.

Ein Dienst der ETH-Bibliothek

ETH Zrich, Rmistrasse 101, 8092 Zrich, Schweiz, www.library.ethz.ch
http://www.e-periodica.ch

L'Enseignement Mathmatique, t 35 (1989), p 217 226

A SIMPLE NASH-MOSER IMPLICIT FUNCTION THEOREM

by Xavier Saint Raymond

This paper is devoted to the so-called "Nash-Moser implicit function

theorem", a very powerful method which during the last decades helped to
resolve several difficult problems of solvability for nonlinear partial differential
equations (see eg Nash [7], Sergeraert [10], Zehnder [11], Hormander [2]
and others'), unfortunately, the proofs that are commonly available
(Moser [6], Schwartz [8], Sergeraert [10], Zehnder [11], Hormander [2, 3, 4],
Hamilton [1]) are very long and technical, and rather frightening for the
uninitiated reader
To correct this impediment, we present here a simple statement and a
simple proof of this type of result, but it should be considered merely as
an introduction to the subject Indeed, the result is neither new nor optimal,
and the interested reader would benefit by studying more elaborate versions
such as that of Hormander [2, 3, 4] However, owing to this simple goal,
we have been able to write a proof which avoids the use of too many
parameters (usually found m such a proof) and shows more clearly the key
ideas

Let us first informally present the problem One wants to solve an

equation
F(u)

where F involves the variables x, the unknown function u(x) and its
derivatives up to the order m If one can construct a solution u0 of
the problem can be rewritten as an
F(uo)
fo fr an /o dose to
implicit function problem by considering <\>(u,
which vanishes
F(u)
at (u0, f0) It is then sufficient to prove that the equation (v, g)
0
defines as a function of gr m a neighborhood of (u0, f0) large enough

/,

-f

f)

to contain
To prove implicit function theorems in infinite dimensional spaces (as
spaces of functions usually are), one commonly uses iterative schemes The
simplest one is known as Picard's iterative scheme \|/ being a right inverse
of (d/du) (u0, f0), one sets
Vk

\|/<|>(Mk,

/)

Uk+1

Uk

Vk

218

X SAINT RAYMOND

To prove the convergence of such a scheme, one needs estimates for the
sequence vk, but if one can estimate s derivatives of vk (symbolically denoted
by tffc L), one gets estimates for only s m derivatives of <\>(uk + 1,f)
since (f> involves the derivatives of uk+1 up to the order m, thus the
convergence will hold by induction only if the right inverse satisfies an
I

estimate

l#P)L<

C|q>L-m

which is true only for a very special type of equations, namely the elliptic
equations
However, it is known that other types of equations are also solvable (for
example hyperbolic equations), and for these, the previous scheme would give
a recurrence estimate of the form

K+l Is ^

\Vk\s + d

with a positive shift d in the number of derivatives that are controlled,

so that this scheme is not convergent To overcome this difficulty, Nash [7]
proposed another scheme involving smoothing operators so that vk\s could
be estimated inductively for a fixed s, since this improvement came at the
cost of introducing large constants, it was also required to find a scheme
with much faster convergence We won't describe here Nash's scheme nor its
improvements by Hormander [2, 3, 4], but only notice that such complicated
schemes are needed if one is interested in optimal results with respect to the
on the right
regularity of the solution without their help, the function
side must be very smooth in order to obtain some smoothness of the
solution u
\

In the theorem stated below, we will establish

a C00 existence theorem

so that the number of derivatives that are used (provided that it is finite)
does not matter For this reason, we are going to use the much simpler

scheme proposed by Moser [5] which consists alternately in using Newton's

scheme and Nash's smoothing operators \|/(u) being a right inverse of

f)

and Sk being a sequence of smoothing operators closer and

(dfy/du) (u,
closer to the identity, one sets
vk

- y\f(uk)<\)(uk,f),

uk+1

uk

Skvk

The key to get the estimates for vk is to have at one's disposal estimates
of linear growth type (see estimate (3) below), which are called "tame estimates"
in Hamilton [1], such estimates are now classical for <\> itself or for its

219

AN IMPLICIT FUNCTION THEOREM

derivatives, but the main problem in the applications of this Nash-Moser

method is to prove them also for the right inverse \|/; here, we will assume
that these estimates hold (cf. (1) and (3)). In the proof we will also use
the simple interpolation formula of Sergeraert [9] who introduced it to prove
that Moser's scheme could lead to C00 results as well.
After the proof of the theorem, we propose a short description of the
classical application of this type of result to the problem of isometric
embedding of Riemannian manifolds, but this is given merely as an illustration
of the method, and we refer to Hormander [3], Sections 2 and 5, for
the details.

To complete this lengthy introduction, we confess that the result stated

here is probably the worst that can be found in the literature on the
subject with respect to the number of derivatives that are used. One reason
is that we have taken all the shifts in the number of derivatives equal to
the maximum, d, to avoid the multiplicity of parameters. In specific
applications however, it is obvious that this can be much improved.
Throughout the paper, we consider the expression <\>(u, f) for various u,
but always the same
e C00 so that it can be written $(u) as well;
finally, we recall that the function u defined in the open subset Q of R"
belongs to the Sobolev space HS(Q) if all its derivatives up to order s
are square integrable over Q (see also the definition with the Fourier transform
in the appendix when Q
Rn). We can now state the result.

Theorem.

Let

<\> :

R";

Hco(Rn)

-* #(Q)

where

is an open subset

of

the norm in Hs(Rn) and by || || s the norm

\s
one denotes by
in HS(Q). One assumes that there exist u0 e Hco(Rn), an integer d > 0,
a real number and constants C1,C2 and (Cs)s^d such that for
\

any

u,v,we
(1)

Hco(Rn),

Vs^d,
u0

3,

<

=> J

||<|>(m)||s< Cs(l + \u\s + d)

||

ty(u)v

||

<|>"(u) (V, W)

||

2d

<
||

C1
2d

<

3d

C2

3d

3d

(when one deals with (nonlinear) partial differential equations of order

these estimates classically hold for

jor

>

+ -). Moreover,

m,

one assumes that

every u e fl(R") such that

u
u0 | 3<J < S, there exists an operator
tyu) Hm(Q) -> H(R") satisfying for any cp e Hm(Q),
\

(2)

(p'(w)\K")<P

(f>

in Q,,

and

220

X. SAINT RAYMOND

(3)

> d,

VS

XKW)<P

Is

<

Cs(||(pL + d+K + d||(pll2d)

is sufficiently small
<$>(u0)
(the so-called "tame estimate"). Then, if
2d
(with respect to some upper bound of I/O, |w0Id and (CS)S^D where
D 16d2 + 43d + 24 -sic!), there exists a function ueH(Rh) such that
\\

(j)(w)

\\

in Q.

This theorem is stated with the Sobolev spaces Hco(Q)

f]s^0Hs(Q) and H(Rn) to be used in local solvability problems for
nonlinear partial differential equations, but one can replace these spaces by
gradations of Banach spaces Bs and Bs respectively with norms \s and
||
||
there exist some smoothing operators (SQ)Q>1:B00 - B^ satisfying
s if
for every ue500,G> 1 and s and t ^ 0
Remark.

^Cs/->|t if s>t;
\\v-sev\s^ cMe->L if s^t
fl SBv\a

(4)

(the construction of such smoothing operators in the case of Sobolev spaces

is given in the appendix); we will also assume that \v\s ^ \v\t whenever
s ^ t. Actually, we will only use the operators SQk where the sequence of
real numbers 9fc is defined in the following way: 60 ^ 2 to be chosen,
0 k /4 ; here are the properties of this sequence that we will use
then dk + 1

0f+1, and

J0fc5

rsv

9(05/4)J^ 9+0'/4), then Ij>0 9/3

+ (j/4) implies 6,90~3/4 when 90 ^ 2.
< eo3(l-eo3/4)_1 < Oo1 since 90~2 < 1
The solution u of the theorem will be obtained as the limit of the
sequence uk that is constructed in the following lemma.

indeed, (5/4)j

Lemma

1.

With the same assumptions

smoothing operators

SQk

vk

(n)k

(iii)k

uk

in the theorem and with the

of the remark,

the sequences

y\f(uk)<\)(uk),

uk+1

are well defined for sufficiently large

precisely, there exist constants (Ut)t^d
that for k ^ 0,

(i)k

as

u0

3d

<

90

if

||

and
and

uk

\\

l%l3d+3< vek3;
Vt^d, (l + K+iL+2d) <

SQkvk

cj>(w0)

II

2d

Qq4;

(independent of k)

<\>(uk)

\\

2d

<

9fc-4

W(l + KL+2d)

more
such

221

AN IMPLICIT FUNCTION THEOREM

Since the property (i) implies that the sequences uk and vk

are well defined (the operator i|/(w) exists by assumption if u u0 3d < 8),
it is sufficient to prove (i), (ii) and (iii), and this is going to be done by

Proof

induction. The property (i)0 is true by assumption.

The tame estimate (3) gives for every t

Proof of (ii).

l^|t<Ct(||(|)(Mk)||t + d+|MkUJ(|)(Mk)||2d).

(6)

For t

d and using (i)k one gets

(7)

< Cd(l + \uk-u0\2d + \u0\2d)

vk \d

||

<\>(uk)

||

2i

<

V O^k-4

where V0
Cd(l + b + \u0\2d). The estimate (ii) will be obtained by inter
polation between (7) and an estimate

\vk\T^

(8)

for

V^

large T. To prove (8), we can use the first assumption (1) to

estimate (uk) in (6); this gives
a

\vk\t< ct{Ct+d(l + \uk\t + 2d) + \uk\t+dC2d(l + \uk-u0\3d + \u0\3d))

^ Ct{Ct+d + C2d(l + h + \u0\3d))(l + \uk\t + 2d).

/9)

We now fix the values N

4(2d+l) and T

3d

(2d + 3)

(N + 3).

0; moreover, if it holds for some

j < k,

we get

The estimate

(i + WT+2d)<(i + lolr+2d)ejr

(10)

obviously holds for

from (iii),. and (5)

(l + K-+ilr + 2d) <

UTQid{l + \u0\T + 2d)Q?2* +

(UTd^)(l + \u0\T + 2d)d^t+1)

that (10) holds by induction for < k if one takes 90 ^ UT. Thus one
gets (8) by replacing \uk\T+2d in (9) by the estimate (10) for
k;
note that V1 depends only on u0 T + 2d and the constants C.
so

With

Qk

9k1/(2d + 3),

KI3d + <
3

the interpolation formula can now be written as

\S~Qkvk\3d + 3

< Cs+sJ?+3

^3d + 3,d^o/T

+ \vk
I

vk \d

S-Qkvk\3d + 3
C3d +

3fTdl^-T \Vk\T

+ C3d+3TV1dk3

because of (7), (8) and our choice of T, and this is (ii)fc.

222

X. SAINT RAYMOND

Proof of (Hi). It follows essentially from the estimate (9) above, if one
observes that uk + 1 \t + 2d can be estimated in terms of vk \t because of the
relations (4) indeed, since uk + 1
uk + SBk vk,
|

\Uk+l\t + 2d

It + 2d +

Uk

^ek

Vk

+ 2d

Uk

whence (iii) with constants Ut depending only on u0

by using (9).
|

Proof of (i).
W

[0, 1],

Since uk
I

Uk

-<k

u0

tSQk Vk

U0

3d

^Qj vj>

^ X

(n)j

SQj VJ

+ ^t + 2d,fik

\t + 2d

j^k

3d

Vk

and the constants C

3d

fr j ^
I

&

allows us to write

C?>d, 3d

X \Vj\ 3d

j^k

j^k
By (5), TV

9r3 <

9n

so

that we have for

W e [0, 1], nk

(H)

90

C3d 3dV/b

+ tSQkvk-u0\3d< 8;

this gives uk + 1 u0 3d < (first part of (i)k + i).

To get an estimate for <\>(uk + 1), we write the following Taylor formula

for t

1,

/*i
<\)(uk)

<\)(uk + 1)

(l-tW(uk + tSQk vk) (SQk vk, SQk vk)dt ;

<\)'(uk)SQkvk

since vk

'(uk)

^(uk)$(uk), (2) gives (\)(uk)

(|)(Mk + 1)

cpi

+ 92

vk)

in Q, whence

with

<Pi

\uk)(S%kvk-vk) and

<p2

(1

-tW(uk + tSQk vk) (SQk vk, SQk vk)dt.

Thanks to (11), we can use (1) to estimate

(px

and (p2: with (4) and (ii)k

one gets
II

<Pl

II

2d

^1

SQkVk

^
II

q>2

II

za

<

C2

Sek p

- Vk\3d ^

C1C3d53d + 39k"3

|%|3d + 3

^l^-3d, 3d + 3^9 -6

< C2CL.3,

h < C2Ci,, 3(lK28t-6

whence

H(^+i)ll2d^c09k-6

only on F and the constants C; for

The proof of the lemma is complete.

(cf. (5)) where C0 depends

we thus get (i)k + 1.

(CoOi-^e^
90

C0,

223

AN IMPLICIT FUNCTION THEOREM

The estimates (i) and (ii) in lemma 1 give the existence of a solution
+
0. But actually, the proof of property (ii)
u e H3d 3(R") of the equation (u)
can be modified to prove an estimate for \vk\s for every s ^ d.

There exist constants (Vs)s>d such that the sequence

satisfies for every k ^ 0 and s ^ d

Lemma 2.

of lemma

vk

\Vk\s^ Vk3.
Proof.

Keeping the value N

(l + k+ilt+2d)ek-/i <

Ufi2kd(l

4(2d+l),

we get from (iii) and (5) that

(L/f9k-1)(l + K|f + 2d)9k-^;

+ \uk\t+2d)dk+\

for any fixed t, 9k ^ Ut for sufficiently large k since 9k tends to infinity,

N is
so that the sequence (1 + \uk\t + 2d)dk
bounded; substituting this into (9),
we get an estimate

vk\t< wM

(12)

where

4(2d+l)

does not depend on t. Now, for any s

can rewrite our interpolation formula with t

9, _ Ql/(s-d)
*>fc

S-Qk Vk

Sek vk

Vk

(s d)(N + 3)

d we

and

^cs,sk-d\vk\d + c^erM^lt
< Cs,dV0dk3 + Cs,tWfik3
where we have used (7) and (12).

Proof of the theorem.

Let

uk

and

vk

be as above. From lemma 2 we have

\SQjVj\s^ CatS\vj\8^ Cs,sVf3

j^0

for any

and

||

<Kttk)

that the sequence uk

5e ,. iis
u0 +
convergent in every #S(R") Glj?09;~3 < co by (5)). Moreover/The'limit
u e Hco(Rn) of the sequence uk satisfies
II

<K")

||

2d

<
<

for any

/c,

so

that

II

<|>(u)

d, so

WMk)

||

II

0 by

2d

2d

"i
V(uk + t(u-Uk)) (U-Uk)dt

II

+ C,

o
I

Uk

3,

taking the limit for k

co.

||

2d

224

x saint raymond

Application to the local isometric embedding

of a rlemannian manifold
(following Hormander [3], Section

2).

Let M be a compact C00 manifold of dimension n and g a smooth

Riemannian metric on M. In local coordinates, we are thus given a positive
definite quadratic form

YdjkdXjdxj,.

J,k

The celebrated theorem of Nash [7], which is at the origin of the method,
states that for some (large) integer N, there is an isometric embedding
u: M -> R^, that is an inj ecu ve map satisfying the system of equations
(13)

(djU, dku)
stands for d/dxj and

gjk

^ j, k ^

for the Euclidean scalar product in

RN ; thus, any compact Riemannian manifold can be thought as a submanifold
of a Euclidean space.
In the proof of this Nash theorem, one first establishes that the set of
metrics g such that the problem can be solved is a dense convex cone in
the set of all C metrics on M, and this leads to the following reduced
problem (see Hormander [3] Section 2): show that the equation (13) can
be solved for every metric in some neighborhood of a fixed metric g.
To illustrate the method described above, let us show how one can use
our theorem to prove this last property locally (and this will give a local
isometric embedding u: M -> RN).
Let Q
{x eR";\ x\ < 1} and choose, near some point x0 e M, local
coordinates such that Q describes a neighborhood of x0; we take a
Cu0: Rn -> R"(" + 3)/2 equal to
where

d3

\(xj)l^j*zn>

<

>

(Xj/2)l^j^, (XjXk)l^j<k^n)

in a neighborhood of Q; this u0 is an isometric embedding for the corres

2
1 +
and gjk
x
ponding metric g in Q, namely the metric gJ3
x}xk
k. Finally, for a metric g close to g, we consider the restriction
if
<\)(u) to Q of the function
\

j/

dp,

(14)

djtuy-gjj^j^n

which is a function in H(Q) valued in R"(" + 1)/2 for any u e H(Rn) valued
in R"(" + 3)/2. Classically, estimates such as (1) hold for s > (n + 2)/2.
The derivative of 4> with respect to u is defined by
(15)

ty(u)v

((djU, dkv} + (dku,

^i^^.

^J

AN IMPLICIT FUNCTION THEOREM

is valued in R"("
^e^P
in R"("
by adding

If

\|/(w)(p

let us consider it
+ 3)/2
0 for 1
n components (pj
as a continuous extension to R" of the function

(16)

+ 1)/25

as a

function valued

^j ^

n,

and define

-^A(u)-^

where A(u) is the n(n + 3)/2 square matrix the rows of which are d-u for
1
^ ^ n and djdku for 1 ^ ^ k ^ n; thanks to our choice of u0, the
matrix A(u0) is invertible on Q, and so is A(u) for any u close enough
to u0. Since Afa)'1 is an algebraic function of derivatives of u up to
order 2, estimates such as (3) are again classical.
Finally, we have to prove that this operator \|/ inverts (j)' (formula (2)).
Applying A(u) to the function v in (16), one gets

--

(dju, v}

--

(djdku, v}

q>j

q>jk

<j

^; ^

The xk derivative of the first equation gives (djdku,v} + (djU, dkv}

0,
and one gets also (djdku, v) + (dku, djV)
0 so that the second equation
and (15) give <\)'(u)v
cp in Q.
Thus all the assumptions of the theorem are fulfilled, and it follows that

if

we can get a solution

(u0) is sufficiently small in some HS(Q) norm;

but according to (14), <fy(u0)
g
g, and the result is that (13) can be
solved for any metric g close enough to g, as required.

Appendix

Construction of the Smoothing Operators in Sobolev

Let us recall that

v e Hs(Rn)

Let

of 0 and vanishing for

if s ^

(l + l^lWf)|2^< oo.

[0, 1] be a C function taking the value

%: Rn ->

Then,

means v e ^'(Rn) and

(2k)-"

v\2s

t,

d+ms

&

12

<

Spaces

^ ^3.

For

v e

in a neighborhood

H(R") and 9

+ |i;/ei2r< xf/e)
^(2Q)2^-<Xl + \tfy\v(Q\2
e2<s-<>(i

>

12(i +1^|2)'

one sets

v& l2

226

X SAINT RAYMOND

and
Ixl^l
estimate
with
since

/8

Similarly, for

this gives the first

(/0) e supp %,

2s ~f

Cst

(4)

< ^/3 for

t,

- 0
- xf/e) +
a Taylor formula gives
xf/9) <
^/9 \k with
k e N since
for
and
for
(i + g\2y v

12

Ck

|cj2)s

2(1

+ k
(1

+ W2)* v&
I

- 0(9

<

C?_s | /9 2(<~s)(l
|

Ck

j>

%0)(0)

%(0)

any

11

+ |cj2)s

v 2,
I

- sosup that
|

x(/c)

0,

v(Q

|/fc'

for

<c?_se2^>(i + i^i2yi%)i2
whence the second estimate (4) with Cs

Cf_s

sup

%(t~s)

\/(t s)]

REFERENCES

[1] Hamilton, R The inverse function theorem of Nash-Moser Bulletin of

AM

S 7

the

(1982), 65-222

[2] Hormander, L The boundary problems of physical geodesy Arch Rat Mech
[3]
[4]
[5]
[6]

Anal 62 (1976), 1-52

Implicit function theorems Lectures at Stanford University, Summer
Quarter 1977
On the Nash-Moser implicit function theorem Annales Acad Sci Fenniae,
Series A I Math 10 (1985), 255-259
Moser, J A new technique for the construction of solutions of nonlinear
differential equations Proc Nat Acad Sci 47 (1961), 1824-1831
A rapidly convergent iteration method and nonlinear partial differential
equations I and II Ann Scuola Norm Sup di Pisa 20 (1966), 265-315

and 499-533
[7] Nash, J The imbedding problem for Riemannian manifolds Ann of Math 63
(1956), 20-63

[8] Schwartz, J T Nonlinear functional analysis, Chap IIA Gordon & Breach,
New York 1969
[9] Sergeraert, F Une generalisation du thorme des fonctions implicites de Nash
C R Acad Sa Pans, 270A (1970), 861-863
Un thorme des fonctions implicites sur certains espaces de Frechet et
[10]
quelques applications Ann Sci Ec Norm Sup Paris 4e serie, 5 (1972),
599-660
[11] Zehnder, E Generalized implicit function theorems with applications to some
small divisor problems I and II Comm in Pure and Appi Math 28
(1975), 91-140, 29 (1976), 49-111

(Reu le 14 juin 1989)

Xavier Samt Raymond

Purdue University and

CNRS