This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
..) = >.. - Ilxll >"llxll-1 and
'l/J(w) = w - f(llxll) wf(llxll) - 1 are holomorphic on .1 (see (5.6)), and the function g(() conditions of Schwarz Lemma: Ig(()1 < 1 and g(O) equivalently
f(>..) - f(llxll)
If(>..)f(llxll) -
1
I
~
I >.. -
=
o.
=
'l/J(f(
Consequently Ig(()1 ~ 1(1, or
Ilxll I >..llxll - 1 .
98
HOLOMORHIC OPERATORS ON COMPLEX BANACH SPACES
Whence, we obtain the inequality
I : : :; If(A)7(lfXlf) If(A)A-- f(llxll) Ilxll Allxll - 1 Taking the limit for A ----+
Ilxll, we
have 1 -If(llxIIW 1 _ IIxl1 2
,
If (1IxlJ)1 : : :; But
1'(A)
=
11·
: : :;
1 1-
Ilx112'
(F'(AXllxll- 1 )xllxll- 1 , x*), so 1'(llxll) = (F'(x)x,x*)llxll- 1 = IJF'(x)xllllxll- 1 ,
hence
,
IIF (x)xll : : :;
Ilxll Ilx112'
1_
§6. Uniformly bounded families of p- holomorphic (holomorphic) operators. Montel property Let
1)
be a p-open set in X. Until now, we did not use the fact that
1)
itself can be
considered as a topological space with the p-topology defined in §2. Recall that in the initial topology of the space X, a p-holomorphic operator can have discontinuities (see Example 1.1.2). However, for the p-topology the next assertion is true.
Any p-holomorphic bounded operator is p-continuous. We prove below a stronger result.
Let 1) be a p-open set and let F = {F"'}"'EQl, F", E 1i p (1) , ~), be a uniformly bounded family (with respect to the norm topology of~), i.e., IJF",(x) II : : :; :::::; M < 00, for all a E 21. Then F is equicontinuous as a family of operators acting from the set 1) with the p- topology into the space ~ with the norm topology.
THEOREM 6.1.
Given an arbitrary c: > 0, take N ~ max{2c:- 1 M, I} and let x and x' be such that p(x, x' - x) > N. Then, setting x' - x = (,h, where Ilhll = 1, (, E C, and using Lemma 2.1, we obtain <J
p(x, (,h) = p(x, h)I('I- 1 >
N
99
Montel property or
1(1 < N-1p(x, h) ::;:; p(x, h). It follows that Fa(x') = Fa (x
+ (h)
has the Taylor representation (3.4). Then, by
Theorem 5.1, we have
IlFa(x') - Fa(x)11 ::;:; IlFa(x + (h) - Fa(x)11 ::;:;
2MI(1 [p(x, h)r 1 .
(6.1)
As h = (-l(X' - x), we actually have
COROLLARY 6.1.
If the set 1) is open, then a uniformly bounded family F =
= {Fa }aE'2l of operators holomorphic on 1), with values in a Banach space~, satisfies a Lipschitz condition, uniformly with respect to a
p(x, h) ;? 8(x) > 0 (see
2.2),
Ilx - x'il : ;:; 8(x).
21.
where hEX,
IlFa(x') - Fa(x)11 ::;:; 2MI(1 [8(X)]-1 ::;:; for all x' such that
E
Ilhll
=
1, (6.1) implies
2M [8(x)r 1 Ilx - x'il
~
Notice that the last relation implies the local uniformly boundedness of the Frechet derivatives of the operators Fa, a E
2(,
too.
We return now to the general case. Let us consider again 1) as a topological space with the p-topology. On the set of all continuous mappings from 1) into
~
we
introduce the topology Tp(1) , ~) of compact convergence (i.e., the uniform convergence on compact subsets in the p-topology of 1), see §5 of Chapter 0). Using Theorem 6.1 and Theorem 0.5.4, we obtain the following analogue of the Montel property.
Let FM be a bounded subset of Hp(1), ~), i.e., FM c::;; {F E : IIF(x)ll::;:; M < 00 for all x E 1)} and assume that for any x E 1),
THEOREM 6.2. E Hp(1),~)
the orbit FM(X) = {y E ~ : y = F(x), FE F M } is a sequentially compact subset of ~. Then the family FM is sequentially compact in the topology Tp (1) , ~). In contradistinction with the classical version of Montel theorem (X = en, ~ = em) we assume in Theorem 6.2 the compactness of the orbits; this compactness follows, for em, from the boundedness of the family F M. The next example shows the necessity of our assumption. Let X = ~ = £2. Consider the REMARK 6.1.
HOLOMORHIC OPERATORS ON COMPLEX BANACH SPACES
100
family {Bd ~ I of linear bounded operators (hence holomorphic operators) defined on the set {x = (0, ... ,0,
£2 : Ilxll < R < oo} by the equalities Bkx = Xn ,·· .). Obviously IIBkXl1 < R for all X with Ilxll < R, k= 1, 2, ....
= (XI,""X n , ... )
Xl,""
E
'--v-'" k
However the sequence {Bd~l is not compact; it is enough, for instance, to consider the values at the point x = (1,0, ... ,0, ... ). Suppose now that
~
is the dual space of some Banach space and that
~
is
metrizable in the corresponding ultraweak topology (this is the case, for instance, when
~
is separable).
0), any bounded set in
Then, by Alaoglu-Bourbaki Theorem (see §1O of Chapter ~
is sequentially compact in the ultraweak topology. Con-
sequently, we can use Theorem 0.5.4 and Theorem 6.1 in order to obtain the next result. THEOREM 6.3.
Suppose that
bounded set FM
~ HpCD,~)
If the space
~
~
is metrizable in the ultraweak topology. Then any
is sequentially compact in the topology Tp (:.D , ~).
is reflexive, the ultraweak topology in Theorem 6.3 can be
replaced with the weak topology of the
space~.
Thus, Theorems 6.1-6.3 provide a
Banach-Steinhaus type theorem (see §8 of Chapter 0). (Note, once more, the similarity between the geometric properties of holomorphic bounded operators and those of linear continuous operators.)
Assume that ~ is equipped with the norm (resp. ultraweak) topology. Let {Fn}~=l' Fn E Hp(:.D, ~), be a sequence of bounded operators. Then the sequence {Fn}~=l is convergent, in the topology Tp(:.D, ~), to a bounded operator F E Hp (:.D , ~) if only if the following two conditions are simultaneously fulfilled: THEOREM 6.4.
1) the family
{Fn}~=l
is uniformly bounded on :.D;
2) the sequence {Fn(X)}~=1 is fundamental (resp. ultraweak fundamental) for all x E :.D' , where:.D' is a subset of:.D, dense in the p-topology. We conclude this section with two generalizations of Vitali Theorem.
Let:.D be a p-open set which is a C-star relatively to a point Xo E :.D, and let {Fn}~=l' Fn E Hp(:.D,~), be a sequence of operators, uniformly bounded on :.D. Assume that for any hEX, Ilhll = 1, the sequence {Fn}~=l is fundamental on the disk Dp(xo, h) ~ 1) in the norm topology of the space~. Then the sequence is fundamental on :.D with respect to the norm topology of~, and, consequently, it is THEOREM 6.5.
Mantel property
101
strongly convergent to some operator FE Ji p ('1) , ~).
Let u be an arbitrary point in '1).
we find an element hEX and A
E
C with
Using the properties of the set '1),
IAI <
1, such that Xo
+h
E '1)
and
U = Xo + Ah. Moreover, p(xo, h) ?: 1. Then the vector-functions fn(() = Fn(xo + (h) are holomorphic on the unit disk 1(1 < 1 and all of them are bounded on this disk, with respect to the norm of~, by a constant M < 1. By our assumptions, there exists d> 0 such that the sequence {fn}~=l is fundamental on the disk 1(1 ~ d. (From now on, the proof reproduces almost verbatim the classical one. However, we may need some estimates established in the sequel, so we present a complete proof.) By virtue of Theorem 5.4 the sequence {fn}~=l is uniformly convergent, i.e., for any given positive 8, there exists N(8) such that whenever n > N(8), we have Ilfn(() - fn+m(() II < 8,
for
1(1 ~ d,
and
m
= 0, 1, ...
From these inequalities and by Cauchy inequality (3.8) we obtain
where 8~m are the Taylor coefficients of the operator Fnm
=
Fn - Fn+m . Note that
we also have
uniformly with respect to k
= 0, 1,2, ....
Then, setting
IAI = r «
1), we obtain
00
IlFn(U) - Fn+m(u) I
=
Ilfn(A) - fn+m(A)11 ~
L
118~mF(xO' h)Ak II·
k=O Using the formula with remainder for the geometric series, we have
=
8(1 - ~)
-1
(1 - (~)
P)
+ 2MrP(1 _
r)-l.
HOLOMORHIC OPERATORS ON COMPLEX BANACH SPACES
102
Now, for any c > 0 and 0 < q < 1, taking p> In
[cC1- r)(l- q)(2M)-lJ
(lnr)-l
(6.2)
and
{) > cq(r - d)dP- 1(r P - dP)-l, we obtain IlFn(u) - Fn+m(u) II <
E
(6.3)
as soon as n > N({)). ~
For the case of ultraweak convergence, the proof of the analogous assertion can be sligthly simplified; it is enough to use Montel property. However, we will be mainly interested in the case of holomorphic operators.
!D is metrizabile in the ultraweak topology and let ~ be a domain in X. Let {Fn}~=l be a uniformly bounded sequence of operators holomorphic in ~. Suppose that {Fn}~=l converges in the ultraweak topology to an operator F E 1{(~I,!D), for any element of a certain open set ~' ~~. Then F admits a holomorphic continuation on the whole ~ and the sequence {Fn}~=l' converges to this continuation in the ultraweak topology, for each element of ~. THEOREM 6.6. Assume that the space
By Theorem 6.3, the sequence {Fn}~=l contains a subsequence {Fnk}k=l which is convergent in the ultraweak topology, for any element of ~, to a certain <J
F E 1{(~, !D). F I ~' = F.
F is a holomorphic continuation of the operator
operator
The operator
F, i.e.,
The uniqueness theorem implies that any other convergent
subsequence {Fn"J~=l has the same ultraweak limit
F.
The proof is complete. ~
In spite of the fact that the previous proof is very simple, it is not at all constructive; it provides no estimates analogous to those in (6.2) and (6.3), which are very useful in many applications (see §2 of Chapter VI below).
Chapter IV Linear operators
This chapter, as well as Chapter 0, is mainly auxiliary. However, compared with Chapter 0, the topics considered here are more special. Some of the subsequent results were presented only in research papers, but not in monographs. Therefore, our exposition in this chapter will be quite detailed, and will include proofs.
§1. The spectrum and the resolvent of a linear operator 1. CLOSED OPERATORS. Let X and ~ be complex Banach spaces and A: X ----+ ~ a linear operator (see Chapter 0, §8 for definition). Recall that the operator A is equally well described as the set of pairs (x, Ax), with x E D(A) ~ X. The operator A is invertible (i.e., A has an inverse operator A-I) if and only if its kernel N (A) = = {x E D(A) : Ax = o} equals {O}. In this case, the operator A-I is defined from R(A) into D(A); moreover, A-I Ax = x for all x E D(A) and AA-Iy = y for all y E R(A).
An operator A is called closed if it is closed as a subset of the topological product X x ~ (i.e., if Xn E D(A), Xn ----+ x and AXn ----+ Y for n ----+ 00, then x E D(A) and Ax = y). If A is closed and invertible, then obviously A-I is also closed. If an operator A is bounded and defined on the whole X, and if B is a closed operator, then the operator A + B ~ defined on D(B) ~ is also closed. THEOREM 1.1. A bounded linear operator A is closed if and only if its domain D(A)
LINEAR OPERATORS
104
is closed. A is bounded then Xn E D(A) and Xn ~ x for n Ax. Therefore A is closed if and only if D(A) is closed. ~
implies that AX n ~
~ 00
In particular, any operator in L(X, ~), i.e., any bounded operator defined on the whole X, is closed. The converse is also true; we give the result without proof. THEOREM 1.2 (Closed Graph Theorem). Any closed linear operator A: X ~ ~ with
the domain of definition D(A)
=
X is closed.
2. SPECTRUM AND RESOLVENT. In this subsection we will consider X =~. Let A be a linear operator on X. A point A E C for which the operator (A - AI)-1 exists and is bounded on the whole X is called a regular point of A. The set of all regular points of A is denoted by p( A). THEOREM 1.3. A linear operator A with at least one regular point is closed.
The complement of the set p(A) in C is called the spectrum of the operator A and is denoted by a(A). According to Theorem 1.3, the spectrum of a non-closed operator coincides with the set C of complex numbers and the set of its regular points is void. This is why, everywhere below is this section, we will consider only closed linear operators. Let A be a linear operator on X. The resolvent TA of A is defined on the set p(A) of its regular points by TA(A) = (A - AI)-1, A E p(A). If no confusion may arise, the subscript A in T A (A) will be omitted. From the relation
T(/L)(A - AI)T(A) - T(/L)(A - /LI)T(A)
= (/L -
A)r(/L)T(A),
(A,/L E p(A))
the so-called Hilbert identity follows: (1.1)
which implies the commutativity of the values of the operator-function T(A). THEOREM 1.4. The set p(A) of the regular points of an operator A is open and the
resolvent T(A) is an analytic function on p(A).
The spectrum and the resolvent of a linear operator which is defined and bounded on the whole X.
105
This means that JL E p(A) and 00
r(JL) = (J - (JL - A)r(A))-l r(A). It follows that r(JL) = L)JL - A)nrn+l(A), where n=O
the series is convergent for all JL satisfying analytic operator-valued function. ~ COROLLARY 1.1. The spectrum
IJL - AI < Ilr(JL)II- 1 .
Thus r(A) is an
(T(A) is closed.
<J The conclusion follows directly from Theorem 1.4 and from de Morgan formula (see §1 of Chapter 0). ~
Note that the conclusion of Corollary 1.1 is also true for non-closed operators, since in this case (T(A) = C. Thus, the spectrum of a closed linear operator A on a complex Banach space is a closed subset of C, which can be void. If the spectrum (T(A) is non-void, it has three parts: 1) (Tl (A) consisting of all those A E (T( A) for which A - AI has a non-zero kernel N(A - AI);
2) (T2(A) consisting of those A E (T(A) for which N(A - AI) = {O}, but there exists a sequence {xn}nEN such that Ilxnll = 1, n E N, and (A-AI)Xn ~ 0 as n ~ 00; 3) (T3(A) consisting of all A E (T(A) such that N(A - AI) = {O} and R(A - AI) = = R(A - AI). It is obvious that (Tl(A), (T2(A) and (T3(A) are mutually disjoint and (T(A) =
i=l
The elements of the set (Tl (A) are called the eigenvalues of A; if A E (Tl (A) then there exists x E X, x =f:. 0, such that Ax = AX. The set X.x of all eigenvectors x corresponding to an eigenvalue A, together with the zero element of the space X, form a subspace of X called the eigenspace of A. The dimension of the eigenspace X.x is called the multiplicity of the eigenvalue A. Consider now the following question: how does the spectrum of an operator looks like. When the operator is unbounded, its spectrum may be the whole complex plane cc. Let us take a bounded closed operator A. In this case, by Theorem 1.1, the domain D(A) is closed in X. From now on we will consider in this section that D(A) = X, i.e., A E L(X) (= L(X, X)). COROLLARY 1.2. The spectrum
(T(A) is non-void and bounded.
<J Assume, on the contrary, that p(A) =
cc.
Using the Hilbert identity for
A = 0 we obtain that r(JL) = JL- 1 r(0)(JL- 1 J - r(O))-l. It follows that IIr(JL) I ~ 0 as JL ~ 00, therefore the resolvent is bounded. Thus, the operator-valued function
LINEAR OPERATORS
106
r: C
L(X) is analytic and bounded everywhere on C. Therefore, the function for: C --t C is analytic and bounded for any f E (L( X)) *. By Liouville Theorem, for is constant for any f E (L(X))*, hence r()..) is constant. Since r( 00) = 0, we obtain r()..) == 0, which contradicts the definition of the resolvent. For 1)..1 > IIAII, using Theorem 0.8.4 we have that the operator A - )..J = --t
= -).. (I -
~A )
maps homeomorphic ally X onto X. Therefore the spectrum a(A) is
contained in the disk
1)..1:::;; IIAII·
~
LEMMA 1.1. The set a3(A) is open.
),,1) is closed and N(A - )..1) = {O}. By Theorem 0.8.5 the operator (A - )..1)-1: R(A - )..1) --t X is bounded. Using Theorem 0.8.4 we get that the operator 1+ ().. - f.L)(A - )..1)-1 is a homeomorphism from X onto X for any f.L E C such that I).. - f.L1 < II(A - )..1)-111- 1. <J According to the definition, if ).. E a3(A) then R(A -
It follows that A - f.LI
= (A - )..1)(1 + ().. - f.L)(A - )..1)-1)
is a homeomorphism from X onto R(A - )..1), i.e., f.L
E
a3(A). ~
For)" E a1(A), we consider the non-decreasing sequence of subspaces
{O}, N(A - ),,1), N((A - )..1)2), ... , N((A - )..1)k) , .... If there exists m ~ 0 such that N((A - )..1)m) = N((A - )..1)m+j) for any j E N, then the subspace N((A - )..1)m) is called the root subspace corresponding to the value ).., and the integer m is called the rank of )... Let P be a bounded projection on X, i.e., a bounded linear operator with the property p 2 = P, and assume in addition that P and A commute: PA = AP. Then the subspaces Xl = PX and X 2 = (1 - P)X are invariant for A, i.e., AX 1 ~ Xl and AX2 ~ X 2· Indeed, setting P 1 = P and P 2 = I - P, for any Xi E Xi, we have
Assume that the spectrum a(A) of the operator A E L(X) is represented as the union of two disjoint closed subsets a1 and a2. Then there exists a rectifiable, simple (i.e., without self-intersections) and closed curve 'Y such that a1 is contained inside the curve 'Y and a2 lies outside. In this case the next result is true. THEOREM 1.5. Suppose that the spectrum a(A) of the operator A admits a partition
(as described above) into two sets a1 and a2. Then the space X decomposes as a
The spectrum and the resolvent of a linear operator
107
topological direct sum X l +X2 of two Banach spaces Xl and X 2 which are invariant for A, and such that the spectrum of the restriction A I Xi coincides with ai, i = 1,2.
Set PI =
The operator PI is bounded and P'f
-~ /r(~)d~. 2m
(1.2)
= PI; indeed:
(The computations above use the Hilbert identity and the Cauchy Integral Theorem). Therefore PI is a projection, and so is P2 = I - Pl. Thus X = X l +X2 where Xi = PiX, i = 1,2. Further, for each A E p(A), we have Plr(A)
= r(A)Pl ,
hence API = PIA. Therefore, the subs paces Xl and X 2 are invariant for A. Set Ai = A I Xi, i = 1,2. One easily checks that rAi(A) = r(A) I Xi. It follows that P(Ai) ;2 p(A). In addition, p(AI) ;2 a2. Indeed, rA , (A)U = r(A)U = r(A)Pl U for all U E Xl, A E p(A). For any A tt. ,,/, using (1.2) and Hilbert identity (1.1), we have
1/
r(A)Pl = - - . 2m
1/
r(A)r(Od~ = - - .
2m
d~ . (r(A) - r(o) \ C A
-
(1.3)
.,
If A lies outside the contour ,,/, then (1.4)
Since the right hand side of (1.4) is holomorphic outside the contour ,,/, then r(A)Pl' and consequently rA, (A), has an analytic continuation, holomorphic outside "/. This analytic continuation is the resolvent of AI. Therefore p(AI) contains the exterior of the contour ,,/, so a(Ad ~ al. Analogously, from (1.3) it follows that r(A)Pl = r(A)
+ 2~i
Jr(~) 'Y
~~
A
LINEAR OPERATORS
108
if A lies inside the contour "y. Thus r(A)(I - Pd (= r(A)P2 ) has an analytic continuation, holomorphic inside T As above, one concludes that a(A2) ~ a2. On the other hand, a point A E a(A) does not belong to peAl) n p(A2). Indeed, if it does, then A E peA) since the operator rA , (A)Pl +rA 2 (A)P2 is an inverse of A - AI. It follows that a(Ai) = ai, i = 1, 2. ~
§2. Spectral radius A useful tool in studying the spectrum of an operator A E L(X) is the Neumann series 00
LA-HlA i ,
where AEC, Ai-O.
(2.1)
i=O
The convergence radius of this power series with operatorial coefficients is given by the formula (2.2) rCA) = lim \lIIAnll· n--->oo
The existence of the limit in (2.2) is based on the next arguments. <3 Set inf II An II;'; n~l
=
r. Let us show that lim II An II;'; ~ r, where the symbol n~~
lim denotes the upper limit. For each E > 0 there exists mEN such that II Am II ~ ~ r+ +E. For an arbitrary n E N we have n = pm + q where 0 ~ q ~ m - 1, pEN. Then
It is easy to see that pmn- l ----> 1 and qn- l ----> 0 as n ----> 00 (m is fixed), therefore lim IIAnll;'; ~ r + E. Since E > 0 is an arbitrary positive number, we have n--->oo
lim IIA n II;'; ~ r. Thus we proved the existence of the limit
n-+oo
lim \lIIAnll = inf \lIIAnll = rCA).
n-+oo
n
~
1
THEOREM 2.1. For any A E L(X), we have
rCA) = sup IAI.
(2.3)
AEa(A)
<3Bytheabovearguments,r(A)? sup IAI. Letusshowthatr(A)~ sup IAI. AEa(A)
Set r
=
AEa(A)
sup IAI. By Theorem 1.4, the resolvent rCA) is analytic outside the disk AEa(A)
Spectral radius
109
of radius r, so it expands there in a Laurent series convergent with respect to the operatorial norm. Using the uniqueness theorem, this Laurent series coincides with the Neumann series. Consequently, lim IIA-nAnll = 0 for IAI > r, hence, for any n-->oo
c > 0, there exists m (= m(c)) EN such that
IIAnl1 ~ (c+r)n
for all n ~ m. It follows
that r(A) = lim IIAnll~ ~ r. n-->oo
COROLLARY 2.1. If
For
IAI < r(A),
then the Neumann series diverges.
IAI > r(A)
the Neumann series is convergent in the operatorial norm and
= r(A).
By Theorem 2.1, Corollary 1.1 and Corollary 1.2, a point from
00
- LA1-iAi i=O
the boundary of the disk of radius r(A) belongs to the spectrum of A. Therefore, the number r(A) is called the spectral radius of the operator A, since r(A) is the radius of the smallest closed disk centered at the origin of the complex plane C, which contains the spectrum of the operator A. From the inequality IIAn I ~ IIAlln, n E N, it follows that r(A) ~ IIAII. It is not difficult to find an example of an operator A for which r(A) < EXAMPLE 2.1.
{el,e2},
IIAII.
Let X be a two-dimensional real or complex space with a basis
Ilelll = IIe211 = 1, and let
A be the operator given by the matrix
A=
(~ ~)
with respect to this basis. Since An = 0 for n
~
2, we obtain that r(A) = 0, while
IIAII ~ 1. The next related result is interesting. THEOREM 2.2 (Ya. B. Rutitzky). Let A E L(X). Then for any c
norm
II . II"
r(A) ~ <J
> 0 there exists a
on X, equivalent with the initial norm, such that
Choose n such that
From the inequalities
IIAII" ~ r(A) + c.
IIAnl1 ~ < r(A) + c,
set r
(2.4)
= r(A),
and let
LINEAR OPERATORS
110
it follows that the norms II . II and II . lie are equivalent on X. Further, IIAlle
=
sup IIAxll e
Ilxll e=l
=
sup ((r + c)n-11IAxll + (r + c)n-21IA 2xll + ... + IIAnxll)· Ilxll e=l
According to the choice of the integer n, it follows that IIAnxll:::;; (r+c)nllxll. Hence, after a rearrangement of the terms, we obtain: IIAlle = (r + c) sup ((r + ct- 11lxll + (r + c)n-21IAxll + ... + IIAn-1xll) = Ilxll e =l
= (r + c) sup Ilxll e = (r + c) II x ll e =l
which proves the right hand side part of inequality (2.4). By (2.3), rCA) does not change if we replace the norm II . lion X with the equivalent norm II . lie. Setting IIAnll e instead of IIAnl1 in (2.3), we obtain that
rCA) :::;; IIAlle. ~ We conclude this paragraph by considering a family (called an analytic operatorial sheaf- see, for instance, A. S. Markus [1]) of linear bounded operators A(A) on X, which depends analytically on a parameter A. Our goal is to study r(A(A)) = rCA) as a function of the parameter A. The next result is true.
Let A(A) be an analytic operatorial sheaf defined C. Then the spectral radius rCA) is a subharmonic function.
THEOREM 2.3 (E. Vesentini [1]).
on a domain
1) ~
Using Cauchy Integral Formula, we have
J 2n
(A(A), I) for any 1 E L(X)*, A E Hence
1)
=
~ 2m
(A(A + ei'Pr, I) dcp,
o
and any circle of radius r centered at A, entirely lying in
1).
J 271"
IIA(A)II
=
1 sup I(A(A), 1)1:::;; -2 Ilfll=l 7f
IIA (A + ei'Pr,
1)11 dcp,
o
i.e., IIA(A)II is a subharmonic function. Since for any n E N the operatorial sheaf An(A) is also analytic, the function IIAn(A)11 is sub harmonic too. Therefore 1
IW(~) II'';; (2~ lIlAn(~ +e"r) II d"') ".; 2~ llIAn(H e"r) II' d",.
Resolvent and spectrum of the adjoint operator Taking the limit for n conclusion. ~
-+ 00
111
in the last inequality, we obtain the required
COROLLARY 2.2 (Maximum Principle). The spectral radius of an analytic operatorial sheaf attains its maximum only on the boundary of the domain :D.
§3. Resolvent and spectrum of the adjoint operator Let A be a linear operator densely defined on the Banach space X. Then (see §9 of Chapter 0) there exists the conjugate operator A * on X*. <J Let ( " .) be the canonical scalar product on X x X* (see §9 of Chapter 0). If A* f = 0 for a certain f E D(A*) then 0 = (x, A* 1) = (Ax,1) for all x E D(A), hence f E R(A)~. Therefore, from R(A) = X it follows that f = O. On the other hand, if y ~ R(A), then by Theorem 0.8.7 there exists a functional f E X* which satisfies the conditions fey) = 1 and f I R(A) = O. Consequently, (Ax,1) = 0 for all x E D(A) so f E D(A*) and A* f = O. Thus we proved the next result. LEMMA 3.1. If A is a linear operator on X with dense domain, then the operator
(A*)-1 exists if and only if R(A) = X. ~ Assume that the operator A is invertible and that X = D(A) = = R(A). Then (A*)-1 = (A- 1)*. In this case, the operator A-I is defined and bounded on the whole Xif and only if the operator A is closed and (A *) -1 is defined and bounded on the whole X*.
THEOREM 3.1.
<J The operator (A*)-1 exists by Lemma 3.1. The existence of the operator
(A-l)* follows from D(A-l)
=
R(A). If y E R(A) and f E D(A*), then
so R(A*H:D((A-l)*) and (A- 1)*A*f=f for all fED(A*). Therefore R((A-l)*)~ ~ D(A*) and A*(A-l)* f = f for all f E D((A- 1)*). Consequently (A-l)* ~ (A*)-I, and, due to the opposite inclusion, (A*)-1 = (A-l)*. If, in addition, the operator A-I is defined on the whole X and is bounded, then (A -1) * is also bounded, hence (A *) -1 is bounded. By Theorem 1.1 the operator A-I is closed and therefore (A*)-1 is closed, too. Conversely, if the operator (A*)-1 is defined and bounded on the whole X*, and if that the operator A is closed, then
LINEAR OPERATORS
112 we have
for all x E R(A) and
f
E
X*, i.e., the operator A-I is bounded.
~
Using these results and Theorem 1.1, we get:
Let A be a closed linear operator on X with D(A) = X. Then C : ~ E p(A)} and TA*(A) = TA(~)*.
THEOREM 3.2.
p(A*)
= {A E
(xo, (A* - AI)!) = ((A - ~I)xo,!) = 0,
for all f E D(A* - AI). But R(A* - AI) = X. Thus, Xo E .lX* = {O}, so the operator A - ~I has an inverse (A - ~I)-1 and both these operators are closed. By Lemma 3.1 we obtain that D ((A - AI)-I) (= R(A - ~I)) = X. This equality and Theorem 3.1 imply that ~ E p(A). ~ Using de Morgan formulas, we obtain:
Under the conditions of Theorem 3.2, the spectrum of the adjoint operator A* is the complex-conjugate of the spectrum of the operator A, i.e.,
COROLLARY 3.1.
O"(A*)
= {A
E C : ~ E O"(A)}.
Let us consider now an operator acting on a Hilbert space X = 5). In this case, the conjugate operator acts on 5), too, so the notion of a selfadjoint operator, i.e., a linear operator A with D(A) = 5), satisfying A* = A, makes sense. Then (Ax,y) = (x,Ay) for all x,y E D(A) (= D(A*)). By Corollary 3.1, the spectrum of a selfadjoint operator A is real, i.e., O"(A) <;;; JR., and its resolvent has the next property: THEOREM 3.3.
Let A be a selfadjoint operator and A =
0:
+ i{3
E
C, {3 =1=
o.
Then (3.1)
We have
Resolvent and spectrum of the adjoint operator
for all x
E
D(A), and thus (3.1) follows.
113
~
A remarkable property of selfadjoint operators is the existence of their spectral decomposition. We state the theorem concerning the spectral decomposition without any proof. First, we need a definition. Let Band C be two bounded selfadjoint operators on Sj. We shall write B ~ C if (Bx, XI ~ (Cx, XI for all X E X. THEOREM 3.4. Let A be a selfadjoint operator on Sj with the domain D(A). Then
there exists a family of orthogonal projections (see §11 of Chapter O)E).., -00 < ). < < 00, satisfying the properties: 1) E).. ~ EJ1 for any)., p with), < p; 2) {E)..} is left hand side strongly continuous, i.e., E)..-ax ---+ E)..x as a ---+ 0, a > 0, for any x E Sj; 3) E- oo = lim E).. = 0, E+oo = lim E).. = I; ).._-00
4) BE)..
)..-00
E)..B for any bounded operator B which commutes with A. (Here, a bounded operator Bon Sj is said to commute with A if x E D(A) implies Bx E D(A) and ABx = BAx.) An element x belongs to D(A) if and only if =
J 00
).2d(E)..x, XI <
00.
-00
For such an element x, we have
J
Ax =
-00
J
J 00
00
)'dE)..x
and
IIAxl12 =
).2d(E)..x, XI·
-00
b
00
The integral
)'dE)..x above is understood as the limit of the integral
a
---+ -00
and b ---+
J
)'dE)..x, for
a
-00
+00.
The family {E)..} mentioned in Theorem 3.4 is called the spectral decomposition of the selfadjoint operator A.
LINEAR OPERATORS
114
§4. The spectrum of a completely continuous operator Let X be a complex Banach space. Consider a completely continuous linear operator A defined everywhere on X. According to Theorem 0.10.6, the complete continuity of A is equivalent with the precompactness of the image AS of the unit ball S of X (={ x E Xj Ilxll ~ I}). As we noticed in §1 of Chapter 0, any linear operator A acting on a finite dimensional space is completely continuous. Its spectrum a(A) coincides with the point spectrum ap(A) = al (A), and contains a finite number (~ the dimension of the space) of eigenvalues, each of these, evidently, of finite multiplicity. The spectrum of a completely continuous operator on an infinite dimensional space has a very similar structure to the one of the spectrum of an operator on a finite dimensional space. First, we establish the next preliminary result: THEOREM 4.1.
The adjoint of a completely continuous operator is completely con-
tinuous. By Theorem 0.10.7, the range of a completely continuous operator A is separable. Let {fn}nEN be a bounded sequence in X*. We show that there exists a subsequence {fnj }j EN such that {A * f nj }j EN is a Cauchy sequence in X*. Let {Vk} kEN be a dense sequence in R(A). For each k, the scalar sequence {(Vk' fn) }nEN is bounded. Therefore, using the diagonal method, we can choose a subsequence {(Vk' fnj )}jEN which is Cauchy for each kEN. Since the sequence {vkhEN is dense in R(A), then {(v,fnj)}jEN is a Cauchy sequence for any v E R(A). The functional lim (v,fn) is linear in v on R(A). By Theorem 0.8.7, there <J
J --+(X)
J
exists an extension of this functional on the whole X, with the same norm. Denote this extension by f. Then, for all v E R(A), we have (v, fnJ ---> (v,!), i.e.,
(Ax,fnj)
--->
(Ax,!),
(4.1)
for all x E X. Let us show that A* fnj ---> A* f as j ---> 00. Assume that this is not the case. Then there exists a subsequence {he} eEN of the sequence {fnj - J} j EN, and a number 8, such that IIA*hell ~ 8> 0, for all e E N. Since IIA*hell = sup I(x, A*he)l, for each Ilxll=l e E N there exists a vector Xe E X, Ilxell = 1, such that I(Axe, he)1 ~ 8/2. From the complete continuity of A it follows that {Axe} eEN contains a Cauchy subsequence, so we may assume that {AxdeEN is a Cauchy sequence itself, i.e., for any e > 0 there exists m = m(e) such that IIAxe - Axsll < e for e, s > m. Then 8 "2 ~ I(Axe, Axs)1 ~ Me + I(Axs, he)l,
The spectrum of a completely continuous operator where M
sup Ilheli. Taking the limit for C ---->
=
eEN ~ ME.
we have 8 /2 contradiction.
Since
E
00
115
and a fixed s, and using (4.1),
is an arbitrary positive number, it follows that 8
=
0, a
~
Let A and B be two completely continuous linear operators on X. Then any linear combination of A and B is a completely continuous operator, too.
THEOREM 4.2. a)
b) If A is a completely continuous linear operator on X and C E L(X) then AC and C A are completely continuous operators.
c) If A = lim An in the norm topology of L(X) and An are completely n--->oo continuous linear operators on X, then A is completely continuous.
a) The assertion is obvious.
b) Since a bounded linear operator transforms bounded sets into bounded sets, the complete continuity of C A and AC follows from the definition of completely continuous operators. c) Let An E L(X), n E N, and An ----> A. If {Xj}jEN is a bounded sequence in X then each sequence {AnXj}jEN is precompact. Using the diagonal method we get the precompactness of the sequence {Axj} jEN' ~ The main theorem of this section deals with the structure of the spectrum of a completely continuous operator. In order to prove it, we need a preliminary result.
A bounded projection on a Banach space is a finite-dimensional operator if and only if it is completely continuous. LEMMA 4.1.
Let dim PX = 00. We define recurrently a sequence {Vn}nEN such that Vn E E PX and Ilvnll = 1;;::: Ilv n - vmll for n -I- m. Assume that the vectors VI, ... , Vk are already defined and set wt = din {VI, ... , Vk}. Then, there exists a vector u E PX such that u 1:- wt. Let wtu = din{vI, ... ,Vk,U}; obviuosly we have < d = inf Ilu - xii.
°
Choose a sequence {Xn}nEN such that Xn E wt and d
= n--->oo lim Ilu-xnll.
xE9J1
Since the space
wt is finite-dimensional, the sequence {Xn}nEN is bounded and, therefore, precompact. We may assume that Xn ----> Z E wt as n ----> 00; then d = Ilu-zll. Put Vk+1 = d-I(u-z). Clearly Ilvk+11l = 1 and }~Jn IIVk+1 - xii = }~Jn
II u -
z - dx d
II = dllu 1 - zll =
1.
LINEAR OPERATORS
116 Further
IIVk+1 - vkll ~ inf
xE9JI
Ilvk+1 - xii
~ 1,
for m = 1, ... , k. Thus there exists a sequence {vn}nEN satisfying all the above mentioned properties. Clearly none of its subsequences is convergent. On the other hand PV n = Vn , n E N, which rontradicts the complete continuity of P. Consequently,
dimPX <
00. ~
THEOREM 4.3. Let A E L(X) be a completely continuous operator. Then (T(A) is an at most countable set, with no limit points different from zero. Each A E (T(A), A -# 0, is an eigenvalue of A of a finite multiplicity, and "X is an eigenvalue of A * of the same multiplicity.
Let A be a limit point for the set of eigenvalues of the operator A. If we assume that A -# 0, then there exists a sequence {An}nEN of eigenvalues of A such that An -# 0 and An ----* A. Let Un be an eigenvector of A corresponding to the eigenvalue An, and let 9Jtn = clin{u1, ... ,un }. Since the vectors U1, ... ,Un , ... are linearly independent, the subspace 9Jt n- 1 is a proper subspace of 9Jt n and dim 9Jt n- 1 = dim 9Jtn - 1. Therefore (see the proof of Lemma 4.1) there exists a vector Vn E 9Jt n with Ilvnll = 1 and such that d n = inf Ilvn - xii = 1. Thus we obtain a sequence
xE9Jln~l
{Vn}nEN with Vn E 9Jt n , Ilvnll = 1 and dn = 1. Let us show that the sequence {A;;-l AVn}nEN contains no Cauchy subsequences. Indeed, for m < n we have A;:;-1 AVn - A;;,1 AVm
=
Vn - (A;;,1 AVm
-
A;:;-I(A - AnI)vn ) .
We can easily prove that the second term in the right hand side of the above equality belongs to 9Jtn - l . Therefore the norm of the right hand side vector is greater than or equal to 1. This means that the distance between any two terms of the sequence {A;;-l AVn}nEN is no less than 1. Consequently, none of the subsequences of this sequence is convergent. This contradicts the complete continuity of A, since the sequence
IIA;;-1vnll
is bounded:
IIA;;-1Vnl ~ (inf IAml) -1 < 00. mEN
Thus A = 0 is the only possible limit point for the set of eigenvalues of A. II. Let us prove now that the lineal R(A-),J) is closed if A -# 0 and A tJ- (Tp(A). Assume that (A - AI)X n ----* y for n ----* 00. We will show first that {xn}nEN is bounded. If this is not so, then we may suppose (choosing eventually a subsequence) that
Ilxnll
----* 00
as n
----* 00.
Set Un
= 11::11. Then Ilunll = 1 and (A - AI)Un
----*
o.
Due to the complete continuity of A we may assume that AUn ----* U, U -# o. Then Au = Au, which contradicts our assumption. Thus Ilxnll ~ c < 00, and we can suppose that AX n ----* x. Then Xn ----* A-I(y - x), hence y = A-I(A - AI)(y - x) E R(A - AI).
Normally solvable operators
117
III. Consider the set a' of all points A E C such that either A E ap(A) or A E ap(A*). By Theorem 3.2, a' ~ a(A). On the other hand C \ cr' ~ peA). Indeed, since R(A - AI) is closed for all A E C \ cr' it is enough to notice that N(A* - AI) = R(A - A1)l... So cr' = cr(A). It remains to show that the eigenspace corresponding to A E a(A) is finite-dimensional. By Lemma 4.1 it is sufficient to establish the complete continuity of the projection P onto this subspace. As in Theorem 1.5, consider the projection P
=
_(27ri)-1
Jr(Od~,
where
I
I is a circle centered at A and with a sufficiently small radius such that 0 is lying outside this circle. We have Ar(~) = I+~r(~), hence ~-1 Ar(~) = ~-1 I+r(o· Since Cld~ =
J
= 0, then P = -(27ri)-1
I
J
C 1Ar(~)d~. By Theorem 3.2, the operator
Ar(~) is com-
I
pletely continuous for any ~ E ,. By the same theorem, the projection P is completely continuous as a uniform limit of integral sums which obviuosly are completely continuous operators. From Theorems 3.2 and 4.1 it follows that cr(A*) \ {O} consists of eigenvalues of A* of finite multiplicity, and, moreover, cr(A*) \ {O} = {X : A E cr(A), A~O}. The range of the projection P* ~ the adjoint of P ~ is the eigenspace of A* corresponding to the eigenvalue A. Since the projections P and P* have the same rank, the multiplicities of A and X coincide. ~ (First Fredholm Alternative). If A is a completely continuous operator on X then, either (A - 1)-1 E L(X), or there exists a non-zero solution of the equation Ax = x.
COROLLARY 4.1
§5. Normally solvable operators Let X, ~ be Banach spaces and A : X --+ ~ be a linear operator with D(A) ~ X, R(A) ~~. Let N(A) (= Ker A) be the kernel of the operator A. Let us recall (see §9 of Chapter 0) that by 9)1l.. and l..9)1* we denote the orthogonal and the *-orthogonal complement of a set 9)1 ~ X, and 9)1* ~ X*, respectively. Assume that the operator A is densely defined, i.e., D(A) = X. Then the adjoint operator A* : ~* --+ X* exists and
THEOREM 5.1.
1)
N(A*) = R(A)l..,
2)
R(A) =
l.. N(A*).
(5.1)
LINEAR OPERATORS
118
If, in addition, D(A) = X and the operator A* is densely defined, then A 3)
N(A)
=-L
R(A*),
E
L(X) and
(5.2)
<J If D(A) = X then (see §9 of Chapter 0) the adjoint A* : ~* ----> X* exists and (Ax, I) = (x, A* I) for all x E D(A), f E D(A*). 1) Let f E N(A*). Then (Ax, I) = (x, A* I) = 0 for all x E D(A), hence f E R(A)-L. So N(A*) S;; R(A)-L. Conversely, let f E R(A)-L. Then (Ax, I) = 0 for all x E D(A), hence f E D(A*). Therefore the equality (Ax, I) = (x, A* I) holds and, because D(A) is dense in X, we obtain A* f = 0, i.e., f E N(A*). So R(A)-L S;; N(A*) and relation 1) is proved. By Theorem 0.9.2 the equality 2) follows from 1) taking the *-orthogonal complements. Assume in addition that D(A) = X and D(A*) =~. Let us show that the operator A is closed. Indeed, let {Xn}nEl":I C X, such that Xn ----> x and AX n ----> Y as n ----> 00. For any f E D(A*) we have lim (Axn' f) = lim (xn' A* f) = (x, A* I) and n---+oo
lim (Axn' f)
n--+oo
= (y, I).
So y
=
n---+oo
Ax and therefore A is closed. By Theorem 1.2 the
operator A is bounded. Consequently, its kernel N(A) is closed. The adjoint operator A* is bounded, too, and everywhere defined (see §4 of Chapter 0). We prove now equality 3). The inclusion N(A) S;; -LR(A*) can be obtained by arguments similar to those in the proof of equality 1). Let x E -LR(A*). Then (x, A* I) = 0 for all f E ~*. Therefore Ax = 0, i.e., x E N(A), hence -L R(A*) S;; N(A). Equality 4) follows from 3) using Theorem 0.4.8. ~ A densely defined linear operator A is called normally solvable if R(A)
=
-L N(A*).
(5.3)
By equality 2) in (5.1) we have: THEOREM 5.2 (Hausdorff). A densely defined operator A is normally solvable if and
only if its range is closed, i.e., R(A)
= R(A).
COROLLARY 5.1. If an operator BE L(X) is completely continuous then A = 1- B is normally solvable. <J If 1 E pCB) then R(A) = ~ (= R(I - B)) is closed. Assume that 1 E a(B). Then, by Theorem 4.3, we have that 1 is an eigenvalue of the operator B of a finite
Noether and Fredholm operators
119
multiplicity, i.e., the corresponding eigenspace Xl is finite-dimensional. Let PI be the projection onto Xl, defined by (1.2), corresponding to the obvious spliting of the spectrum: UI = {l} and U2 = u(B) \ UI· Then Xl = PIX and X = PIX + (I - PI)X. Setting X 2 = (1 - PdX and B2 = B I X 2 we obtain that 1 E p(B2). Therefore (I - B 2)X2 = X 2 and, consequently, R(A) = (I - B 2)X = (1 - B 2)X2 = X 2 is a closed subspace of X. ~ The next result follows from (5.3). THEOREM 5.3. If an
operator A : X
----+ ~
is normally solvable then the equation
Ax=y
has a solution, if and only if (y, 1)
=
0 for all solutions j of the equation A*j = O.
COROLLARY 5.2. If equation (5.1)
is solvable for any y E
(5.4)
(5.5)
has the only solution j = 0, then equation (5.4)
~.
§6. N oether and Fredholm operators The following classes of operators play an important role in the theory of linear equations with normally solvable operators. A normally solvable operator A : X ----+ ~ is called a Noether operator, or an N-operator, if both the spaces N (A) and N (A *) are finite-dimensional. For a Noether operator A, both n(A) = dimN(A) and m(A) = dimN(A*) are non-negative integers, called the number of zeros and the defect of the operator A, respectively. The integer X(A) = n(A) - m(A) is called the index of the operator A. A Noether operator A satisfying X(A) = 0, i.e., n(A) = m(A), is called a Fredholm operator, or an F -operator. Let A : X ----+ ~ be a Noether operator. If {gdk=l is a basis of IJ!(A*) then the condition of solvability for equation (5.4) can be written as a system of equalities:
(y,gk)
=
0,
k = 1,2, ... ,m.
If, in addition, A is a Fredholm operator then, using Corollary 5.2, we obtain the next
result, usually referred to as the Second Fredholm Alternative.
LINEAR OPERATORS
120
A : X ----t ~ is an F-operator then, either the equation (5.4) has a unique solution for any vector in the right hand side of that equation, or equation (5.5) has only the zero solution.
THEOREM 6.1. If
(5.3) the solvability of the equation (5.4) for any vector in its right hand side is equivalent with the equality ~ = J.. N(A*), which leads to N(A*) = {a}. From dimN(A) = dimN(A*) we obtain N(A) = {o}. On the other hand, let us assume that equation (5.5) has only the zero solution, i.e., N(A) = {a}. Then N(A*) = {a}. By (5.3) we obtain R(A) = ~, i.e., equation (5.4) has a solution for any vector y in its right hand side. Since N(A) = a such a solution is unique. ~ <J By
Suppose now that A E L(X, ~). If A is a Fredholm operator and, moreover, N(A) = {a}, then by Theorem 6.1 its range R(A) coincides with the whole space ~. By the Banach Inverse Mapping Theorem, the operator A is a linear homeomorphism between the spaces X and~. Assume, on the contrary, that N(A) =f. {a}. In this case the operator A can be perturbed by a finite rank operator K such that the corresponding perturbation B = A + K becomes a homeomorphism between the spaces X and ~. The construction of such a perturbation, which is given in the sequel, was obtained for the first time by E. Schmidt [1] in the case of integral operators. Let A E L(X,~) be a Fredholm operator, with n = n(A) ~ 1. Let {adk=l be a basis of N(A), and let {gdk=l be a basis of 5Jt(A*). By the Hahn-BanachSuchomlynoff Theorem we can choose two collections of elements Udk=l and {bdk=l in the spaces X* and ~, respectively, such that
(6.1) Define a linear operator K by the formula n
Kx
= 2)x, fi)bi ,
x
E
X.
(6.2)
i=l
The operator K acts from the space X onto the finite-dimensional subspace £ = clin{b 1 , ... , bn } ~ ~ and satisfies the conditions
Kai=bi,
i=l, ... ,n.
(E. Schmidt). Let A E L(X,~) be a Fredholm operator with n(A) ~ 1. Then the operator B = A + K, where K is defined by (6.2), has a bounded inverse
LEMMA 6.1
Projections. Splitable operators
121
operator B- 1 defined on the whole ~, i.e., B is a homeomorphism of the spaces X and~.
Assume that Bx = 0 for some x E X. Then, by (6.2), we have
<J
n
Ax
= -
(6.4)
2.Jx, fi)bi. i=1
From (6.4) and using the equality N(A*) = R(A).l (see Theorem 5.1) we obtain that (Ax, gk) = 0 for any gk E N(A*), k = 1, ... , n. Consequently n
O=L(x,fi)(bi,gk)
=
(X,fk),
(6.5)
k=l, ... ,n.
i=1
n
From (6.4) and (6.3) it follows that Ax = 0, i.e., x E N(A), and therefore x = L D:iai. i=1
By (6.5) we find D:k = (x, fk) = 0, hence x = O. So we conclude that N(B) = {O}. Let us prove now that R(B) =~. Let y E ~ be fixed. Consider the element n
f} = y - L(y,gi)bi . By (6.1) we have (f},gi) = 0, hence f} E .IN(A*) = R(A). This i=1
means that there exists an element x E X which satisfies f} = Ax. Set n
x=x+ L((y,gi)-(x,fi))ai. i=1
Then Ax
=
Ax, and using (6.3) we obtain
n
= f}
+L i=1
n
(x, fi)bi + L( (y, gi) - (x, fi) )bi = f} i=1
n
+L
(y, gi)bi = y.
i=1
Thus the equality R(B) = ~ is proved. Now the conclusion of our lemma follows from the Banach Inverse Mapping Theorem. ~
§7. Projections. Splitable operators 1. PROJECTIONS. We have already dealt with projections when we defined the direct sum of spaces (see §7 of Chapter 0). Recall that by a projection on a Banach space X
122
LINEAR OPERATORS
we mean a linear operator P defined on the whole space X, with the property p 2 = P. We will consider below only continuous projections which, for brevity, will be simply called projections.
If P is a projection then Q = I - P is a projection, too, since obviously Q is linear, is defined on the whole X, is continuous, and Q2 = (I - p)2 = 1- P = Q. The restriction of the projection P on its range R(P) = PX is the identity operator P on this subspace, so that, if P =1= 0 then IIPII ~ 1. Indeed, for any U E R(P) we find an element x E X such that u = Px and, therefore, Pu = P Px = p 2 x = Px = u. Thus R(P) is closed, i.e., it is a subspace in X. Any element x E X has a unique representation x = u + v where u = Px, v = Qx. Therefore the space X decomposes into the direct sum of subspaces R(P) and R(Q), i.e., X = R(P)+R(Q). By Theorem 0.4.2 this direct sum is topological. Thus any projection P gives rise to a decomposition of the space X into the topological direct sum X = X l +X 2 , where Xl = R(P) is the subspace onto which P acts, and X 2 = R(I - P) is the subspace along which the operator P projects the elements of X. It is clear that R(I - P) = N(P) and R(P) = N(I - P). The projections P and Q = I - P are called mutually complementary. According to the previous considerations, a subspace £ of a Banach space X is called topologically complementable if there exists a projection P on X such that R(P) = £. Generally speaking, not all the subspaces of a Banach space are topologically complementable. There are a lot of papers and monographs devoted to this topic - the existence of topological complements (see, for instance, M. Kadec, B. Mityagin [1]). However, in the case of a Hilbert space X = 5), any subspace £ has not only a topological complement, but also an orthogonal complement (see Theorem 0.11.2). More precisely, any proper subspace £ can be obtained as the range of a projection P, with minimal norm IIPII = 1, such that the subs paces R(P) and R(I - P) are orthogonal. Let us return now to the general case of a Banach space X. A projection P on X will be called proper if IIPII = 1. Since the finite-dimensional subs paces of a Banach space have always topological complements, they play an important role in some problems concerning projections. To be more specific, let X be a Banach space and (!; a finite-dimensional subspace of X, with dim (!; = n. Choose a basis {ai}f=l in (!; and define the linear functionals Ii, 1 ~ i ~ n, as follows. For u = alaI + ... +anan E (!; we put Ii(u) = ai, so that (ai, Ij) = Dij, i,j = 1, ... , n. Obviously the functionals Ii are well-defined on the subspace £ and are bounded. By Theorem 0.8.7 they admit extensions on the
123
Projections. Splitable operators whole X, denoted also by
Ii, i = 1, ... , n, with
the same norms. For x E X we set
n
Px
=
L(x, Mai.
(7.1)
i=l
Then the operator P is a projection on X, with R(P)
=
IE. Indeed,
n
=
L (x, li)ai = Px, i=l
for any x E X, hence p 2 = P. Further, it is clear that R(P) <;;;: IE. On the other hand, for each u = a1a1 + ... + ana n we have Pu = u. Therefore IE <;;;: R(P) and, consequently, IE = R(P). 2. SPLITABLE OPERATORS. Assume that P1 and E1 are two given projections on the Banach spaces X and ~, respectively. We denote
Then, any linear operator A : X
----> ~
defined on the whole X has the representation (7.2)
According to this representation, it is convenient to write the operator A in the matrix form (7.3) where Aij : £) ----> 9)1i is the restriction of the operator EiAPj on the subspace £j, = 1,2. The matrix representation (7.3) of the operator A is called non-degenerate if the operator A22 is a homeomorphism from the space £2 onto 9)12. Provided that this condition is fulfilled and, in addition,
i,j
(7.4) then the representation (7.3) is called regular.
124
LINEAR OPERATORS
An operator A : X -+ ZJ defined on the whole space X is said to be splitable if it admits a regular representation (7.3). THEOREM 7.1. A normally solvable operator A E L(X, ZJ) is splitable if and only if its kernel N(A) and its range R(A) have topological complements.
<J Suppose that the subs paces N(A) and R(A) have topological complements, i.e., there exist the projections Pi on X and El on ZJ such that R(Pd = N(A) and R(E1) = R(A). Set P 2 = I -Pi, E2 = I -E 1. Then formula (7.2) is true for all x E X. Since PiX E N(A), then E 1AP1x = 0, E 2AP1x = and Ax = E 1AP2x + E 2AP2x. Moreover, y = AP2x E R(A), so E 2y = 0, hence Ax = E 2AP2x. Thus, in our case, representation (7.3) has the particular form
°
A=
(0° 0) A22
(7.5) '
where A22 is the restriction of the operator E 2AP2 on R(I - Pd - the topological complement of the subspace N(A). It follows that R(A 22 ) = R(A), N(A 22 ) = {O} whence by the Banach Inverse Mapping Theorem we obtain that A22 is a homeomorphism. Obviously equality (7.4) is satisfied. So we proved that the corrditions of the theorem are sufficient. Conversely, assume the existence of the projections Pi on X and El on ZJ such that representation (7.3) is regular. Then the equation Ax = is equivalent to the system of equations { Al1P1X + A 12 P2X = (7.6) A 21 P 1X + A 22 P 2X = 0.
°
°
The second equation in (7.6), on its turn, is equivalent to (7.7) From (7.7) and (7.4) it follows that the first equation in (7.6) is an identity. Thus the kernel N(A) consists exactly of those elements which satisfy (7.7). Let us define a linear operator P on X by (7.8)
We have
125
Projections. Splitable operators
i.e., P is a projection (recall that A2"l : £2 -+ 9J1 2 and P l 9J1 2 then by (7.7) we obtain Px = (Pl + P 2)x = x, whence N(A) show that R(A) <;::; N(A). For each x E x we have
= {O}). If x <;::;
N(A), R(A). Next, let us E
Therefore, if we set Y = Px, then PlY = Plx and by (7.8) we find
From here and based on (7.7) it follows that Y E N(A) and, consequently, R(A) <;::; <;::; N(A). Due to the inverse inclusion we obtain that P is a projection onto N(A), and therefore N(A) has a topological complement. Assume now that Y E R(A), i.e., the equation Y = Ax has solutions. Equivalently, the system of equations {
Al1P1X + A 12 P2X = ElY A 21 P 1X + A 22 P 2X = E 2y
(7.9)
has solutions. The second equation in (7.9) gives
Replacing P2 x in the first equation of system (7.9) by the right hand side of the last equality, and according to (7.4), we obtain that R(A) consists exactly of those elements Y E ~ for which the relation ElY = A12A2"2l E 2y is fulfilled. Put E = = El - A12A2"l E l . As above, we get that E is a projection onto R(A). Consequently R(A) has a topological complement. ~ COROLLARY 7.1. Any Noether operator and, in particular, any Fredholm operator
A E
L(x,~)
is splitable.
At the end of this section we will establish, for a given operator A E L(x, ~), the relationship between the operator B = A + K considered in Lemma 6.1, where K is defined by (6.2), and the representation (7.3). Assume that A E L(x,~) is a Fredholm operator with n = n(A) ~ 1 and that {adk=l' {gdk=l are fixed bases for N(A) and N(A*), respectively. Let {fdk=l and {bdk=l be the biorthogonal systems corresponding to {ak}k=l and {gk}k=l (see §6). Set £n = clin{bl, ... , bn } and define the projection P l on x by (7.1), and the
LINEAR OPERATORS
126
n
projection El on !D by ElY
= L (y, gi)b i . Then the projection E2 = I - El maps !D i=l
onto R(A). Indeed, for any Y E !D we have n
ElY = L (y, gk)bk, k=l n
whence E2y
= Y - L(Y,9k)bk and k=l n
(E 2y,gj) = (y,gj) - L(y,9k)(bk,gj) = O. k=l
It follows that E 2 y E R(A). Moreover, from (5.3) and by the biorthogonality of the systems {bk}k=l and {gk}k=l' we have'cn n R(A) = {O}. Therefore!D = 'cn+R(A). Analogously, X = N(A)+9J1, where 9J1 consists exactly of those vectors x E X for which (x, Ii) = 0, i = 1, ... , n. By Theorem 0.4.7, the n-dimensional spaces 'cn and N(A) are homeomorphic. Consider now the operator B = A+K, where K is defined by (6.2). By (6.3), the operator KIN (A) yields a homeomorphism between the spaces N (A) and 'cn. It is easy to see that, relatively to the four projections P l , P2 , E l , E 2 , the operator B has the matrix representation
B =
(K~l A~2 )
,
(7.10)
where Ku = K I N(A) and A22 = A I 9J1 (see (7.5)). It follows that the operator B is an extension of the operator A22 on the whole space X. By Schmidt Lemma the operator B is invertible. Its inverse B- 1 has the matrix representation (7.11) Notice that representations (7.10) and (7.11) are not regular, since condition (7.4) is not fulfilled.
§8. Invariant subspaces In this section we consider operators A E L(X), i.e., bounded linear operators A : X --> --> X with D(A) = X. A subspace ,c ~ X is called invariant relatively to A if A,C ~ 'c.
Invariant subspaces
127
1. Let A E L(X). Assume that the kernel N(A) has a topological complement, i.e.,
X = N(A)+9J1, where 9J1 is a subspace of X. Let us consider the following question: when is the topological complement 9J1 of N(A) an invariant subspace relatively to A? If, in particular, the operator A is splitable, then, by Theorem 7.1, the existence of two decompositions of the space X follows: X
= N(A)+9J1,
X
= 'c+R(A).
When these two decompositions coincide, i.e., N(A) = ,c, R(A) = 9J1, then the subspace 9J1 is invariant relatively to the operator A and X = N(A)+R(A). Let now A be an arbitrary operator in L(X). Consider the following two sequences of lineals: (8.1) {O}, N(A), N(A 2 ), ... ,N(Ak ), ... , X, R(A), R(A 2 ), ... , R(A k ), ... .
(8.2)
Relatively to inclusion, the sequence (8.1) does not decrease and the sequence (8.2) does not increase, i.e., N(Ak) <:;;; N(Ak+l), and R(Ak) ~ R(Ak+l), for any k = 0,1, .... If there exists mEN such that
then we say that the operator A has a finite ascent, equal to m. If there exists r E N satisfying the conditions R(Ar)
= R(A r + 1 ),
and
R(Aj)
= R(Aj+l),
j
= 0, ... , r - 1,
(8.4)
then we say that the operator A has a finite descent, equal to r. It is easy to verify that condition (8.3) implies the equality N(Am) = N(Ak) for all kEN, k ~ m, and condition (8.4) implies the equality R(AI) = R(Ar) for all lEN, l ~ r. Moreover, as we will prove below, if a normally solvable operator has simultaneously a finite ascent m and a finite descent r then these integers coincide, i.e., m = r.
Assume that a normally solvable operator A E L(X) has a finite ascent m and a finite descent r. Then 1) the subspace R(Ar) is invariant relatively to A and the restriction A I R(Ar) is a homeomorphism; 2) the space X decomposes into the direct sum THEOREM 8.1.
LINEAR OPERATORS
128
3) the operator A has the same ascent and descent, i.e., m
=
r.
1) From R(N) = R(AT+l) and AR(AT) = R(N+ 1) we infer AR(N) = R(N). Let us show now that N(A I R(AT)) = {O}. Set A I R(N) = A. Let x E E R(AT) with Ax = o. Choose n E N such that n ~ max{m, r}. Then x E R(An), i.e., x = An z , where z E X. But An+l z = Ax = Ax = 0, whence z E N(An+l) = N(An). Therefore x (= An z ) = O. By Theorem 5.2 it follows that R(AT) is a subspace of X whence, by using Theorem 0.8.5, we obtain that A I R(AT) is a homeomorphism. 2) Let x be an arbitrary element of X. Put v = A ~T AT x and u = x-v. Then v E R(AT) and since ATu = ATx - ATv = ATx - AT A~T ATX = ATx - ATx = 0 we have u E N(AT). If we find other elements u' E N(AT) and v' E R(AT) such that x = u' + v' then <J
and since v' E R(AT) it follows that ATv'
= ATV', i.e., v' = A~T ATX = v. Conse-
= x - v = u and assertion 2) is proved. 3) Let us prove now that m = r. According to 2) any element x E N(AT+l) has the form x = u + v, where u E N(AT) and v E R(AT). Then 0 = AT+lX = = AT+lU+AT+lV and from AT+1U = 0 it follows that AT+lv = 0 hence, by virtue of 1), we obtain v = O. This means that x = u E N(AT) and therefore N(AT+l) = N(AT), i.e., m ~ r. On the other hand, if y E R(Am), i.e., y = Am x for some x E X then, representing x as x = u+v where u E N(AT), v E R(AT), we obtain y = Amu+Amv . Since N(Am) = ... = N(AT) then u E N(Am) hence Amu = 0, i.e., quently, u'
y
So m
~
= Amv =
r and the proof is complete.
Am+lA~lv E R(Am+l). ~
REMARK 8.1. If an operator A satisfies the conditions in Theorem 8.1 and A = I -B, then the space N(Am) is the root subspace for the operator B corresponding to the eigenvalue A = 1 (see §1) and the integer m itself is the rank of this eigenvalue. COROLLARY 8.1. Let B E L(X) and assume that the operator I - B is normally solvable. If A = 1 is an eigenvalue of rank 1 for the operator B and an eigenvalue of finite rank m* for the operator B* then the operator I - B is splitable and
x=
N(I - B)-t-R(I - B).
(8.6)
<J By Remark 8.1 the operator A = 1- B has the finite ascent m = 1 and the operator A* = 1- B* has the finite ascent m*. Therefore N ((A*)m*) =
129
Invariant subspaces =
N ((A *)m' +1). This equality and (5.1) imply that R (Am') = R (Am' +1). Conse-
quently, the operator A has the finite descent r ~ m*. By Theorem 8.1, the equality m = r = 1 holds. The decomposition (8.6) follows now from this last equality and (8.5) . • To conclude this subsection we give an example of a normally solvable operator with a finite ascent, but without a finite descent. EXAMPLE 8.1. Let SJ be a Hilbert space with an orthonormal basis {ei}~O' Define the linear operator A on SJ by the equalities: Aeo = 0, Aei = eH1, i)! 1. Then A E L(SJ), N(Ak) = clin{eo} and R(Ak) = clin{ek+1, ek+2, .. . }. Thus the operator A has the finite ascent m = 1 but it has not a finite descent. At the same time the operator A * has the finite descent r* = 1 but it has no finite ascent.
2. In this subsection we use a different approach to study the problem of decomposing the space X into a direct sum of the form (8.6). A key role in this approach is played by the next condition: (VB) The family of the iterations of the operator B is uniformly bounded, i.e.,
IIBkl1 ~ M < 00,
k = 1,2, ....
As a preliminary step let us prove the following result.
Assume that the operator BE L(X) satisfies condition (VB). Then the intersection of the kernel and the range of the operator A = I - B is equal to {O}, i.e.,
LEMMA 8.1.
R(A) n N(A) = {O}. <J
(8.7)
Consider the operators n
Bn
=n- 1 LB k ,
n=1,2, ....
(8.8)
k=l
From the equality BnA = n- 1
(~Bk - ~Bk+1)
=
n- 1(B - Bn+1) and by con-
dition (VB) it follows that for any y E R(A) the relation lim BmY
n--->CXl
=0
(8.9)
LINEAR OPERATORS
130
holds. If, on the contrary, y E N(A), then BnY = Y and equality (8.9) fails for O#YEN(A). ~ COROLLARY 8.2. If the operator BE
A
= J - B is Fredholm, then equality
L(X) satisfies condition (UB) and the operator (8.6) is true.
Since A is a Fredholm operator, the kernel N(A) is a finite-dimensional subspace of X. Let dimN(A) = n and let {adk=l be a basis of N(A). Using Hahn-Banach-Suchomlynoff Theorem and Lemma 8.1 we may find a system of vectors
Udk=l
~
x* with the properties: (8.10)
(y, !k) = 0,
k
= 1, ... ,n,
(8.11)
for any y E R(A). Condition (8.10) implies that the system of vectors Udk=l is linearly independent and from (8.11) together with (5.1) it follows that Udk=l is a basis of N(A*) (recall that dim N(A*) = n). The operator P defined by (7.1) is a projection from X onto N(A), and J - P is a projection onto R(A) = J..N(A*). ~ Let us return now to the more general case when the operator A is only normally solvable. For such an operator the next result is true.
L(X) satisfies condition (UB) and the operator A = J - B is normally solvable. Then equality (8.6) holds if and only if the limit in the strong operator topology (see §8 of Chapter 0) of the sequence of operators defined by (8.8) exists. The operator THEOREM 8.2. Assume that the operator B E
n
P
= n~oo lim Bn = n---+oo lim
is a projection onto N(A) and condition (UB).
IIPII : :; M,
n- 1 ' "
L.....t
Bm
(8.12)
m=l
where M is the same upper bound as in
Necessity. From (8.6) it follows that any element x E X can be represented
as
x = Y + z, where Y E R(J - B) = R(A) and z E N(J - B) = N(A). By this equality and from (8.8), (8.9) and condition (UB) it follows that IIBnl1 : :; M, n E N, and n
lim Bnx
n-+oo
= n-+oo lim n- 1 ' " (I - A)m z = z. ~ m=l
(8.13)
131
Invariant subspaces
As a consequence of the Uniformly Boundedness Principle, the operator P defined by (8.1) is a bounded linear operator with IIPII ~ M. According to (8.13) P is a projection onto N(A). Sufficiency. Assume that the operator P defined by equality (8.12) exists. Again as a consequence of the Uniformly Boundedness Principle, the operator P is linear, bounded and IIPII ~ M. For each Z E N(A) we obtain (see (8.13)) P(z) = z, hence N(A) ~ R(P). On the other hand, if u E R(P) then there exists v E X for n
which u
= Pv = lim
n- 1
n---i'OO
""'
~
Bmv. Then, by condition (UB), we have
m=l
Au = (I - B)u =
nl~ n-
1
(t,
Bmv -
t,
Bm+lv) =
lim n-1(Bv - Bn+lv) = 0,
=
n--+oo
hence u E N(A). Thus we obtain that R(P) = N(A) and that P I R(P) is the identity operator. Consequently, P is a projection onto N(A). Let us show that N(P) = R(A), a relation which clearly will complete the proof of (8.6). Let y E N(P). Then, by (8.12), for any c > 0 there exists n E N such that IIBnYl1 < c. We have y - BnY E R(A); indeed
y-BnY= (I-n- 1 tBm)y=n-1 t(I-Bm)y= m=l
m=l
n
= n-1(I - B)
L (I + B + B2 + ... + Bm-1)y = (I -
B)Yn,
m=l
n
where Yn = n- 1
L (I + B + B2 + ... + Bm-1)y. m=l
Since y - BnY E R(A) and R(A) is closed, taking into account that c was an arbitrary positive number, we obtain y E R(A), i.e., N(P) ~ R(A). The converse inclusion R(A) ~ N(P) follows directly from (8.9). ~ The arguments in our proof of Theorem 8.2 above are, actually, suitable modifications of the arguments used by K. Yoshida in the proofs of the Statistical Ergodic Theorem and its consequences (see, for instance, K. Yoshida [1]). This theorem asserts the following: REMARK 8.2.
Let B E L(X) be an operator satisfying condition (UB). Consider the operators B n , n E N, defined by (8.8) and assume that the sequence {BnX}nEN is weakly compact for x E X. Then this sequence is convergent in the norm topology of the space X.
132
LINEAR OPERATORS This theorem and Theorem 8.2 lead to the next simple but important result.
COROLLARY 8.3. Let X be a reflexive space. Assume that an operator B E L(X)
satisfies condition (UB) and the operator 1- B is normally solvable. Then the equality (8.6) is true. Some authors (see, for example, F. Riesz [1]) refer to Corollary 8.3 as the Statistical Ergodic Theorem.
Chapter V Nonlinear equations with differentiable operators §1. Fixed points. Banach principle 1. Let 1) be a topological space. Assume that F is an operator defined from 1) into 1),
i.e., (1.1)
DEFINITION 1.1. A point
z E
1)
is called a fixed point for the operator F if F(z)
= z.
Condition (1.1), which is refered to as the invariance of the set 1) for the operator F, enables us to define the iterations Fn of F, by the recurrent relations Fl = F, F n+1 = Fn 0 F, where "0" denotes the composition of operators. DEFINITION 1.2. 1) A fixed point z E 1) ofthe operator F is called locally attractive if there exists a neighborhood U ~ 1) of the point z, such that the iterations Fn(x) converge to z for all x E U. If U = 1) we shall say that z is an attractive (or globally attractive) fixed point. 2) The point z is called an s-fixed point (successful approximative fixed point) for the operator F in 1) if it is attractive uniformly on each closed and bounded subset of 1), i.e., the iterations Fn(x) converge uniformly on each closed and bounded subset of 1). 3) A fixed point z is called repulsive if the sequence Fn(x) converges to z if and only if x = z. 4) A fixed point z will be called neutral if it is neither locally attractive nor repulsive.
NONLINEAR EQUATIONS
134
DEFINITION 1.3. A fixed point z of F is called isolate if there exists a neighborhood
of this point which does not contain other fixed points for the operator F. REMARK 1.1. A locally attractive point z is an isolate fixed point. Moreover, a globally attractive point is unique. Indeed, assume that on a neighborhood ti of the point z the iterations Fn(x) converge to z for all x E ti. If we suppose that the operator F has another fixed point y in ti, then Fn(y) ----) Z. But y is a fixed point for the operator Fn for any n, since Fn(y) = FoFn-l(y) = F(y) = y. Hence z = y. We notice also that for a continuous operator F the convergence of the iterations implies
the existence of fixed points. A lot of work has been done to obtain criterions for the existence of fixed points and to study the properties of these points (uniqueness, isolation, attractiveness, and so on) for various operators. These criterions are frequently keystones in the theory of nonlinear operator equations. The theorems which establish conditions for the existence of fixed points are usually called fixed point principles. As a rule, the formulation of a fixed point principle starts with the condition of invariance (1.1). However, in many cases this condition is insufficient for the existence of a fixed point, and, even more, it is insufficient for the existence of an s-fixed point. Notice that the presence of an s-fixed point guarantees not only its uniqueness, but also the possibility of approximating it (by approximative computations). Let us remark that the development of the computer techniques enables one to approximate a fixed point by iterations even for rather complicated operators. At the same time the computer methods do not insure against accidental computational errors, which practically do not matter as soon as we know in advance that the iterative proccess converges uniformly.
One of the most effective fixed point principle is the Banach principle for contractive operators. Let (X, p) be a complete metric space with the metric p( " . ) (see §4 of Chapter 0). An operator F defined from a subset ~ <;;;; X into X is called a contractive operator or a contraction, if it satisfies a Lipschitz type condition with a constant q < 1, i.e.,
p(Fx, Fy)
~
qp(x, y),
x, y
E ~.
(1.2)
Obviously a contractive operator is continuous. THEOREM 1.1 (Banach principle for contractive operators). Let ~ be a closed
subset of (X, p) and let F be an operator defined on ~ which is a contraction and satisfies the invariance condition (1.1). Then the operator F has an s-fixed point in
135
Fixed points. Banach principle
::D. The rate of convergence for the iterations of the operator F to this fixed point is described by the relations
p(Fnxo, Fmxo) where m
~
~ Lp(xo, Fxo), 1-q
(1.3)
nand Xo is an arbitrary element in ::D.
<J As a matter of fact it is sufficient to prove inequality (1.3). This inequality implies that the sequence {xn}nEN, where Xn = Fn xo , is fundamental, and this property is uniformly fulfilled relatively to Xo in any closed and bounded subset of ::D. Therefore, by the completeness of the space X and the closedness of the set ::D, the sequence {x n }nEN converges to a point Z E ::D, uniformly relatively to Xo E ::D. Indeed, by (1.2) it follows that
Analogously,
By induction we get
p(xk,xk+d ~ qkp(xo,xd, for all k = 1,2, .... From these relations, using the triangle inequality and the formula for the goemetric series, we have
+ P(Xn+l' Xn+2) + ... + P(Xm-I, Xm) ~ ~ qn p(XO, xd(l + q + ... + qm-n-I) ~ 1 qn p(xo, xd.
p(xn' x m ) ~ p(xn, xn+d
-q
Setting now z = lim Fnxo, where Xo is an arbitrary element of ::D, we obtain that z n-+oo is an s-fixed point for the operator F in::D. ~ If we weaken slightly the conditions required for the operator F in Theorem
1.1, we obtain the next version of the Banach principle.
closed subset of a complete metric space X and let F be an operator satisfying the in variance condition (1.1). Assume that there exists an integer p ~ 1 such that the operator FP is a contraction on ::D. Then the operator F has a globally attractive, and, consequently, a unique fixed point in ::D. COROLLARY 1.1. Let::D be a
<J By Theorem 1.1 the operator FP has an s-fixed point z. Choose an arbitrary element Xo E ::D and consider the sequence {xn : n EN}, where Xn = Fnxo and
NONLINEAR EQUATIONS
136
N denotes the set
{O, 1,2, ... ,n, ... }. Let us show that this sequence converges to z. Indeed, the sequence breaks up into a finite number of subsequences
{FkP XQ
:
kEN},
{FkP X1
:
kEN}, ... ,
{FkP xp_ 1
:
kEN}.
Each of these subsequences is convergent to z. Now we show that z is a fixed point for F. (Notice that this does not follow from the convergence of the sequence {xn : n E E N}, since under our assumptions the operator F is not necessarily continuous.) We have FP F z = F FP Z = F z, whence it follows that F z is a fixed point for the operator FP. According to the uniqueness of the fixed point for FP we obtain that Fz = z. ~ REMARK 1.2. In the general case it is not true that an s-fixed point of the operator
FP, for certain p > 1, is an s-fixed point for F, too. However, as we shall prove below, this is true for holomorphic operators. In order to approximate a fixed point by iterations under the assumptions of Corollary 1.1 it is advisable, if it is possible, to start with an explicite computation of the operator FP and then to use its iterations. REMARK 1.3. The condition q < 1 in (1.2) is not necessary for the existence of s-fixed points. For example, consider the operator Fx = ax2, where 1/2 < a < 1, which maps the interval [0, 1] into itself and has the s-fixed point 0, though q = 2a > 1. Nevertheless, in the general case condition q < 1 can not be weakened by setting q ~ 1. For example, the operator Fx = x+a, a -I- 0, which maps IR into IR and satisfies condition (1.2) with q = 1, does not have fixed points. In this example the diameter of the set on which the operator F acts is infinite. However, there are examples of operators acting on domains with bounded diameter, in infinite dimensional spaces, which satisfy condition (1.1), or condition (1.2) with q = 1, and do not have fixed points (for more details see §2 below).
Let us return now to the concept of s-fixed points and state, without proof, a remarkable theorem due to P. R. Meyers [1], which explains the substance of this notion and represents the most complete converse of the Banach principle. (In connection with this result see also A. Levin and E. Lifsic [1].)
(P. R. Meyers). Let X be a complete metric space and 1) a closed subset of X. Assume that an operator F satisfies the invariance condition (1.1) and has an attractive fixed point z E 1), which is an s-fixed point in a neighborhood of z. Then z is an s-fixed point in 1) and, moreover, for any q with < q < 1 there exists a metric on the space X equivalent with the original metric, such that F satisfies the contraction condition (1.2) on :D with respect to this new metric. THEOREM 1.2
°
Although Meyers' result is merely an existence theorem, it shows that in some
Non-expansive operators
137
concrete problems, if we know a piori about the existence of an s-fixed point, then, for estimating the rate of convergence of the iterative process, it is convenient to try to find an equivalent metric, with respect to which F is a contraction. 2. One of the most important fixed point principles is the well known principle due to J. P. Schauder, which is a generalization of the finite-dimensional fixed point principle of Brouwer. THEOREM 1.3 (J. P. Schauder). Let F be an operator which maps a closed convex subset:D of a Banach space X into itself and is completely continuous on:D. Then F has at least one fixed point in :D.
All the proofs of this theorem we know about rely on subtle topological and geometric considerations. As a rule, these proofs use Brouwer Theorem, finite-dimensional approximations, and also some results from the degree theory for mappings and the vector fields theory (see, for instance, V. A. Trenogin [1], M. A. Krasnoselskii [1]). This is why we stated Schauder Theorem without proof. Obviously, neither Theorem 1.1, nor Theorem 1.3, are related to each other. Let us remark also that Schauder principle has no constructive features; it fails to indicate the number of fixed points, as well as methods for their approximation. In the following section we will consider a larger class of operators, which includes the class of contractive operators. At the same time, under the assumptions of Schauder principle, we will indicate some approximation methods for the fixed points.
§2. Non-expansive operators Let (X, p) be a metric space and :D a closed subset of X. DEFINITION 2.1. An operator F : :D strictly non-expansive, if the inequality
---+
p(Fx, Fy) is fulfilled for all x, y
for any x, y
E
:D, x
E
=1= y.
~
X is called non-expansive, respectively
p(x, y)
(2.1)
p(Fx, Fy) < p(x, y),
(2.2)
:D, respectively, if
NONLINEAR EQUATIONS
138
A large amount of work is devoted to the theory of non-expansive operators, either on metric, or on Banach spaces (see for example Z. Opial [1]). Here we treat only those aspects which are related to the fixed points of differentiable operators. As we have already noticed in Remark 1.2, the operators satisfying condition (2.1) may have no fixed points. In addition, let us present a few other examples.
2.1 (T. Hayden, T. Suffridge [1]). Let Co be the space of all sequences x = (Xl> ... , X n , ... ) of real or complex numbers which converge to zero, with the norm Ilxll = max IX n I· Define an operator F on the unit ball of this space by the equality
EXAMPLE
n
Fx =
(!, Xl, X2,·· .), for
any X = (Xl, ... , Xn , .. . ). Clearly IIFxII ,,;; 1 if IIxil ,,;; 1, i.e., F satisfies condition (1.1), and IIFx - Fyll = 11(0, Xl - YI,···, Xn - Yn," .)11 = IIx - YII, Le., F satisfies condition (2.1), too. At the same time, if we assume that there exists an element Z = (ZI, ... ,Zn, ... ) such that Z = Fz then, necessarily, Zl = Z2 = ... = = Zn = ... = i.e., Z ~ Co, a contradiction.
!,
We can show that condition (2.2), as well, is not enough for the existence of fixed points. EXAMPLE
2.2. Let X =
~
and :D
= [1, (0).
The operator F defined by Fx = X +
dearly does not have fixed points, while F(:D) jFx - Fyi
for all x, Y ~ 1,
X
= Ix -
Y+
~-
tl
=
~
Ix -
1
-
X
:D and
YII x Yx ; 11 < Ix - YI,
i=- y.
Related to the example above we easily notice that for any q, with 0 < q < 1, there exist x, Y E [1,(0) such that jFx - Fyi> qlx - YI. An analogous example can be produced on a ball with a finite radius in an infinite dimensional space, too. EXAMPLE 2.3 Let
F defined by Fx
1
+-n sin n.
=
X = Co and let :D be the unit ball in X. Consider the operator n-1
(YI, ... , Yn," .), where X
=
(Xl, ... , x n , ... ) and Yn
=
- - xn+
n We can easily prove that the operator F satisfies both conditions (1.1) and
(2.2) on:D. At the same time, F has no fixed poins in:D. Assume, on the contrary, n -1 1 that Z = Fz, where Z = (Zl"'" zn) E:D. This means that Zn = --Zn + - sin n. n n Therefore, Zn = sin n, which contradicts the definition of the space Co. All these examples show that the existence of a fixed point for a non-expansive operator requires some additional conditions.
Let (X,p) be a complete metric space, :D a closed subset of X, and F a strictly non-expansive operator on :D satisfying the invariance condition
THEOREM 2.1.
139
Non-expansive operators
(1.1), i.e., F(1») <:;: 1). If there exists at least one point Xo E 1) such that the orbit ~ Xo = {x E 1) : x = F n Xo, n E N} is precompact, then the operator F has a unique fixed point z E 1). Moreover, the sequence {xn : Xn = Fnxo, n E N} is convergent to z.
~ is continuously invertible; 3) there exists a number L > 0 such that .) = (F(>.x), I) = (A (h) ')II < 1, for all >. E c. with 1>'1 < 1, and .x) = 0 for all >. E c. with 1>'1 < 1. Hence I(Fx,1)1 = 1. But, on the other hand, by Cauchy-Bunyakowsky inequality and by (V.1.5) we have I(Fx, 1)1 ~ IlFxll . 11111 < 1, a contradiction which proves the theorem. ~ (U(A ~ defined by the equality
COROLLARY 5.1.
1) -> ~
be an
Implicit and invertible operators for all Xl,X2 E::D. Then, on the ball
p
. { = mIn
m = {y E ~ 1,
mL( VI
:
Ily - yoll ~ p}, where
1
} ,
+ m + y'm)2
there exists a unique Frechet differentiable operator p-l : condition p-l(yO) = Xo and is an inverse ofp. Moreover, we have
IIp-l(y)11 ~ min{R, y'm(mL( for all y
E
167
m---; X which satisfies the
vm + 1 + y'm))-l},
m.
As we have already noticed, the transition from an inverse operator problem to an implicit operator problem is also possible. We illustrate this by an example involving analytic operators, for which we will indicate some other estimates. We will use the Cauchy-Goursat method of majorant series.
lPx
=
lPxo
+L
8k lP(xo, h),
k=l
where h = x - Xo, lPxo = A-lyo, 8l lP(xo, h) = A-lp'(xo)h = h, and 8k lP(xo, h) = = A- 1 8k P(xo, h) are homogeneous forms of order k. The equation Px = y is obviously equivalent to the next equation 00
h
+L
8k lP(xo, h) = z,
(5.16)
k=2
where z = A-l(y - Yo) E X and Ilhll ~ R. We will try to find a solution for this equation of the form 00
h
=
P(z)
=
L 8kP(O, z),
(5.17)
k=O
where 8kP(O, z) are homogeneous forms of order k. Substitute series (5.17) in equation (5.16) and equate the forms of the same order. As a result we obtain an infinite recurrent system from which we get all the forms 15kP(O, z), succesively. (Explicit
NONLINEAR EQUATIONS
168
formulas for 8k P(0, z) can be found, for example, in J. Orava and A. Halme [1J and A. P. Yuzhakov [2J.) If we show that the formal series (5.17) converges normally on some neighborhood {z EX: Ilzll < 8} of the point z = 0, then this will prove the existence of an analytic solution of equation (5.16), and the existence and the analyticity of the operator tfJ- I = l}/-I A-I: I{) --t X for all y with Ily - yoll < p = 8m-I, where m = lIa-III, as well. For the proof let us consider the numerical series 00
'P(()
=
L (kI1 8kl}/(XO, h)11 k=2
which converges for all (with 1(1 ~ R, ( E IK, Ilhll = 1. By Abell Theorem the series is convergent for all ( lying in the disk of radius R of the complex plane. Consequently, the function 'P(() is analytic on this disk. Let sup I'P(() I ~ M < 00. Then, by
1(1 ~ R,
(EC
Cauchy inequalities, we have
(5.18) and, consequently,
I'P(() I ~ M
f
1(l k R- k
~ M I~:
(1 _I~)
-1
= q(I(I).
k=2
Consider the auxiliary numerical quadratic equation
t=I(I,
t=1]+q(t), which, obviously, has the unique solution
satisfying the condition teO)
= O.
It is clear that for
11]1 ~ 8 =
1] with
R2 4(M
+ R)
the discriminant of this equation is greater than or equal to 11]IR ~ 0 and, therefore, the function t(1]) is analytic on 1] in the disk 11]1 ~ 8. Hence it can be represented as a convergent power series 00
t(1])
=
L
k=O
Ck1]k.
169
Implicit and invertible operators
Since the coefficients Ck can be also found from the recurrent relations obtained substituting the last series in equation (5.19), using relation (5.18) it is not difficult to get the estimates
From these estimates it follows that the series (5.17) converges normally for all z with Ilzll < D. This means that the operator 1[/-1 is analytic on the ball Ily - yoll < p = = om-I. Thus we proved the next result.
Let X, ~ be arbitrary Banach spaces over the field lK and
THEOREM 5.3.
p M
= R2(4(M + R)IIA- 1 11)-1,
= IIA- 1 11
sup 11f:(kOk
(Ee,
Ilhll=1
there exists an analytic operator
Let X, ~, A be Banach spaces over the field lK and let F : X x A be an operator analytic in the domain
COROLLARY 5.2. ~ ~
{(x, >.) :
II(x, >.) -
(xo, >'0)11
Assume that the operator A = F~(xo, >'0) : X on the ball {>. E A : II>' - >'011 ~ p}, where
~ ~
~
~
R}.
is continuously invertible. Then
(5.20)
( A B = 0
F~ (xo, fA
Uo
00
>'0)) '
=
(xo, >'0),
h
=
(x - Xo, >. -
>'0),
and fA is the identity operator on A, there exists a unique implicit operator G(>.) satisfying the identity F(G(>'), >.) == 0 and the condition G(>'o) = Xo.
170
NONLINEAR EQUATIONS
REMARK 5.1. The method of constructing the inverse operator by formula (5.17),
described in the proof of Theorem 5.3, is called the method of undetermined coefficients. It was used more than once by different authors, in order to solve some concrete integral equations (see, for example, N. N. Nazarov [1] and A. I. Nekrasov [1]). It seems that this method in its general form was used for the first time by K. T. Ahmedov [1] for an implicit operator depending on a numerical parameter, and by J. Orava and A. Halme [1] for the inverse operator problem. Formulas (5.13)-(5.15) were established by V. A. Trenogin (see [1]). Using the same approach we can obtain some other, yet equivalent, estimates for T, 'f] if we set II(x,A)11
= max{llxll, IIAII}
or
In §5 of Chapter VI below we will establish some more precise - and more convenient for computations - estimates for the numbers p, 8, in the case of complex spaces.
Chapter VI Nonlinear equations with holomorphic operators Throughout this chapter we will consider complex Banach spaces only. In this case, many local features related to the solvability of equations with operators that are differentiable in the complex sense, turn out to be global.
§1. s-fixed points for holomorphic operators. A converse of Banach principle In this section we deal with differentiable operators and their fixed points that have the property of "succesful approximation", i.e., s-fixed points (see Chapter V). Let f> be an arbitrary bounded domain in X and F an operator holomorphic in f> (i.e., F is Frechet differentiable in the complex sense on a neighborhood of each point of f». Assume that F satisfies the invariance condition
F(f» <;;; fl.
(1.1 )
Suppose further that z E f> is a fixed point for F, that is, Fz = z. As we have already noticed, if F E C(f>, X), i.e., F admits a continuous extension on of> - the boundary of the domain f> - , and if the point z is an attractive fixed point for the operator F (see §1 of Chapter V), then, as soon as z is a local s-fixed point, it is a global s-fixed point too. Indeed, by Meyers Theorem (Theorem V.1.2), the next assertion is a consequence of our assumptions:
(*) for any q with 0 < q < 1 there exists a metric p on the set f>, equivalent to the metric given by the norm of X, relatively to which the operator F satisfies the
NONLINEAR EQUATIONS
172
Lipschitz condition with constant q, that is, p(Fx, Fy)
~
qp(x, y),
x, y E 1:>.
(1.2)
Thus the assumptions in Theorem V.1.1 - the Banach principle - are fulfilled. If inequality (1.2) is satisfied on some closed subset II ~ 1:> and F(ll) ~ ll, then we will say that the operator F is a q-contraction on ll. The following result holds.
Let F be a holomorphic operator on 1:>, satisfying condition (1.1), let zE1:> be a fixed point for F, and let A=F'(z) be the Fhkhet derivative of Fat z. a) The following assertions are equivalent: 1) Fn x -+ Z for any x E 1:> and the operator e iO I - A is normally solvable (see §5 of Chapter IV) for all () E [0, 27rJ, where I is the identity operator on X; 2) rCA) < 1, where rCA) denotes the spectral radius of the operator A; 3) there exist a number r > 0 and a norm II . II * equivalent to the original norm of X, such that F is a ql-contraction on the ball II,. = {x E 1:> : Ilx - zll* ~ r} for a certain ql with 0 ~ ql < 1; 4) there exist a number q2 with 0 ~ q2 < 1 and a metric equivalent to the metric given by the norm of X relatively to which the operator F is a q2-contraction on some neighborhood of the point z. b) If, in addition, F E C(1:>, X) and Fn x -+ Z for any x E 81:> then conditions 1)-4) are equivalent to condition (*). THEOREM 1.1.
EXPLANATIONS: 1. Under condition b) the implication 1) =} (*) is a global converse of the Banach principle with respect to a metric equivalent to the metric given by the norm of X. The implication 1) =} 3) is a local converse of that principle, but this time, with respect to a norm equivalent to the norm of X. In A. A. Ivanov's book [1] a construction of a metric such as in condition 4) above is given. In many concrete problems this construction turns out to be rather difficult. If we know already the spectral radius rCA), then it is very simple to find the norm II . 11* and the number r appearing in condition 3) (see Lemma V.3.1). This enables us to estimate the rate of convergence to z of the iterations {Fnx }nEN, starting with an arbitrary x E II,.. 2. The implication 2) =} 1) means that condition rCA) < 1 is, in fact, a global feature of the s-fixed point z (see §3 of Chapter V) and, consequently, it guarantees the uniqueness of the fixed point z on the whole 1:> and also the convergence of the iterations, for any x E 1:>.
Proof of Theorem 1.1. The implication b)&1)&2)=} (*) follows by the
173
s-fixed points for holomorphic operators
already mentioned Meyers Theorem. The implication (*) =} 4) is obvious. Therefore it is sufficient to prove only the equivalence of conditions 1)-4). Let us prove that 1) =} 2). Denote by An the Frechet derivative of the operator F n at the point z, i.e., An = (Fn)'(z), n ~ 1 (then Al = A). From condition Fn x - z -+ 0 and by Cauchy inequalities for the holomorphic operator Fnx - z (see Chapter III) it follows that lim An h
n--->oo
= 0,
for all hEX
(1.3)
(here the limit is considered in the strong sense). Moreover, An = An, where An is the n-th iteration of the operator A (see identity (3.5) in Chapter V). Further, it will be enough to show that the set [2 = {A E C : IAI ~ I} is contained in the resolvent set of the operator A. This will clearly imply that r(A) < 1. By (1.6) the set [2 does not intersect the point spectrum Ul (A) of the operator A (the set of eigenvalues of the operator A). Indeed, if A E [2 n ul(A) then there exists y E X, y =1= 0, such that Ay = AY, and consequently, Any = Any, whence lim Any =1= 0, which contradicts n--->oo
(1.3). Moreover, by the Uniform Boundedness Principle it follows that
(1.4) Therefore the sequence n=O
converges uniformly to the operator (AI - A)-l for all A with IAI have (AI - A)Smx = x - A-(m+l) AX, x E X.
> 1. If IAI = 1 we
It follows that the set R(AI - A) - the range of the operator AI - A - is dense in X. Taking into account condition 1) once again, we conclude that R(AI -A) = X. Since, on the other hand, A tI- Ul (A), it follows that the operator AI - A is continuously invertible. Thus r(A) < 1. The implication 1) =} 2) is proved. The implication 2) =} 3) follows from Theorem V.3.3, and the implication 3) =} 4) is trivial. It remains to prove that 4) =} 1). By 4) it follows that the iterations {Fnx} nEN converge to z uniformly on some neighborhood of the point z. Since the set 1) is simply connected, then, by Vitali Theorem, the above mentioned sequence converges to z for any x E 1). Let ll,.1 = {x E 1) Ilx - zll < rt}, where rl is sufficiently small. For any c with 0 < c < rl we can find n such that IlFn x - zll < rl - c for all x E ll,.1' Hence and by Cauchy inequalities it follows that
NONLINEAR EQUATIONS
174
or, equivalently, IIAnl1 < 1, which means that the operator eilt I - A is continuously. invertible for all () E [0, 27r]. Thus condition 1) is fulfilled. ~ COROLLARY 1.1. If for some p;:: 1 the operator FP has the s-fixed point z E 1),
then z is an s-fixed point for F, too. REMARK 1.1. By Theorem IIL8.4, in condition 1) above it is enough to require the convergence of the iterations {Fnx} nEl'l for x in a dense subset of the domain 1).
If for some pEN the operator FP is completely continuous, then, by Theorem 1.8.1, the operator Ap = AP is completely continuous, too. It follows that the spectrum of the operator A coincides with the point spectrum (}l (A). Thus we get the next
result.
n C(1), X) be (1.1). Assume that for some pEN the operator FP iterations {Fn x }nEl'l converge to a certain point z E is a dense subset of1), then z is an s-fixed point for COROLLARY 1.2. Let F E H(1), X)
an operator satisfying condition is completely continuous. If the 1) for all x E 1)' u 81), where 1)' the operator F.
Recall that there exist examples of operators that are not completely continuous but have completely continuous powers (see, for instance, L. V. Kantorovic and G. P. Akilov [1]). REMARK 1.2. Actually, inequality (1.4) follows from the invariance condition (1.1), and the boundedness of the domain 1) only. Indeed, fix a ball of radius R and center z completely contained in 1), and a ball of radius M with center at the zero element of X, completely containing 1). Then for all x with Ilx - zll ~ R we have
and, by Cauchy inequalities, it follows that
for all
h with Ilhll
~
R. Therefore
But this means (according to the proof of implication 1) =} 2)) that the spectrum of the operator A lies completely in the unit disk. Thus we have proved the next important result.
s-fixed points for holomorphic operators
175
LEMMA 1.1. Let F be an operator satifying condition (1.1) and let z E 1) be a fixed
point for F. If A = F'(z), then r(A) ~ 1, where r(A) is the spectral radius of the operator A. The equality r(A) = 1 holds if and only if z is not an s-fixed point. This assertion can be also considered as an analogue of Schwartz Lemma (see §5 of Chapter III). Assume further that the operator F is continuous on the closure of the domain 1) and satisfies the stronger invariance condition (1.5) In this case, as it will be shown below, the structure of the spectrum of the operator A = F' (z) is relevant only on the boundary of the unit disk of the complex plane, and not inside this disk. For simplicity we will consider that 1) is a ball in X centered at the point z = o. THEOREM 1.2. Let us assume, in addition to condition (1.5), that F(O) = 0 and
where al (A) is the point spectrum of the operator A (i.e., the continuous and residual spectra of the operator A lie completely inside the open disk .:1 = {>. E C. : I>' I < 1}). Then z = 0 is an s-fixed point for the operator F. Suppose, on the contrary, that z = 0 is not an s-fixed point for the operator F. Then, by Lemma 1.1, we get r(A) = 1, and, consequently, there exists a number () E [0, 27r] such that e iO is an eigenvalue of the operator A. It follows that we can find an element x E 81) satifying Ax = e iO x. Choose a linear functional 1 E X* such that (x, I) = 1, 11111 = 1, and consider the analytic function
= >.eiO(x, I) +
(Q(>.x), I)
+ Q(>.x), I) =
= >.e iO +
(Q(>.x),I),
where Q = F - A and>' E C. with 1>'1 < 1. From (1.5) it follows that 11
NONLINEAR EQUATIONS
176
If the operator F satisfies (1.5), F(O) = 0, and for a certain 1, the operator FP is completely continuous, then 0 is an s-fixed point for
COROLLARY 1.3.
pEN, P ~ the operator F.
Indeed, the spectrum of the operator AP
=
(FP)'(O) satisfies the condition
In particular we have
Therefore 0 is an s-fixed point for the operator FP. By Corollary 1.1 it follows that 0 is an s-fixed point for the operator F, too. ~ Thus, we see that the geometric conditions of invariance and strong invariance enable us to infer from some local features of an operator at a fixed point, certain global properties of that point such as uniqueness, succesful approximation, and so on. In addition, the condition of strong invariance allows us to weaken slightly the requirements imposed on the spectrum of the operator A. Let us note here that the strong invariance condition alone is not enough to guarantee the attractiveness of a fixed point. Indeed, in Example V.2.4 we considered a linear, therefore holomorphic, operator A satisfying the strong invariance condition
IIAxl1 < Ilxll,
x =1= 0,
and such that its iterations Anx converge to 0 if and only if x = O. A sufficient condition for the above mentioned matter is the so-called condition of uniform or strict invariance: F(::O) ~ ::0', (1.6)
-,
where::O is strictly contained in::O. Moreover, as we will show in the sequel (Theorem 2.2), condition (1.6) is also sufficient for the existence of a fixed point (and, at the same time, for its uniqueness and for its uniform approximation by iterations of the operator F at any element x E ::0).
Criterions for the existence of an s-fixed point
177
§2. Criterions for the existence of an s-fixed point and its extension with respect to a parameter Let X and A be complex Banach spaces. Assume that 1) is an open bounded set in X and fl is a finitely connected p-open set in A (see §2 of Chapter III). Consider an operator 1> : 1) x fl ---> 1) defined for any (x,..\.) E 1) x fl, such that 1>(.,..\.) E 1i(1), X) for all fixed..\. E fl, and 1>(x, .) E 1ip(fl, X) for all fixed x E 1) (Le., the operator 1> is holomorphic on x and p-holomorphic on ..\.). A solution of the equation
x(..\.) = 1>(x(..\.) , ..\.) is called an s-solution, if for any operator Xo : fl the sequence {xn(..\.)}nEN defined by Xn+l (..\.) =
1>(xn(..\.), ..\.),
---> 1)
(2.1) such Xo (= Xo (..\.)) E 1ip(fl, 1»),
n = 0, 1,2, ... ,
(2.2)
converges in the strong topology of X to x(..\.) uniformly on each compact subset of fl, and, for a fixed..\. E fl, uniformly relatively to those xo(..\.) with values inside 1); in other words, x(..\.) is an s-fixed point of the operator 1>(., ..\.), for each..\. E fl. Now let us state the main result of this chapter.
Let 1> : 1) x fl ---> 1) be an operator which satisfies the above mentioned conditions. If, for at least one value ..\.0 E fl there exists an s-fixed point x* for the operator 1>(., ..\.0), then there exists an s-solution for equation (2.1). Moreover, for any class of operators IU s:;; 1ip(fl,1») which is closed in the topology of uniform convergence on the compact subsets of fl, and contains the orbit ~ = {v n : V n +1 = = 1>(vn' ..\.), n> N ~ 1, Vo E 1ip(fl, 1»)} of a certain element Vo, we have x(..\.) E IU. In particular, it is always true that x(..\.) E 1ip(fl, 1»). THEOREM 2.1.
u (= u(· )) E 1ip (fl, 1») the composed operator 1>(u(-),·) belongs to the class 1ip(fl, 1»), too. For any x E 1), ..\. E fl, TEA and ( E
00
(2.3) n=O
where, for any fixed..\. and x, an is a homogeneous form in T of order n, and the series in the right hand side of (2.3) converges normally for all ( E
an
1
= -2 . (P) 7l"1
J
1<:I=p
(2.4)
NONLINEAR EQUATIONS
178
On the other hand, since the operator t[J is holomorphic on x, it follows that for any fixed ( with 1(1 = p and any element hEX with a sufficiently small norm, a representation of the following form
= t[J(x, A + (T) + t[J~(x, A + (T)h + Q(x, h, A + (T)
(2.5)
holds, where IIQ(x, h, A + (T)II = o(llhll). Moreover, by the generalized Schwartz Lemma (see §5 of Chapter III) we get IIQ(x, h, A + (T)II ~ Mllhl1 2 , where, due to the boundedness of the set :D, the constant M can be chosen independently of A Eiland x E :D. In its turn, the linear part in (2.5) can be represented in the integral form
t[J~(x, A + (T)h = ~ 2m
J
(P)
C 2 t[J(x
+ th, A + (T)dt,
Itl=6
where the number 8 is small enough, and t E C. Hence it follows that the vector-function t[J~(x, A + (T)h is continuous on ( with 1(1 = p, for any fixed A E il, x E :D, hEX, TEA. Consequently, from (2.4) and (2.5) we obtain
where bn(x, A, T)h
J t[J~(x,
= ~ (P) 2m
A + (T)hc(n+l)d(
1(I=p
is a linear operator on h (b n E L(X)), and
Ilcn(x, h, A, T)II =
~ 2m
J
(P)
c(nH)Q(x, h, A + (T)d(
~
1(I=p
~ 2~ . MllhI1 2 27rp-n = o(llhll), which clearly means that the operator an (·, A, T) is Frechet differentiable on x for any fixed A E il, TEA. From the above proof it follows that the vector-functions an(u(A+(T), A, T) are holomorphic in the disk 1(1 < p(A, T) for any n = 0,1,2, .... Then, by Weierstrass Theorem (see B. V. Shabat [1]), we obtain that for any functional I E X* with 00
11111 = 1, the vector-function rp(() =
L (an(u(A + (T), A, T), f) is holomorphic in the n=O
Criterions for the existence of an s-fixed point
179
same disk. Notice that the last series converges absolutely by virtue of the normal convergence of the series (2.3), and
+ (T), A + (T)), I).
By Theorem III.2.1 the vector-function <1>(U(A + (T), A + (T)) is also holomorphic in the disk 1(1 < p(A,T). Hence it follows that the operators {Xn(-)}nEl"l defined by (2.2) are p-holomorphic in f2, i.e., if Xo E Hp(f2, 1», then x n (·) E Hp(f2, 1», for any n = 1,2, .... Let us show now that the sequence {x n (· )}nEl"l converges uniformly on any compact subset of f2. Taking into account the structure of the set f2 and by the generalized Vitali Theorem it is sufficient to prove the convergence on any disk f2p(AO, T) <;;:; f2 centered at a point AO and with a non-zero radius p. Since the point x* is an s-fixed point for the operator <1>(., AO) : 1> ~ 1>, then, by Theorem 1.1, there exists a neighborhood II of the point x* on which the operator <1>( . , AO) is a contraction with a certain constant q (0 ~ q < 1), relatively to a norm II . 11* equivalent to the original norm of the space X. Moreover, we can find a ball of radius M with respect to the norm II . 11* which completely contains the domain 1>. Then, by (2.3) and the Cauchy inequalities (III.3.8), it follows that
Since the operators a n (·, AO, T) are holomorphic in 1>, using the estimates in Corollary III.6.1 we get
for all x, y in the ballllr with a sufficiently small radius r point x*. Then
>
0 and centered at the
(Xl
n=O
~
2Mr- I KJ,
qllx - yll* +
1 _ KJ,P
Ilx - yll*·
p
Clearly, for any ql with q < ql < 1 we can choose a number PI with 0 < PI < p, such that the right hand side of the last inequality does not exceed qIilx - yll*. Moreover, by Schwartz Lemma, taking into account that <1>(x*, >'0) = x*, we obtain
NONLINEAR EQUATIONS
180
for 1(1 ~ P2 = min{p(2M)-1, pd· r(1 - ql). Hence it follows that
+ (7) - x* II ~ ~ 11<1>(x, >'0 + (7) - <1>(x*, >'0 + (7)11 + 11<1>(x*, >'0 + (7) ~ qllix - x* II + r(1 - qd ~ r. 11<1>(x, >'0
x*11 ~
for any ( with 1(1 ~ P2 and x E 1.4. Thus, for all >. E Dp2 (>'0,7), i.e., >. = >'0 + (7, where 11(11 < P2, the operator <1>( " >.) maps the ball 1.4 centered at x* into itself, and <1>(.,>.) is a ql-contraction on 1.4. Consequently, the sequence {Xn(>')}nEN converges to a certain element x(>.) E ~, for all >. E D p2 (>'0, 7). Since Ilxn(>')11 ~ M, for all n = 0,1,2, ... , it follows that the sequence converges uniformly, so our claim is proved. By Theorem III.6.4 we have x n (-) E Hp(D, ~). It is clear that x(·) E sn whenever sn is a subset of Hp(D,~) closed in the topology of uniform convergence on the compact subsets of D, and containing the orbit ~ = {vn+l = <1>( V n , >.) : n > N} of a certain element Vo E Hp(D, ~). ~ REMARK 2.1. The last assertion of our theorem turns out to be interesting in the
following situation. Consider an open set Dl ::J D. The class sn = H(Dl'~) is contained in Hp(D, ~). Although the operator <1>(x, .) E Hp(D,~) is not necessarily holomorphic in D l , it is possible for the orbit ~ of an element Vo E H( D l , ~) to be contained in H(Dl' ~). Then, taking into account that the class sn = H(Dl'~) is closed in the topology of uniform convergence on compact subsets, we obtain that the solution x( >.) of equation (1.1) is holomorphic in D l . Some related examples will be presented below (see §6). As a consequence of Theorem 2.1 we obtain the next criterion for the existence of an s-fixed point for a holomorphic operator. THEOREM 2.2. Let ~ be a bounded domain which is (>star-shaped relatively to
oE
X. An operator F E H(~, X) satisfying the condition F(~) S;;; ~ has an s-fixed point Z E ~ if and only if there exist a subset i> s;;; ~, a number pEN, and a number c > 0 such that dist(FP(i», ai» =
inf _ Ilx - yll > c.
(2.6)
yEFP(1J) xE8i>
<J
Necessity. Let z E
~
be an s-fixed point for F in ~ and set r
=
inf Ilx-yli.
xE81J
Then, by the uniform convergence of the iterations F (x) to z it follows that for any c with 0 < c < r and any set i> completely contained in ~ together with its boundary, n
Criterions for the existence of an s-fixed point
181
there exists a number p such that IIFP(x) - zll < r - E, for all x E i). Clearly (2.6) is fulfilled for such E and p. Sufficiency. Consider the operator
IlFxll ~ r, for all x with
Ilxll < R.
Then the operator F has an s-fixed point z E :D.
It seems that this result was obtained for the first time by M. Helve [1] in the finite-dimensional case and by C. Early and R. Hamilton [1] in the general case of a Banach space, using the generalized Poincare metric (see also L. Harris [4], T. Hayden and T. Suffridge [1], and K. Goebel and S. Reich [1]). The results in this section have various applications which will be listed in the sequel. Now we return to an example of an integral equation studied in the previous chapter. Let us consider equation (V.4.20):
J b
x(t)
=
K(u, t)xm(u)du + y(t).
a
As it was proved in §4 of Chapter III, for y (= y(t)) with relation
IIF(x)11
~
K,(b - a)rm
+ lIyll < r,
Ilyll <
m - 1 the mJmK,(b - a)
NONLINEAR EQUATIONS
182
is fulfilled for all x such that Ilxll ~ r < r*, where r* is the largest rooth of the equation r = Nr m + Ilyll, where N = K(b - a) and K = max IK(a, t)l· a,tE[a, bJ By Corollary .2.1 the integral equation above has a unique solution x(t), satisfying the condition max Ix(t)1 < r. [a, bJ Moreover, for any function xo(t) continuous on [a, bJ and such that max Ixo(t)1 ~ r, [a, bJ the sequence b
Xn+l(t) =
J
K(a, t)x~(a)da + yet),
a
n = 0,1,2, ... , converges to x(t) uniformly on [a, bJ. Let us emphasize once more that for r* < Ilxll, where r* is the smallest positive rooth of the equation Nr m + lIyll = r, the operator fails to be a contraction (see Subsection 4 in §4 of Chapter IV).
§3. Regular fixed points. Geometric criterions Let 1) be a bounded domain in X. In this section we will assume that the operator F, holomorphic in 1), admits a continuous extension on the boundary f)1) of the domain 1), i.e., FE H(1),X)nC(1),X). The invariance condition (1.1) is in this case equivalent to condition F(1)) ~ 1). (3.1) Example V.2.1 considered in the previous chapter shows that condition (3.1) above is not enough for the existence of a fixed point for the operator F (the operator in Example V.2.1 is holomorphic and, moreover, non-expansive). Therefore in the subsequent considerations it will be necessary to impose supplementary restrictions. 1. Let us start with a simple assertion.
Assume that the domain 1) is
THEOREM 3.1.
°
Ilxt -
F(xt) II = IltF(xt) -
F(Xt)11 ~ IIF(xt)11 It-1I,
Regular fixed points. Geometric criterions
183
by the boundedness of the operator F on :D, we obtain that the sequence {yt} = {(I - F) (Xt)} converges for t ---> 1. Then, under our assumptions, it follows that the set {xt} is compact. Since the operator F is continuous on :D we infer that any limit point of 'the set {Xt} is a fixed point for the operator F. ~ In particular, if the operator F is completely continuous, then the operator I - F is proper. Therefore, Theorem 3.1 represents an analogue of the Schauder principle for holomorphic operators acting on a ((>star-shaped domain (recall that the classical Schauder principle is stated for convex domains). However, like the Schauder principle, Theorem 3.1 above does not guarantee the uniqueness of a fixed point. Moreover, neither the isolation, nor the succesful approximation of a fixed point follows. Note also that, if we deal with assumptions weaker than the strict invariance condition (V.2.6), then, naturally, we have to replace the succesful approximation property with a weaker condition. A suitable substitute is the property of regularity. DEFINITION 3.1. Let z be a fixed point for an operator F which is Frechet differentiable in a neighborhood of z. The point z is called a regular point (or a proper point) for the operator F if 1 is not an eigenvalue of the operator A = F'(z). Otherwise the point z is called singular.
By Theorem 1.1 an s-fixed point for a differentiable operator on a complex Banach space is regular. For a completely continuous differentiable operator the regularity of a fixed point implies its isolation. Indeed, the equation
x
=
F(x)
can be written as z+ h
= z + Ah + Q(z, h),
where h has a sufficiently small norm and IIQ(z, h) II = o(lIhll). Since 1 is not an eigenvalue of the compact operator A, the operator I - A is continuously invertible and the last equation is equivalent to h
= (I - A)-lQ(Z, h) = t[J(h).
Taking into account the equality t[J'(O) =0 we obtain that h=O is an s-fixed point for the operator t[J on a certain ball with a sufficiently small radius. Hence the considered equation has the unique solution X=Z in some neighborhood of the point z.
NONLINEAR EQUATIONS
184
We will show below that in some special cases a regular fixed point for a not necessarily completely continuous holomorphic operator may be isolated as well. The next result introduces a geometric condition which guarantees not only the isolation of a fixed point for a non-compact operator, but also its uniqueness.
n C(1), X) be an operator which maps 1) into itself. Assume that for some p ~ 1 the operator FP is compact and has no fixed points on the boundary [)1). Then the operator F has a unique fixed point inside 1) and this point is regular.
THEOREM 3.2. Let F E H(1), X)
<J Let J be the set of all fixed points for the operator FP. Under our assumptions there exists a domain 1) 1 such that 1) 1 :J J and 1) 1 C 1). Consider on 1) 1 the vector field 1- tFn which, for t E (0, 1) and t close to 1, is homotopic to the field 1- Fn and has no zeroes on [)1)1 (see M. A. Krasnoselskii and P. P. Zabreiko [1]). By Theorem 2.2, for such t's the operator Ft = tFP has an s-fixed point Xt E 1). It follows that these points are regular. Since the winding number 'Y of two homotopic non-degenerate vector fields is the same, then
Therefore
Xt
E 1)1, for any
t
E (0, 1) close to 1, and, moreover
Hence the equality
is equivalent to the fact that the set J has but one point z which is regular. By uniqueness, the point z is a fixed point for the operator F, too (see §1 of Chapter V). Moreover, if we assume that z is not a regular point for the operator F, then there exists an element u =1= 0, U E X, such that Au = u, where A = F'(z). But then APu = (FP)'(z)u = u, a contradiction. ~ Thus we obtained an analogue of the extension to the boundary principle for a non-discrete analytic set (see M. Herve [1]): if the set of fixed points of a compact holomorphic operator has more than one point inside a domain, then this set contains points from the boundary of the domain, too. 2. For operators that are holomorphic in domains of a strongly convex space we may give up the compactness condition. Moreover, for a holomorphic operator defined on
Regular fixed points. Geometric criterions
185
a ball of a strongly convex space and mapping that ball into itself, the set of all its fixed points has an affine structure.
Let ~ be the unit ball - centered at 0 - of a strongly convex Banach space X, and F E H(~, X) n C(~, X) an operator satisfying condition (3.1). Let ~ be the set of all fixed points for the operator F. IfO E ~ then ~ n ~ = Q; n ~, where Q; is the set of all fixed points for the linear operator A = F' (0). THEOREM 3.3 (W. Rudin [1]).
<J Let z E ~, z -=I- O. Choose a point u E 8~ with Ilull = 1, such that z = tu for some t E (0, 1). Consider now the function 'P(A) = (F(AU), I), holomorphic on the unit disk .:1 = {A E C : IAI < I}, where f E X* satisfies the conditions (u, I) = 1 and I f I = 1 (such a functional does exist by a consequence of the Hahn-Banach Theorem). The function 'P satisfies, obviously, the conditions: I'P(A)I < 1, for all A E .:1, 'P(O) = 0, and 'P'(O) = (Au, I). Since 'P(t) = t -=I- 0, by a consequence of the one-dimensional Schwartz Lemma we obtain 'P(A) = A, for all A E .:1, hence 'P'(A) = (Au, I) = 1. Since the space X is strongly convex it follows that Au = u, and, therefore Az = z, i.e., z E Q;. Thus we proved the inclusion ~ n ~ S;;; Q;. Conversely, let z E Q; n ~, z -=I- O. Then, for some function 'P as above, we have
'P(A) = (A(Au), I)
+ (Q(AU), f) = A(U, I) + (Q(AU), f),
where Q(x) = F(x) - A(x) and Q'(O) = O. Then 'P'(O) particular, 'P(t) = t, or (C 1 F(tu), I) = 1.
=
1, hence 'P(A) == A. In
By the strong convexity of X we get that C 1 F(tu) = u, or Fz = z. ~ From the above proof it follows that if 0 is a regular point for an operator F satisfying the conditions of the theorem, then 0 is the unique fixed point inside ~. However, it is necessary to note that from Rudin's theorem we can not infer any relationship between the boundary points of the sets Q; and ~; there are examples when ~ n 8~ -=I- Q; n 8~. Nevertheless, an assertion analogous to Theorem 3.2 is also true in this case. Namely, the absence of fixed points on the boundary of the domain ~ implies the existence of a unique fixed point inside ~. COROLLARY 3.1. ~
n 8~
=
Let the domain
~
and the operator F be as in Theorem 3.3. If
0, then the point 0 is the unique fixed point for the operator F.
Consider the holomorphic operator
NONLINEAR EQUATIONS
186
for any x E 8'1). From Theorem 1.2 it follows that A = 1 is not an eigenvalue of the operator P'(O), and this means that 0 is the unique fixed point for the operator P. But the operators F'(O) and p'(O), as well as the operators F and P, have the same fixed points, hence the assertion follows. ~ 3. In this subsection we consider strongly non-expansive operators, that is
IIF(x) -
F(y)11 <
Ilx - yll,
for all x -1= y E '1),
(3.2)
or non-expansive operators, that is
IIF(x) -
F(y)11
~
Ilx - yll,
for all x, y
E
'1),
(3.3)
acting on a complex Banach space X. THEOREM 3.4. Let '1) be the unit ball of a complex Banach space X, centered at
the origin, FE 1{('1), X) n C('1), X) an operator satisfying condition (3.1), and z E '1) a fixed point for F. If the condition a) the intersection between the spectrum O'(A) of the operator A = F'(z) and the unit disk of the complex plane
are fulfilled, then z is an s-fixed point for the operator F, and the sequence
xn
=
OXn-l
+ (1 -
converges to z uniformly on '1) for any 0
E
(3.4)
O)F(xn-d
[0, 1) and Xo
E
'1).
(See the theorems in §2 of Chapter V.) The operator F(O) = 01 + (1- O)F satisfies condition a), or both the conditions a) and b), simultaneously with the operator F. Indeed, from (3.2) or b) it follows that IIAII ~ 1, hence rCA) ~ 1. By Lemma IV.1.1 there are no points from 0'3(A) on the unit circle of the complex plane. Therefore, by a), on that circle we can find points from O'l(A) only. Let B(O) = (F(O))'(O), and choose an arbitrary t E [0, 211'J. <J
.
Then e 1t I - B(O)
= (1 - 0)
(e1 _ 00) e 0 I - A . Setting = l-=-e we obtain IAI it -
A
it -
~ 1. If
F satisfies condition a), then the operator AI - A is either a homeomorphism of X, or has a non-zero kernel. In the first case the operator B(O) - eit I is a homeomorphism
Regular fixed points. Geometric criterions
187
of X, hence eit is a regular point of it, and in the second case p, = eit is an eigenvalue of B(B). Thus it is enough to prove the uniform convergence to z of the sequence of iterations {Fn(XO)}nEN' Xo E :D, only. Let hEX be such that Ilhll = 1 and z = th for a certain t E [0, 1). Consider the vector-function A('\) = F'('\h). By Lemma 1.4.2, it is analytic in the open disk .:1 of the complex plane C. From conditions (3.2) or b) the inequality sup
IIF'(x)11
~1
XEiJ
follows for any closed convex subset jj <;;;:D. Therefore r(A('\)) ~ 1, for all ,\ According to the maximum principle for the spectral radius, if
r(A('\o)) < 1,
E
.:1.
(3.5)
for some '\0 E .:1, then r(A('\)) < 1, for all ,\ E .:1. Thus reACt)) = r(F'(z)) < 1. The uniform convergence of the sequence of iterations {Fn(xO)}nEN, Xo E :D, follows from this inequality and Theorem 1.1. Thus, it remains to prove that condition a) together with one of conditions (3.2) or b) imply (3.5). We show that we can set '\0 = 0, i.e., rCA) < 1. 1. Assume that conditions a) and (3.2) are fulfilled. Then the operator G defined by G(x) = F(x) - F(O) satisfies the inequality
IIG(x)11 < Ilxll,
0 -1= x E :D.
(3.6)
Observe that G'(O) = A and let us show that the operator A has no eigenvalues on the unit circle of the complex plane. Suppose, on the contrary, that the equality Au = eitu is fulfilled for a number e it , where t E [0, 2nJ, and a certain u E :D with u -1= O. By the Hahn-Banach-Suchomlinoff Theorem we can choose a functional f E X* such that Ilfll = 1 and (u,1) = Ilull = r < 1, and let us consider the function
1) = = '\e it + r- 1(H('\u), 1).
g('\) = (r-1G(,\u), =
(r- 1A('\u), 1) + (r- 1H('\u), f)
(3.7)
This function is analytic on the closed disk .:1, and, by (3.6), the inequality Ig('\)1 < 1 holds for all ). E .:1. From (3.7) it follows that g'(O) = e it . According to Schwartz Lemma we obtain g(,\) = '\e it . Then l(r-1G(u), 1)1 = Ig(l)1 = 1, an equality which contradicts (3.6). Hence the unit circle consists of regular points for the operator A, so that, by Corollaries IV.1.l and IV.2.2, we obtain rCA) < 1.
NONLINEAR EQUATIONS
188
2. Assume that conditions a) and b) are fulfilled. Then, for all x E :D and t E [0,27rJ, the inequality IleitG(x)11 :::;; Ilxll is true. By Theorem 3.3 this inequality and the strong convexity of the space X imply that the set of all fixed points for the operator e-itG which lie inside :D coincides with the set of all fixed points for the linear operator e-itA = e-itG'(O). Therefore, if eit E O"I(A), i.e., there exists v EX with Ilvll = 1 such that v = e-itAv, then AV = eitG(Av), for all A E .d. It follows that F(AV) = G(AV) + F(O) = AW + s, where W = e-itv and s = F(O). As in the first part of the proof, we choose f E X* with Ilfll = 1 such that (w, f) = Ilwll = 1, and consider the function 'ljJ(A) = (F(AV),I) which is analytic in.d. It is easy to see that 1'ljJ(A) I :::;; 1, for all A E .d. On the other hand
'ljJ(A) = A(W, I)
+ (s, I)
=
A + (s, I).
Hence IA + (s, f) I :::;; 1, for all A E .d, and, therefore, (s, I)
(w + s, I)
= (w, I) = 1.
= 0.
But then (3.8)
Since IIAw + sll = IIF(Av)11 :::;; 1, for all A E .d, taking the limit as A ~ 1 we obtain Ilw + sll :::;; 1. By Theorem 0.4.7, this inequality, (3.8) and the strong convexity of the space X imply that s (= F(O)) = 0, which contradicts b). As above, we conclude that r(A) < 1. ~ If the operator F is completely continuous, then, by Theorem 1.8.1, the operator A = F'(O) is completely continuous, too; so, by Theorem IV.4.3, condition a) is fulfilled. The converse is not true. Some examples of non-completely continuous operators for which the Fnkhet derivative is completely continuous are presented, for instance, in L. V. Kantorovic and G. P. Akilov [1]. Note also - as Example V.2.4 shows - that condition a) is essential. In the case of a reflexive space X it is sufficient to assume that the operator ei7) I - A is normally solvable for all "7 E [0, 27rJ, in order to accomplish condition a). This follows from the Statistical Ergodic Theorem of K. Yoshida (see §4 of Chapter IV). REMARK 3.1.
It is not difficult to observe that the assertions in Theorem 3.4 remain true if one replaces the operator A with the operator An, where An is a certain power of the operator A, n = 1, 2, .... In many cases this may be useful, as the next example shows. REMARK 3.2.
Let X = Tn be the space of all bounded sequences x = (Xl, X2, ... , of complex numbers, with the norm Ilxll = sup IXkl, and let :D be the open
EXAMPLE 3.1.
Xk, ... )
k
A priori estimates unit ball of X. Consider the operator F : :D where
1
2
--+
189
:D given by Fx = (Zl' Z2, ... , Zk, ... ) 3
= "2 X1 + 8' Zk = 0, for k = 2m + 1, m = 1,2, ... , = (k - 2)(k -1)- l xk_l for k = 2m, m = 1,2, ... Zl
Zk
It is easy to see that F satisfies condition (3.3) on :D and condition (3.2) inside :D.
Moreover, obviously, the point x* = (~,O,O, ... ) is a fixed point for the operator F. Further, the operator F is Frechet differentiable and the operator A = F'(O) is defined by Ax
=
(Pl,P2,'" ,Pk, .. .), where
Pk
=
°
= 2m + 1, m = 0,1,2, ... , = (k - 2)(k - 1)- l xk_l for k = 2m, m = 1,2, ... Pk
for k
The operator A is not compact but it is nilpotent, since A2 = 0; hence condition a) is fulfilled for the operator A2. Consequently, the sequence (3.4) converges to x* for any Xo E :D.
§4. A priori estimates and the extension of an s-solution to the boundary of the domain The criterions for the existence of a fixed point considered in the previous section were based, essentially, on the following approach. Given the operator F(·) : :D --+ :D, we first construct an operator pC, .) : :D x L\ --+ :D, where L\ = {A E
(4.1)
This second step is possible because of additional properties of the operator F such as, for instance, the properness of I - F or the compactness of F, and so on. In the theory of linear equations the method of extending a solution with respect to a parameter is well-known. Roughly speaking, the method consists in establishing a priori estimates of the possible solutions for equation (4.1) - with P a linear operator - , for all >. E [0, 1) (let say, upper bounds for the norm Ilx(>')II) and then proving that such a solution can be extended on the interval [0, 1] (see, for instance, V. A. Trenogin [1]).
190
NONLINEAR EQUATIONS
We will use an analogous approach for equation (4.1) when if> is a holomorphic operator. To this end, it is enough for our purposes, to establish a priori estimates of the solutions for a linear equation associated to the given operator if>. THEOREM 4.1. Under the conditions in Theorem 2.1 let [l be the open disk ,1 =
= {>. E C : IAI < I} and assume that the operator if>(., A) : 1)
--+ 1) is continuous on for any fixed A E ,1. Moreover, suppose that the partial derivatives of the operator if> satisfy the conditions: 1) 11if>~(x, A)II ~ L < 00, (x, A) E 1) x ,1; 2) there exists 0: with 0 ~ 0: < 1 such that for all possible solutions of the linear equation
1)
(I - A(A))z
=
y,
z, Y E X,
where A(A) = if>~(x(A), A), and x(A) is a possible solution for equation (4.1) on ,1, the a priori estimate
Ilzll ~ M(1-IAI)-"'llyll
(IAI < 1)
is true. Then the vector-function X(A) ~ a solution of equation (4.1) ~ is defined and continuous on the closed disk ,1 = {A E C : IAI ~ I} and satisfies on this disk the Holder-type condition
Ilx(A) - x(A')11 ~ KIA _ A' 11 -""
(4.2)
where
and R is the radius of a ball centered at the origin and containing the domain
1).
x'(A)
= [J -
A(A)l-lif>~(X(A), A),
and, from conditions 1), 2), we obtain
IIX'(A) II ~ ML(l-IAI)-"'.
A priori estimates Therefore
1
(P)
191
1
J
x' (rei'P)dr
:::; M L
o
J
(1
~rr )<>
ML I-a'
0
and, consequently, the limit 1
lim x (re i6 ) = (P)
r-+1
J
x' (rei'P) dr
o
exists and is finite for all 'P E [0, 27r]. Thus the vector-function X(A) is defined for all A E r = aLl. Let us show first that (4.2) is fulfilled for all A, A' E r, with a suitable constant K. This will prove also the continuity of X(A) on r. Without losing the generality, we may assume that larg A-arg A'I :::; 7r. Consider the vector-function X(A) = (2R)-l x (A). Using the relation
'P 2 - 'P' 7r Ie1'P . - e1'P. I , I'P - 'P, I :::; 7r Isin - I ="2 I
for all I'P - 'P'I :::; 7r, we obtain that it is enough to prove the inequality Ilx(A) - x(A')II:::; K1 largA - argA'1 1 -<>
(4.3)
with a suitable constant K 1. To this end we may assume that Iarg A - arg A'I < 1, since, on the contrary, (4.3) with K 1 = 1 is obvious. Let us represent the left hand side of (4.3) as
X(A) - X(A') = (P)
J
x'(()d(,
I
where l is the piece-wise smooth curve consisting of the line segments [A, tAl and [tA', A'], and of the arc which connects tA and tA' along the circle 1(1 = t = 1-Iarg A - arg A'I < 1. Then
Ilx(A) - x(A')11 :::;
J
(1Ix'(rA)11
+ Ilx'(rA')II)dr + a/Alt Ilx' (tei'P) II d'P argA argA '
J
argA
MLt(l_t)-<>d'P 2R
:::;
NONLINEAR EQUATIONS
192
Thus (4.3) is fulfilled for all A, A' E
r, with
KI =maX{1,
ML 2R
(_2 +1)}. 1- IX
This means that (4.2) is fulfilled, for the same A, A', with
We fix now the points A, X E r and consider the vector-function Xl (() analytic in the disk 1(1 < 1 and continuous on the closure of this disk. Since Xl (0) = 0, by Schwartz Lemma we obtain
= X((A) - X((A'),
It follows that for any r E [0, 1] we have
Ilx(rA) - x(rX)11 ~ r sup Ilx(~A) ~
rK 2 sup I~A
-
-
x(~X)11 ~
1~1=1
~A'II-<> =
K2r<>lr(A - A'W-<> ~
(4.4)
1~1=1
Finally, let us prove inequality (4.2) - with an appropriate constant K -, for all A, A' E ,1 lying on the same ray, that is, arg A = arg A'. In this case
Ilx(A) - x(A')11 ~
1>.'1
JIlx'
(te iarg >.) II dt
~
1>'1 1>.'1
J
(4.5)
MLdt
(1 - t)<>
1>'1
Now let A and X be arbitrary elements of ,1. Choose a point A" such that and arg A" = arg X. It is easy to show that in this case we have
IA"I = IAI
IA - Alii + IA" - AI ~ 31)' - )"1·
A priori estimates
193
From inequalities (4.4) and (4.5) it follows that Ilx(A) - x(A')11 ~ Ilx(A) - X(A")II
~ K21A -
+ Ilx(A") -
+ _2a1 MLIA"
A"1 1 -a
x(A')11 ~
_ A'1 1-a
-0:
~ 31- a 2a max {K2' 1 ~a 0: ML} IA ~ 3 (~) a K21A _
A'1 1-a
~
~
A'11-a.
~
Setting K = 3(2/3)a K2 we get (4.2).
COROLLARY 4.1. Let F E H(::D, X) n C(::D, X), where::D is a C-star-shaped domain in X, and assume that for all possible solutions of the equation
B(A)Z where B(A)
= y,
AE
[0,
1], Y E X,
(4.6)
= 1- AF'(x) and x is a fixed element of::D, the a priori estimate Ilzll ~ 1'IIYII
(4.7)
is true. Then the operator F has a fixed point x* which can be approximated by the sequence Xt of the s-fixed points for the operators tF with t E [0, 1), and the rate of convergence is determined by the estimate
Ilx* - xtll ~ ~(1 - t), where ~ = (3/2)71" R . max{2, 31'} and R is the radius of a ball centered at the origin and containing ::D. Moreover, if the point x* belongs to ::D, then it is regular.
Indeed, the existence of the point x* E ::D and the estimate of the convergence rate follow straightforwardly from Theorem 4.1 if we set tP(x, A) = AFx, L = R, M = l' and 0: = o. If x* is an interior point of ::D then we can define the operator B = B(l) = 1- F'(x*). By estimate (4.7) and by the extension with respect to a parameter principle for the solutions of a linear equation (see V. A. Trenogin [1]), equation (4.6) has a unique solution for x = x*, A E [0, 1] and y E X. Consequently, the operator B is continuously invertible and A = 1 is not an eigenvalue of the operator F'(x*). ~
As an example, let us consider the integral equation
JL 1
x(t) =
0:
00
o n=2
Kn(t, s)xn(s)ds
+ (1 -
o:)y(t),
(4.8)
NONLINEAR EQUATIONS
194
where Kn(t, s) and y(t) are continuous complex-valued functions on [0, 1] satisfying the conditions ly(t)1 ~ t,
°t ~
~ 1, and 00
L
IKn(t, s)1 ~ 1.
n=2
We write equation (4.8) in the operatorial form
x = aF(x)
+ (1 -
a)y,
where F is the operator defined by the integral in the right hand side of (4.8), satisfying, obviously, the conditions
FE H('IJ, x) n C('IJ, x),
F('IJ) where
x = 0[0,
1] and 'IJ =
<;;:;
{x
'IJ,
F(O)
Ex: Ilxll
Let "Y
=
=
(sup Ilxll::;; 1
f
°
and F'(O) = 0,
= max Ix(t)1 < I}. O::;;t::;;l
n IKn(t, S)X n (S)I)-l
n=2
It is obvious that for all a E [0, "y) the operator t:Px = aF(x) + (1 - a)y acts from 'IJ into 'IJ and satisfies 11t:P'(x)11 ~ a"Y- 1 < 1. According to a consequence of the Banach principle it follows that equation (4.8) has a unique solution x(t) with IIx(t)11 < 1. Let us also show that - independently of the value of"Y - if a E (~, 1], then equation (4.8) has a unique solution. Indeed, using Schwartz Lemma, we obtain 11t:P(x)11 ~ allF(x)11
+ (1 -
a)llyll ~ allxl1 2
+ (1 -
a)llyll,
for any x with Ilxll ~ r ~ 1. Note that for any r E [(1 - a)a- 1 , 1] with 1/2 < a ~ 1, there exists q E [0, 1) such that ar2 + (1 - a) < qr. Hence 11t:P(x) II < qr, for all x with Ilxll ~ r. So our assertion follows straightforwardly from Theorem 2.2. In addition, the estimate Ilx* II < (1 - a)a is true. Thus we proved that equation (4.8) has solutions for a E [0, "Y) u (~, 1]. Finally we show that when "Y = 1/2 there exists a solution for all a E [0, 1]. To this end we consider the operator G(·, .) defined by G(A, x) = aF(x) + A(l - a)y, where A E L1 = {>. E C : IAI < I}. Obviously, G(L1 x 'IJ) <;;:; 'IJ and the operator G(O, .) = aF has the s-fixed point 0
195
Local inversion of holomorphic operators
for any a E [0, 1]. Let us show that if "I = 1/2, then the operator G ( " .) and the ssolution X(A) of equation (4.1) satisfy the conditions in Theorem 4.1 for any a E [0, 1]. Condition 1) is obviously fulfilled. We prove now condition 2). Together with the operator G consider the real-valued function f(s, t) = as 2 + (1 - a)t, s, t E [0, 1]. One of the solutions for the quadratic equation s = f(s, t) is the function
s(t)
=
1 - J1 - 4a(1 - a)t 2a '
relatively to which we assert that
Ilx(A)11 ~ s(t), Indeed, set XO(A)
= A(l - a)xo
IAI ~ t < 1.
(4.9)
and
Xn +l(A)=G(A,X n (A)),
n=0,1,2, ... ,
where Xo is an arbitrary element of :D. Obviously Ilxo(A)11 ~ s(t), for IAI ~ t, and assuming that Ilxn(A)11 ~ s(t) for IAI ~ t, then by induction we obtain
Taking the limit for n
~ 00
we get (4.9). Hence it follows that for IAI
=
t we have
III - A(A)II = III - G~(A, x(A))11 ?: 11 - IIG~(A, x(A))111 = =
?:
11 - allF'(x(A))111 ?: 11 - 2allx(A)111 ?:
11-1 + J1- 4a(1-
since 4a(1 - a) ~ 1, for a E [0, 1). By Theorem 4.1, the limit lim X(A) 'x->l
of equation (4.8). For a
=
a)tl
= x*
?:~,
exists and, obviously, x* is a solution
1, equation (4.8) has the solution x*
=
0.
§5. Local inversion of holomorphic operators and a posteriori error estimates Let us consider the equation
4>(X) =
°
(5.1)
NONLINEAR EQUATIONS
196
associated with an operator P : X -+ X, and assume that Xo E X is an approximate solution of this equation. There are a lot of methods for determining the value of xo, but many of them involve computational difficulties. Therefore in practice it is important to know - as accurate as possible - a posteriori estimates for the errors appearing in the approximate computation of Xo. (See M. A. Krasnoselskii [2], [3], and L. V. Kantorovic and G. P. Akilov [1].) The notion of error estimate, needs a more precise definition. We will state a definition introduced by M. A. Krasnoselskii.
error estimate for an approximate solution Xo of equation (5.1) is an arbitrary domain m c X containing xo, provided that this equation has at least one solution x* in m. Let J be the set of all solutions for equation (5.1). The error of an approximate solution Xo is defined by DEFINITION 5.1. An
8(xo) = inf Ilxo - xii· xEJ
Let us assume that the operator P is Frechet differentiable in a neighborhood of the point Xo. Without restricting the generality, we may set Xo = 0; on the contrary, equation (5.1) can be always replaced by the equivalent equation p(xo + h) = 0, and one has to look for the exact solutions of the last equation in a neighborhood of the approximate solution ho = O. The value P(O) E X is called the unbinding of equation (5.1). In many computing proccesses the point Xo = 0 is considered to be a sufficiently "good" approximate solution of equation (5.1) if the norm IIp(O)11 = E of the unbinding is small enough. Generally speaking, this works well in two cases: either in connection with a certain problem, when we are interested only in an approximate equation p(x) ~ 0 rather than equation (5.1), or, when the error 8(0) of the approximate solution Xo approaches 0 as E approaches O. Let us note that for a continuous operator P the converse is always true, namely, E -+ 0 for 8(0) -+ O. Therefore we are interested in estimating the value of the error 8(0) in terms of the norm E = IIp(O) II. To this end we represent the operator P as
p(x) = p(O)
+ p'(O)x + w(x),
where w(x) = o(llxll). In this section we are dealing with the so-called regular case, when the operator A = p(O) is continuously invertible. Then equation (5.1) is equivalent to
x=Q(x)+y, where Q(x)
= -A-1w(x) and y = _A-lp(O).
(5.2)
Local inversion of holomorphic operators
197
Since the operator F = I - Q satisfies the condition F'(O) = I, then, by the Local Inversion Theorem, there exists an operator G defined for all y E X with a sufficiently small norm, which is a solution for equation (5.2). Our aim is to obtain estimates - as precise as possible - for the domain of existence and continuity of the operator G. These estimates will enable us to find solutions for the above considered problem. Throughout this section we will assume again that X is a complex Banach space. To begin with, let us suppose that the operator Q is defined and differentiable on the ball 1) centered at Xo and with radius R (clearly the operator F = I - Q has the same properties). The next theorem is useful in obtaining the above mentioned estimates.
Let
THEOREM 5.1.
IIQ(x)11 : :;
x E X.
Suppose that the function 'Ij;(t) = t -
Go(Y) == 0,
Gn+1(y) = y + QG(y),
n
= 1,2, ....
Assume, by induction, that the operator G n is defined and differentiable on the ball QJ and satisfies the relation
(5.3) By the recurrence rule, the operator G n +1 is also defined and differentiable on QJ and satisfies an analogous estimate. Indeed, since the function
IIGn+1(y)11 : :; Ilyll + IIQGn(y)11 : :; Ilyll +
NONLINEAR EQUATIONS
198 point y
= O. By Vitali Theorem the sequence converges to a differentiable operator
G : QJ --+ ~ on the whole domain QJ. Thus assertion 3) is proved and by (V.5.3) assertion 2) follows immediately. Further, by continuity, the operator G satisfies the identity G(y) = y + QG(y) for all y E QJ. Hence FG(y) = y and so assertion 1) is proved, too.
~
If we set now M = sup IIQ(x)11 < 00, then, by Schwartz Lemma, we get xeD IIQ(x)11 ~ MR- 21IxI1 2. Therefore we may choose cp to be the function cp(t) = MR- 2t 2. It is easy to see that the corresponding function 'ljJ(t) = t - cp(t) provides a homeomorphism from the interval [0, R- 2(2M)-lj onto the interval [0, R- 2(4M)-lj. Then, by Theorem 5.1, the next result follows. COROLLARY 5.1. Let Q be an operator defined and differentiable on the ball ~ = = {x : Ilxll < R} and satisfying the conditions Q(O) = 0, Q'(O) = o. Let F = 1- Q.
Then the operator G radius r, where
F- 1 is defined on the ball
=
QJ
with center at the origin and
R R2} r = min { -, M and M = sup IIQ(x)ll· 2 4
Ilxll
Moreover,
~ min { R, ~ } . IIYII, Ilyll < r,
IIG(y)11
and the sequence {Gn}nEN, with Go(Y) == 0 and Gn+l(Y) = y+QGn(y), n = 1,2, ... , converges to the operator G for all y E QJ. In the finite-dimensional case this result was obtained by A. P. Yuzhakov using the theory of logarithmic residues (see [1], [2]).
then
R
R
2
2
R2(2M)-1,
·"(R) >- R - M >- R - - = -.
"'"
"'"
Let us return now to the initial equation (5.1) and assume that the operator lJ> is differentiable on the ball ofradius R with center at the origin, and let m = IIA -111. From the already obtained results it follows that if the unbinding lJ>(0) satisfies the estimate E: =
11lJ>(0) II ~ 2~ min{1, R(2M)-1},
where
M
=
m· sup 11lJ>(x) - lJ>' (O)x - lJ>(0) II, IIxll
Local inversion of holomorphic operators
199
then the error of the approximate solution Xo satisfies the relation 8(0) ~
E .
This obviously implies that 8(0)
m . R· min{l, R(2M)-1}.
-+
0 as
E -+
o.
REMARK 5.1. It is worth to emphasize the qualitative aspect of the estimate found in Corollary 5.1. In the usual estimates obtained for arbitrary Banach spaces (see, for instance, M. A. Krasnoselskii [2], [3]) the constants rand M are related - as a rule - by a condition of the form ar + (3M ~ R, a, (3 ~ O. Whence it follows that for the existence of a unique solution of the equation (5.2) both rand M are necessarily bounded (at least by the number R). In our case it is clear that if M is sufficiently small, then r may be chosen as close to R/2 as we want, and, conversely, if r is sufficiently small, then the solution of equation (5.2) exists for arbitrary large values of M. We illustrate these remarks by an example.
Consider the Hammerstein equation b
x(t) =
J
K(s, t)f(s, x(s))ds + yet),
(5.4)
a
where: a) K(s, t) is a continuous function on the rectangle [a, b] x [a, b], and b) the function f(s, t) is analytic with respect to u in the disk lui < R of the complexe plane for a fixed s E [a, b], is continuous in s E [a, b] for a fixed u, and satisfies the conditions
f(s,O)
=
f~(s,
0,
0)
=
o.
Some conditions for the existence of a unique solution of equation (5.4) were obtained in §4 of Chapter V, using the Banach principle. They were {
liM(b - a) + Ilyll ~ Ro LIi(b - a) < 1,
where
Ro < Ii
=
max
a ~ s,
t
R,
~ b
IK(s, t)l,
NONLINEAR EQUATIONS
200 M =
max
a~s~b
If(s, u)l,
and
L
=
sup
If~(s, u)l·
a~s~b lui ~Ro
lul~R
These inequalities lead to the next relations
Ro M< ",(b-a)'
L
<
1 ",(b - a ).
Since equation (5.4) is a particular case of equation (5.2), from Corollary 5.2 we obtain a different condition for the existence of a unique complex-valued solution of equation (5.4), i.e.,
IlyIIM",(b -
a) < Rmin
{~, ~},
(5.5)
where, as a matter of fact, the constant L does not appear, and the constant M may be as large as we want if the norm Ilyll of y is sufficiently small. Let us note also that condition (5.5) is simpler than condition (V.5.20) obtained in §5 of Chapter V using the Cauchy-Goursat method of majorant power series. To conclude this section, we state a consequence of the already obtained estimates, based on the so-called covering theorems. These theorems play an important role in the finite-dimensional complex analysis (see Goluzin [1], Shabat [3], Lebedev
[1]). COROLLARY 5.2. Let F be a bounded analytic operator defined on some domain 1)
and such that F'(xo) is continuously invertible at least for one point Xo E 1). Further, let {Fn}nEN with Fn E H(1), X), be a sequence of analytic operators which converges uniformly to the operator F on some neighborhood of the point Xo. Then there exist a domain mand an integer no such that
§6. Single-valued small solutions in some degenerate cases Let X, A and !il be complex Banach spaces. Consider an operator F, with values in !il, defined and Frechet differentiable on a neighborhood of the point (0,0) E X x A. Assume that
F(O, 0)
=
0.
(6.1)
Single-valued small solutions If on some neighborhood fl by the equation
c A
201
of 0 E A there exists an operator x = X(A) defined
F(X(A), A)
= 0,
(6.2)
and such that X(A) --> 0 as A --> 0, then x is called a small solution of equation (6.2). Whenever the linear operator F~(O, 0) is continuously invertible (the nondegenerate case), in view of the Implicite Function Theorem, a local implicit operator x = X(A) exists and, obviously, it is a small solution due to its continuity. In the degenerate case, when 0 is a point in the spectrum of the operator F~(O, 0), some other possibilities occur also: either there are no solutions at all, or there is a solution but it is not single-valued (in this case we have to deal with branches of the solution), or, finally, there exists a single-valued solution, but it is not small (as, for instance, when x(O) =J 0). These situations are generically called critical. They will be considered in detail in Chapter VIII. Here we will be concerned with criterions which give some sufficient conditions to avoid critical situations when the operator F~(O, 0) is degenerated, i.e., criterions for the existence of a single-valued small solution in the degenerate case. Unfortunately the criterions which will be presented below have merely a particular feature: they provide neither sufficient nor necessary conditions to handle a general non-critical situation in a degenerate case. However they are extremely important in specific problems, since any small solution is stable. 1. SMALL SOLUTIONS IN A GENERALIZED SENSE. THE FINITE-DIMENSIONAL CASE.
Let us note that, as long as we are concerned with stability, in many problems it is enough to check it only along all linear one-dimensional directions, i.e., to show that small perturbations of the parameter A along anyone-dimensional linear subspace yield small changes in the solution X(A). In this respect, the next notion of "generalized smallness" is useful.
X(A) of equation (6.2), where the operator F satisfies condition (6.1), will be called small in the generalized sense if for anyone-dimensional subspace IE <;;; A there exists a neighborhood II C IE of 0 E IE on which the vectorfunction X(A) is defined and satisfies the condition X(A) --> 0 as A --> 0, A E ll. DEFINITION 6.1. A solution
Obviously if an operator x( .) is Gateaux differentiable on a neighborhood of the point 0 and satisfies equation (6.2) and condition x(O) = 0, then it is a small solution in the generalized sense. (Therefore a small solution in the generalized sense may not be always continuous on a whole neighborhood of 0.) Of course when the space A is one-dimensional, a small solution in the generalized sense is also small
NONLINEAR EQUATIONS
202
in the usual sense and therefore the considered set-up is interesting only when the dimension of A is greater than l. The above remarks lead to the conclusion that a possible approach towards some criterions for the existence of small solutions in the generalized sense, in a degenerate case, is to look for such solutions not in the class of Frechet differentiable operators but in the class of Gateaux differentiable operators (or in the class of pholomorphic operators if X, A, ~ are complex spaces). In this respect, Theorem 2.1 turns out to be useful. Until now we used neither that the operator 1>(., .) is p-holomorphic in the variable oX, nor the p-holomorphicity of its fixed point x = x(oX) (see §2 of this chapter). In this subsection we will consider only the finite-dimensional case to illustrate how Theorem 2.1 can be used. Let 1 be an arbitrary function, holomorphic in a neigborhood of 0 E en, with 1(0) = O. For any subspace ~ ~ en, we will denote by !1o(f I ~) the multiplicity of 0 for the function 1 I ~ - the restriction of 1 to ~ (see L. A. Aizenberg and A. P. Yuzhakov [1]). Let A = en, X = ~ = e and assume that F is a function which is holomorphic in a neighborhood li l x li2 of the point (0,0) E en+! and which admits the following representation (6.3) F(x, oX) = P(oX)x + Q(x, oX), where P and Q are holomorphic in li2 ~
P(O) = Q'(O, 0) = Obviously, THEOREM
en
and lil x li2 ~
e
n
+ l , respectively, and
o. F~(O,
0)
= 0 and so we find
ourselves in a degenerate case.
6.1. Assume that the functions P and Q in (6.3) satisfy the condition
Then, for anyone-dimensional subspace <E c en there exist a neighborhood m~ <E of <E and a unique function x(oX), holomorphic in m, implicitely defined by equation (6.2) and such that x(O) = o. Thus x(oX) is a small solution of equation (6.2) in the generalized sense. Moreover, if the function
oE
l(oX) = Q(O, oX) [p(oX)]-2
(6.4)
is holomorphic on li2, then there exists a neighborhood E ~ en of the point 0 E en in which the function x(oX) is holomorphic, and, consequently, x(oX) is a small solution in the usual sense.
Single-valued small solutions <J Consider the function
tJ>(x, A)
tJ> : e n +1
=
---+
{ x0,
203
e defined by the equality
F(x, A) peA) , A i= 0 A = o.
Fix an arbitrary element h E en with Ilhll = 1. Taking into account the assumptions of the theorem we obtain the representation
tJ>(x, (h) = (kg (x, (h),
(6.5)
for all ( E e with 1(1 < p(h), where p(h) is sufficiently small, k = /-Lo(Q I em) - /-L;;::' 1, and the function g(x, (h) is holomorphic in ( for any fixed x E lit, i.e., the function tJ> is p-holomorphic in some p-open subset [l :3 0 (we can choose [l = (ti2 \ {A E en : : peA) = O}) U {O E en}). Thus equation (6.2) is equivalent to
x = tJ>(x, A) considered on the set [l. Further, from (6.5) it is obvious that for any h with Ilhll = 1 we may choose p(h) such that tJ>(x,(h) E ti2 for all ( with 1(1 < p(h) and any x E til. Therefore there exists a p-open subset [ll :3 0 in [l having the property that tJ>(x, A) E til for (x, A) E til x [l2. Since the point x = 0 is an s-fixed point for the operator tJ>(., 0) : e ---+ e (recall that tJ>(., 0) == 0), by Theorem 2.1 it follows that the last equation has a unique s-solution X(A) which is p-holomorphic in [ll. So the first assertion of the theorem is proved. For the proof of the second one let us remark that, by condition (6.4), the function Vi (A) = tJ>(0, A) = P(A)f(A) is holomorphic in some neighborhood of the point 0 E en. Therefore the functions Vk+1(A) = tJ>(VdA) , A) are also holomorphic in a neighborhood E of the point 0 E en. Thus the orbit of the element Vo == 0 (vo E H(E, C)) is completely contained in H(E, C). Hence x = x(·) E H(E, C) (see Remark 2.1). ~ and F(x, A) = A~x2 + (Ai - A2) x + A~. -A1A2, where (Al,A2) = A E e Obviously the function F admits representation (6.3) with peA) = Ai - A2, F~(O, 0) = 0, and satisfies the first requirement in Theorem 6.1. Therefore equation (6.2) has a unique small in the generalized sense solution X(A) which is p-holomorphic in some set [l <;;;; (:2, [l :3 O. We deduce easily this conclusion by a straightforward computation. Indeed, solving the quadratic equation F(x, A) = 0 EXAMPLE 6.1. Let A
= e2 ,
X
=!D = e,
2.
NONLINEAR EQUATIONS
204
and setting A = (h, where hE ((:2 with Ilhll = 1, we obtain two solutions, X2(A). One of these solutions is given by
Xl
(A) and
and, therefore, it is small in the generalized sense, while the second one satisfies the condition lim x2((h) = 00 and, consequently, it is not small. Let us note (->0
that Xl (A) is a continuous function on some p-open set [l E 0, whose intersection with ~2 is the set fl represented in Figure 2 and having the form of a helix.
Figure 2 EXAMPLE 6.2. For the same A, X and ~ let now F be the function given by
where g(A) is an arbitrary holomorphic function. Obviously F(O, 0) = 0 and F~(O, 0) = = 0, and, consequently, in this case, we may not use the classical Implicit Function Theorem. However, the function F satisfies both the first and the second requirements in Theorem 6.1. Hence the equation F(x, A) = 0 has a unique holomorphic solution X(A) such that x(O) = o. Indeed, solving straightforwardly this equation we obtain 00
X(A) = -(Ai - A2)
L
2n Ayn[g(A)t+l,
n=O
where the series in the right hand side converges uniformly on some neighborhood of the point A = o.
Single-valued sma,l1 solutions
205
2. SMALL SOLUTIONS FOR AN OPERATOR LINEAR IN THE PARAMETER. THE INFINITE DIMENSIONAL CASE. We return once more to a complex Banach space X. Now we will suppose that equation (6.2) has the next special form
Bx
=
AG(X)
+ y,
(6.6)
where: a) B : X ----+ X is a linear normally solvable operator such that (*) the kernel N(B) = ker B of the operator B is isomorphic with the linear space Sj = X~R(B) where R(B) is the range of the operator B, b) y is an arbitrary fixed element in R(B), c) G is an operator anaytic in the ball:D = {x EX: Ilx - zll < R}, where z is a solution of the equation Bz = y, and d) A E C. In this setting we have to find conditions for the existence of an analytic solution X(A) of the equation (6.6) satisfying the relation x(O) = z, and to indicate the domain of analyticity for this vector-function. Equations as (6.6) above arise in connection with a series of problems in nonlinear mechanics (see for example P. P. Rybin [1], [2], M. M. Vainberg [1] and M. M. Vainberg and V. A. Trenogin [1]). Moreover, equation (6.6) with an integral operator G, is related to the well-known Duffing equation, which has applications in electrotechnics. When N(B) = {O}, the operator B is continuously invertible and, consequently, we find ourselves under the conditions in the Implicit Function Theorem. When N(B) -I- {O}, it is convenient to assume that B is finitely degenerated, i.e., ind B = dim N(B) - dim N(B*) = 0 (B* is the adjoint of B). In this case, conditions for the existence of a unique analytic solution X(A) have been obtained for the first time apparently by K. T. Ahmedov [1]. To be more specific, according to the Nekrasov-Nazarov method, one looks for a solution represented by a power series whose convergence is established using a majorant numerical series. This approach enables one to indicate an estimate for the analyticity domain of the solution X(A) and this turns out to be important in many applications. Our aim in what follows is to provide conditions for the existence of a unique solution for equation (6.6) in the case of an arbitrary space N(B) which is isomorphic with the space Sj (this property is obvious if the operator B is finitely degenerate). These conditions are obtained by reducing our problem to finding an s-solution for an equation of type (2.1). This approach enables us to point out a different way for approximating X(A), and also to establish a more convenient and more precise estimate for the domain of analyticity.
NONLINEAR EQUATIONS
206
First of all we observe that by the Banach Theorem the operator B I X~N(B) - the restriction of B on the subspace X~N(B) - has a continuous inverse operator r : R(B) -> X~N(B). Setting x = z + h we write G(x) as G(z
+ h) = G(z) + Ah + Q(z, h),
(6.7)
where A is the Fnkhet derivative of the operator G at the point z and IIQ(z, h)11 =
= o(llhll)· Let us introduce the notations M(G) = sup IIG(x)ll; xE:D
M(Q)
= sup
IIQ(z, h)ll;
Ilhll
P : X -> Sj -
A=
A I N(B) -
the projection from X onto Sj;
the restriction of the operator A on N(B).
THEOREM 6.1. If A is an isomorphism from N(B) onto Sj, and G(z) E R(B), then
equation (6.6) has a unique analytic solution X(A) for all A such that
R . { dR d Rd 2 IAI < 8 = 211T1IM(G) mm 1, 2M(Q) , 211AII' 8M(Q)IIAII
}
'
(6.8)
where d = IIA-1 PII- 1. Moreover, X(A) can be approximated in the norm topology of X by the sequence {Xn(A)}nEN which is defined by the relations Xn(A)
= z + ArG(X n -1(A)) + Vn -1(A)),
Vn(A) = -[AAPArG(X n _1 (A))
+ APQ(z, Xn(A)
(6.9) - z)],
n = 1,2, ... , where XO(A) is an arbitrary element of X such that
Ilxo(A) - zll
=
min { R,
2M~Q)} ,
and VO(A) is an arbitrary element of N(B) such that
Ilvo(A)1I <
~f.
(6.10)
207
Single-valued small solutions
Setting x =·u+v, where u E X':""N(B), v E N(B), and projecting equation (6.6) on the subspaces Sj and R(B), respectively, we obtain a system of equations
(6.11)
PG(u+v)=O, Bu
= >"G(u + v) + y,
(6.12)
r
to the both sides of which is equivalent to equation (6.6). Applying the operator (6.12) and taking into account the equalities u = x - v and ry = z we obtain PG(x) x
= 0,
(6.13)
= z + >..rG(x) +v = FI(>..,x,v).
(6.14)
We perform now a few formal transformations, the substantiation of which will become apparent below. Let us replace h in (6.7) by >..rG(x) +v. Based on (6.14) we get G(x)
= G(z) + >"ArG(x) + Av + Q(z, >..rG(x) + v).
Applying the operator P to the both sides of the last equality, and taking into account the equalities (6.13), PG(z) = 0 and PAv = Av (recall that Av E Sj) we obtain 0= Av
+ >"PArG(x) + PQ(z, >..rG(x) + v),
or Av = -[>"PArG(x) +PQ(z,>..rG(x)
+ v)],
whence v
= _[>"A- I PArG(x) + A-I PQ(z, >..rG(x) + v)] = F2(>'" x, v).
Consider now the Banach space ~
= max{llxll, Ilvll}, for any w = (x,v)
E~.
(6.15)
=
X-t-N(B) with the norm given by Ilwll = Denoting F = (FI ,F2) we may rewrite the
system (6.14), (6.15) as w = F(>",w).
(6.16)
We show that the mapping F is defined and analytic in some domain in ~, and for all >.. E fl = {>.. E IC : 1>"1 < b} (see (6.8)) it satisfies the conditions in Theorem 2.l. The mappings Fl and F2 are defined and analytic if Ilhll (= II>..G(x)
+ vii) < r:::;; R.
(6.17)
NONLINEAR EQUATIONS
208
Consider the domain ::Dr = {w : IIx - zll < r, Ilvll < rc} included in !C. Inequality (6.17), for all w = (x, v), is obviously fulfilled under the following conditions: 0< c < 1 and
r(1 - c)
(6.18)
IAI < IlrIIM(G)'
Let us establish the existence of such an r = f > 0 and of a function 8(r) : [0, R] ----+ ----+ [0, 8] so that the mapping F leaves the domain ::Df invariant for A with IAI < 8(f). From (6.14) and (6.18) it follows that IIF1 (A,w)11 < r for any r with 0 < r < Rand IAI < p(r). Consider the inequality (6.19) It is not difficult to convince ourselves that inequality (6.19) follows from
IAI <
drc -
r2 M(Q)R- 2
IIArIIM(G)
=
tp(r),
for any r such that tp(r) > O. A direct computation shows that
{Rdc
(Rdc)2}
O~~R tp(r) = tp(r) = mm 21IArIIM(G)' 41IArIIM(G)M(Q) A
•
where f
= min { R,
,
2M~ Q) } .
Thus the mapping F is analytic and F(A, w) E ::Df for all w E ::Df and all IAI < 8(f) = min{p(f), tp(f)}. Clearly 8(f) depends also on c, and, as one can easily see, it attains its maximum value for c = 2- 1 , whence (6.8) follows. From (6.14) and (6.15) it is not difficult to remark that the operator F(O, . ) is a q-contraction on some neighborhood of the point Wo = (z,O) for a certain q with 0 < q < 1. By Theorem 2.1 we obtain an s-solution W(A) = (X(A), V(A)) of equation (6.16) and the iterative sequence Wn(A) = F(A,W n-1(A)) = (Xn(A),Vn(A)) which appears in formulas (6.9), (6.10). ~ To conclude this section let us show what the conditions in the theorem above mean in the case of a finitely degenerate operator B. Assume that N (B) has dimension m and ind B = O. Futher, let {V1, . .. , v m } be a basis of N(B) and let SJk = clin{Av1, ... ,Avk-1,Avk+1, ... ,Avm} be the linear hull of the elements AV1"'" AVk-l, AVk+1,"" Avm. Consider the spaces 'ck = = R(B)+-SJk which are proper subspaces of the space X.
Single-valued small solutions
209
The following conditions are equivalent: 1) the operator A = A I N(B) is an isomorphism from N(B) onto Sj; 2) d k = P(AVk' ~k) = inf IIAvk - xii> 0, k = 1, ... , m;
PROPOSITION 6.1.
XE£k
3) the equalities
are fulfilled for a certain basis
{'Ij;d~l
of the space N(B*) (B* is the adjoint of B).
<J From 1) it follows that the system of elements {Avd~l forms a basis of Sj. Assume that condition 2) fails for some k = 1, ... , m. Then there exists a non-zero element Y E R(B) such that
AVk = Y +
L if-k
1:::;; i
or
~
(XiAVi,
(Xi
E
m
m
Y = L;JiAvi, i=l
where ;Ji = -(Xi if i i- k, 1::;; i ::;; m, and ;Jk = 1, whence Y E Sj. This is impossible since Y i- 0 and Sj n R(B) = {o}. The implication 1) =} 2) is proved. Assume now that condition 2) is fulfilled. By a consequence of the HahnBanach Theorem (see, for instance, Chapter 0) there exist the elements 'lj;k E X*, k = 1, ... , m, such that (6.20)
('Ij;k, AVk)
=
1,
II'Ij;kll = d-,;l,
(6.21) (6.22)
for all k = 1, ... , m. From (6.20) it follows that 'lj;k E N(B*). Indeed, since ~k ::J R(B) then for any x E X the equality
holds, whence - in view of the fact that x is arbitrary - we have B*'Ij;k = O. Moreover, taking account of equality (6.21) we obtain that
NONLINEAR EQUATIONS
210
In order to prove 2) =? 3) it remains to establish that the system of elements {'¢d~l is linearly independent. Assume that
for some complex numbers 'Yk
E
C. Then for each 1 ~ i
~
m we obtain
Let us prove now implication 3) =? 1). To this end it is sufficient to show that, if condition 3) holds, then the system of elements {Avd~l is a basis of Sj. Indeed, from 3) it follows that the system {Avd~l is linearly independent. It remains to show that the elements AVi, i = 1, ... , m, lie in Sj. Assuming the contrary, we obtain that for some 1 ~ j ~ m the element AVj lies in R(B). From the Fredholm alternative it follows that AVj is orthogonal to the space N(B*), which is impossible since
for all j = 1, ... ,m.
~
Chapter VII Banach manifolds §1. Basic definitions Let Sj be a Hausdorff tolopogical space and let !D be a Banach space. DEFINITION 1.1. The space Sj is called a manifold modelled on !D, if for any point x E Sj there exist an open neighborhood ilx s;:;; Sj of x and a homeomorphism i.px : llx ----t s;:;; !D, where is an open set in !D.
mx
mx
If the space !D is the n-dimensional Euclidean space OC n , then Sj is called simply an n-dimensional manifold.
A pair (llx, i.px) as above is called a chart (sometimes, a local chart) of the manifold Sj. A family of charts {(llx,i.px)}x such that Ullx 2 Sj is called an atlas of x
the manifold Sj. We say that two charts (llx, i.px) and (lly, i.py) satisfy the conm dition of e _ compatibility, m = 0,1,2, ... ,00 (respectively, of O-compatibility), if whenever llx n lly f. 0, the mappings i.px 0 i.p:;/ : i.py(llx n lly) ----t i.px(ilx n lly) and i.py Oi.p;;l : i.px(llxnlly) ----t i.py(llxnlly) are of class em, m = 0,1,2, ... ,00 (respectively, of class 0 ~ the class of analytic mappings). DEFINITION 1.2.
If all the charts belonging to a given atlas are em-compatible, m = 0, 1,2, ... , (respectively, O-compatible), where m does not depend on the charts, then we say that the manifold is of class em (respectively, of class 0). A manifold of class eO is called a topological manifold. A manifold of class 0 is called analytic (if!D is a complex space, the manifold is called complex-analytic). 00
BANACH MANIFOLDS
212
The manifolds of class C m , m ) 1, are also called smooth manifolds. REMARK 1.1.
In connection with the above introduced definitions the following
questions arise. 1) Can we speak about an unique manifold if the space SJ is endowed with two different atlases? An answer to this question depends, of course, on the convention we agree upon. The advisability of a convention is determined by the applications which may occur later on. Usually two atlases on the same space SJ are called equivalent if any chart of one of the atlases is compatible (relatively to a fixed class C m , respectively 0) with any chart of the other atlas. In this case we say that both atlases define one and the same structure of a Banach manifold on the space SJ. 2) Are there manifolds which coincide as topological spaces but have different models? In other words, if SJ is a manifold with the model space !D, can SJ be a manifold with a model space !D' =I- !D? An answer to this question will be given in the sequel. However, even at this moment, it is obvious that, if this is the case, the spaces !D and !D' must contain some homeomorphic open sets. Any open subset of some Banach space X (in particular the space X) offers a simple example of a manifold. Indeed, for such a subset il S;;; X, the model space is the space X itself, and the atlas consists of one chart defined by the whole set il and the identity homeomorphism. These manifolds are smooth (of class C=) and even analytic. If X is the n-dimensional Euclidean space, then these manifolds are n-dimensional. Any subset of a space X consisting of isolated points is a zero-dimensional manifold. A more interesting example is offered by the circle {(x, y) E]R2 : x 2 +y2 = 1}. Indeed, let us choose the open cover of the circle given by the four semicircles {i4}[=1 obtained intersecting the circle with the half-spaces y > 0, x > 0, y < 0, x < 0, respectively. Take a point (x, y) with y > on the semicircle ill. The mapping 'PI which associates to (x, y) E ill the point x E]- 1, 1[ S;;; ]R1 is a homeomorphism. The inverse mapping 'Pi1 is given by 'Pi 1(x) = (x, v'1- x 2). The pair (il1,'Pd defines a chart. The charts (il2' 'P2), (il3, 'P3), (il3, 'P3) are obtained analogously. Thus the circle is a manifold with the model space ]R1, or, a manifold of dimension 1. Further, let us remark that the indicated atlas gives a smooth structure on the circle as a manifold. Indeed, consider, for instance, the neighborhood il12 = ill n il2 consisting
°
213
Smooth mappings
of that part of the circle which lies in the positive quadrant y > 0, x > 0. There are two homeomorphisms defined on ilj2, namely, 'PI : (x, y) ----* x and 'P2 : (x, y) ----* y. The overlap mapping 'PI 0 'P2I : Y f---7 X is given by the equality x = and is a homeomorphism from the intervaljO, 1 [ onto itself. Analogously we can show that the sphere is a smooth manifold of dimension 2. In the above mentioned examples it is worth to remark that the circle and the sphere are subsets of the spaces JR2 and JR3, respectively, which, on their turn, are also manifolds. We will be concerned with such aspects in §3.
+JT=Y2
§2. Smooth mappings The main reason for introducing the notion of manifolds consists in the possibility they offer to differentiate and integrate functions or mappings defined on "nonlinear" spaces. This is done by identifying locally these spaces with theirs models. Thus, for example, a function defined on a circle can be considered locally as a function given on an interval of the real line. Therefore it is convenient to think of a circle not as a set of points in the two-dimensional space JR2, but as a set of points which is identified locally with subsets of the one-dimensional space JR I . Let us give the precise definitions.
em,
m ~ 1, modelled on the space !D, and let f be a mapping from Sj into a Banach space 3. The mapping f is called m-times continuously Frechet differentiable at a point a E Sj if there exists a chart (lla, 'Pa) with a E lla <:;;; Sj and 'Pa : lla ----* l!1a <:;;; !D, such that the mapping f 0 'P;;I : : l!1 a ----* 3 is m-times continuously Frechet differentiable, in the usual sense, at the point b = 'Pa(a) E !D. DEFINITION 2.1. Let Sj be a manifold of class
For the correctness of this definition it has to be shown that the differentiability of f is independent of the choice of a compatible chart. Indeed, let (ll~, 'P~) with a E ll~ be another chart, compatible with (lla, 'Pa) for which the mapping f 0 'P;;I is m-times differentiable at the point b = 'Pa(a) E!D. Then the mapping f 0 'P';;I = = f 0 'P;;I 0 'Pa 0 'P';;I is m-times differentiable at the point b' = 'P~(a) E !D as the composition of two m-times differentiable mappings, since the mapping 'Pa 0 'P';;I is m-times differentiable by the compatibility of the charts. In particular, if!D = K n , then the differentiability of f means the existence of the partial derivatives of the mapping f 0 'P;;I, with respect to the variables (XI,X2, ... ,Xn ) E Kn, at the point Xa = ('PIa(a)"",'Pna(a)). Therefore the homeomorphism 'Pa is often called a local system of coordinates for the manifold Sj at the
BANACH MANIFOLDS
214
point a, and the corresponding neighborhood lla ~ Sj is called a local coordinate neighborhood. In other words, a chart in an atlas of the manifold Sj consists of a local coordinate neighborhood and a local system of coordinates. , m ~ 1, modelled on spaces lD and 3, respectively, and let 9 be a mapping from Sj into 9)1. The mapping 9 is called m-times Fn§chet differentiable at a point a E Sj if there exists a local system of coordinates '¢b at b = g(a) E 9)1, such that the mapping f = '¢b 0 9 : Sj ---> 3 defined on some neighborhood of the point a E Sj is m-times differentiable at a in the sense of Definition 2.1.
DEFINITION 2.2. Let Sj and 9)1 be two manifolds of class C m
In this case it is also easy to show that the differentiability of g, in the indicated sense, does not depend on the choice of the local system of coordinates '¢b. Moreover, we have the next obvious result. PROPOSITION 2.1. A mapping 9 : Sj ---> 9)1 is m-times differentiable at a point a E Sj
- in the sense of Definition 2.2 - if there exist two local systems of coordinates, 'Pa at the point a E Sj and '¢b at the point b = g(a) E 9)1, such that the mapping '¢b 0 9 0 'P;; 1 : lD ---> 3 is m-times Frechet differentiable at the point 'Pa (a) ElDin the usual sense. Analogously we introduce the notion of continuously differentiable mapping and the notion of analytic mapping (if the manifold is analytic) at a point, or, on some open subset, of a manifold Sj. DEFINITION 2.3. Let Sj and 9)1 be two manifolds modelled on the spaces
lD and 3,
respectively, and let 9 be an one-to-one mapping from Sj onto 9)1 such that both 9 and g-l are smooth mappings at any point. Then 9 is called a diffeomorphism between
the manifolds Sj and
9)1,
and the manifolds Sj and
9)1
are said to be diffeomorphic.
Let us note that the model spaces of some diffeomorphic manifolds coincide, up to a linear isomorphism. In particular, two finite-dimensional diffeomorphic manifolds have the same dimension.
§3. Submanifolds Let
9)1
be a manifold modelled on the space X and let 5j be a subset of 9)1.
DEFINITION 3.1. The set 5j is called a submanifold of 9)1 if for any a E 5j there
Submanifolds
215
exists a chart (f)a, a), f)a ~ 9J1, such that a(a) = 0 and a(f)a n Sj) = lU a n Ita, where lU a = a(f)a) and Ita is a linear subspace of X. A sub manifold Sj is called a direct submanifold if each of the spaces Ita has a direct complement in X, i.e., X =
!D.
Assume that
!D
is a subspace of X with a direct complement in X, i.e.,
X where
3
= !D+3,
(3.1)
is a linear subspace of X. Then the following conditions are equivalent:
a) for any point a E Sj there exists a chart (ita, CPa) such that the homeomorphism'ljJa = cp;;-l : lU a ----; ita, lU a = CPa(ita ), is continuously Fhkhet differentiable as a mapping from !D into X and the linear operator P'ljJ~ (b) : !D ----; !D is continuously invertible, where P is the projection from X onto !D, and 'ljJ~ (b) is the Fhkhet derivative of'ljJa at b = CPa(a) E!D; b) for any point a E Sj there exist a neighborhood it~ ~ Sj of the point a, a neighborhood IU~ of the point c = Pa E !D, and a Fnkhet differentiable mapping f : IU~ ----; 3, such that
{XEX: x=u+f(u), uEIU~}=it~.
is sufficient to prove the implication a) =} b) only, since the inverse implication b) =} a) is obvious. Indeed, the chart (ita, CPa) where ita = ll'a and CPa is the inverse of the mapping 'ljJa(u) = U + f(u) (i.e., CPa is the projection mapping
BANACH MANIFOLDS
216
from il'a into~) satisfies all the requirements in condition a). More precisely we have PljJ~(b) = I I ~, where I is the identity mapping on X and b = c = Pa. Assume that condition a) is fulfilled. By the Inverse Mapping Theorem it follows that P'l/Ja : QJ~ ----+ ~ is a diffeomorphism on some neighborhood QJ~ s;;: ~ of the point b E ~, i.e., for all x sufficiently close to the point a and y sufficiently close to the point b, the equation
has the unique solution
y = g(Px) = g(u),
u = Px,
and the mapping 9 is differentiable on a neighborhood of the point c = Pa. Further, let ll~ = 'l/Ja(QJ~). Then the points of the set ll~ are exactly the points x E X for which the equality (3.2) Qx = Q'l/Ja(g(Px)) is fulfilled, where Q = I - P. Setting x = u + v, where u f(u) = Q'l/Ja(g(u)) the desired conclusion follows. ~
Px, v
Qx and
Assume that the hypotheses of Theorem 3.1 and one of the conditions a) or b) are fulfilled. Then S) is a direct submanifold of X. THEOREM 3.2.
----+
X defined on some neighborhood 1)a C X
h = tfJx = P(x - a) where
+ Qx -
f(Px),
(3.3)
f is given by the right hand side of (3.2). Setting u = Px, v = Qx, we have tfJ~(a) = P -
f~(Pa) and tfJ~(a) = Q. Since both the partial derivatives are continuous on some neighborhood of the point a, tfJ is differentiable on that neighborhood. Moreover, tfJ is a diffeomorphism, since its inverse mapping IJt = tfJ- 1 : X ----+ 1)a, defined by the equality x = h + f(P(h + a)) + Pa, is also differentiable on some neighborhood of the point h = O. Further, by (3.2) and (3.3) we obtain tfJ(a) = 0 and tfJ(x) = P(x - a) E ~ for all xES) n 1)a, i.e., by (3.1), S) is a direct submanifold of X. ~
The mapping lJta : ~ ----+ S) satisfying condition a) in Theorem 3.1 is called a local parametrization of the manifold S). The mapping lJta(u) = u+ f(u) appearing in condition b) of Theorem 3.1 gives also a parametrization of the manifold Sj, which, due to its special form, is called an explicit or a direct parametrization.
Submanifolds
217
In the case of a finite-dimensional space X any submanifold admits obviously a local parametrization given by a suitable permutation of the coordinates. Therefore, from Theorems 3.1 and 3.2 the next result follows.
Let X = OC N and let Sj t;;; X be a smooth manifold of dimension n ~ N. Then Sj is a submanifold of X if and only if one of the following conditions is fulfilled: a) for any point a E Sj there exists a chart (ita, <Pa) such that the homeomorphism 'l/Ja =
Xl =XI,X2 =X2,···,Xn =Xn ,
In many mathematical theories the notion of submanifold is more convenient and more useful than the abstract notion of manifold. Moreover, in the mathematical literature, the term "manifold" is often used to call a set which, according to our definition above, is a submanifold (see, for instance, H. Cartan [1] and J. Milnor [2]). Nevertheless, in applications, both notions are equally justified. To be more specific, let us notice that a mechanical system with a finite number of degrees of freedom can be considered as a finite-dimensional manifold, and for this description it is not necessary to use any ambient space. For example, a system with four degrees of freedom, which is given as a set of points in the three-dimensional space, can not be considered as a sub manifold of that space. However, it is possible to "embed" such a system in a certain abstract space, or a manifold, of a greater dimension using some one-to-one mappings, and this enables us, sometimes, to obtain a series of new properties of the considered system.
9J1 be two manifolds. An embedding of Sj into 9J1 is any mapping f : Sj --> 9J1 such that the image f(Sj) is a submanifold of 9J1, and f induces a diffeomorphism between Sj and f(Sj). DEFINITION 3.2. Let Sj and
For our purposes it is not necessary to develop here the theory of embeddings. We mention only, without proof, that any smooth finite-dimensional manifold can be smoothly embedded in some finite-dimensional Euclidean space.
BANACH MANIFOLDS
218
§4. Complex manifolds and Stein manifolds 1. As we have already mentioned, a manifold 5) is called complex if its model space is a complex Banach space. Of course, any smooth complex manifold 5) modelled on ~ can be considered as a real smooth manifold modelled on ~lft - the realification of the space ~ (in the finite-dimensional case, a complex manifold of dimmension n is a real manifold of dimension 2n). Further, it is clear that if the mappings
Note that in the finite-dimensional case it is enough to require the analyticity of one of the indicated mappings only, since its inverse is also analytic, by a theorem of Osgood.
In what follows we will be concerned mainly with complex manifolds, i.e., manifolds of class 0, modelled on complex Banach spaces. The notion of a Fn3chet differentiable mapping f from a complex analytic manifold 5) into a complex analytic manifold [)1 is introduced with respect to the previous Definitions 2.1 and 2.2. If such a mapping f is Frechet differentiable, in the complex sense, on neighborhoods of any point in 5), then the mapping f is called analytic (or holomorphic). The term "holomorphic mapping" is mostly used in the theory of complex manifolds. In what follows we will use it in order to keep the tradition.
In particular, if [)1 = C is the complex field, then f is called a complex-valued analytic function on 5). The set of all complex-valued analytic functions on 5) will be denoted by A(5)). Let F(5)) be an arbitrary set of functions defined on 5) with values in C. DEFINITION 4.1. For an arbitrary compact set IB
Q;F(j») =
{x
E 5) :
If(x)l::(
sup
c
If(y)1
5) the set
for all
f
E F(5))}
yEI!3
is called the F(5))-convex hull of the set lB. In particular, the A(5))-convex hull of a compact set IB holomorphic convex hull of the set.
~ 5)
is called the
219
Complex manifolds and Stein manifolds
( 4.1)
A manifold Sj is called F(Sj )-convex if i5 F(Sj) is a compact subset of Sj for any compact subset ® S;;; Sj. In particular, the A(Sj)-convex manifolds are called holomorphically convex. DEFINITION 4.2.
In order to give some examples of holomorphically convex manifolds let us consider first the case of a complex Banach space X. For any subset ® S;;; X its geometric closed convex hull co ® can be also defined as co® =
{x
EX:
Il(x)l::;:
sup li(y)1 for alli E yErB
x*}.
If ® is compact, then by a theorem of Mazur, co ® is compact, too. Since X* = = L(X,q S;;; A(X), by (4.1) we obtain that i5 A (Sj) C co®, hence i5 A (Sj) is compact for all compact subsets of X, and, consequently, X is a holomorphically convex manifold. Let now Sj S;;; X be a complex-analytic manifold modelled on ~, where ~ S;;; S;;; X is a subspace of X. Assume that the hypoteses in Theorem 3.1 and one of the conditions a) or b) are fulfilled. According to Theorem 3.2, Sj is a submanifold of X. Since the restriction to Sj of any analytic function on X is an analytic function on Sj, then using again (4.1) we have (4.2)
for any compact subset ® S;;; Sj, where i5 A(X) denotes the A(X)-convex hull of ® considered as a compact subset of X. Clearly, by (4.2) and since i5 A(X) is compact, it follows that if Sj is a closed subset of X, then Sj is a holomorphically convex manifold. Actually this last assumption can be replaced by a weaker one, taking into account the fact that the geometric convex hull of a compact subset of the Banach space X is closed. The easy proof of the next result is left to the reader. Let X and Sj be as above, with Sj not necessarily closed in X. Let :D S;;; X be a domain, convex in the geometric sense, such that Sj S;;; :D. If Sj is closed in :D with respect to the induced topology on :D, then Sj is a holomorphically convex manifold. LEMMA 4.1.
We will now define a new class of manifolds.
BANACH MANIFOLDS
220
DEFINITION 4.3. A complex-analytic manifold Sj of a finite dimension n, which is
countable at infinity, is called a Stein manifold if a) Sj is holomorphically convex; b) Sj is holomorphically separate, i.e., for any a, bE Sj with a I- b there exists IE A(Sj) such that I(a) I- I(b); c) for any a E Sj there exists a collection of n functions II,··· , In E A (Sj ) which provide a local coordinate system at a. Let us recall that a Hausdorff topological space Sj is said to be countable at infinity if there exists a countable family <5 1 , <5 2 , ... of compact subsets of Sj such that each compact subset of Sj is contained in some <5 j, j = 1, 2, .... THEOREM 4.1. Any complex-analytic manifold Sj which is a closed submanifold of
a Stein manifold M is a Stein manifold.
Conditions a) and b) are obvious, since the restriction to Sj of any function which is analytic in the whole manifold M is an analytic function in Sj. Let us prove condition c). Take an arbitrary point a E Sj. Since a E M, there exists a system of functions II, ... , 1m E A(M), m = dim M, which provide a local coordinate system for M at a. On the other hand, since Sj is a submanifold of M of dimension n ~ m, there exists a local chart CD, 'P) for M with 'P = ('PI, ... , 'Pm) and a E ~, such that
~
n Sj = {x
E
M
'Pn+dx) = ... = 'Pm(x) = O}.
Since the Jacobian det(8 I i
1-0,
)
8'Pj
i,j=l,
,m
then for a suitable choice of 1 ~ iI, ... , in ~ m we have
det (
:~: )
/L,v=l,. ,n
I- O.
The last relation shows that the restrictions of Ii! , ... ,lin to Sj provide a local system of coordinates for Sj at a. This completes the proof. • Since the space em, as well as any convex domain in em are Stein manifolds we have: COROLLARY 4.2. Every complex-analytic manifold Sj which is a closed submanifold
of em is a Stein manifold of dimension n ~ m.
Complex manifolds and Stein manifolds
221
Assume now that the ambient manifold for 5j is a convex domain 1) in an infinite dimensional Banach space X. In this case, we can not use Theorem 4.1 since 1) is not a Stein manifold (in particular 1) is not finite-dimensional). However the next generalization of Corollary 4.2 is true. Let 1) be a convex domain in a Banach space X (in particular, X), and let 5j be a finite-dimensional countable at infinity complex manifold, which is a submanifold of1) and is closed with respect to the induced topology on 1). Then 5j is a Stein manifold.
THEOREM 4.2. 1) =
The holomorphic convexity of the manifold 5j follows directly from Lemma 4.1. For the holomorphic separation of 5j it is enough to recall that X* = L(X,
where (el, ... ,en) is a fixed basis in \Ca, and consider the restrictions of these functionals to 5j. Since rank(II, ... , in) = dim5j, then Udi=l (~ A(5j)) provides a local system of coordinates for 5j at the point a. The theorem is proved. ~
Chapter VIII Non-regular solutions of nonlinear equations §1. Ramification of solutions.
Statement of the problem Let 1) and [l be domains in the Banach spaces X and A, respectively, and let F be an operator acting from the topological product Sj = 1) x [l into X. Assume that, for a given Ao E [l, the equation x = F(x, A) (1.1 ) has the solution Xo E 1), i.e., Xo = F(xo, Ao). The point (xo, Ao) E Sj is called a regular point of equation (1.1) if, for any A in some neighborhood it <;;;; [l of the point Ao, there exists a unique solution x(A) E 1) of equation (1.1) such that x(Ao) = Xo. Otherwise the point (xo, Ao) is said to be non-regular. There are two kinds of non-regular points. A non-regular point (xo, Ao) of the first kind is characterized by the following property: there exists a neighborhood it of the point Ao such that equation (1.1) has no solutions for any A E it with A i- Ao, and for A = Ao it has the unique solution Xo. Such points are called non-prolongation points. A non-regular point (xo, Ao) of the second kind has the property that in any neighborhood of the point Ao there exists A for which equation (1.1) has more than one solution. Such points are called ramification points. If on some neighborhood it of the point Ao the identity
Xo == F(xo, A) is true, and the point (xo, Ao) is a ramification point, then Ao is called a bifurcation point.
NON-REGULAR SOLUTIONS
224
Bifurcation points play an important role in different applications. For instance, in perturbation theory for elastic systems, bifurcation points define the so-called critical loadings; in wave theory the bifurcation points define characteristic flows in which the waves occur, and so on.
Let us consider some examples. EXAMPLE 1.1. Let
x=
A = C. Consider the equation
whose solutions are
X(A) =
~±
v'f=4):.
2 2 The point (0,0) is a regular point for this equation, since for all A sufficiently close to Aa = 0 the equation has the unique solution x(A) = (1 - VI - 4A)/2 close to Xa = 0 and satisfying the condition x(O) = o. At the same time (~, is a ramification point since for all A close to Aa = there exist two solutions
i)
i
and satisfying the condition EXAMPLE 1.2. Let
Xi
(i)
= ~, i = 1,2.
x = e[O, 1] and A = 1Ft Consider on x the equation 1
x(t) = j (AX(S) a
+ x2(s))ds.
We can easily see that this equation has exactly two solutions: xdt) == 0 and X2(t) == == I-A. If we consider as (xa, Ao) the point (0,0), then this point is not a ramification point since in the domain :D of all x E x such that Ilxll = max Ix(t)1 < ~, the [a, 1)
considered equation has, for any A < ~, the unique solution x( t) == o. But if we consider the point (0,1), then this is a ramification point: the solutions X1(t) and X2(t) "branch out" from it. Moreover, the point AO = 1 is a bifurcation point; setting 1
F(x, A) = j(AX(S) +x2(s)ds we obtain F(O,A) == 0 for all A E 1Ft a An important particular case of equation (1.1) is offered by an equation of the form x = AP(x), (1.2)
Equations of ramification
225
where the operator
When equation (1.2) has a non-trivial solution x* i- 0, x* E 1), for some A E OC, then, by analogy with the theory of linear operators, the number AO 1 is called an eigenvalue of the operator
(xo, Ao) with Xo
=
F(xo, AO)'
Then, by Theorem V.5.1 about implicit operators, the condition for A = 1 to be a regular point of the operator A = F~(xo, Ao) gives a possible answer to Problems 1 and 2. This condition is but a sufficient one. In §6 of Chapter VI we gave some examples which showed that the above mentioned condition is not a necessary one; in those examples the point (xo, Ao) was a regular point for equation (1.1), but the number A = 1 was a point of the spectrum of the operator A. By the same Theorem V.5.1, this last condition is necessary for solving Problems 3 and 4.
§2. Equations of ramification 1. THE GENERAL SETTING. Let (xo, Ao) be a given point satisfying equation (1.1), i.e., Xo = F(xo, Ao). Without loosing the generality, we may assume, and we will,
NON-REGULAR SOLUTIONS
226
= (0,0). (If this is not the case, then we can make the following changes: y = X-Xo, f-l = A- Ao, p(y, f-l) = F(y+xo, f-l+ Ao) -Xo, and can consider the equation
that (xo, Ao)
y
1)
= p(y, f-l)
which is equivalent to equation (1.1) and, moreover, satisfies p(O, 0) = 0.) Further, let us assume that the following conditions are satisfied: i) the operator F is Frechet differentiable in the domain Sj = 1) x il, where
<; X, il <; A; ii) 1 is an eigenvalue of the operator A = iii) the operator B
=
F~(O,
0); I - A is normally solvable and splitable (see §7 of
Chapter IV). Then the operator B admits a regular matrix representation
obtained using a quadruple of projections Pi, P2 = I - H, Ql, Q2 = I - Ql. We denote (as in Chapter IV) lEi = PiX, 1E2 = P2X, IE~ = Q1X, lEg = Q2X. Then equation (1.1) can be written as a system of equations
{
BllU + B 12 V = Q1P(U + V,A) B21 U + B 22V = Q2P( U + v, A),
where U = PiX, V = P2x, P(-, A) of (2.1) as
= F(-, A) -
(2.1)
A. Let us rewrite the second equation
and consider the point (u, v) E lEi xA as a parameter. Since the operator H~(O, 0, 0) (= Q2P~(0, 0, 0) = B 22 ) is continuously invertible, by Theorem V.5.1 about implicit operators the last equation has a unique solution
= B22 -
v
= R(u, A)
(2.2)
defined on some neighborhood of the point (0,0) E lEi X A and satisfying R(O,O) = O. By substituing the right hand side of (2.2) in the first equation of (2.1) we obtain the equation (2.3) feu, A) = 0, which defines the implicit operator u = U(A) on the subspace PiX. If there exists a (single- or a multiple-valued) solution U(A) of equation (2.3), then U(A) provides a solution X(A) of the initial equation (1.1) given by the formula
X(A) = U(A)
+ R(U(A), A).
(2.4)
227
Equations of ramification
Equation (2.3) is called the equation of ramification for equation (1.1). Using formula (IV.7.4) it is easy to see that the operator f has the form
and acts from the space lEI into IE~. Equation (2.3) like the initial equation (1.1) is degenerate, i.e., f~(O, 0) = 0 (this follows easily by a direct computation). Therefore the Implicit Function Theorem can not be applied in this case. However, sometimes this equation can be studied using methods from finite-dimensional analysis. Namely, if the subspaces lEI and IE~ are finite-dimensional, then equation (2.3) can be represented as a system with a finite number of equations and a finite number of unknowns. There are many well-known different methods for the search and the study of solutions for such equations: elimination methods, Newton diagrams, algebraic methods, and others (see, for instance, M. M. Vainberg and V. A. Trenogin [1]). The above described general setting allows us to find different equations of ramification, depending on the properties of the operator B and on the choice of the subspaces lEI, IE~, 1E2, lEg, and of the projections PI, P2 , QI, Q2 relatively to which the regular matrix representation of the operator B is obtained. LIAPUNOV EQUATION OF RAMIFICATION. Let us consider the case when lEI N(B) and lEg = R(B) (recall that B is a normally solvable operator). Then the operator B admits the matrix representation
2.
=
B _ -
(0 0) 0
B22
'
and system (2.1) becomes {
0 = (I - Q)p(u + v,),) B 22 V = Qp(u + v, ),),
where Q is a projection onto R(B), P is a projection onto N(B) and u v = (1 - P)x. Using the first of these equations, the second one can be written as
If v
= R( u,),)
Px,
is a solution of this equation, then the equation of ramification (2.3) is
f(u,),)
= (I - Q)p(u + R(u, ),),),) = o.
(2.5)
228
NON-REGULAR SOLUTIONS
Assume, in addition, that B is a Fredholm operator (see §6 of Chapter IV), i.e., dimN(B) = dimN(B*) = n. Let {'Pdf=l and Ndf=l be bases in N(B) and N(B*), respectively, and let bdf=l C X* and {zdf=l C X be two systems of elements biorthogonal to these bases, respectively. Moreover, we consider that bdf=l was chosen such that (v, "Ii) = 0, i = 1, ... , n, for any v E X~N(B). This can be done by Hahn-Banach Theorem. Then every element U E N(B) can be represented as n
U=
L (i'Pi,
(i E lK
i=l
(lK is, as usual, the field of scalars). At the same time the equality n
(1 - Q)x
=
L(x, ~i)Zi
i=l holds for the projection 1 - Q. Taking into account the linear independence of the elements {zi}f=l' equation (2.5) can be rewritten as a system of n equations with n unknowns (1,"" (n:
k
=
1,2, ... ,no Equations as in (2.6) above, corresponding to integral equations with analytic functions, have been considered for the first time by A. M. Liapunov in connection with the equilibrium figures of a fluid mass in a revolving motion. Therefore, equations (2.6) are usually called the Liapunov equations of ramification. The above described approach, which reduces the initial equation to the system (2.6), is called the Liapunov procedure. This procedure can be also effective in the study of the solutions for equations with fixed values of the parameter A (see §4 below). 3. SCHMIDT EQUATION OF RAMIFICATION. With the notations and under the assumptions of the previous subsection, we infer other equations of ramification using essentially Lemma IV.6.1 (Schmidt Lemma). According to this lemma, the Fredholm operator B can be represented as
B=C-K,
Equations of ramification
229
where C is a continuously invertible operator on X, and K is a finite rank operator of the form n
Kx
= L(X,!'i)Zi. i=1
Then equation (1.1) can be rewritten as n
Cx
= <.p(x,>..) + LWiZi,
(2.7)
i=l
where
Wi
=
(X,!'i),
i = 1, ... ,no
(2.8)
If in equation (2.7) we consider (WI, ... , Wn , >..) E OC n x A as a parameter, then, using the Implicit Function Theorem, we obtain a solution for this equation of the form
x = S(W, >..),
(2.9)
where W = (WI, ... , wn ). To determine w, we substitute (2.9) in (2.8). Thus we obtain the following system of equations
Wi=(S(W1, ... ,Wn ,>"),!'i),
i=1,2, ... ,n,
(2.10)
which is called the Schmidt system of equations of ramification. This procedure, in contrast to the ones described in the previous Subsections 1 and 2, has the advantage that the considered operators are defined on the whole space X, and it is not necessary to split the operator B or to use the projections PI, P2, Q1, Q2. At the same time, there exists a direct connection between the Liapunov and the Schmidt equations of ramification.
In order to establish this connection, we rewrite first equation (2.10) into n
a slightly different form. Set x
= U
+ v, where U
=
L (i'Pi i=l
E
N(B), and remark that
by the equalities ('Pi,!'j) = 8ij , i,j = 1, ... , nand (v,!'j) = 0, j = 1, ... , n, we have
Further, C'Pi
=
B'Pi + K'Pi
=
Zi and, therefore
NON-REGULAR SOLUTIONS
230
Consequently, equation (2.7) can be written as (2.11) Let v = R((l, ... , (n, A) be a solution of equation (2.11), considered as an implicit n
function of the parameter ((1, ... , (n, A). Substituting x =
L (iifi + R((l, ... , (n, A) i=l
in (2.8) we obtain for (j the equations (j
=
(~(iifi + R((l, ... , (n, A), 'Yj )
,
j
=
1, ... , n,
which are equivalent to the equations in (2.10). According to (ifi, 'Yj) = 8ij , i,j = 1, ... , n, we finally get
( R (~ (iifi, A)
,'Yj )
=
0,
j
=
1, ... ,n.
(2.12)
Let us show now that system (2.12) represents nothing else but the Liapunov system of equations of ramification. In the literature (see, for instance, V. A. Trenogin [1]) (2.12) is often called the Liapunov-Schmidt system of equations of ramification.
Indeed, since C* = B* + K*, then C*'ljJi the system (2.12) can be written as
=
"Ii, i
= 1, ... ,n, and, consequently,
It is easy to see that the space x~N(B) is the kernel of the operator K, i.e., Kv = 0 for all v E x~N(B). Therefore equation (2.11) is equivalent to the next equation
BV=P(U+V,A), where B is the restriction of the operator B to the subspace x~N(B). Hence it follows that R(u,A) = B22Ip(u+R(u, A), A) = R(U,A), where B22 and R(-,·) are the same as in Subsection 2. From (2.11) and (2.12) we get the system of equations
Equations of ramification for an analytic operator which coincides with the Liapunov system of equations of ramification (2.6).
231 ~
As we have noted above, the Liapunov system is degenerated, i.e., det
II ~~; (0, 0) II = o.
Therefore, if it has a solution ((A) = ((1 (A), ... , (n(A)), then the functions (i(A) can be either single-valued or multiple-valued. Their feature determines the feature of the vector-function U(A), and, consequently, the feature of the solution X(A) of the initial equation (2.1).
§3. Equations of ramification for an analytic operator. The problem of coefficients 1. Let us assume that the operator F in equation (1.1) is analytic with respect to
the arguments (x, A) in some neighborhood of (0,0) E X x A. As before we will use the notations A = F~(O, 0), B = I - A, and we will suppose that the operator B is splitable. We consider again the system of equations
Bv = (u + v, A)
(3.1)
feu, A) = (I - Q)(u + R(U(A), A)) = 0
(3.2)
equivalent to equation (1.1), where u = Px, v = (I - P)x, B is the restriction of the operator QB(I - P) to the subspace ~2 = (I - P)X, P is a projection onto N(B), Q is a projection onto R(B), and = F - A. By the Implicit Function Theorem, equation (3.1) has a solution v = R(u, A) on some neighborhood U x Q of (0,0), where U ~ N(B), U 3 0, Q ~ A, Q 3 O. This solution is analytic with respect to the variables (u, A) in the whole neighborhood U x Q. Consequently, the operator f (-, .) that defines the left hand side of the equation of ramification (3.2), is also analytic in U x Q. Moreover, from ~(O, 0) == 0 it follows that R~(O, 0) == 0, hence f~(O, 0) == o. This means that the operator f admits on the domain U ~ N(B) the representation 00
feu, A) = feu, 0) +
L
Li(U, A),
(3.3)
i=O
where Li(U, A) are homogeneous forms of degree i = 0,1,2, ... in u for a fixed A E Q, and analytic operators in A for a fixed u E U. In addition Li(u,O) == o.
NON-REGULAR SOLUTIONS
232 On the other hand
00
J(u,O) =
L
Hk(u),
(3.4)
k=2
where Hk are homogeneous forms of degree k = 2,3, .... The series in (3.3) and (3.4) have non-zero radii of absolute convergence. The operators Hk, k = 2,3, ... , and L i , i = 0, 1,2, ... , are called the coefficients of ramification for equation (3.2). If all the coefficients of ramification are zero, then Problem 3 in §1 about the number of "branches" has an obvious answer: this number is infinite. Indeed, equality (3.2) is identically fulfilled on il. Replacing u(>.) in equality (2.4) by an operator u : il -4 il, we obtain different small solutions for equation (1.1). The features of these solutions are determined by the feature of the chosen operator u( . ). In particular, for >. = 0, the point x = 0 is a non-isolated solution for equation x = F(x,O). Moreover, according to R~ (0,0) = 0 we obtain that the operator G = = I I N(B) + R(·, .) is a bijection from the set il <;;; N(B) onto the set J n 1), where 1) is a neighborhood of 0 in X and J is the set of all fixed points for the operator F(·,O) (see Problem 4 in §1). If B is a Fredholm operator, and not all the coefficients of ramification Hk, k = 2,3, ... , are equal to zero, then it can be shown that the equation of ramification has a finite number of small solutions, determined by the first operator Hk which is different from zero (see, for instance, M. M. Vainberg and V. A. Trenogin [1]). In practice, it turns out frequently that both the construction of the equation of ramification and the computation of its coeficients are either rather cumbersome, or lead to difficult to solve problems. That is why the equation of ramification is not easy to handle. In the above cited book there are examples in which the first two or three coefficients of ramification are computed by reccurent methods, but these computations require a lot of work. In §4 below we will show that if the operator F(·, 0) leaves invariant some neighborhood of the point x = 0, then all the coefficients Hk in the decomposition of J(u,O) are equal to zero.
§4. The description of the set of fixed points for an analytic operator In this section we are mainly concerned with Problem 4 of §1, in the case of a complex Banach space X. More precisely, we will study the set of fixed points for the analytic operator T = F(·, 0) acting from X into X, or, equivalently, the set of solutions of
The description of set of fixed points
233
equation (1.1) for a fixed value of the parameter. We show that under some natural conditions this set has a quite complete description. Let us assume that in the set where T is defined there exists a bounded domain :D ~ X which is invariant for T, i.e., T(:D)
~
:D,
(4.1)
and let J be the set of all fixed points for T in :D. If :D is a ball in a strictly convex space, then the set J has a very simple structure: either J is void, or it is an affine variety in X (see Rudin Theorem Theorem 3.3. in §3 of Chapter VI). We emphasize that the condition of strict convexity is essential in this respect. For example, the operator F acting on the space X = CC 2 and defined by F(x) = (Xl, O.5(XI + x~)), X = (XI,X2), leaves invariant the unit polydisk :D = {x
E
CC 2
:
IXil < 1, i = 1, 2},
but the set J is not linear. Moreover, the intersection of 1)1 and J, where 1)1 is the set of all eigenvectors of the operator F'(O) corresponding to the eigenvalue 1, has but one point, X = O. Thus, in the absence of the strict convexity, Rudin Theorem is not true. Nevertheless, as we will see below, the assertions analogous to some consequences of Rudin Theorem such as the regularity of an isolated point, the continuousness of the set of non-regular points, its connectivity, the approach to the boundary, and others, are also true for a larger class of spaces. Let T be an analytic operator on an arbitrary domain :D of a complex Banach space X satisfying condition (1.1), and let J be the set of all fixed points for T in :D.
Let 9]1 be a connected component of the set J containing a given point x* E J. If the operator cP = I - T is Fredholm on some neighborhood of the set 9]1, then: 1) 9]1 is a direct complex-analytic manifold of dimension n, where n = dimker(I - T'(x*)); 2) if the set :D is convex and J contains more than one point, then n > O. THEOREM 4.1.
Recall that a nonlinear operator cP is called Fredholm at a point x* if the linear operator cP' (x*) is Fredholm.
By assertion 1) in Theorem 4.1 it follows, in particular, that whenever x* is an isolated point in J, then n = 0 = dimker(I - T'(x*)), a conclusion which -
REMARK 4.1.
NON-REGULAR SOLUTIONS
234
by Banach Theorem - implies the continuous invertibility of the operator 1- T' (x*). This means that an isolated fixed point is regular. Moreover, from assertion 2) it follows that, in this case, x* is the only fixed point for T in :D. Thus, for an operator T which leaves invariant some convex domain :D, and has a Fredholm complement cp up to the identity operator, the next alternative is true: either a fixed point for T (if it exists) is unique and regular, or there is an infinity of fixed points and all of them are non-regular and non-isolated. THEOREM 4.2. Let:D be a convex domain in X. Assume that the operator T admits
a continuous extension on :D and the operator cP = I - T is proper. Then: 1) the closure J of the set J in :D is connected; 2) if cP is Fredholm on some neighborhood of J, then J is a Stein manifold,
and, moreover, if J contains more than one point, then
J n 8:D =I=- 0.
The proofs of both theorems are based on the following preliminary results. Let :D be an arbitrary bounded domain in X, J =I=- 0, x* E J, and 91 = = ker(I - A), where A = T'(x*). Assume that for some subspace IE <;;; 91 there exists a linear projection P onto IE, commuting with the operator A: PA
=
AP.
(4.2)
1.1 <;;; IE of the point 0 E IE, with values in ker P such that 1IK,(u)11 = o(llull). Then the identity
LEMMA 4.1. Let K, be an analytic operator defined on some neighborhood
PT(x*
is true, for all u E 1.1 such that x* <J
+ u + K,(u» == Px* + u + u + K,(u)
(4.3)
E :D.
Without loosing the generality we will consider that x*
=0E
X and that
1.1 is a ball centered at the point 0 E IE, such that u + K,( u) E :D for all u E 1.1. Let us fix an arbitrary u E 1.1 and consider the vector-functions
1(1::::;: I}, Tn = ToT n - 1 , n = 1,2, ... , and TO = I. Since where (E L1 = {( E C the operator T satisfies the condition T(:D) <;;; :D and the domain :D is bounded, then all the functions 'Pn are defined on L1 and there exists M < 00 such that II'Pn(()1I ::::;: M,
n
=
1,2, ....
235
The description of set of fixed points The functions 'Pn have the power series representations
where hn E X is the first non-zero coefficient - if such a coefficient exists representation of the nonlinear part of 'Pn, and mn :? 2. By the Cauchy inequalities (see §2 of Chapter III) we have
Ilhnll:::::::M,
in the
(4.4)
n=1,2, ...
From the conditions APx = Px and (4.2) we get
PAnx == Px,
n
1,2, ...
=
By these relations, a direct computation of 'Pn(() , based on an induction argument, gives
'Pd()
= PT((u + K((U))
= PA((u) + PAK((U) + PS((u + K((U)) where S
=
= (u
+ h1(m + ... , 1
= T - A, and 'P2(() = PT2((U
+ K((U))
= PT(A(((u
+ K((U)) + S((u + K((U)))
=
= PA 2((u)+PAK((U)+PAS((U+K((U))+PS((u+AK((U)+S((U+K((U))) = =
(u+2h 1 (m 1
+ ...
Thus we conclude that mn = ml and h n = n· hl' n = 1,2, .... By (4.4) this is possible if and only if hl = O. Consequently
and setting (
= 1 we
obtain (4.3). ~
LEMMA 4.2. Let 1> be a convex domain in X and let VJt be a connected component of the set~. Let us assume that for some neighborhood 11 ~ 1> of VJt n 1>, such that 11 n (~\ VJt) = 0 the operator clJ = 1- T is uniformly non-degenerate on the set r = 811 \ 81>, i.e., there exists p > 0 such that
IlclJ(x)II :? p,
(4.5)
NON-REGULAR SOLUTIONS
236 for all x E T. Then 9J1
=
J, i.e., J is connected.
Let y E :r be an element which runs through some compact set ~ S;;; 1). For any e: E (0,1] we consider the operator T (= T(e:,y)) defined by the equality
T(x)
=
(1 - e:)T(x)
+ e:y.
By the convexity of 1) and by condition (4.1) the operator strictly inside 1). Consequently, the equation
x = T(x)
(4.6)
T
maps the domain 1)
(4.7)
has a unique solution x (= x( e:, y)) which depends continuously on y E ~ and satisfies the relation (4.8) Ilx(e:,y) - Tx(e:,y)11 < e:M, where M is the radius of a ball which contains completely the domain 1). Choose now two arbitrary points a, b E J and set y( 0:) = (1 - o:)a + o:b, where 0: E [0, 1]. Then the vector-function x(e:, y(o:)) defined by equation (4.7) is continuous in 0: on the interval [0, 1] and satisfies the conditions
x(e:,y(O)) = a,
x(e:,y(l)) = b.
Now if we assume that J =I- 9J1, then choosing b E J \ 9J1 and a E 9J1 n 1) we obtain that there exists 0:0 E [0, 1] for which x(e:, y(O:o)) E T. Setting e: = pM- 1 , by (4.8) we get the inequality II
Proof of Theorem 4.1. We will consider that x* = 0 E:r. Let A = T'(O). From the relations Am = (Tm)'(O), by the boundedness of the domain 1) and from Cauchy inequalities it follows that (4.9) for a suitable constant M <
00.
From Corollary V.8.2 we obtain the decomposition
:r =
»1+91,
(4.10)
where »1 = ker(I - A) and 9l = Im(I - A). We use now the Liapunov procedure. Let P and Q be linear projections from :r onto the subspaces »1 and 9l, respectively. Then the equation
x = T(x),
The description of set of fixed points
237
for x of a sufficiently small norm, is equivalent to the system of equations
u=PT(u+v)
(4.11)
v=QT(u+v),
(4.12)
where u = Px and v = Qx. By the Implicit Function Theorem equation (4.12) has a unique solution v = t£(u) E 9l (= ker P), which is analytic in some neighborhood ti of the point 0 E 1)1. A direct computation shows that 11t£(u)11 = o(llull). By substituing this solution in (4.11), according to Lemma 4.1 we obtain an identity which is satisfied for all u E ti. In other words, the point
x=u+t£(u)
(4.13)
is a fixed point for the operator T, for all u E ti ~ 1)1. Now, if we choose a neighborhood D ~ X of the point 0 E X such that P D ~ ti ~ 1)1, then the homeomorphism 'ljJ: D ----+ X defined by the equality 1Jr(x) = X-t£(Px) gives a local chart of the domain ~ as a manifold, which satisfies the condition 1Jr(0) = 0 and, by (4.13), the condition 1Jr(9J1 n D) ~ 1)1, too. Thus, by (4.10) it follows that 9J1 is a direct sub manifold of X of dimension n = dim 1)1. The assertion 1) in Theorem 4.1 is proved. We prove now assertion 2). Assume that 0 E ~ is an isolated point in this set (n = 0). Then the operator 1> = 1- T is locally invertible, i.e., we can find the numbers ri,r2 > 0 such that for all y with Ilyll ~r2 the equation 1>(x) = y has a unique solution x = 1>-l(y) in the ball Ilxll ~ ri, which depends analyticaly on y and satisfies the condition x(O) = O. Then for all x with Ilxll = ri inequality (4.5) is fulfilled for a suitable p. By Lemma 4.2, we obtain that ~ = {O}, which contradicts the assumption in 2). ~
Proof of Theorem 4.2. 1) Let 9J1 be a connected component of the set ~. Choose a neighborhood ti c ~ which contains 9J1n~ and such that tin (~\9J1) = 0. Then 1>(x) =I- 0, whence, by the fact that 1> is proper, inequality (4.5) follows with a suitable p. Our assertion is a consequence of Lemma 4.2. 2) If we assume in addition that 1> is a Fredholm operator, then, according to Theorem 4.1, it follows that 9J1 is a finite-dimensional complex-analytic submanifold of ~. Again by the fact that 1> is proper we conclude that ~ is countable at infinity and thus, by Theorem VII.4.2, it is a Stein manifold. Finally, let us suppose J n a~ = 0. Then there exists a neighborhood ~i (~ ~) of the set ~ such that all the values of the operator 1> (= 1- T) on a~i are
NON-REGULAR SOLUTIONS
238
different from zero. Consequently, for
t
In Theorem 4.1 we used that
decomposition X
=
lJ1-i-91 (*). During the translation of this book the paper of
P. Mazet and J.-P. Vigue [1] appeared in which it is also proved that under condition
( *) the set
~ of the fixed points of the operator T can be represented locally as a
complex-analytic manifold. Some versions of this theorem have been proved earlier by E. Vesentini [2], [3] and J.-P. Vigue [2] in the n-dimensional case, by M. AbdAlla [1] for the product of Hilbert balls, and by D. Shoikhet [2] in the general case of Banach spaces (see also M. Herve [2] and D. Shoikhet [3]). In addition let us note that P. Mazet and J.-P. Vigue established also some other conditions which are equivalent to condition (*).
By Theorem IV.8.2 such a decomposition holds too in the case when the operator
Chapter IX Operators on spaces with indefinite metric §1. Spaces with indefinite metric Let us consider a complex Banach space 23 endowed with a norm 11·11, and decomposed in a topological direct sum
The bounded mutually complementar projections generated by this decomposition are denoted by P±, such that 23± = P±23, Pl = P± and P+ + P_ = I. We define a functional J" on 23 by the formula
J,,(x) =
Ilx+II" -llx-II",
1/
> 0, x
E
23, x± = P±x.
(1.1 )
We say that the functional J" introduces an indefinite metric on the space 23, which is also called a J,,-metric. The space 23 endowed with a J,,-metric is called, for brevity, a J,,-space. J,,-spaces appeared as a natural generalization of Hilbert spaces with indefinite metric, which we will briefly recall in what follows. Let us consider a complex Hilbert space 5) with the inner product (., .) (see §11 of Chapter 0). Let G be a bounded self-adjoint invertible operator on 5) (i.e., the operator G is defined on the whole space 5) and the inverse operator G- 1 exists). We define a sesquilinear form [., .] on 5), called the G-metric, by the formula [x, y] = (Gx, y), x, y E 5). If the operator G is indefinite, i.e., its spectrum on the real line lies on both sides of zero, then the G-metric is also indefinite: for a non-zero vector x E 5) the "scalar square" [x, x] can be positive, negative or equal to zero. Assume now that the operator G- 1
240
SPACES WITH INDEFINITE METRIC
is bounded. Then the space Jj endowed with the G-metric is called a Krein space. Denoting by E).. the spectral resolution of the operator G we set
o
P- =
J
dE)..,
-00
Then P_ and P+ are mutually orthogonal projections: pl = P± = PJ., P+ +P_ = I. Setting Jj± = P±Jj we obtain a decomposition of the space Jj in the orthogonal direct sum Jj = Jj+ EB Jj_. Let us introduce on Jj a new inner product (., . ) by the formula
(x, y)
=
[p+x, yj - [p_x, y],
x, Y
E Jj.
This new inner product is equivalent to the initial one. Indeed, if x E Jj, then
(x,x) = (G(P+ - P_)x,x)::;;
IIGII(x,x),
and, conversely,
(x, x)
= (GP+x, x) - (GP_x, x) ;?
;?Al(P+X,X) +A2(P-X,X);? min{..\I,A2}(X,X), where Al = inf{..\ : A E u(G I Jj+)} > 0 and -A2 = sup{..\ : A E u(G I Jj_)} < Relatively to the inner product ( ., .), the form [., .j can be written as
[x, yj = (Jx, y),
J = P+ - P_.
o.
(1.2)
The operator J satisfies the properties J* = J- 1 = J and is called the canonical symmetry of the considered Krein space. The form [., .j given by (1.2) is called the J-metric of the space, and the space Jj with the inner product (., . ) and the J-metric [ ., . J is called a J -space. Setting Ilxll = (x,x)! for each x E Jj, we have (1.3)
hence a Krein space is a Hilbert h-space, in which P+, P_ are orthogonal projections. Let us return now to an arbitrary Jv-space 23. A vector x E 23 is said to be positive, negative or neutral if it satisfies the condition Jv(x) > 0, Jv(x) < 0 or Jv(x) = 0, respectively. A vector x such that Jv(x) ;? 0 (respectively, Jv(x)::;; 0) is
Spaces with indefinite metric
241
called non-negative (respectively, non-positive). The set of all non-negative (respectively, non-positive) vectors in lB will be denoted by it+ (respectively, it_). By it++ (it __ ) we will denote the set of all positive (respectively, negative) vectors in ~, and by ito we will denote the set of all neutral vectors. Before introducing a new notion, let us recall that by a lineal in lB we mean any subset L ~ ~ such that AX + f..Ly E L, for all X, y ELand A, f..L E C. The term subspace is constantly used to call a lineal that is closed in the norm topology. A lineal L is said to be non-negative if the inequality Jv(x) ~ 0 is true for all X E L. Analogously we define the non-positive, positive, negative or neutral lineals. The non-negative or non-positive lineals are also called semi-definite lineals, whereas the negative or positive lineals are said to be definite. Among the definite lineals we distinguish the uniformly definite ones, i.e., those definite lineals L with the property IJv(x) I ~c(L)llxIIV,
for all X E L, where c( L) > 0 is a suitable constant. By the continuity of the functional J v , the closure in ~ of any semi-definite (uniformly definite) lineal is still semi-definite (uniformly definite). For definite but not uniformly definite lineals a similar property is no longer true, as the next example shows. We need a simple preliminary result. LEMMA 1.1. Let X be a Banach space with the norm II . II, and let M be a lineal in X on which a bounded linear operator T is defined. If there exists a projection operator PM from X onto the lineal M, with the norm IIPM II = 1, then the operator T can be extended on the whole space X, without increasing the norm, by setting the extension equal to 0 on (I - PM)X.
Recall (see Chapter 0) that a linear operator T defined on a lineal '1)(T) is said to be an extension of the operator T defined on a lineal '1)(T) if'1)(T)
Tx
=
c '1)(T) and
Tx, for all x E'1)(T).
Let T be the operator defined by the equalities Tx = Tx, x EM, and Ty = 0, Y E (1 - PM)X. Then T is an extension of T on the whole space X and for all z = x + Y E M-i-(I - PM)X = X we have
IITzl1 = IITxl1 = IITPMzl1 ~ IITII ·llzll, i.e., IITII ~ IITII· Since the inequality IITII ~ IITII is obvious, we get IITII = IITli· ~ EXAMPLE 1.1. Let lB be a Jv-space with lB+ infinite dimensional, and let x+ E lB+,
x_ E lB_ be two fixed elements such that IIx+1I = Ilx-11 = 1. We define the operator
242
SPACES WITH INDEFINITE METRIC
K+ : lin{x+} - lin{x_} by the equality K+x+ = x_. Then IIK+II = 1. It is wellknown (see V. P. Odinec [1]) that for the one-dimensional subspace M = lin{x+} there exists a projection operator PM from ~ onto M, with the norm IIPM II = 1. We extend K+ to an operator K+ defined on the whole space ~+ by K+ I (I -PM)~+ = = O. According to Lemma 1.1 we have IIK+II = 1. By a direct verification we obtain that the lineal £+ = (P+ + K+)~+ is a non-negative subspace of~. Moreover, £+ is infinite dimensional. Let [ c £ + be a lineal which is dense in £ + and such that x+ + x_ ~ [ (for example we can choose [ = ker'P, where 'I' is a discontinuous linear functional on £ +, with 'P(x+ + x_) =1= 0). Then the lineal [ is positive, whereas its closure "£ = £ + contains a neutral vector x+ + x_ =1= O. ~
§2. Angle operators We introduce now the angle operator of a semi-definite lineal, one of the most important objects of study in this chapter. Let £ + be a non-negative lineal of ~. Since £ + C R+, it follows that liP_xii:::; IIP+xll, for all x E £+. Consequently, there exists a well-defined linear operator K+: P+£+ - p_£+, given by K+(P+x) = p_x, x E £+. Equivalently, for all x = x+ + x_ E £+ with x± = P±x, we have x_ = K+x+. It is clear that IIK+II :::; 1. The operator K+ is called the angle operator of the lineal £ +. We define analogously the angle operator K_ : P_ £ _ - P+ £ _ of a non-positive lineal £_. Further we will study mainly the non-negative lineals. The corresponding results for non-positive lineals can be obtained analogously. If the lineal £+ is positive (i.e., £+ \ {O} C R++), then IIK+x+11 < Ilx+ll, for all 0 =1= x+ E P+ £ +. If, in addition, £ + is a uniformly positive lineal, then IIK+II:::; 1-,),(£+), where 0 < ')'(£+):::; 1 is a suitable constant. Indeed, by the definition of the uniform positivity of the lineal £ + we have Ilx+ Ilv -llx_llv ~ c( £ + )llxII V , for all x = x+ + x_ E £+, where c(£+) > o. Hence, and by the obvious relation Ilx+11 = IIP+xll:::; IIP+llllxll, it follows that
Angle operators
243
A semi-definite lineal is said to be maximal semi-definite provided that it is not a proper subset of another semi-definite lineal. The maximal definite, maximal uniformly definite and maximal neutral lineals are introduced analogously. Considering on the set of all semi-definite lineals the order relation corresponding to the inclusion, and using Zorn Lemma (see §1 of Chapter 0), we conclude easily that any semi-definite lineal is contained in a maximal semi-definite lineal. The analogous conclusions for definite or neutral lineals are also true. Let us consider the class 9Jt+ of all non-negative lineals L + such that P+ L + =
= 123+. Analogously, we define the class 9Jt_ of non-positive lineals. It is obvious that alllineals L + E 9Jt+ are maximal non-negative subspaces. The converse is, generally speaking, false. Namely, as Lemma 1.1 shows, the question whether a maximal nonnegative subspace L + belongs to the class 9Jt+ is related to the question about the existence of a projection operator of norm 1 from the space 123+ onto the subspace P+ L +. It is well-known (see §11 of Chapter 0) that in the class of all Banach spaces, the Hilbert spaces are the only ones which have the property that any subspace is the range of a projection operator of norm 1. Therefore, when 123+ is a Hilbert space, the set of all maximal non-negative subspaces coincides with 9Jt+. In the general case of a Jv-space 123, this is no longer true. (For more information about projections of norm 1 on a Banach space see V. P. Odinec [1].) However, as Lemma 1.1 shows, if dim 123+ ~ 2, then we have dim L + ~ 2 for any maximal non-negative subspace L + of 123. In what follows we will denote by 9Jt~ (9Jt~) the set of all uniformly positive (respectively, negative) subspaces L+ E 9Jt+ (L_ E 9Jt_). Later on we will use the next result. PROPOSITION 2.1.
Every vector x E.R+ (.R++) is contained in a subspace L+ E 9Jt+
(9Jt~). Let K+ be an angle operator corresponding to the one-dimensional subspace lin{x+} c .R+, where x+ = P+x. If x E .R++ then IIK+II < 1. As we have noted above, anyone-dimensional subspace is the range of a projection of norm 1. Therefore the operator K+ can be extended, preserving the norm, to an operator K+ defined on the whole space 123+. The subspace L+ = (P+ + K+)123+ satisfies the desired condition. ~
Let us consider a subspace L + in the class 9Jt+. For any such subspace L + we have an angle operator K+ : 123+ -> 123_, with IIK+II ~ 1, defined on the whole space 123+, such that
(2.1)
SPACES WITH INDEFINITE METRIC
244
or C+ = (P+ + K+)IB+. In order to determine the operator K+ we observe first the next simple fact. LEMMA 2.1. If
C is a non-negative lineal then the projection P+ maps C homeo-
morphically onto P+ C.
IIP+xll ~ lIP-xii· Therefore Ilxll IIP+xll ~ IIP+II Ilxll· ~
<J For any x E C we have
~
211P+xll.
On the other hand,
IIP+xll+llP-xll ~
~
Now let C+ E !.m+ be fixed. By Lemma 2.1 the operator P+ I C+, i.e., the restriction of the projection P+ to C +, has a bounded inverse (P+ I C +) -1; for all x in C+ the equality x = (P+ I C+)-l x + is fulfilled. Then x_ = P_(P+ I C+)-l x +. Consequently, the operator K+ in (2.1) has the form (2.2) We denote by K+ the closed unit ball of the space C (IB+, IB_) of all bounded linear operators acting from IB+ into IB_, and by K~ its interior. By (2.1) any K+ E K+ determines a subspace C+ in the class !.m+. Summing up all the considerations above we obtain the next result.
There exists a bijective correspondence between the sets !.m+ and K+ given by the relations (2.1) and (2.2).
THEOREM 2.1.
§3. Plus-operators We consider a linear operator A defined on a lineal ::D(A) of a Jv-space lB. DEFINITION 3.1. The operator
A is called a plus-operator if A(.R+ n ::D(A)) c .R+.
A is said to be strict if Jv(Ax) n ::D(A) , where p,(A) > 0 is a constant.
DEFINITION 3.2. A plus-operator
all x E ~
DEFINITON 3.3. The operator
~
p,(A)Jv(x), for
A is called J;; -expansive if (3.1)
for all x E ::D(A) n .R+. If the inequality (3.1) is fulfilled for all x operator A is said to be Jv-expansive. DEFINITION 3.4. The operator
E
::D(A), then the
A is called uniformly J;; -expansive if (3.2)
Plus-operators
245
°
for all x E :D(A) nR+, where ,(A) > is a constant. If the inequality (3.2) is fulfilled for all x E :D(A), then A is said to be uniformly Jv-expansive. DEFINITION 3.5. The operator A is said to be a Jv-isometry if
for all x E :D(A). A Jv-isometry is called a Jv-semi-unitary operator if :D(A) = 113, and a Jv-unitary operator if :D(A) = 113 and, in addition, R(A) (= A:D(A)) = 113.
In the general case, a plus-operator on a Jv-space 113 is unbounded. We give next an example of an unbounded Jv-expansive plus-operator A with :D(A) = 113. EXAMPLE 3.1. Let 113 be a Jv-space with 113+ infinite dimensional. Relatively to the decomposition 113 = 113++113_ we define the linear operator A on 113 by the matrix
° 0)
A = (A11
I
'
where A11 is an unbounded expansive operator on 113+, i.e., IIA11X+11 )! Ilx+11 for all X+ E 113+. Then the operator A is unbounded and
Our next aim is to establish a criterion for the continuity of a plus-operator. We need first a preliminary result. Let A be a plus-operator such that :D(A) n R++ -=I=- 0. Then the boundedness of the operator A is equivalent to the boundedness of the operator P+A.
LEMMA 3.1.
The necesity of the condition is obvious. In order to prove the sufficiency we assume, on the contrary, that A is unbounded. Then there exists a sequence {xn : n EN}, Xn E :D(A), such that Xn ----> and IIAxnl1 ----> 00 for n ----> 00. Since P+A is bounded we have P+Axn ----> 0, hence IIP-Axnll ----> 00. For any sequence {zm : mEN}, Zm E 113, with the property Zm ----> for m ----> 00, we clearly obtain that IIP-Axn + P-zmll ----> 00, for n,m ----> 00. Consequently, Jv(Axn + zm) ----> -00 for n, m ----> 00. Set Zm = m- 1 Ay, where y E :D(A) n R++ is fixed. Then there exists pEN such that Jv(Axn + zm) < 0, for all n, m)! p. But <J
°
°
SPACES WITH INDEFINITE METRIC
246
By the continuity of Jv there exists lEN, l ~ p, such that Jv(xn + (l/p)y) > 0, for all n ~ l. Put n = l. Then Jv(Xl + (l/p)y) > 0, and Jv(AXl + A(l/p)y) = = Jv (A(Xl+(l/p)y)) < 0, which contradicts the assumption that A is a plus-operator. Thus the bounded ness of P+A implies the boundedness of A. ~ THEOREM 3.1. Let
C
~
A be a plus-operator with the property that A(1J(A)
n Jt++)
\ {O}. If 1J(A) contains a uniformly positive lineal £ such that P+A12
C =
= P+A1J(A), then the boundedness of the operator A is equivalent to the boundedness of the operator P+A I £ : £
--+
IB+.
The boundedness of P+A I £ follows obviously from the boundedness of A. Conversely, let us assume that P+A I £ : £ --+ IB+ is a bounded operator. By Lemma 3.1 it is sufficient to establish the boundedness of P+A. Let d = IIP+A I £ II. Since P+A12 = P+A1J(A) it follows that for any z E P+A1J(A) with Ilzll > d there exists y E £ with Ilyll > 1 such that P+Ay = z. By the uniform positivity of £ we find E > 0 such that Jv(u) > 0 for any u in the E-neighborhood of the set M = {x : x E 12, Ilxll ~ I}. Let w E1J(A) be an arbitrary element with Ilwll < E. Then Jv(y - w) > 0 for any y EM. Since A(1J(A) n Jt++) C Jt+ \ {O} we obtain o i= Jv(Ay-Aw) = Jv(A(y-w)) ~ 0, hence P+Ay i= P+Aw, for all elements y EM. Consequently, there exists a number 8 > 0 such that IIP+Awll ::;; 8, for all w E 1J(A) with Ilwll < E, a condition which clearly implies the boundedness of P+A. As we have already noted, by Lemma 3.1 the proof of Theorem 3.1 is complete. ~ <J
Actually, under our assumptions, we can avoid the use of Lemma 3.1. Indeed, take as above w E 1J(A) with Ilwll < E, and y EM. Since Jv(y + w) > 0 we have y + wE 1J(A) n Jt++, hence 0 i= Jv(Ay + Aw) = Jv(A(y + w)) ~ o. It follows that
On the other hand we have
IIP_Awll = II(P_Ay + P_Aw) - p_AYII ::;; IIP_Ay + P_Awll
+ IIP_Ayll ::;; IIP+Ay + P+Awll + IIP_Ayll ::;; IIP+Ayll + IIP+Awll + IIP_AYII·
Therefore, taking into account that
IIP+Awll ::;; 8, we obtain
IIAwl1 ::;; IIP+Awll + IIP_Awll ::;; ::;; 211P+Awll + IIP+Ayll + IIP_AYII ::;; 28 + IIP+Ayll + IIP_AYII, for all w
E1J(A)
with
Ilwll < E,
and the boundedness of A follows.
::;;
Plus-operators
247
Theorem 3.1 enables one to obtain a simple and natural condition for the bounded ness of a strict plus-operator defined on the whole space 113. This very class of operators will be, mainly, the object of our further investigations.
Let A be a strict plus-operator defined on 113. If there exists £ ~ E E 9)1~ such that A £ ~ E 9)1+, then A is bounded.
THEOREM 3.2.
We show the boundedness of the operator P+A I £~ : £~ -.., 113+. Since £~ E 9)1~ we have Jv(x) ~ c(£~)llxIIV, for all x E £~, where c(£~) > O. Therefore
for all x E £~, where fJ(A) > O. Hence it follows that P+A(£~) is closed in 113+, P+A I £ ~ is one-to-one, and (P+A I £ ~)-l maps continuously the Banach space P+A(£~) onto the Banach space £~. By Theorem 0.8.5 the operator P+A I £~ is continuous, too. Applying Theorem 3.1 we obtain the bounded ness of A. ~ We will state next an obvious consequence of Theorem 3.2.
Let A be a strict plus-operator defined on the whole space 113. If All (= P+AP+) maps 113+ onto 113+, then the operator A is bounded. COROLLARY 3.1.
Further we will consider bounded plus-operators defined on the whole space 113. We need the following result.
A strict plus-operator A maps a uniformly positive subspace onto a uniformly positive subspace.
LEMMA 3.2.
Let £ + be a uniformly positive subspace. Then
where x E £+ and c
= fJ(A)c(£+) > O.
If z E A£+, then z
= nlim Axn , where ..... oo
Xn E £+. Since IIAxn - Axmll ~c~llxn - xmll, the sequence {xn : n E N} is fundamental in £ + and therefore Xn -.., x for n -.., 00, where x E £ +. This means that z (= lim Ax n ) = Ax E A£+, hence A£+ is a subspace. Hence and from n ..... oo
Jv(Ax)
~
cllxll v the uniform positivity of the subspace A£+ follows.
~
We consider now the action of a strict plus-operator on the subspaces belonging to 9)1+. By the defect number of a subspace £ (c J\+) we mean the cardinal number def £ = dim(SB+I P+ £) (note that by Lemma 2.1 P+ £ is a subspace of £ +i 113+1 P+ £ is the quotient of 113+ by P+ £).
SPACES WITH INDEFINITE METRIC
248
THEOREM 3.3. If A is a strict plus-operator, then the defect number def A £ + is
the same for all £ +
E
9J1~. .
--
0
1
2
By Lemma 3.2 we obtam A£+ = £+, for all £+ E 9J1+. Let £+, £+ E E 9J1~ and let K+, K~ be their angle operators. By the convexity of K~ the operator Kt = (1 - t)K+ + tK~ is the angle operator of some £~ E 9J1~ for any t E [0, 1]. We associate to the family {Kt} a new family of operators {WA(Kt) : WA(Kt) = = P+A(P+ + Kt), t E [0, I]}. Since A takes non-zero values on the set of positive vectors, all the operators W A (Kt) are P + -operators (we are using here the terminology in I. C. Gohberg and M. G. Krein [1]). By Theorem 7.1 in that book, for any t E [0, 1] we can find T > 0 such that for all s E [0, 1] with Is - tl < T the operators WA(K+) have the same defect index. We use the Heine-Borel Lemma and cover the interval [0, 1] with a finite number of open intervals, each of them corresponding to a constant value of the defect index. Therefore def A £ ~ = def A £ ~
!.
The possibility of extending Lemma 3.2 and Theorem 3.3 to the whole class 9J1+ is related to the values of the constant 1/. Namely, - as the next example shows - for 1/ ~ 1 this is impossible. EXAMPLE 3.2. Let 1)3
=
1)3++1)3_ be a two-dimensional Jv-space, where
1/
~ 1 and
1)3± = lin{e± : Ile±11 = I}. With respect to the basis {e+,e_} we define a linear operator A by the matrix
Then
i- 0
i.e., the operator A is Jv-expansive. But A( e+ - e_) = 0 and the vector e+ - e_ is neutral. When 1/ > 1 it is possible to extend Theorem 3.3 to the whole class 9J1+. We need the next result. LEMMA 3.3. Let A be a strict plus-operator acting on a Jv-space 1)3, with
Then there exists 8(A) > 0 such that IIAxl1 ;? 8(A)llxll, for all x
E
1/
>
l.
Jt+.
Assume the conclusion is false. Then there exists a sequence {xn : xn
E
Jt+,
= I} such that Axn ---; 0 for n ---; 00. We have Ilx+11 = IIP+xnll;? ~llxnll = ~. 1 l l 1 Set yn = (1 + An)vX+ + A;{x~, where A;{ = IIAx n ll- 2 , for Axn i- 0, and An = n, Ilxnll
otherwise. Then Jv(yn) ;? 2- V and
Jv(Ayn)
=
Jv (A ((1
+ An) ~
- At) x+
+ At Axn)
---; 0,
Symmetric properties for n ----)
00,
249
which contradicts the fact that A is a strict plus-operator.
~
Now using Lemma 3.3 we can extend Theorem 3.3 on the whole 9Jt+. 3.4. Let A be a strict plus-operator acting on a Jv-space Q3, with v > 1. Then the defect number def A L + is the same for all L + E 9Jt+.
THEOREM
<J
The proof is almost the same as the one of Theorem 3.3.
~
COROLLARY 3.2. If a strict plus-operator A on a Jv-space Q3, with v > 1, maps at least one subspace from 9Jt+ onto a subspace in 9Jt+, then A L + E 9Jt+, for all L + E 9Jt+.
§4. Symmetric properties of a plus-operator and its adjoint In the dual space Q3* of the Jv-space Q3 we define the sets oft:±: by
where P± are the adjoint operators of the projections P±. Clearly P± are projections on Q3* and Q3* = F+. Q3* + P': Q3* = Q3~ +Q3~. We introduce also Jt(; = oft~ n oft~. As in the case of the space Q3, for every lineal M ± C oft:±: there exists the angle operator Q ± :
: P±M± ----) P~M± such that IIQ±II ~ 1 and M± = {xl + Q±xl : xl E P±M±}. By 9Jt:±: we denote the class of all subspaces M ± C oft:±: such that P± M ± = Q3:±:. We single out also the class 9Jt:±:o of all uniform subspaces in 9Jt:±:. As in §2 above (Proposition 2.1) we establish easily that any vector x* E oft~ (x* E oft~) is contained in some subspace M + E 9Jt~ (M _ E 9Jt~), and any positive (negative) vector x* is contained in a subspace from 9Jt~o (9Jt~o). We are interested to describe the structure of the orthogonal complements of the subspaces in the classes 9Jt+ and 9Jt_. All the considerations below will be explicitly developed for subspaces in the class 9Jt+. The corresponding results for subs paces in the class 9Jt_ can be establish analogously.
If L + E 9Jt+ (9Jt~), then L ~ E 9Jt~ (9Jt~o). When the space Q3+ is reflexive, then for any positive subspace L + E 9Jt+ its othogonal complement L ~ is negative.
THEOREM 4.1.
<J If L+ E 9Jt+, then L+ = (P+ + K+)Q3+, where K+ E IC+. Further, we easily get that L~ = (P': + Q_)Q3~, where Q_ = -K.+, Q_ : lB~ ----) Q3~, and IIQ-II ~ 1. Hence L~ E 9Jt~. If L+ E 9Jt~, i.e., IIK+II < 1, then IIQ-II (= IIK+II) < 1, therefore L ~ E 9Jt~o. Assume now that lB+ is reflexive and let .c + E 9Jt+ be a
SPACES WITH INDEFINITE METRIC
250
positive subspace, i.e., IIK+x+11 < Ilx+11 for all 0 =1= x+ E lB+. Suppose that there exists x~ E lB~, with Ilx~11 = 1, and such that IIQ_x~11 = 1. Then IIK~x~11 = 1. Since the space lB+ is reflexive, we find a vector x+ E lB+ with Ilx+ II = 1 and such that (x+, K~x~) = 1, that is (K+x+, x~) = 1. From this relation and the inequality I(K+x+,x~)1 ~ IIK+x+11 it follows that IIK+x+11 = 1, which contradicts the assumption that C + is positive. ~ REMARK 4.1. The condition in Theorem 4.1 for lB+ to be a reflexive space can not
be dropped, as the following example shows.
x = (... , x - k, O}, where
.),
with the standard basis {en: nEZ, n
=1=
... ,
en = (... ,O, ... ,~, ... ,O, ... ). n
= £ I -+ £ I such that lB ± = £ I, and define an angle operator K + by 1 the equalities K+e;; = ~ne;:;+l' where ~n = 1 - ~' nEZ, n =1= 0, and e; are the elements in the standard bases of the spaces lB±. We easily see that IIK+x+1I < Ilx+11 for all 0 =1= x+ E lB+, hence the subspace C+ = (P+ + K+)lB+ is positive. Further we have lB* = m-+m, where m = £ i, and lBi: = m. Let z E lB~ be the vector with the coordinates Zk = 1, k E Z, k =1= O. Then, using the equalities K~e;; = ~ne~_l' we obtain IIK~zll = Ilzll· Therefore the subspace M_ = = (P~ - K~)lB~ contains We consider lB
Ct
the neutral vector z -
K~z =1=
O.
~
The next result follows from the equality -L£-L subspace £ ~ lB*.
= £, which holds for any
COROLLARY 4.1. If in Theorem 4.1 ~e replace the space lB by the space lB*, and instead of the orthogonal complements we consider the *-orthogonal complements, then the conclusions remain true.
We return now to the study of plus-operators. Recall that we are dealing only with bounded plus-operators defined on the whole space lB. DEFINITION 4.1. A plus-operator A is said to be a double plus-operator if A * Jt~
c
c
Jt~.
Of course, not every plus-operator is a double plus-operator. It is sufficient to consider on the two-dimensional Jv-space lB = lB+-+lB_, where lB± = = lin{e± Ile±1I = I}, the plus-operator A given - with respect to the basis
251
Symmetric properties { e+, e_} -
by the matrix
the operator A * is not a plus-operator. In the considered example the closure of A.c + for some subs paces .c + E 9J1~ does not belong to 9J1+. Our next aim is to formulate conditions under which a plus-operator is a double plus-operator. We need a preliminary result. For
10:1 < 1;31
Let x* be a negative element oIIB*, i.e., 1!Pi-x*11 there exists a subspace .c + E 9J1~ such that (.c +, x*) = {O}.
LEMMA 4.1.
<
IIP~x*ll.
Then
.c += N( x *) n:.o++ I·III { (U
Then p+.c + = IB+ and Let x
nlB+, 0: E
C.
•
(.c +, x*) = {O}.
(y+, x*) } y+- (y_,x*)y- .
We show that
.c + E
9J1~.
= x+ + x_ = Z + 0: (y+ - ~~~: ::~ y_ ), where Ilx+11 = 1, We have I(x+, x*)1 = 10011(y+, x*)1 ~ IIPi-x*ll. Thus Ilx-11 = 10:11 ~~~:::~ I
THEOREM 4.2. If a plus-operator A
Z
E N(x*)n
~ IIP~x*ll· (y_,X*)-l < 1.
on
IB
satisfies the condition (4.1)
then A is a double plus-operator.
Let us consider first a positive element x* E Jt~ and assume that A * x* rt. rt. Jt~. Then IIPi-A*x*11 < IIP':'A*x*lI, i.e., A*x* is a negative element in IB*. By Lemma 4.1 there exists .c + E 9J1~ such that (.c +, A*x*) = {O}. Then (A.c +, x*) = = {O}, a relation which, by (4.1), contradicts a conclusion of Theorem 4.1. Therefore our assumption is false, hence A*x* E Jt~ for all positive elements of IB*.
Take now an arbitrary vector 0 =I- y* +P':'y* we have IIP+y~1I
>
IIP':'y~ll,
EJt~. Then for all y~ = (1 + ~) P+y* +
i.e., each y~ is a positive element. From the first
SPACES WITH INDEFINITE METRIC
252
part of the proof it follows that A *y~ E Jt~. Taking the limit as n
A*y*
----* 00,
we obtain
E Jt~. ~
If a plus-operator A has the property that A L + E 9J1+ for all L + E 9J1+, then, generally speaking, the adjoint operator A * fails to have the analogous property. Let 1)3 = 1)3++1)3_, where the space 1)3+ is reflexive and has a Schauder basis {ei : i E N}. We define a plus-operator A on 1)3 in the following way: on 1)3+ we put Ael = 0, Aei = ei-l, i = 2,3, ... , and on 1)3_ we set A == o. Then the operator A is bounded, and A L + E 9J1+ for all L + E 9J1+. For A * we have A*ei = ei+l, i = 1,2, ... , i.e., A* M + tf. 9J1~ for M + E 9J1~ However, if we add the next condition: "A maps the non-zero vectors from 1)3+ into positive vectors", then the property "A L + E 9J1+ for all L + E 9J1+" is symmetric. More precisely, we have: EXAMPLE 4.1.
1)3+ be a reflexive space and assume that A(I)3+ \ {O}) C Jt++. 9J1+ for all L+ E 9J1~, then A* M+ E 9J1~ for all M+ E 9J1~.
THEOREM 4.3. Let
If AL+ E
Let us consider a Jp-space 1)3, with p > 1. We introduce on 1)3 a new norm, equivalent to the initial one, by
1
Correspondingly, we introduce on 1)3* the conjugate norm Ilx* Ilq = (1Ix~ Ilq + Ilx~ Ilq) q where q
= ~1' and define a p-
,
Jq-metric on 1)3* by the relation
DEFINITION 4.2. A strict plus-operator A on the Jp-space 1)3 is called double-strict if A* is a strict plus-operator on the Jq-space 1)3*. DEFINITION 4.3. A strict plus-operator A is said to be focusing, if
for all x
E
Jt+, where i(A) >
o.
(Compare with Definition 3.4.)
Symmetric properties
253
A focusing operator A is said to be double-focusing, if A * is also a focusing plus-operator in the Jq-space 1l3*.
In the case of a Hilbert J-space Sj any strict plus-operator A with the property A .c ~ E 9J1+ is double-strict and JL(A) LEMMA 4.2.
= JL(A*)
for some
.c ~
E
9J1+
(4.2)
(see M. G. Krein and Y. A. Shmulian [1]).
In a Jp-space Il3 every strict plus-operator A with property (4.2) satisfies
the condition
(4.3) Assume that the conclusion is false. Then we find a sequence {x~ n E N} such that x~ E 1l3_, Ilx~11 = 1 and IIAl/A12X~11 -+ 1 for n -+ 00. Set an = = IIA1/A12x~11 and Xn = a~lAl/A12X~ -x~. Then Xn E.Ito (C Jt+) and
for n
-+ 00.
On the other hand, since the operator A is strict, by Lemma 3.3 it follows that IIAxl1 ~ 8(A)lIxll for all x E Jt+. Hence
So we obtained a contradiction.
~
Every focusing strict plus-operator A with property (4.1) on a reflexive Jp-space Il3 is double-focusing and double-strict.
THEOREM 4.4.
sentation
With respect to the basis {1l3+, 1l3_} the operator A has the matrix repre-
(~~~ ~~~),
where All = P+AP+, A12 = P+AP_, A21 = P_AP+, A22 = P_AP_. By Theorem 3.4, from condition (4.2) it follows that All : 1l3+ -+ 1l3+ is a homeomorphism. Relatively to the basis {1l3 ~, 1l3:'} the operator A * has the matrix representation
SPACES WITH INDEFINITE METRIC
254
Let us show that for the operator A * the condition
(4.4) analogous to (4.3), is fulfilled. Indeed, since A is focusing it follows that ::;; ,B11P+Axll, for all x E .It+, where ,B < 1. Hence we have
lIP-Axil ::;;
for all x+ E 113+ and any angle operator K+ E K+. Setting K+ = 0 and considering the adjoint operators we obtain IIAh II ::;; ::;; ,BIIAi111, a relation equivalent to (4.4). By the reflexivity of 113, the angle operators of the subspaces in OO1~ are exactly the adjoints of the angle operators of the subspaces in 001_. Next we show that
(4.5) where K _ E K _. Let us suppose, on the contrary, the existence of a sequence {K~ : n EN}, such that K~ E K_ and Ilk~*11 ---f 1 as n ---f 00, where k -n*
=
(A*12
Then we find elements y~
E
+ A*22 Kn*) -
(A*11
+ A*2 K1 -n*)-l .
113_ such that IIY~ II
= 1 and 8n = IIK~y~ II
---f
1 as
n ---f 00. Setting xn = k~y~ - 8ny~ we obtain Ilx+. II = Ilx~ II, i.e., xn E Jlo (c .It+). Hence - using the fact that A is focusing - and from (4.3) it follows that IIP_Axnll::;; (1 - i'(A))IIP+Axnll and IIP+Axnll ~ yn = k~y~ - y~, we obtain xn - yn ---f 0, whence
E
> 0, for n
~ no.
Setting
(4.6) Let M~ = (p~ (Ayn,
+ K~*) 1J3~.
M~) =
Then A* M~ = (p~
(yn, A* M~)
=
((p_
+ K~*) 1J3~.
Further
+ kr:.) yr:., (p~ + kr:.*) 1J3~) = {O},
which, by (4.6), contradicts condition M~ E OO1~; so (4.5) is proved. It remains to show that the plus-operator A* is strict. For any x* E .It~ we have: Jq(A*x*)
= Ilp~A*x*llq
-llp':A*x*ll q ~ (1- ,B*q)
II(Ai1 + A;lK:') x~llq =
= (1- ,B*)q IIAi1 (I + (Ai1)-1 A;lK:') x~llq ~ ~ Ilx~llq ~ ~Jq(x*),
Symmetric properties
255
Theorem 4.4 assures that any focusing strict plus-operator on a reflexive space, which satisfies condition (4.2), is collinear to a double uniformly -expansive operator, i.e., to an operator which is uniformly -expansive simultaneously with its adjoint. To conclude this section we take into consideration some conditions under which a Jp-expansive operator A is double Jp-expansive, i.e., both A and A * are J p-expansive.
J:
J:
Let A be a strict plus-operator on the Jp-space 23. Then the operators P_ ± AP+ and P_ ± P+A are continuously invertible.
LEMMA 4.3.
We consider the operators P_ ± P+A. Assume that Ilxnll = 1 and (P_± n ±P+A)x ----; 0 as n ----; 00. Then x~ ----; 0 and P+Ax n ----; O. Since Ilxnll = 1 we have Ilx+11 ;;:: a > 0 for n;;:: no. Further, P+Ax n = Anx+ + A12x~ and, since A12X~ ----; 0, then Anx+ ----; 0, too. On the other hand, by Lemma 3.3, we have IIA l1 x+ll;;:: 811x+11 ;;:: 8a > 0, for n;;:: no, a contradiction. Hence it follows that II(P-± ±P+A)xll ;;::~±llxll, for all x E 23, where ~± > O. By Theorem 0.4.5 we conclude that the operators (P- ± p+A)-l exist and II(P- ± p+A)-lll ~ ~±l. Assume now that (P_ ± AP+)yn ----; 0 as n ----; 00, where Ilynll = 1. Then y~ + P_Ay+ ----; 0 and P+Ay+ ----; O. As above, from Lemma 3.3 it follows that y+ ----; o. This means that P_Ay+ ----; 0, whence y~ ----; 0, too. Thus we obtain again that II(P+ ± AP+)yll ;;:: '/]±llyll for all y E 23, where '/]± > O. Hence the operators (P- ±AP+)-l exist and II(P- ±AP+)-lll ~'/]±l. ~ <J
LEMMA 4.4. If A
is a double-strict plus-operator on a Jp-space 23, then P+AP+23+
= 23+.
(4.7)
We notice that for a strict plus-operator on a Jp-space 23 condition (4.7) is equivalent to condition (4.2). <J By Lemma 4.3 the operator P_ + AP+ is a homeomorphism and therefore the lineal (P_ + AP+)23 is closed. If (P_ + AP+)23 =I- 23, then we find an element x* =I- 0 in 23* such that ((P- + AP+)23,x*) = {O}, that is (P::" + PtA*) x* = O. But A* is a strict plus-operator on 23*, hence, by Lemma 4.3, the operator P::" + P-i'-A* is invertible. It follows that x* = 0 ~ a contradiction. Thus (P_ + AP+)23 = 23, a relation which clearly leads to (4.7). ~
LEMMA 4.5.
Let A be a Jp-expansive operator on lB. Then the operator (4.8)
SPACES WITH INDEFINITE METRIC
256
is well-defined and
IIUxll p
~
Ilxll p
for all x E :D(U).
= (P_-P+A)y, where y E lB, and Ux = (P+-P_A)y. Hence and by Jp(Ay)? Jp(Y)
it follows that IIUxll~
=
IIP+yll~ =
+ IIP_Ayll~ ~
II(P- -
IIP-yll~
+ IIP+Ayll~ =
P+A)yll~ = ilxll~·
LEMMA 4.6. If a strict plus-operator A on the Jp-space lB satisfies condition (4.7),
then the operator U given by (4.8) is defined on the whole lB and the following inversion formulas are true: U A
= (P+ - P_A)(P- - p+A)-l
= -(P+ + P_U)(P- + p+U)-l
=
=
-(P_
+ AP+)-i(p+ + AP_),
(P- - U p+)-i(p+ - U P_).
(4.9) (4.10)
+ AP+)-i(p+ + AP_) + (P+ - P_A)(P_ - p+A)-l = = (P- +AP+)-i((p+ + AP_)(P_ -P+A) + (P- +AP+)(P+ -P_A))(P__ p+A)-l = = (P_ + AP+)-l(AP_ - P+A - P_A + AP+)(P- - p+A)-l = o. (P-
We establish now the first equality in (4.10), i.e., A By (4.8) we have
Ux
=
= -(P+ + P_U)(P_ + P+U)-i.
(P+ - P_A)(P- - p+A)-lX,
i.e., x = (P_ - P+A)z, z E lB and Ux relations
= (P+ - P_A)z. Hence we get the following
U(P_ - P+A)z
=
(P+ - P_A)z,
-P+U(p- - P+A)z
=
-P+z,
Invariant semi-definite subspaces
257
and
where J = P+ - P_. As it was already proved above (P_ - P+A)IB = lB. By Lemma 4.3 the operator (P_ - p+A)-l exists and is bounded. Therefore, by (4.8), P+(P_ - p+A)-lP+IB+ = P+UP+IB+ = IB+. We establish, as above, that (P_-P+U)fJ3 = IB and that the operator P_ -P+U is continuously invertible. Therefore (P_ - P+A)z = -(P_ - P+U)Jz and -P+Az = -(P_ - p+U)-lJz - P_z. On the other hand -P_Az = U(p- - P+A)z - P+z = -U(P_ - p+U)-lJz - P+z. Then
But
+ I)(P- - p+U)-l J = 1- (U + I)(P- + p+U)-l = + P+U) - (U + I))(P- + p+U)-l = -(P+ + P_U)(P- + p+U)-l.
1+ (U
= ((P-
So the equality A = -(P+ + P_U)(P_ + p+U)-l is established. Finally, the second equality in (4.10) can be proved analogously with the second equality in (4.9). ~
A Jp-expansive operator A is double Jp-expansive if and only if it satisfies condition (4.7).
THEOREM 4.5.
The necessity of condition (4.7) was established in Lemma 4.4. We prove its sufficiency. By Lemma 4.6 the inversion formulas (4.9)-(4.10) are true. By Lemma 4.5 it follows that IIUxil p ~ Ilxllp, x E lB. This means that IIU*x*llq ~ Ilx*llq, x* E IB*. Taking the adjoint operators in (4.9)-(4.10) we obtain
= (p~ - A * p t )-1 (pt - A * p~) = = - (pt + P~A*) (p~ + p t A*)-l ,
(4.11)
= (pt - P~U*) (p~ - p+.U*)-l = = _ (p~ + U*p+.)-l (P+' + U*P~).
(4.12)
U*
A*
From (4.11) we get x* Therefore
=
(p~
+ P+.A*) y*,
and, consequently, Jq(A*y*) ;?: Jq(y*).
~
y*
E
IB*, U*x*
= - (P+' + P~A*) y*.
258
SPACES WITH INDEFINITE METRIC
§5. The problem of invariant semi-definite subspaces In 1943 S. L. Sobolev concluded his work [1] "About the motion of a symmetrical top whose cavity is filled with a liquid" . The motion of the top is described by an equation of the form dR di=iBR+Ro,
R(t=O)=R(O),
(5.1)
where B is a linear operator and R is a vector in the phase (complex infinite dimensional) space. The solution of this equation is given by
J t
R
= e iBt R(O)
+
eiB(t-t,j R o (t1) dt 1,
o since R(t = 0) = R(O). In this case the operator B is considered to be bounded, so that, for sufficiently large absolute values of A E C, there exists the resolvent r A = (AI - B)-1. Then e iBt =
_1_. 27f1
J
eiAt r A dA,
I
where 'Y is a circle of a sufficiently large radius centered at 0 E C. A Hermitean form Q(R 1 ,R2 ) can be constructed on the phase space {R}. Its behaviour is determined by the quantity
where: a) k is the rotational momentum of the gravitational force and depends on the angle which expresses the deviation of the z-axis relatively to the vertical direction; b) w is the angular speed of the rotation of the top around its axis (the z-axis coincide with the symmetry axis of the top and the origin of the coordinate axes is considered at the fixed point); c) A1 is the innertial momentum of the shell relatively to the x- and y-axis associated to the top, and A2 is the innertial momentum of the liquid relatively to the same axes; d) C 1 and C 2 are the innertial momenta of the shell and of the liquid relatively to the z-axis; e) it and 12 are the distances between the fixed point and the centers of mass of the shell and of the liquid, respectively;
Invariant semi-definite subspaces
259
f) g is the gravitational acceleration; g) M
1
and M
2
are the masses of the shell and of the liquid, respectively.
If L > 0, then the form Q is positive definite, and if L < 0, then Q is an indefinite form with one negative square. The operator B is self-adjoint relatively to the form Q. Therefore, for L > 0 its spectrum is real and e iBt is a bounded operator. If L < 0 then the operator B can have a non-real spectrum. In this last case there exists a pair of complex-conjugate eigenvalues, coresponding to a pair of neutral eigenvalues. The presence - or the absence - of these non-real eigenvalues has a determining influence in the stability of the solution for equation (5.1). Extending Sobolev's investigations, L. S. Pontryagin published in 1944 his work [1], devoted to the study of self-adjoint operators relatively to a Hermitean form with K, < 00 negative squares. Let A be such an operator. Using some subtle algebraic methods the existence of a pair (1:- +, 1:- _) of invariant subspaces for A can be established, where 1:- + is a maximal non-negative subspace, and 1:- _ is a maximal non-positive subspace, with the following properties concerning the spectra of A I 1:- + and A I 1:- _: the eigenvalues in O"(A I 1:- +) and O"(A I 1:- _) are two by two complex-conjugate, and the imaginary parts of these eigenvalues are non-negative for 1:- + and non-positive for 1:- _. As a recognition of Pontriagyn's contributions, the Hilbert spaces with an indefinite metric having a finite number K, of positive or negative squares are called Pontriagyn spaces. Usually they are denoted by IIt<. By the end of the 40's and the beginning of the 50's, I. S. Iohvidov [1] using the Caley-Neumann transform, established the relationship between 1f-self-adjoint and 1funitary operators and proved that for a 1f-unitary operator U there exists a pair of invariant maximal semi-definite subspaces (1:- +, 1:- _) such that the spectra of U I 1:- + and U I 1:- _ lie outside and inside the unit disk, respectively. In 1950 M. G. Krein's paper [1] appeared, where a theorem on the existence of an invariant subspace for a plus-operator on a space IIK, was proved by an ingenious use of Schauder's fixed point principle. Later on the Iohvidov-Krein Theorem was extended by M. L. Brodskii [1] to the case of a plus-operator acting on a Banach J",-space ~ with dim ~+ < 00 Afterwards, the invariant subspace problem for the general case of an operator acting on a space with indefinite metric, which has both the number of positive and negative squares infinite, was considered. The axioms of such spaces were elaborated by M. G. Krein [3], and, in his honour, they are called Krein spaces. M. G. Krein achieved also an important step in solving the invariant subspace problem. In 1964 his paper [2] was published, in which, using once again a fixed point principle, he es-
SPACES WITH INDEFINITE METRIC
260
tablished the existence of an invariant maximal non-negative subspace £ + for a plusoperator A with a completely continuous "corner" P+AP_, and studied the spectrum of the restriction A £ +. On the other hand, in 1960 R. S. Phillips [1] considered the problem of extending a pair of semi-definite subspaces which are invariant relatively to some commutative operator algebra ~, to a pair of maximal semi-definite subs paces , by preserving the invariance relatively to~. To be more specific, this problem occurred in connection with the Cauchy problem for a dissipative system 1
under the condition that the real part of the matrix (a n i n j ) is non-positive for any finite collection {nd of natural numbers and where 00
L
laijl2
<
00
for every i E N,
laij 12
<
00
for every j
j=l 00
L
E
N.
i=l
This type of problems arise in the theory of small perturbations of the dissipative systems with infinitely many degrees of freedom. The precise statement of Phillips' problem is the following. Let 5) be a Krein space with the J-metric [', .], and let ~ be a commutative algebra of operators acting on 5), closed in the weak operator topology, and such that, together with any A E ~ the algebra ~ contains the operator AC, the adjoint of A relatively to the J-metric. Let (12 +, £ _) be a pair of subspaces which are invariant relatively to all operators in ~, such that £ ± C Jt± and [x, y] = 0 for all x E 12+, y E 12_ (i.e., 12+ and 12_ are J-orthogonal). The problem is to find out whether there exists a pair (12 ~ax, £ ~ax) of invariant relatively to ~ subspaces such that 12±ax E 9Ji±, 12± ~ 12±ax, and [12~ax, 12~ax] = {o}. Phillips reformulated this problem for a commutative group of J-unitary operators. In his paper [1] he stated the conjecture that the indicated extension of the subspaces is always possible. He proved that the conjecture is true only in the case of a finite-dimensional space 5). Soon after that, M. A. Naimark [1] proved that Phillips conjecture is true for commutative groups of J-unitary operators on a Pontriagyn space IlK' Afterwards, some advances were reached in a series of works (as, for example, V. S. Shulman [1], E. A. Larionov [1], [2], H. Langer [1], [2], V. A. Khatskevich [1]). On one hand the extension problem for invariant subspaces was solved for larger (even non-commutative)
Invariant semi-definite subspaces
261
families of operators acting, as above, on a Pontriagyn space IlK (and, sometimes, on IId. On the other hand, some restrictions were found for commutative groups of J-unitary operators on an arbitrary Krein space for which Phillips conjecture has a positive answer. In 1970 the work of J. W. Helton [1] appeared. There the existence of a maximal semi-definite invariant subspace was proved for any commutative group it of J-unitary operators on a Krein space which satisfies the next condition: there exists an operator U in it represented - relatively to the decomposition fJ = fJ+ EB fJ- by a matrix of the form
U=
(U~l ~2) +C,
where C is a completely continuous operator, and the spectra of Ul1 and U22 are disjoint, i.e., a(Ul1 ) n a(U22 ) = 0. Every group of 7r-unitary operators on a Pontriagyn space ilK satisfies this condition since it contains the identity operator
o) =
I-
(1+ 0
o -I-
)
+2P_
=
(-h 0
and either P + or P _ is completely continuous.
Based on Helton's result, V. A. Khatskevich [2] proved that Phillips conjecture is true for such a group it. Improving Helton's ideas, T. Ya. Azizov considered the so-called 1i and J( (1i) classes of double J -expansive operators and solved the invariant subspaces extension problem. His results in this direction are summed up in T. Ya. Azizov and 1. S. Iohvidov's book [1]. Other papers, known to us, in which the mentioned problem was solved for some families of operators acting this time on a Jv-space are: 1. S. Iohvidov [1], V. A. Khatskevich [3], H. Langer [3]. In its general form Phillips' problem is not yet solved. Moreover, the problem of the existence of a maximal semi-definite invariant subspace for an arbitrary Junitary operator on a Krein space fJ with infinite dimensional subs paces fJ+ and fJis unsolved too. In the next section we will apply the results of Chapters V and VI about fixed points for holomorphic operators to the problem of invariant subs paces for a plus-operator on a Jv-space.
SPACES WITH INDEFINITE METRIC
262
§6. An application of fixed point principles for holomorphic operators to the invariant semi-definite subspace problem Let 23 = 23++23- be a Banach space with the Jp-norm
where x = x+ + x_ E 23, X± E 23± and p > 1. We consider a bounded strict plusoperator A defined on the whole space 23 and satisfying condition (4.7): P+AP+23+ = = 23+. By Corollary 3.2 the operator A maps any subspace I: + E 9)1+ onto the subspace A I: + which belongs to 9)1+, too. Let now K+ E K+. Then the subspace 1:+ = (P+ +K+)23+ belongs to 9)1+, therefore AI:+ E 9)1+. By Theorem 2.1 we know that AI:+ = (P+ +K+)23+, where ~ -1 K+ = P_(p+ I AI:+) E K+. We have:
= (All + A 12 K+)23+, P_A23+ = (A21 + A22K+)23+.
P+AI:+
Therefore, setting K+
= FA(K+)
we obtain: (6.1)
Thus to each strict plus-operator A with property (4.7) it corresponds the linear fractional transformation FA: K+ ....... K+ given by (6.1). From Lemma 4.2 it follows that IIAll A1211 < 1. Therefore
FA(K+)
= (A2l +A22 K+) (I + A IlA 12 K+)-1 All =
= (A21 +A22K+)
C~o(-l)n (AIlA12K+t) All,
(6.2)
i.e., the mapping FA is holomorphic on K +. From (6.1) it follows that any fixed point K+ E K+ for the transformation FA is the angle operator for a subspace 1:+ = (P+ + K+)23+ E 9)1+, invariant relatively to A, and, conversely, if A I: + c I: + for some I: + E 9)1+ then the angle operator of the subspace 1:+ is a fixed point for FA. Hence, based on Corollary VI.2.1, we obtain: THEOREM 6.1. Let A be a focusing plus-operator with property (4.7). Then A has
in 9)1+ a unique invariant subspace I: ~ E 9)1~ satisfying A.c ~ = .c ~, and there exists a number r < 1 such that IAI :::; rlJ.t1 for all J.t E a(A I .c~) and A E a(A) \ a(A I .c ~).
An application of pixed point principles
263
Consider the transformation FA given by (6.1). From Definition 4.3 of a focusing plus-operator, it follows that
IJFA(K+)II ~ r, where r
= 1- 8(:)' 8(A)
is the number appearing in Lemma 3.3, and "'(
~ min{8(A),
"'((A)}. Thus 0 ~ r < 1. By Corollary VI.2.1, the transformation FA has an s-fixed point K~ E K ~. Therefore L ~ = (P+ + K~) 23 + (E 9J1~) is the unique invariant subspace of the operator A in 9J1+. The result on the spectrum of the operator A was established by A. V. Sobolev and V. A. Khatskevich in [1]. Their proof uses other methods than those in the present book. ~ According to Corollary 3.2 we have:
Let 23 be a Jp-space, such that 2L is the dual space of a normed space X, and let A be a strict plus-operator on 23 satisfying (4.7), such that its "corner" A12 is completely continuous. Then A has an invariant subspace L + E 9J1+, such that A L + = L + and COROLLARY 6.1.
Let tP A(Z, K+) = zFA(K+), where Z E L1 (L1 is the unit disk in the complex and K+ E K+. Obviously, tPA(Z, K+) ~ K+ for any Z E L1. If Z E L1 then tP A(Z, .) satisfies all the conditions in Corollary VI.2.1. Therefore there exists a unique fixed point K+ in K+, i.e., tP A(Z, K+) = K+. By Alaoglu-Bourbaki Theorem and by Theorem 0.2.19, the property of 2L to be the dual space of a normed space implies that the set K+ is compact in the ultraweak operator topology. The complete continuity of the operator A12 implies - by Theorem 0.4.15 - the convergence in the strong operator topology of the sequence (I + Ai} A 12 K.+) -1 (see (6.2)) for any sequence {K.+ : n E N} c K+ which is convergent in the ultra-weak operator topology. Therefore tPA(Z, .) is continuous in the ultra-weak operator topology. Let {zn : n E N} be a sequence of complex numbers which converges to Z = 1 and such that the sequence {K~n : n E N} is convergent in the ultra-weak operator topology to K+ E K+. Then tP A (Zn' K~n) = K~n ---.. K+ for n ---.. 00. But tPA (zn,K~n) ---.. FA(K+), whence K+ = FA(K+). By Theorem 6.1, for any n E N there exists rn < 1 such that IAI ~ rnlfll for all fl E O'(An I L n ), A E O'(An) \ O'(An I L n ), where Ln = (P+ + K~n)23+ and An = P+A + znP_. Hence, taking the limit for n ---.. 00, we obtain (6.3). ~
plane
q
SPACES WITH INDEFINITE METRIC
264
DEFINITION 6.1. We say that a subspace £ ±
exists £
±ax E 9J1± such that £ ± C;;; £ ±ax .
c .Yt± belongs to the class T± if there
COROLLARY 6.2. If an operator A satisfies all the conditions in Corollary 6.1, then
any subspace £ E T +, such that A £ = £, is contained in some subspace /:, + E 9J1+
which is invariant for A. <J Let us denote by K+ the set of all those K+ E K+ for which £
c (P+ +
+K+)!B+. Under our assumptions we have K+ =I 0. We can easily prove that K+ is convex and closed in the weak operator topology. From £ = A £ it follows that FA(K+) C K+. Repeating - for the set K+ - the corresponding arguments in the proof of Theorem 6.1 and Corollary 6.1, we obtain the existence in K+ of a fixed point K+ for the transformation FA. Then /:,+ = (P+ + K+)!B+ is the desired invariant for A subspace which extends £. ~
c .Yt± is called a dual pair (d.p.). If £± E 9J1± then (£+, £_) is called a maximal dual pair (m.d.p.).
DEFINITION 6.2. A pair (£ +, £ _) where £ ±
A dual pair is said to be invariant for an operator A if A £ ± C;;; £ ±. THEOREM 6.2. Any focusing strict plus-operator A on a reflexive Jp-space!B which
satisfies property (4.7) has a unique invariant m.d.p. (£~, £~). Moreover, £~ E E 9J1~, A £ ~ = £~, £ ~ + £ ~ = !B, there exists a number r < 1 such that 1)..1 ~ rlMI for all)" E u(A I £~), ME u(A I £~), and u(A) = u(A I £~) U u(A I £~). <J By Theorem 4.4, the operator A * is a focusing strict plus-operator satisfying condition (4.7). By Theorem 6.1, the operators A and A* have in 9J1+ and 9J1~ the unique invariant subspaces £ ~ E 9J1~ and £ ~* E 9J1~*, respectively, such that A£~ = £~ and A* £~* = £~*. Setting £~ = l..£~* we obtain A£~ c £~. By Theorem 4.1, we know that £ ~ E 9J1~. We show now that £ ~ + £ ~ = !B. Let K ± be the angle operators of £~. We have P++K~ +P_+K~ = I+K, where K = K~ +K~ and IIKllp < 1. By Theorem 0.4.6, the operator 1+ K is a homeomorphism from !B onto !B, whence (I + K)!B = !B, i.e., £ ~ + £ ~ = !B. From the last equality it follows that u(A) = u(A I £~) Uu(A I £~). Hence, by Theorem 6.1, the conclusions in Theorem 6.2 concerning the spectrum u(A) are true. ~
Let A be a double strict plus-operator on the reflexive Jpspace !B, such that A12 and A21 are completely continuous. Then any invariant d.p. (£+, £_) for which A£± = £± and £± E T± can be extended to an invariant m.d.p. (£~ax, £~ax). COROLLARY 6.3.
<J By Lemma 4.4, the operator A satisfies condition (4.7). By Corollary 6.2,
An application of pixed point principles the subspace
.c + can
265
.c r;'ax E 9J1+. + Q~.)I"2L ~ .c_. Since
be extended to an invariant for A subspace
R:: of those Q+ E K~ for which (P+ .c _ E T_, the set R:: is non-void, and from A.c _ = .c _ it follows that FA R::) c c R::. Hence, as in the proof of Corollary 6.2, we obtain the existence of a fixed Consider the subset
* (
point Q+,i.e.,FA*(Q+) = Q+. Setting -K_ = Q+ and get that A.c Illax c .c Ill ax and .c _ c .c Ill ax . ~~
.c Illax =
(P_ + K_)12L we
To conclude this section we establish a characteristic spectral property of a plus-operator with a focusig power.
Let A be a strict plus-operator on a reflexive Jp-space IB, satisfying property (4.7). Then a power A rn of A is a focusing operator if and only if there exists a m.d.p. (.c +, .c _) like the one in Theorem 6.2.
THEOREM 6.3.
Necessity. Let Am be a focusing operator for some m?': 1. It is easy to
see that FAm =
PA.
By Theorem
VI.2.2,
the transformation FA has an s-fixed
point K~ E K~. Therefore the subspace .c~ = (P+ + K~)IB+ is invariant for A. By Theorem 4.4, the operator A *m is focusing, too. Arguing as above we find an invariant for A * subspace .c ~* E 9J1~o. Setting .c ~ = l...c ~* we obtain the invariant m.d.p. (.c~, .c~). By Theorem 6.2, the spectrum O"(A) satisfies all the mentioned conditions. Sufficiency. Assume the existence of an invariant m.d. p. (.c ~, .c ~) such that 9J1~ and IAI ~ rlILI for all A E O"(A I .c ~), IL E O"(A I .c ~), where 0 ~ r < 1. Then IB = .c ~ +.c ~ and therefore any vector x E IB can be represented as x = y + z, where y E .c~, z E .c~. Put
.c ~
E
We have
+ 211P-11 IIAnzl1 IIAnyl1 211P+11 IIAnzl1 IIAnyl1
IIP_Anyll IIP+Anyll 1_
By Theorem
IV.2.2, we find
ml
EN such that
SPACES WITH INDEFINITE METRIC
266 where r ~ rl
< 1. Therefore sup
IIAnzl1
zE£':..,
Ilzll=l inf
---"----'-:---,;---,--:::----:-;- ----+ yE £':.. ,
IIAnYl1
0,
for n
----+ 00.
IIYII=l
Hence there exists a number m;;:: ml such that lm(x) ~ 'Y is focusing.
~
< 1, i.e., the operator Am
References
M.
ABO-ALLA,
[1] L'ensemble des pointes fixes d'une application holomorphe dans un produit fini de boules-unite d'espaces de Hilbert est une sous-variete banachique complexe, Ann. Mat. Pum Appl. (4), 153(1988),63-76.
D. [1] On injective holomorphic Fredholm mappings of index 0 in complex Banach spaces, Comment. Math. Univ. Carolin., 21:3(1980), 513-525.
ABST,
AHMEOOV,
K. T.
[1] The analytic method of Nekrasov-Nazarov in non-linear analysis (Russian), Uspehi Mat. Nauk, 12:4(1957), 135-153. ArZENBERG,
L. A.; YUZHAKOV, A. P.
[1] Integral representations and residues in multidimensional complex analysis (Russian), Nauka, Novosibirsk, 1979.
T. [1] Linear operators in Krein spaces, Sapporo, Japan, 1979.
ANOO,
T. YA.; IOHvIDov, 1. S. [1] Foundations of the theory of linear operators in spaces with indefinite metric (Russian), Nauka, Moscow, 1986.
Azrzov,
C. [1] On the converse of the Banach fixed point principle, Colloq. Math., 7:1(1959), 41-43.
BESSAGE,
REFERENCES
268
M. L. [1] On properties of an operator mapping the non-negative part of a space with indefinite metric into itself (Russian), Uspehi Mat. Nauk, 14:1(1959), 147-152.
BRODSKII,
CARTAN, H.
[1] Differential calculus. Differential forms (Russian transl.), Mir, Moscow, 1971.
N. [1] Uniformity in linear spaces, Trans. Amer. Math. Soc., 44(1938), 304-356.
DUNFORD,
N.; SCHWARTZ, J. [1] Linear operators. I, Intersciencc, New York, 1958.
DUNFORD,
EARLY, C.; HAMILTON, R.
[1] A fixed point theorem for holomorphic mappings, Proc. Sympos. Pure Math., 16(1970),61-65. EDELSTEIN, M.
[1] An extension of Banach contraction principle, Proc. Amer. Math. Soc., 12(1961), 7-10. I. M.; RAIKOV, D. A.; SHILOV, G. E. [1] Commutative normed rings (Russian), Fizmatgiz, Moscow, 1960.
GELFAND,
A. E. [1] Theorems on implicit abstract functions (Russian), Dokl. Akad. Nauk SSSR, 132:3(1960), 501-503.
GELMAN,
GOEBEL,
K.;
REICH, S.
[1] Uniform convexity, hyperbolic geometry and nonexpansive mappings, in Pure and Appl. Math. Series of Monographs and Textbooks, Marcel Dekker Inc.,1984. I. C.; KREIN, M. G. [1] Fundamental aspects on defect numbers, root numbers and indexes of linear operators, Uspehi Mat. Nauk, 12:2(1957), 43-48.
GOHBERG,
M. [1] Geometric theory of functions of a complex variable (Russian), Nauka, Moscow, 1966.
GOLUZIN, G.
HARRIS,
L.
[1] Schwarz's lemma in normed linear spaces, Prod. Nat. Acad. Sci. USA, 62(1969), 1014-1017. [2] A continuous form of Schwarz's lemma in normed linear spaces, Pacific J. Math., 38(1971),635-639.
REFERENCES
269
[3] The numerical range of holomorphic functions in Banach spaces, Amer. J. Math., 93(1971), lO05-1019. [4] Schwarz-Pick systems of pseudometrics for domains in normed linear spaces, Advances in Holomorphy, North Holland, 1979, pp. 345-406. T.; SUFFRlDGE, T. [1] Fixed points of holomorphic maps in Banach spaces, Proc. Amer. Math. Soc., 60(1976),95-lO5. [2] Biholomorphic maps in Hilbert spaces have a fixed point, Pacific J. Math., 38( 1971), 419-422.
HAYDEN,
HELTON,
J. W.
[1] Operators unitary in an indefinite metric and linear fractional transformations, Acta Sci. Math., 32(1971), 261-266. HERVE,
M.
[1] Several complex variables, Tata Institute of Fundamental Research, Bombay and Oxford Univ. Press, 1963. [2] Analyticity in infinite dimensional spaces, in Studies in Math., 10, De Gruyter, 1989. E.; PHILLIPS, R. S. [1] Functional analysis and semigroups, A.M.S. Providence, 1957.
HILLE,
IOHVlDOV, I. S.
[1] On spectra of hermitean and unitary operators on spaces with indefinite metric (Russian), Dokl. Akad. Nauk SSSR, 71:2(1950), 225-228. [2] Banach spaces with a J-metric, J-non-negative operators (Russian), Dokl. Akad. Nauk SSSR, 169:2(1966), 259--26l. [3] Banach spaces with a J-metric and certain classes of linear operators in these spaces (Russian), Bul. Acad. Stiince RSS Moldoven., 1(1968), 60-80. [4] Studies in the theory of linear operators on spaces with a J-metric and the theory of Toeplitz forms (Russian), Ph. D. Thesis, Odessa, 1976.
A. A. [1] Fixed points for mappings in metric spaces (Russian), in Studies in topology, 66(1976), Nauka, Leningrad, pp. 5-lO2.
IVANOV,
M. I.; MITYAGIN, B. S. [1] Complement able subspaces in Banach spaces (Russian), Uspehi Mat. Nauk, 28:6 (1973), 77-94.
KADEC,
L. V.; AKILOV, G. P. [1] Functional analysis (Russian), Nauka, Moscow, 1977.
KANTOROVIC,
REFERENCES
270
T. [1] Perturbation theory for linear operators, Springer-Verlag, Berlin-HeidelbergNew York, 1966.
KATO,
KHATSKEVICH, V.
A.
[1] On an application of the contraction principle in the operator theory on spaces with indefinite metric (Russian), Funct. Anal. i Pril., 12:1(1978),88-89.
[2] On J-semi-unitary operators on Hilbert J-spaces (Russian), Mat. Zametki, 17:4 (1975),639-647. [3] On invariant subspaces for some classes of operators on normed spaces with indefinite metric (Russian), Mat. Issled., 6:3(1971), 133-147.
A.; SHOIKHET, D. M. [1] Fixed points for analytic operators on Banach spaces and applications (Russian),
KHATSKEVICH, V.
Sib. Mat. J., XXV:1(1984), 188-200.
M. A. [1] Some problems of non-linear analysis (Russian), Uspehi Mat. Nauk, 9:3(1954),
KRASNOSELSKII,
57-114. [2] On some new fixed point principles (Russian), Dokl. Akad. Nauk SSSR, 208:6 (1973), 1280-128l. [3] Topological methods in the theory of non-linear integral equations (Russian), Gostehizdat, Moscow, 1956. [4] Integral operators on spaces of summable functions (Russian), Nauka, Moscow, 1966.
M. A.; SOBOLEV, A. V. [1] On finite rank cones (Russian), Dokl. Akad. Nauk SSSR, 225:6(1975), 12561259.
KRASNOSELSKII,
(et al) [1] Approximative solutions of operatorial equations (Russian), Nauka, Moscow, 1969.
KRASNOSELSKII, M. A.; VAINIKKO, G. M.; ZABREIKO, P. P.;
KRASNOSELSKII, M. A.; ZABREIKO, P. P.;
[1] Geometric methods in nonlinear analysis (Russian), Nauka, Moscow, 1975. M. G. [1] On an application of a fixed point principle in the theory of linear mappings in spaces with an indefinite metric (Russian), Dokl. Akad. Nauk SSSR, 5:2(1950), 180-190.
KREIN,
REFERENCES
271
[2] An introduction in the geometry of indefinite J-spaces and the theory of operators on these spaces (Russian), in Second Summer School in Mathematics. I, Kiev, 1965, pp. 15-92. [3] On a new application of a fixed point principle in the theory of operators on spaces with indefinite metrics (Russian), Dokl. Acad. Nauk SSSR, 154:5(1964), 1023-1026.
M. G.; RUTMAN, M. A. [1] Linear operators having an invariant cone in a Banach space (Russian), Uspehi Mat. Nauk, 3:1(1948), 3-95.
KREIN,
KREIN, M. G.; SHMULIAN, Yu. L.
[1] On plus-operators in spaces with indefinite metric (Russian), Mat. Issled., 1:2 (1966), 172-210. [2] On functional-linear transformations with operatorial coefficients (Russian), Mat. Issled., 2:3(1967), 64-96.
H. [1] Invariant subspaces for a class of operators in spaces with indefinite metric, J. Funet. Anal., 19(1975), 232-24l. [2] On invariant subspaces for linear operators acting in spaces with indefinite metric (Russian), Dokl. Akad. Nauk SSSR, 169(1966), 12-15. [3] Remarks on invariant subspaces for linear operators in Banach spaces with indefinite metric (Russian), Mat. Issled., 4:1(1969), 27-3l.
LANGER,
A. [1] Extension of dual subspaces (Russian), Dokl. Akad. Nauk SSSR, 176:3(1967), 515-517. [2] Extensions of dual subs paces invariant relatively to an algebra (Russian), Mat. Zametki, 3:3(1968), 253-260.
LARIONOV, E.
N. A. [1] Area principle in the theory of single-valued functions (Russian), Nauka, Moscow, 1975.
LEBEDEV,
A. Ju.; LIFSIC, E. A. [1] On the generalized construction principle of M. A. Krasnoselskii (Russian), in Problems of the mathematical analysis of compound systems, 1, Voronets State University, 1967.
LEVIN,
A. M. [1] On the equilibrium figures, slightly different from ellipsoids, of a homogeneous
LIAPUNOV,
REFERENCES
272
liquid mass in a revolving motion, in Collected Works, vol. 4, Izdat. Akad. Nauk SSSR, Moscow, 1959 [2] Sur les figures d'equilibre peudifferentes des ellipsoides d'une masse liquide homogene donee d'un mouvement de rotation. Premiere partie, in Zap. Acad. Nauk, 1, St. Petersburg, 1906.
A. S.; MACAEV, V. I. [1] On spectral properties of holomorphic operator-functions in Hilbert spaces (Russian), Mat. Issled., 9:4(1974), 79-9l.
MARKUS,
P.; VIGUE, J.-P. [1] Points fixes d'une application holomorphe d'un domain borne dans lui-meme, Acta Math., 166(1991), 1-26.
MAZET,
P. R. [1] A converse to Banach's contraction theorem, J. Res. Nat. Bur. Standarts, 716 (1967), 73-76.
MEYERS,
J. [1] Critical points of complex hypersurfaces (Russian transl.), Mir, Moscow, 1971. 2] Topology from the differentiable viewpoint, University Press, Virginia, 1965.
MILNOR,
L.R [1] Recent developments in infinite dimensional holomorphy, Notas fis. Cent. brasil. pesquisas fis., 20:13(1973), 243-275.
NACHBIN,
NAIMARK, M.
A.
[1] On commuting unitary operators in the space II", Dokl. Akad. Nauk SSSR, 146:6(1963), 1261-1263. NAZAROV, N. N.
[1] Integral equations of the Hammerstein type (Russian), Trudy SACU, 33, Tashkent, 1941. N EKRASOV,
A. I.
[1] On steady waves, Izv. Ivanobo- Voznesenskovo Polytech. Inst., 6(1922), 155-17l.
P. [1] Minimal projections in Banach spaces. Uniqueness and existence problems and applications (Russian), WSP, Warsaw, 1985.
ODINEC, V.
Z. [1] Lecture notes on nonexpansive and monotone mappings in Banach spaces, Center for Dinamical Systems, Brown University, Providence, USA, 1967.
OPIAL,
REFERENCES ORAVA,
J.;
HALME,
273
A.
[1] Inversion of generalized power series representations, J. Math. Appl., 45(1974), 136-141. PETTIS,
B
[1] On integration in vector spaces, Trans. Amer. Math. Soc., 44(1938), 277-304. PHILLIPS, R.
[1] The extension of dual subs paces invariant under an algebra, in Proc. Internat. Sympos. Linear Spaces, Ierusalim, 1960, Ierusalim Acad. Press, Oxford - London - New York - Paris, 1961, pp. 366-398. PONTRYAGIN,
L.
S.
[1] Hermitean operators in spaces with indefinite metric (Russian), Izv.Akad. Nauk SSSR, Ser. Mat., 8(1944), 243-280. RIESZ,
F.
[1] Some mean ergodic theorems, J. London Math. Soc., 13(1938), 274-278.
W. [1] Function theory in the unit ball of en (Russian transl.), Mir, Moscow, 1984. [2] The fixed-point sets of some holomorphic maps, Bull. Malaysian Math. Soc., 1(1978),25-28. [3] Functional analysis (Russian transl.), Mir, Moscow, 1975.
RUDIN,
RYBIN, P.
P.
[1] Analytic study of a perturbed linear integral equation, Uchenye Zapiskii Kazansk. Univ., 117:2(1957), 75-78. V. A.; KHATSKEVICH, V. A. [1] On normed J-spaces and some classes of linear operators in these spaces (Russian), Mat. Issled., 8(1973), 56-75.
SENDEROV,
SCHMIDT,
E.
[1] Zur Theorie linearen und nichtlinearen Integralgleichungen. 3. Teil: Uber die Auflosung der nichtlinearen Integralgleichungen und die Verzweigung iher Losungen, Math. Ann., 65(1908), 370-399.
L. [1] Analysis. II (Russian transl.), Mir, Moscow, 1972.
SCHWARTZ,
B. V. [1] Introduction in complex analysis (Russian), Nauka, Moskow, 1969.
SHABAT,
REFERENCES
274
[2] Introduction in complex analysis (second edition) (Russian), Nauka, Moscow, 1976. SHOIKHET,
D. M.
[1] Remarks on fixed points for non-expansive analytic operators, in Complex analysis and mathematical physics, Krasnoiarsk, 1988, pp. 145-150. [2] Invariance principle in the theory of fixed points for analytic operators, preprint, Institute of Physics, Siberian Branch, Akademy of Sciences of USSR, no. 33M, Krasnoiarsk, 1986. [3] Some properties of Fredholm mappings in Banach analytic manifolds, Integral Equations Operator Theory, 16(1993),430-450.
C. [1] On the fixed points of fractional-linear transformations (Russian), Funet. Anal. i Priloj., 14:2(1980), 93-94.
SHULMAN, V.
SOBOLEV, A. V.; KHATSKEVICH, V. A.
[1] On definite invariant subspaces and the structure of the spectrum of a focusing plus-operator (Russian), Funet. Anal. i Priloj., 15:1(1981),84-85.
L. [1] About the motion of a symmetrical top whose cavity is filled with a liquid, J. Appl. Math. Theoretical Physics, 3(1960), 20-55.
SOBOLEV, S.
B.; FOIA§, C. [1] Analyse harmonique des operateurs de l'espace de Hilbert, Akademiai Kiad6, Budapest, 1967.
SZ.-NAGY,
TRENOGIN, V. A.
[1] Functional analysis (Russian), Nauka, Moscow, 1980.
M. M. [1] Variational method and monotone operator method in the theory of non-linear operators (Russian), Nauka, Moscow, 1972.
VAINBERG,
VAINBERG, M. M.; AIZENGENDLER, P. G.
[1) Methods of investigation in the theory of the ramification of solutions, in Itogy nauky, Mat. Anal., VINITI, Moscow, 1965, pp. 4-70. VAINBERG, M. M.; TRENOGIN, V. A.
[1) Ramification theory of the solution of non-linear equations, (Russian), Nauka, Moscow, 1969.
REFERENCES
275
E. [lJ On the subharmonicity of the spectral radius, Boll. Un. Math. Ital., 4(1968), 427-429. [2J Complex geodesics, Comp. Math., 44(1981), 375-394. [3J Complex geodesics and holomorphic maps, Sympos. Math., 26(1982), 211-230.
VESENTlNI,
VIGUE, J.-P.,
[1 J Points fixes d'applications holomorphes dans un domaine borne convexe de Trans. Amer. Math. Soc., 289(1985),345-353. WITTSTOCK,
en,
G.
[lJ Uber invariante Teilraume zu positiven Transformationen in Raumen mit indefiniter Metrik, Math. Ann., 3(1972), 167-175. YOSIDA, K.
[lJ Mean ergodic theorem in Banach spaces, Proc. Imp. Acad. Tokio, 14(1938), 292294.
A. P. [lJ On the application of multiple logarithmic residue for the decomposition of implicit functions in power series (Russian), Mat. S., 97:2(1975), 177-192. [2J On estimates of the convergence domain and the remainder of the multiple Lagrange series for systems of implicit functions (Ryssian), Sib. Mat. J., 21:5(1980), 176-179.
YUZHAKOV,
A. P.; TSIKH, A. K. [lJ On the multiplicity of zeros of systems of holomorphic functions (Russian), Sib. Mat.J., 1:3(1978), 693-697.
YUZHAKOV,
M. A. [lJ On a method for obtaining new fixed point principles (Russian), Dokl. Akad. Nauk SSSR, 176:6(1967), 1233-1235.
ZABREIKO, P. P.; KRASNOSELSKII,
M. G.; KREIN, M. G.; KUCMENT, P. A.; PANKOV, A. A. [lJ Banach bundles and linear operators (Russian), Uspehi Mat. Nauk, 30(1975), 101-157.
ZAIDENBERG,
List of Symbols
8wF(xo, h) 8F(xo, h) DF(xo, h) dF(xo, h) Vw(xo, ~)
the first weak variation of the (nonlinear) operator F at a point Xo the first variation of the (nonlinear) operator F at a point Xo the Gateaux differential of the (nonlinear) operator F at a point Xo the Frechet differential of the (nonlinear) operator F at a point Xo the class of all (nonlinear) operators acting from a normed space x into a normed space ~ and having first weak variation at a point Xo the class of all (nonlinear) operators acting from a normed space x into V(xo, ~) a normed space ~ and having first variation at a point Xo Pettis integral (weak integral) the class of all p-holomorphic (holomorphic with respect to p-topology) operators F : 'j) ---> ~, where 'j) is a p-open set in x and x, ~ are normed spaces the class of all holomorphic operators F : 'j) ---> ~, where 'j) is an open 1t('j),~) set in x and x, ~ are normed spaces the spectrum of a linear operator A a(A) the spectral radius of a linear operator A r(A) an indefinite metric on a Banach space Jv(x) the set of all non-negative (non-positive) vectors in a Banach space with st+ (st_) indefinite metric st++ (st __ ) the set of all positive (negative) vectors in a Banach space with in definite metric 9)1+ (9)1_) the class of all maximal non-negative (non-positive) subspaces of a Banach space with indefinite metric
Subject Index
A
Freche differential 36, 213
Aizenberg, L. 202
Fredholm operator 120, 208, 233, 234
analytic manifold 233 analytic operator 53 Azizov, T. 261 B
G Gateaux differential 35, 53 G-metric 239 H
Banach principle 134 Banach Steinhaus type theorem 100 bifurcation point 233 Bijection 4
holomorphic function 83 holomorphic convex hull 218 Hausdorff space 12, 14 Hahn-Banach-Suchomlynoff theorem 27
C
Herve, M. 181, 184
Cauchy inequalities 87
HOlder condition 190
completely continuous operator 21, 30, 55, 144 converse of Banach principle 136, 173
I
Implicit Function Theorem 160, 164, 189, 203
D
invertible operator 169, 197, 198
defect number 247
injection 4
E
K
ergodic theorem 131 F
Krasnoselskii, M. 57, 137, 157, 170, 196, 199
focusing operator 252
Krein space 240, 259
280
SUBJECT INDEX
Krein, M. 240, 253, 259 L Liapunov equations of ramification 227 local chart 212
relativity compactness 12 Rudin, W. 185,233
s Schmidt equation of ramification 228 Schmidt lemma 120
M mapping 3 maximum principle for spectral radius 111 metric space 13 metrizability 15, 21, 200 Montel property 18, 23, 99
semi-definite lineal 241 Shabat, B. 200 Schauder principle 137 Schwarz lemma 93 small solution 201 smooth manifold 212
N
split able operator 129, 226
negative vector 240
Stein manifold 220, 234
neutral vector 240
strictly convex space 28
non-expansive mapping 136, 186, 226
submanifold 216
normally solvable operator 118
Suffridge, T. 138
o
T
operator
topological
- proper 182
- isomorphism 12
- g-analytic 53
- product 10
- 6-analytic 53
- space 5
- F-analytic 54
Trenogin, V. 137, 157, 170, 189
- Hammerstein 42, 45, 149, 157
V
- Nemytski 42
Vainberg, M. 137, 157
- p-holomorphic 202
Vesentini, E. 110, 182
- Urysohn 41
u
p
unbinding of an equation 196
positive vector 240
y
R
Yoshida, K. 131
regular fixed point 182
Yuzhakov, A. 200
Titles previously published in the series OPERATOR THEORY: ADVANCES AND APPLICATIONS BIRKHAuSER VERLAG
1. H. Bart, I. Gohberg, M.A. Kaashoek: Minimal Factorization of Matrix and Operator Functions, 1979, (3-7643-1139-8) 2. e. Apostol, R.G. Douglas, B.Sz.-Nagy, D. Voiculescu, Gr. Arsene (Eds.): Topics in Modern Operator Theory, 1981, (3-7643-1244-0) 3. K. Clancey, I. Gohberg: Factorization of Matrix Functions and Singular Integral Operators, 1981, (3-7643-1297-1) 4. I. Gohberg (Ed.): Toeplitz Centennial, 1982, (3-7643-1333-1) 5. H.G. Kaper, e.G. Lekkerkerker, J. Hejtmanek: Spectral Methods in Linear Transport Theory, 1982, (3-7643-1372-2) 6. e. Apostol, R.G. Douglas, B. Sz-Nagy, D. Voiculescu, Gr. Arsene (Eds.): Invariant Subspaces and Other Topics, 1982, (3-7643-1360-9) 7. M.G. Krein: Topics in Differential and Integral Equations and Operator Theory, 1983, (3-7643-1517 -2) 8. I. Gohberg, P. Lancaster, l. Rodman: Matrices and Indefinite Scalar Products, 1983, (3-7643-1527-X) 9. H. Baumgartel, M. Wollenberg: Mathematical Scattering Theory, 1983, (3-7643-1519-9) 10. D. Xia: Spectral Theory of Hyponormal Operators, 1983, (3-7643-1541-5) 11. e. Apostol, e.M. Pearcy, B. Sz.-Nagy, D. Voiculescu, Gr. Arsene (Eds.): Dilation Theory, Toeplitz Operators and Other Topics, 1983, (3-7643-1516-4) 12 H. Dym, I. Gohberg (Eds.): Topics in Operator Theory Systems and Networks, 1984, (3-7643-1550-4) 13. G. Heinig, K. Rost: Algebraic Methods for Toeplitz-like Matrices and Operators, 1984, (3-7643-1643-8) 14. H. Helson, B. Sz.-Nagy, F.-H. Vasilescu, D.Voiculescu, Gr. Arsene (Eds.): Spectral Theory of Linear Operators and Related Topics, 1984, (3-7643-1642-X) 15. H. Baumgartel: Analytic Perturbation Theory for Matrices and Operators, 1984, (3-7643-1664-0) 16. H. Konig: Eigenvalue Distribution of Compact Operators, 1986, (3-7643-1755-8) 17. R.G. Douglas, C.M. Pearcy, B. Sz.-Nagy, F.-H. Vasilescu, D. Voiculescu, Gr. Arsene (Eds.): Advances in Invariant Subspaces and Other Results of Operator Theory, 1986, (3-7643-1763-9) 18. I. Gohberg (Ed.): I. Schur Methods in Operator Theory and Signal Processing, 1986, (3-7643-1776-0)
19. H. Bart, I. Gohberg, M.A. Kaashoek (Eds.): Operator Theory and Systems, 1986, (3-7643-1783-3) 20. D. Amir: Isometric characterization of Inner Product Spaces, 1986, (3-7643-1774-4) 21. I. Gohberg, M.A. Kaashoek (Eds.): Constructive Methods of Wiener-Hopf Factorization, 1986, (3-7643-1826-0) 22. VA. Marchenko: Sturm-Liouville Operators and Applications, 1986, (3-7643-1794-9) 23. W. Greenberg, C. van der Mee, V. Protopopescu: Boundary Value Problems in Abstract
Kinetic Theory, 1987, (3-7643-1765-5) 24. H. Helson, B. Sz.-Nagy, F.-H. Vasilescu. D. Voiculescu, Gr. Arsene (Eds.): Operators in Indefinite Metric Spaces, Scattering Theory and Other Topics, 1987, (3-7643-1843-0) 25. G.S. Litvlnchuk, 10M. Spitkovskii: Factorization of Measurable Matrix Functions, 1987, (3-7643-1883-X) 26. N.Y. Krupnik: Banach Algebras with Symbol and Singular Integral Operators, 1987, (3-7643-1836-8) 27. A. Bultheel: Laurent Series and their Pade Approximation, 1987, (3-7643-1940-2) 28. H. Helson, C.M. Pearcy, F.-H. Vasilescu, D. Voiculescu, Gr. Arsene (Eds.): Special
Classes of Linear Operators and Other Topics, 1988, (3-7643-1970-4) 29. I. Gohberg (Ed.): Topics in Operator Theory and Interpolation, 1988, (3-7634-1960-7) 30. Yu.!. Lyubich: Introduction to the Theory of Banach Representations of Groups, 1988, (3-7643-2207 -1) 31. E.M. Polishchuk: Continual Means and Boundary Value Problems in Function Spaces, 1988, (3-7643-2217-9) 32. I. Gohberg (Ed.): Topics in Operator Theory. Constantin Apostol Memorial Issue, 1988, (3-7643-2232-2) 33. I. Gohberg (Ed.): Topics in Interplation Theory of Rational Matrix-Valued Functions, 1988, (3-7643-2233-0) 34. I. Gohberg (Ed.): Orthogonal Matrix-Valued Polynomials and Applications, 1988, (3-7643-2242 -X) 35. I. Gohberg, J.W. Helton, L. Rodman (Eds.): Contributions to Operator Theory and its Applications, 1988, (3-7643-2221-7) 36. G.R. Belitskii, Yu.!. Lyubich: Matrix Norms and their Applications, 1988, (3-7643-2220-9) 37. K. Schmudgen: Unbounded Operator Algebras and Representation Theory, 1990, (3-7643-2321-3) 38. L. Rodman: An Introduction to Operator Polynomials, 1989, (3-7643-2324-8) 39. M. Martin, M. Putinar: Lectures on Hyponormal Operators, 1989, (3-7643-2329-9) 40. H. Dym, S. Goldberg, P. Lancaster, MA. Kaashoek (Eds.): The Gohberg Anniversary Collection, Volume I, 1989, (3-7643-2307-8) 41. H. Dym, S. Goldberg, P. Lancaster, MA. Kaashoek (Eds.): The Gohberg Anniversary
Collection, Volume II. 1989. (3-7643-2308-6)
42. N.K. Nikolskii (Ed.): Toeplitz Operators and Spectral Function Theory, 1989, (3-7643-2344-2) 43. H. Helson, B. Sz.-Nagy, F.-H. Vasilescu, Gr. Arsene (Eds.): Linear Operators in Function Spaces, 1990, (3-7643-2343-4) 44. C. Foias, A. Frazho: The Commutant Lifting Approach to Interpolation Problems, 1990, (3-7643-2461-9) 45. J.A. Ball, I. Gohberg, L. Rodman: Interpolation of Rational Matrix Functions, 1990, (3-7643-2476-7) 46. P. Exner, H. Neidhardt (Eds.): Order, Disorder and Chaos in Quantum Systems, 1990, (3-7643-2492-9) 47. I. Gohberg (Ed.): Extension and Interpolation of Linear Operators and Matrix Functions, 1990, (3-7643-2530-5) 48. L. de Branges, I. Gohberg, J. Rovnyak (Eds.)" TopIcs in Operator Theory. Ernst D. Hellinger Memorial Volume, 1990, (3-7643-2532-1) 49. I. Gohberg, S. Goldberg, M.A. Kaashoek: Classes of Linear Operators, Volume I, 1990, (3-7643-2531-3) 50. H. Bart, I. Gohberg, M.A. Kaashoek (Eds ): Topics in Matrix and Operator Theory, 1991, (3-7643-2570-4) 51. W. Greenberg, J. Polewczak (Eds.): Modern Mathematical Methods in Transport Theory, 1991, (3-7643-2571-2) 52. S. PrOssdorf, B. Silbermann: Numerical Analysis for Integral and Related Operator Equations, 1991, (3-7643-2620-4) 53. I. Gohberg, N. Krupnik: One-Dimensional Linear Singular Integral Equations, Volume I, Introduction, 1992, (3-7643-2584-4) 54. I. Gohberg, N. Krupnik: One-Dimensional Linear Singular Integral Equations, Volume II, General Theory and Applications, 1992, (3-7643-2796-0) 55 R.R. Akhrnerov, M,I. Kamenskii, A.S. Potapov, A.E. Rodkina, B.N. Sadovskii: Measures of Noncompactness and Condensing Operators, 1992, (3-7643-2716-2) 56. I. Gohberg (Ed.): Time-Variant Systems and Interpolation, 1992, (3-7643-2738-3) 57. M. Demuth, B. Grarnsch, B.W. Schulze (Eds.): Operator Calculus and Spectral Theory, 1992, (3-7643-2792-8) 58. I. Gohberg (Ed.): Continuous and Discrete Fourier Transforms, Extension Problems and Wiener-Hopf Equations, 1992, (ISBN 3-7643-2809-6) 59. T. Ando, I. Gohberg (Eds.): Operator Theory and Complex Analysis, 1992, (3-7643-2824-X) 60. P.A. Kuchrnent: FloquetTheory for Partial Differential Equations, 1993, (3-7643-2901-7) 61. A. Gheondea, D. Timotin, F.-H. Vasilescu (Eds.): Operator Extensions, Interpolation of Functions and Related Topics, 1993, (3-7643-2902-5)
62. T. Furuta, I. Gohberg, T. Nakazi (Eds.): Contributions to Operator Theory and its Applications. The Tsuyoshi Ando Anniversary Volume, 1993, (3-7643-2928-9) 63. I. Gohberg, S. Goldberg, M.A. Kaashoek: Classes of Linear Operators, Volume 2, 1993, (3-7643-2944-0) 64. I. Gohberg (Ed.): New Aspects in Interpolation and Completion Theories, 1993,
(ISBN 3-7643-2948-3) 65. M.M. Djrbashian: Harmonic Analysis and Boundary Value Problems in the Complex Domain, 1993, (3-7643-2855-X) 66. V. Khatskevich, D. Shoiykhet: Differentiable Operators and Nonlinear Equations, 1993, (3-7643-2929-7)
BIRKHAU5ER
I. Gohberg I S. Goldberg
Basic Operator Theory 1981. 304 pages. Softcover. 3rd printrun ISBN 3-7643-3028-7 Basic Operator Theory provides an introduction to functional analysis with an emphasis on the theory of linear operators and its application to differential and integral equations, approximation theory, and numerical analysis. A textbook designed for senior undergraduate and graduate students, Basic Operator Theory begins with the geometry of Hilbert space and proceeds to the spectral theory for compact self-adjoint operators with a wide range of applications. Part of the volume is devoted to Banach spaces and operators acting on these spaces. Presented as a natural continuation of linear algebra, Basic Operator Theory provides a firm foundation in operator theory, an essential part of mathematical training for students of mathematics, engineering, and other technical sciences.
PI .... order through your boo~_ or wrhe to:
Birkhiiuser
Birkhauser Verlag AG POBox 133 CH·4010 Basel/ Switzerland
FAX ••41/61/2717666
Blrkhiuser Verlag AG Baed Boston Berlin
For order. orl .....tIng In _ USA or CMade: BirkhiUler
44 HartzWav Secaucus, NJ 07096-2491/ USA
If you would like regular ~~e infama1lon from Bukhauser please write to the follOWIng address for your personal copy aI the BIrkhIu_
,.,,,.,,,./k;. Qw"eriy.
MATHEMATICS
i
Victor Khatskevich and David Shoiykhet Differentiable Operators and Nonlinear Equations
The need to study holomorphic mappings in infinite dimensional spaces, in all likelihood, arose for the first time in connection with the development of nonlinear analysis. A systematic study of integral equations with an analytic nonlinear part was started at the end of the 19th and the beginning of the 20th centuries by A. liapunov, E. Schmidt, A. Nekrasov and others. Their research work was directed towards the theory of nonlinear waves and used mainly the indetermined coefficients and the majorant power series methods, which subsequently have been refined and developed. Parallel with these achievements, the theory of functions of one or several complex variables was gradually enriched with more significant and subtle results. The present book is a first step towards establishing a bridge between nonlinear analysis, nonlinear operator equations and the theory of holomorphic mappings on Banach spaces. The work concludes with brief exposition of the theory of spaces with indefinite metrics, and some relevant applications of the holomorphic mappings theory in this setting. In order to make this book accessible not only to specialists but also to students and engineers, the authors give a complete account of definitions and proofs, and also present relevant prerequisites from functional analysis and topology.
a
Birkhauser Verlag Basel · Boston · Berlin
ISBN 3-7643-2929-7 ISBN 0-8176-2929-7