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!
LINEAR FUNCTIONS
21
For normed linear spaces, this theorem has the following corollaries : 1 . If a continuous linear functional f (x) is defined on the space M which is a linear manifold in the normed linear space E, then there exists a continuous linear functional J (x), which is an extension off (x) to the entire space E, such that its norm is equal to the norm of the functional f ( x) : 11 / 11 = 11 /11 · 2. For any element x0, different from zero, of the normed linear space
E, a continuous linear functional f(x) on E exists such that
11/ 11 = 1 and f (xo)
=
ll xo ll .
3. Let T be a nonempty convex open set in a normed linear space E and let M be a linear manifold not intersecting T. A closed hyperplane
exists containing M and not intersecting T. 4. Examples of linear functionals.
Mathematical analysis presents many examples of linear functionals. The functional N
f (x) = L ak<';k k= 1 i s a continuous linear functional on the spaces IP, m, c and c0 whose clements are sequences x = { 1;1 , <'; 2, . . . , <';m . . . }. lf the sequence {an} is bounded, then the functional 00
f (x) = L ak<';k k= 1
00
will be a continuous linear functional on the space /1 . If L l ak l < oo, then k= 1 this linear functional is continuous on the spaces IP and m and on the \Uospaces c and c0 of the space m. The functional f (x) = lim <';k k+ ., continuous on the space c. Its norm is equal to one. The integral 1 00
f (x)
= I x (t) dt 0
22
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
can be used on the function spaces LP (O, 1), M(O, 1), C(O, 1) and others as an example of a continuous linear functional. In this connection, the Lebesgue integral, defined on L 1 (0, 1 ), can be considered as an extension of the functional generated by the Riemann integral on the space C(O, 1) c L 1 (0, 1 ) . The linear functional 1
P (x) =
f0 x (t) cp (t) dt
has a more general form. If cp (t ) is a bounded measurable function, then P(x) also is continuous in the spaces LiO, 1 ), M(O, 1) and C(O, 1 ). In the space C(O, 1 ), the value of the function x(t) at a fixed point t0 will be a continuous linear functional whose norm is equal to one : f (x) = x (t0) ,
\\! II = 1 .
Analogously, the value of the kth derivative x
23
CONJUGAT E SPACES
respect to each of the variables x, y ; that is,
B (ct1X1 + ct 2x 2 , y) = ct 1 B (x 1 , y) + ct 2B (x 2 , y) , B (x, ct1y 1 + ct 2Y2 ) = ct1 B(x, y 1 ) + ct 2 B (x, J2 ) . We say that the bilinear functional B(x, y) puts the linear systems E and F in duality or that E and Fare in duality (with respect to B) if the following two conditions are satisfied : 1 . For every x ¥ 8 in E, there exists an y E F such that B(x,y) #0; 2. For every y # 8 in F, there exists an xEE such that B(x, y) # 0. Let E be a linear system over the field of real or complex numbers and let E be the set of all linear functionalsf(x) defined on E. The operations of addition of elements f and g and multiplication of an element f by a number A. can be introduced in the following manner in E : 1) h f+ g is the functional on E such that
h (x) = f (x) + g (x)
(x E E) ;
2) f1 = A.f (A. is a number, fE E) is the functional on E such that
ft (x) = A.f (x) .
The set E becomes a linear system which is called the algebraic conjugate space of the space E. If we define the bilinear functional B(x,J) f (x) on E x E, then E and l are placed, in duality, by means of this bilinear functional. 2. Conjugate space to a normed linear space. The set E' of all con t i n uous linear functionals defined on the normed linear space E is a linear manifold in the algebraic conjugate space E since the sum of two con t i nuous linear functionals and the product of a continuous linear func t ional and a number are continuous linear functionals. If we take as the n o r m of the element fEE' the norm l l f ll of the functional f(x), then E' becomes a normed linear space which is called the conjugate space of the ' pace E. The space E' is complete, so the conjugate space of a normed linear ' pace is always a Banach space. The concept of a continuous li near functional for metric linear spaces l·a 1 1 he defined in exactly the same way as was done for normed linear ' pan�s. H owever, there is no assertion of the type of corollary 2 of no.
24
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
3, § 3 which guarantees the existence of a nontrivial continuous linear functional. Moreover, examples of metric linear spaces E exist for which the conjugate space E' consists solely of the continuous linear functional which is identically equal to zero on E. The presence in E of a convex neighborhood of zero different from E is a necessary and sufficient conditionfor the existence of a nontrivialfunctional in E. For concrete normed linear spaces, the problem of the description of their conjugate spaces often arises. The statement of this problem is made more precise. It is a fact that the definition of a conjugate space gives its description. In a narrower sense, by the description of the conjugate space we mean the designation of a concrete Banach space which is isometric to E', the conjugate of the given space E, and a method of computing the values f(x) of a functional /E E' on the elements xEE. However the posed problem does not have one specific answer since concrete spaces, isometric to the space E ' can be constructed in different ways. For example, let the norm ,
n 1
be defined in En. If the linear functionals on En are represented in the form n1
f (x) = a1�1 + kL ak + l (�k + 1  �k) , =1
then the space E! with the norm
11/ 11 = max l ak l 1 � k �n
will be the conjugate space to En . If the linear functionals are represented in the usual form by n
then the space E; with the norm
11/ 11 = max L b , 1 �k�n i= k will be the conjugate space. n
CONJUGATE SPACES
25
The spaces E! and E; are isometric. The correspondence between n
them is given by the relation ak�L b1. i=k
The problem of the description of the conjugate space is welldefined if a method of computing the values of a functional is given analytically in advance, or, as we say, in the form of a functional. Moreover, a linear functional is usually sought, by analogy with linear forms, in the form of a sum of products or in the form of an integral of a product.
Examples 1. The space IP (p > 1). An arbitrary continuous linear functional de fined on the space IP is representable in the form 00
i=
1
The space conjugate to the space IP is isometric to the space lq, where I 1 += 1.
p
q
2. The space /1 • Every continuous linear functional on /1 is representable
in the form
00
f (x) = I ftC i= 1
where II/ II
=
sup l /;1 < oo .
l � i < oo
The conjugate space of /1 is isometric to the space m. 3 . The space c 0 • A continuous linear functional on c 0 can be given by the equality 00
00
where II/ II =
L1 l ftl < oo .
i=
The conj ugate space of c0 is isometric to the space /1 • 4. The space LP(O, I ), p > I . An arbitrary continuous linear functional
26
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
on the space LP (O, 1) is representable in the form 1
f (x)
=
f x(t) a (t) dt , 0
p where a (t) E Lq (O, l), q = p1 The norm of the functional f is defin ed by the formula 1 1 II/ II = I a (t)l q dt .
{J 0
r
The conjugate space of LP (O, 1) is isometric to the space L q (O, 1) = 1 . +
) � (� The space L1 (0, 1). A continuous linear functional on L1 (0, 1) is 5.
representable in the form
1
f (x)
=
f x (t) a (t) dt , 0
where a (t) is a function bounded almost everywhere on the interval [0,1] and 11/ 11 = vrai max l a (t)l . O � t� 1 The conjugate space of the space L 1 (0, 1) is isometric to the space M(O, 1). 6. The space C(O, 1 ). Every continuous linear functional on C (0, 1) is representable in the form of a Stieltjes integral 1
f (x) =
J x (t) dg (t) , 0
where g (t) is a function of bounded variation. The functional f(x) is not changed if an arbitrary constant is added to the function g (t) ; therefore we set g(O) = O. However, even with this condition, different functions g(t) can generate the same functional. These functions can be distinguished by their values at the points of discontinuity lying inside the interval [0, 1]. If, for example, only functions g (t) are considered
27
CONJUGATE SPACES
for which
g (t + O) + g (t  O) for t e (O, l) , g (t) = 2 then the correspondence between the functionals f (x) and the functions g (t ) becomes onetoone. In this connection 1
II ! II = V(g) . 0
The conjugate space of the space C(O, 1 ) is isometric to the subspace V0 (0, I) of the space V(O, I) consisting of all functions g (t)e V(O, I ) satisfying the conditions : g (O) = O and g (t) =t[g (t + O) + g (t  0)] for te(O, 1 ) .
3 . Weak and weak* topology. In the conjugate space E' of a normed linear space E, in addition to the topology induced by the metric of E' (the strong topology), we also define the socalled weak* topology. For every o: > O and every finite number of elements X; (i= I, 2, , n) of E, we denote by W(x1, , Xn, o:) the set ofallf eE' such that I f(x;)l � o:. The topology for which the sets W(xl> . . . , x"' o:) form a fundamental system of neighborhoods of zero is called the weak* topology a (E', E ) in the conjugate space E', where x1, x2, , Xn range over all possible finite collections of elements of E and o: ranges over the set of all positive numbers. Analogously, the weak topology is defined in the space E. We denote by W(/1, ,/,, o:) (/1, JneE', o: > O) the set of points x in E such that I /; (x)l � o:. The topology for which the sets W(f1, , J"' o:) form a fundamental system of neighborhoods of zero, where /1, , jn range over all possible finite collections of elements of E' and o: ranges over the set of all positive numbers, is called the weak topology a (E, E') . Thus, every set containing a set W(x1, , Xn, o:) ( W(/1, Jn, o:)) is taken for a neighborhood of zero V�a) in the weak* (weak) topology ; every neighborhood V� of the element x is obtained from some neighbor hood of zero v: by means of its "displacement" by the element x : .
.
.
. . •
.
• • •
•
.
• • •
•
•
.
.
• • •
.
.
.
.
•
The weak* and weak topologies are locally convex topologies since the 'cls W(x 1 , x", et) and W( /1 , , fn, et) are convex. ' • • •
,
.
•
•
28
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
In the case of an infinitedimensional space E, the weak topology u(E, E') is an example of a nonmetrizable locally convex topology ; the weak* topology u(E', E) will also be nonmetrizable if the space E is infinitedimensional and a Banach space. A sequence of continuous linear functionals /1 ,/2 , . . ,/,., . . . is called weakly convergent to the functional fo if it converges to fo in the weak* topology u(E', E) of the space E'. In order for the sequence/1,/2 , . . . ,fn, . . . to be weakly convergent to the functional /0, it is necessary and sufficient that fn (x) n t /0 (x) for every element xeE. We say that a sequence x l > x2 , . . . , xn, . . . of elements of E is weakly convergent to the element x0 if it converges to x0 in the weak topology u(E, E') of the space E. The criterion for weak convergence of the sequence x 1 , x 2 , . . . , x"' . . . to the element x0 is that the equality limf (xn) = f (x0) nbe satisfied for every functional/ E E'. In contrast to weak convergence of elements (functionals), the convergence of a sequence of elements (func tionals) with respect to the norm of the space E (E') is called strong convergence. Strong convergence of elements (functionals) always implies weak convergence; however, the converse statement is not true, generally speaking, in the case of an infinitedimensional space. Strong and weak convergence of elements and functionals are equivalent in a finite dimensional Banach space. Strong and weak convergence of elements are also equivalent in the space 11 which is an infinitedimensional Banach space; however, the weak topology in an infinitedimensional normed linear space is always weaker than the initial topology (see [5]). Never theless, the following assertion is valid : a convex set T in a normed linear space E has the same closure both in the initial topology of the space E and in the weak topology u (E, E'). In particular, if the sequence {xn} converges weakly to x0, then there m exists a sequence of linear combinations { l: A. lm) x;} converging in the i= 1 norm to x0 • .
OCJ
4.
Properties of a sphere in a conjugate Banach space. A set S in a topological space E is called compact if from every covering of S by open sets of the space E a finite subcovering can be extracted.*)
*) A set which is compact in a separated (Hausdorff) topology is often called
bicompact in Russian mathematical literature.
CONJUGATE SPACES
29
For a metric space, this definition of a compact set is equivalent to the definition of a closed compact set given in no. 7, §2. Every closed sphere is compact in the weak* topology u(E', E) in the space E' conjugate to the normed linear space E. Every closed sphere in the space E', the conjugate to a separable normed space and equipped with the weak* topology u (E', E), is a compact metriz able space. Let T be a convex set in the linear system E. We say that the point xe Tis an extremalpoint of the set Tifthere is no open segment containing x in T. In the Euclidean space Rn, every boundary point of the unit sphere is extremal ; the unit sphere contains only two extremal points x(t) = 1 and x(t)=  1 in the space C(O, 1), and the unit spheres do not contain extremal points in the spaces c0 , L1 (0, 1 ). Every compact convex set in a locally convex linear topological space E is the closed convex hull of the set of its extremal points. (M. G. Krein, D. P. Milman) It follows from this that the unit sphere S of an infinitedimensional conjugate Banach space E' contains an infinite set of extremal points ; therefore, the spaces C(O, 1 ), c0, L1 (0, 1) are Banach spaces which are not isometric to any conjugate Banach spaces. 5. Factor space and orthogonal complements. Let M be a (closed !) subspace of the Banach space E. The collection of elements X = x + M, where x is a fixed element of E, is called a residue class with respect to the subspace M. The collection of all residue classes will be a linear system if the class X+ Y, constructed from the element x + y where x and y are elements of the classes X and Y, is understood to be the sum of the classes X and Y. Analogously, the class A.X is constructed from the element A.x where xeX. With this introduction of the operations of addition and multiplication by numbers, the collection of all residue classes is called the factor space E/M of the space E with respect to the subspace M. The norm II X II = inf ll x l l xeX
is introduced in the factor space E/M. The factor space E/M is a Banach space with respect to this norm. The collection of all continuous linear functionals on E which are zero
30
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
on M is denoted by Ml_ . The collection Ml_ is a weak* closed (that is, closed in the topology a(E', E)) subspace of the conjugate space E' and is called the orthogonal complement of the subspace M. Conversely, every weak* closed subspace M' c E' is the orthogonal complement of the subspace M c E which consists of all elements on which every functional from M' becomes zero. The relation f (x) = F (X) (x e X) establishes a onetoone correspondence between the functionals feMl_ and the continuous linear functionals F on E/M (Fe(E/M)'). In this correspondence, the space conjugate to the factor space E/M is isometric to the space Ml_. Every functional of E' , naturally, generates a continuous functional on M ; the functionals of Ml_ generate the null functional on M. Conversely, every continuous functional on M can be extended to all of E without increasing its norm. The space M', conjugate to the space M, is isometric to the space E'/M l_ . 6. Reflexive Banach spaces. Let E be a normed linear space and E' the space conjugate to E. Since E' is a Banach space, it makes sense to speak about the space E" = (E')' conjugate to E'. Every element x0eE generates a continuous linear functional Fxo ( f ) on E which is defined by the equality Fx0 Cf ) =f (x0). Thus, a onetoone correspondence is established between the elements of the space E and some subset n (E) of the space E". This correspondence is an isometry between the spaces E and n (E) and is called the natural mapping of the space E into the space E". A Banach space E is called reflexive if it is isometric, with respect to the natural mapping, to its second conjugate space E". '
Since the correspondence which gives the isometry has a special form in this case (the element Fx( f)eE" corresponds to the element xeE), the existence of an isometry between the spaces E and E" still does not allow us to conclude that the space E is reflexive. Thus, socalled quasireflexive Banach spaces E exist for which E" = n (E)EBEn where En is ndimensional. Such spaces E are isomorphic to the spaces E" but are not reflexive. Examples of nonreflexive quasireflexive spaces E exist which are isometric to the spaces E".
LINEAR OPERATORS
31
Every subspace M of a reflexive (quasireflexive) space E is reflexive (quasireflexive) ; the factor space E/M is reflexive (quasireflexive). Criteria for the reflexivity of a Banach space 1 . In order that the space E be reflexive, it is necessary and sufficient that every continuous linear functional f(x), defined on E, attain its supremum on the unit sphere of the space E; that is, that an element x1 exists such that ll x111 = 1 andf (x1) = 11 f ll . 2. In order that the Banach space E be reflexive, it is necessary and sufficient that its unit sphere S(e, I) be compact in the weak topology
u (E, E '). 3. In order that the Banach space E be reflexive, it is necessary and
sufficient that its unit sphere be a closed set in an arbitrary normed topology comparable (see [5]) with the initial topology of the space. A Banach space E is called uniformly convex if for every e > 0 there exists a (J > O such that ll x ll = l , II Y II = l, and x + y > 1  J implies 2 ll x y l < e. 4. Every uniformly convex Banach space is reflexive.
The class of uniformly convex Banach spaces does not coincide with the class of all reflexive Banach spaces : an example of a reflexive Banach space which is not uniformly convex can be given. The spaces lP , LP (O, 1 ) where p > 1 are uniformly convex and therefore reflexive. Every finitedimensional Banach space is reflexive. All the remaining spaces considered in no. 5, § 2 are nonreflexive Banach spaces. § 5. Linear operators
I . Bounded linear operators. Let E and F be two linear systems. We say that an operator A with values in F (an operator acting from D to F) is defined on a set D c. E if to every element X E D there corresponds ; an element y =AxE F. The set D is called the domain of the operator and i s denoted by D (A). The collection of all elements y of F, representable in the form y = Ax (xED (A)) is called the range of the operator A and is denoted by R (A). The operator squaring: A x (t ) = x 2 (t ) is an example of an operator in the space C (0, I ). The entire space C (O. I ) serves as the domain of /
32
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
this operator ; the collection of all nonnegative functions of C (0, 1) is its range. This same operator, considered on the space L2 (0, 1 ), will map it onto the collection of nonnegative functions of L1 (0, 1 ). The operator A is called linear if D(A) is a linear manifold in E and
for arbitrary elements x 1 , x2 eD(A). The following serve as examples of linear operators in an arbitrary linear system E: the identity operator I, setting in correspondence to each element of E this same element : lx= x ; the operator of similitude : Ax = A.y (xe E, A. is a fixed number). In a finitedimensional space E" ' the linear transformations of the space serve as examples of linear operators. Such operators can be given by means of a square matrix (a;k) ; if x= g I > � 2 , , �n} and y = {1J 1 ,1] 2 , ,1Jn} , then • • •
• • •
n
The integral operators
1
J
y (t) = Ax (t) = K (t, s) x (s) ds 0 are the analogues of such operators in function spaces. If, for example, the kernel K(t, s ) is continuous, then this linear operator is defined on the entire space C (0, 1) and maps it onto some part of the space C (0, 1 ). The linear operator of differentiation, Ax(t) = x' (t ), defined on the continuously differentiable functions, D(A) = c< 1 ) (0, 1 ), can be considered in the space C(O, 1 ) . The entire space C (0, 1) will be the range of this operator. If this operator is extended to the collection of absolutely continuous functions, then its range will be the space L 1 (0, 1 ). In the theory of generalized functions (see chapter VIII) the operator of differentiation is extended to the entire space C (0, 1 ), and, in this connection, it maps the space C(O, 1 ) onto some space of generalized functions. Now let E and F be two normed linear spaces. An operator A is said to be continuous at the point x0eD(A) if Axn +Ax0 whenever Xn +x0 (xn ED(A)). If the operator A is defined and continuous at every
LINEAR OPERATORS
33
point of the space E it is called simply a continuous operatorfrom E into F. A linear operator, defined on E, is called bounded if where C does not depend on the element xeE. In order that a linear operator, acting from E into F, be continuous, it is necessary and sufficient that it be bounded. The smallest of the numbers C in the last inequality is called the norm of the operator A and is denoted as follows : I I A IIE ....F · If F coincides with E, then it is written simply as I I A II · It follows from the definition that II Ax ii F = sup II AxiiF . II AII E+F = sup x e E llxiiE J l x Ji e = l x ,O O
2. Examples of bounded linear operators. Integral operators. Interpo lation theorems. 1) Operators in finitedimensional spaces. Every linear operator A given by a matrix (a1k ) in a Banach space En is bounded. Its norm depends on the norm which is introduced in the space. If the norm ll x ll = max l �d i
is introduced, then If
I I A II = max L la,kl · 1 � i �n k= 1
ll x ll = then If the Euclidean norm
n
n
L =
i 1
1�,1 , n
I I A II = max L la1kl · 1 :!fk:!fn i= 1 /
is introduced, then I I A ll = J�1 where J.lt is the largest eigenvalue of the matrix A A * here (A* = (ak1)). If the matrix (a1k ) is symmetric, then )Jt1 = A.1, where A.1 is the largest e ige nva lue of the matrix A.
/
34
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
2) Integral operators. If a linear integral operator with a continuous kernel K(t, s) is considered as an operator from C(O, 1) into C(O, I), then it is bounded and 1 JJA JJc .... c = max
O :!f t :!f l
f IK (t, s)J ds . 0
This operator as a bounded operator from L 1 (0, 1) into L1 (0, I) has the norm 1 JJAJJL, .... L, = max
O � s :!f l
f JK (t, s) J dt. 0
If the operator A is considered as an operator from LP (O, 1) into LP(O, I), then the inequality 1 1 � 1  P� P (t, t JK I I A I I Lv >Lv � ��: JK 0 1 1 0 0 is valid for its norm. This last assertion follows from a general fact : if the linear operator A is simultaneously a bounded operator from C(O, I) into C(O, 1) (or from M(O, I ) into M(O, I)) andfrom L 1 (0, I) into L 1 (0, 1), then it is bounded as an operator .from LP(O, I) into LP(O, I) and 1 1 JJ A JJ Lp>Lp � I JAIIM>PM I I A l l£. >Lt .
�
f (t, s)J ds � t�:: f
s)l d �
I
Analogously, if a linear operator is bounded as an operator from Lp, (0, 1) into LP� (0, I) and as an operator from LP2 (0, I) into LP; (0, 1 , then it is bounded as an operator from LP(O, I ) into LP, (O, I ) where I I  �t It I I �t It + and , = , + and It is an arbitary number in [O, I]. = P P1 Pz P P1 Pz Furthermore
)



,
(M. Riesz) The last assertions are called interpolation theorems in the theory of operators and allow wide generalizations to other classes of Banach spaces.
LINEAR OPERATORS
35
3) Operators of potential type. The class of integral operators with discontinuous kernels of the form K (t, s )
1 =
i t  sl
;.
,
where 0 < A < 1, are called operators ofpotential type. 1 If 1  A. > , then the operator with the kernel K(t, s) can be considered p as a bounded operator from the space LP(O, 1 ) into the space C(O, 1 ) 1 If 1 A <  , then the operator will not act from LP(O, 1 ) into C(O, 1 ) , p and it can be considered as a bounded operator acting from LP (0, 1 ) 1 1 into Lp, (0, 1 ) where  =  ( 1  A) . .
P1
P
Analogous facts are valid for operators of potential type defined for functions on a bounded ndimensional region G. In this connection, we understand the quantity i t  si as the distance between the points t and s; the number A should be included in the interval (0, n). The operator A 1 A. 1 acts from LP(G) into C(G) if 1   > , and into Lp 1 (G) if 1   <  where n p n p 1 1 A  =   ( 1  ) . It is bounded in both cases. n P1 P All these assertions are also valid for operators with kernels of the A (t, s) form K(t, s ) = where A (t, s) is a continuous kernel. I t  s l" 4) Singular integral operators. In the previous example the kernel of the integral operator had a summable singularity. The kernels of singular 1 . The integral operators have nonsummable singularities of the form itsi Hilbert transformation x (s) Ax = ds , J ts 
00
I
/
 oo
where the integral is understood in the sense of a principal value, provides the simplest example of such an operator. It can be shown that this operator is a bounded operator acting fro m 1.1, (  oo , oo ) into LP(  oo , oo ) for I
36
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
A manydimensional (ndimensional) singular operator is an integral operator with a kernel of the form K (t,
s) = Qit(t,t s!" s) '
where Q (t, 1"), as a function of its second argument r , is homogeneous of degree zero and has an integral equal to zero on the unit sphere S. If the integral of I Q (t, 't" ) I P' on the sphere S is uniformly bounded with respect to t, then the singular operator is a bounded operator from
G ;= )
1 . (CalderonZygmund) LP (Rn ) into LP (Rn) for 1 < p < oo + 5) Hilbert and Hardy integral operators. The operator ,
s) x ( Ax = J t + s ds 00
0
is bounded from LP (O, oo ) into LP (O, oo ) for 1 < p < oo , and its norm is 1C equal to . " . (Hilbert inequality). sm �p The operator
t
Ax = : I x (s)ds 0
is bounded from LP (0, oo ) into LP (0, oo ) for p > 1 , and its norm is equal to p . (Hardy inequality) p1 6) Differential operators. Linear differential operators, considered as operators taking a space into itself, are not bounded as a rule. In parti cular, the derivative operator is not bounded in the space C (O , 1 ) ; if it is considered as an operator from c< l ) (O, 1) into C (O, 1 ) , then it is bounded and its norm is equal to 1 . Analogously, a linear differential operator of order l with continuous coefficients can be considered as a bounded operator from c<0 (0, 1) into C (0, 1 . For the study of linear partial differential operators, generally, either a Holder space (classical approach) or the spaces W! are used. Thus, an
)
LINEAR OPERATORS
37
elliptic operator of second order n n OX o 2x + L bi (t)  + c (t) x , Ax =  L aii (t) oti otiotj i = l i, j = l defined in an ndimensional region G, is considered as a bounded operator from the space W� (G) into the space L 2 (G).
3 . Convergence of a sequence of operators. Let {An} be a sequence of bounded linear operators acting from a normed linear space E into a normed linear space F. The sequence {An} is said to be convergent in the norm to the bounded linear operator A0 from E into F if limii A0  An11E ....F. = 0. n > oo The sequence {An} is said to be strongly convergent to the operator A0 if limll A0x  AnxiiF = 0 for all x E E. n > oo The sequence {An} is said to be weakly convergent to the operator A0 if the sequence {Anx} is weakly convergent to A0x for all xEE. Convergence in the norm implies strong convergence, and strong convergence implies weak convergence. The converse assertions are, generally speaking, not true. If the sequence {An} converges strongly to A0 and the sequence of norms of the operators An is bounded : II AniiE .... F � M (n = 1 , 2, . . . ), then the operator A0 is also a bounded linear operator and n > oo In the case when E is a Banach space, the last assertion is considerably strengthened : if a sequence of bounded linear operators An acting from a Banach space E into a normed linear space F converges strongly to the operator A0, then the sequence of norms of the operators is bounded and, hence, the operator A0 is also bounded. In order for a sequence of bounded linear operators, acting from a Banach space E into a Banach space F, to converge strongly to sonie bounded linear operator, it is necessary and sufficient that : I) the sequence of norms of the operators An be bounded; 2) the sequence {Anx'} be convergent for every element x' of some everywhere dense set D e:. E. The last theorem has many applications to problems connected with convergence and summability of series and integrals, the convergence of
38
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
interpolation processes, of processes of mechanical quadratures, and so on (see [19]). The facts mentioned are valid, of course, for the case when the space F coincides with the collection of all real or complex numbers. In this case the word "operator" is replaced by "functional" and "strong convergence of operators" is replaced by "weak convergence of functionals" in all formulations.
4. Inverse operators. Let a linear operator A map a linear system E
into a linear system F. If the operator A has the property that Ax = (J only for x = 8, then for every y in the range R (A) of the operator A there corresponds only one element x for which y = Ax (the solution of the equation y = Ax is unique). This correspondence can be considered as an operator B defined on R (A) with values filling E . The operator B is linear. By definition, BAx=x ; therefore the operator B is called a left inverse of A . If R (A) =F, that is, the operator A establishes a onetoone corre spondence between E and F, then the operator B, defined on all of F, is called simply the inverse operator of A, and is denoted by A t . By definition, A  1 Ax = x (xeE) and AA  1 y = y (yeF). One of the profound facts of the theory of Banach spaces is the follow ing assertion. If a bounded linear operator A, mapping a Banach space E onto a Banach space F, has an inverse A  l , then the operator A  t is bounded. (s. Banach) . This theorem ceases to be valid if the completeness of one of the spaces E or F is relinquished. It is generalized to some classes of linear topo logical spaces, in particular, to complete metric spaces. The theorem about the inverse operator means, in other words, that from the existence and uniqueness of the solution of the equation
Ax = y for every y in F it follows that the solution x = A  ty depends in a con tinuous manner on the right hand side y. If a bounded linear operator A from a Banach space E to a Banach space F has an inverse, then bounded linear operators close to it also
39
LINEAR OPERATORS
have inverses : if II B  A ll E
+
F
< II A  t1il
then the operator B has an inverse B  t.
F + e
'
5. Space of operators. Ring of operators. The operations of addition and multiplication by a number are introduced in a natural manner for linear operators mapping a linear system E into a linear system F. By definition, A = ()(tA t + ()(2 A 2 is the operator for which (x E E) .
If E and F are normed, then all the bounded linear operators from E into F form a linear system L (E, F) which can be normed by means of the norm II A IIE F · If F is complete, then L (E, F) is a Banach space. If we consider operators taking a space E into itself, then an operation of multiplication can also be introduced for them : by definition, A = A t A 2 if .....
Multiplication, generally speaking, is noncommutative : it is possible that A t A 2 # A 2 A t · If A tA 2 = A 2 A t , then we say that the operators A t and A 2 commute. If A t , A 2 e L (E, E), then A eL(E, E) where If an operation of multiplication x · y is introduced in a normed linear space so that the space becomes a ring (more precisely, an algebra), and
ll x · Y ll :::; l l x ll II Y II , then it is called a normed ring (algebra) (see ch. VI). The space L (E, E) is a normed ring. The ring has an identity, thy identity operator I. If E is a Banach space, then the collection of operators having inverses forms an open set in this ring. /
6. Resolvent of a bounded linear operator. Spectrum . Let E be a complex Banach space and  A be a bounded li near operator acting on it.
40
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
The complex number Jl is called a regular point of the operator A if the operator A  )J has an inverse (A  Jll)  1 • In the opposite case, Jl is called a point of the spectrum of the operator If Jl is a regular point of A, then the bounded linear operator (A  Jll)  1 is called a resolvent and is denoted by R;.. The regular points form an open set in the complex plane ; the spectrum is closed. All points lying outside the circle of radius IIAII with center at the origin are regular. All points of the spectrum are in the disk l il l � II A ll . For il l > I I A ll , the series expansion for the resolvent
A.
I
,
R "'
=
_
1 Jl
(J
A A� An · · . + Jl + + · + .. + 2 Jl Jln
)
is valid, where the series converges in the operator norm. The radius of the smallest disk with center at the origin containing the spectrum of the operator A is called the spectral radius rA of the operator A. The formula (I. M. Gelfand)
rA = lim n +
oo
j u An II
is valid ; moreover, the limit always exists. It follows from the preceding that rA � IIA 11 . Moreover, rA � ::; II An II · On the basis of the Cauchy criterion, the series for the resolvent will converge if rA < I ill and diverge if rA > ill . In particular, the series
I
(I  A)  1 =  R 1 = I + A + A 2 + · · · + An + · · ·
I
converges if rA < and diverges if rA > I . The spectrum of an arbitrary bounded operator is a nonempty set. If
rA = lim j ii An ll = 0 , n +
oo
then the spectrum consists of one point, Jl = 0. The Volterra integral operator t
Ax = J K(t,s)x(s)ds, 0
where K(t, s) is a function continuous in the triangle O� s � t � 1 , is an example of such an operator. If this operator is considered as a bounded
41
LINEAR OPERATORS
operator in the space C(O, 1) or L2 (0, 1), then its spectrum consists of the point Jl = O. This corresponds to the statement that the equation
x (t)  J1
t
f K (t, s) x (s) ds = y (t) 0
has a unique solution for an arbitrary right hand side and arbitrary The Hilbert identity
Jl.
RA  R,. = (Jl  Jl) RAR,.
is valid for any two regular points Jl and 11 · The limit in the sense of the operator norm RA  R . 2 I1m . ,  RA p,+ A Jl  J1
is called, naturally, the derivative with respect to Jl : d RA
dJl
= R 2A .
An operatorvalued function of Jl is called analytic at Jl 0 if it can be expanded in a neighborhood of Jl 0 in a series of positive integral powers of (JlJl 0 ) which converges with respect to the operator norm. The resolvent is an operatorvalued function of Jl which is analytic in the region consisting of regular points of the operator A. For xeE, the function RAx, whose domain consists of regular points of A, is an analytic function with values in E. Let Jl 0 be a pole of the analytic function RA .Then any element RAx has an expansion in a Laurent series
e0
em _ 1 . . . RAx = + + + + Jl  ilo (Jl  ilo)m (Jl �o)m 1 + fo + /1 (Jl  ilo ) + · · · + fn (Jl  llot + e1
·
·
·
The element e0 = Jle0 (x) satisfies the equation and is called an eigenvector of
the operator A
corresponding to the
42
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
eigenvalue A.0• The elements e 1 , . . . , em _ 1 satisfy the relations Ae 1 = A.0e 1 + e0 , Ae2 = A.0e2 + e1 , . . . , Aem _ 1 = A.0em_ 1 + em _ 2 and are called associated vectors to the eigenvector e0. A subspace L of the space E is called invariant with respect to the operator A if xeL implies that AxeL. The finitedimensional subspace which consists of all possible linear combinations of the elements e0, e1, . . . , em _ 1 is invariant with respect to the operator A and is called a radical subspace. The elements ek form a basis for it in which the matrix of the operator A has the form of a Jordan cell (see SMB "Higher Algebra" ch. 2, § 1 , no. 1 1). If A.0 is a simple pole of the resolvent R;. (m = 1 ) then to the eigenvalue ,
A.0 there correspond only eigenvectors of the operator A. Associated vectors are absent. By means of the resolvent, the concept of a function of a bounded operator is introduced. If f (A.) is a function, analytic in a region con taining the spectrum of the operator A, then we define the function f(A) to be the operator
f (A) =  � Jf (A.) R;. dA. , 2nz 'j T
where the contour r contains in its interior the spectrum of the operator A. This integral is a generalization of the Cauchy integral. It does not depend on the choice of the contour r. 7. Adjoint operator. Let E and F be normed linear spaces and let A be a bounded linear operator acting from E into F. If g ( y) is a continuous linear functional on F (g e F' ), then the functional
f (x) = g (A x)
will be a continuous linear functional on E ( f e E'), where I I! liE· � Ji g II F' II A II E>F .
Thus, to every functional g E F' there corresponds a functional fEE' ; i.e. an operator A'g f is defined . This operator A' is called the adjoint of the operator A. =
LINEAR OPERATORS
43
The adjoint operator is a bounded linear operator ; moreover, II A' II = II A II . If A, BeL (E, F) , then (A.A)' = A.A' and (A + B)' = A' + B'. If A,BeL(E, E ) , then (AB)' = B' A'. If an integral operator with continuous kernel K( t, s) is considered, for example, as a bounded operator from LP(O, 1) into LP(O, 1), then the integral operator with kernel K ' ( t, s) =K(s, t ) , that is, the operator 1
A'x =
f K (s, t) x (s) ds , 0
(
)
considered as an operator from Lq(O, 1) into Lq (O, 1) � + � = 1 . is the p q adjoint to it. If E and F are Banach spaces, then the existence of the operator (A' t 1 where (A' t 1 = (A 1 ) ' is a necessary and sufficient condition for the existence of the inverse operator A  1 . It follows from the last assertion, applied to the operator A  AI and its adjoint A'  AI, that the spectra of the operators A and A' coincide. A more important relation between the properties of an operator and its adjoint is studied in no. 9. 8 . Completely continuous operators. Let E and F be Banach spaces. A linear operator, acting from E into F, is called completely continuous if it maps every bounded set of the space E onto a (relatively) compact set of the space F. Complete continuity of a linear operator implies continuity. The converse, generally speaking, is not valid. For example, the identity operator I is continuous, but in the case of an infinitedimensional space E it is not completely continuous. It suffices for the complete continuity of a linear operator that it map the unit sphere of the space E into a compact set of the space F. The range of a completely continuous operator is separable. A completely continuous operator maps every weakly convergent sequence of elements into a sequence which is convergent with respect to the norm. The limit with respect to the norm of a sequence of completely contin uous operators is again a com pletely continuous operator. A strong, and hence a weak, li mit of a seq uence of completely continuous operators
44
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
need not be a completely continuous operator. Let, for example, E be the Banach space /1• The projectors PN, which set in correspondence to every x = g1, �2, , �n • · · · } the element PNx = g 1, . . . , �N• 0, 0, . . }, are completely continuous, but their strong limit is the identity operator which is not completely continuous. A linear combination of completely continuous operators is a com pletely continuous operator. The product of a completely continuous operator with a bounded operator is a completely continuous operator. The set of all completely continuous operators of L (E, E ) forms a closed ideal in the normed ring L (E, E ) . The adjoint operator of a completely continuous operator is completely continuous. A onedimensional linear operator of the form .
• • •
where Y1 is a fixed element of F and /1 (x) is a fixed functional from E', is the simplest example of a completely continuous operator. The onedimensional operator is abbreviated as follows : A = /1 ® y 1 . An arbitraryfinitedimensionallinear operator has the more general form A=
m
L1 /; ® Y; ,
i=
where y1eF and f1eE ' . By definition, Ax
m =
(x) y; . 1 / 1
"[.
i=
A finitedimensional operator is completely continuous. It is not known whether every completely continuous operator is representable as a limit with respect to the norm of finitedimensional linear operators. A linear operator is called atomic if it is representable in the form 00
Ax = "[. J1 (x) y1 , i= 1
L 00
where y1 e F, f1 e E' and
I= 1
llftli E' IIY;II F < co .
LINEAR OPERATORS
45
Atomic operators constitute an important subclass of the class of completely continuous operators. Let A be a completely continuous operator, defined on and acting into a Banach space E. If it is finitedimensional, then its spectrum coincides with the finite set of its eigenvalues. In the general case, its spectrum consists of no more than a denumerable number of points. In an infinite dimensional space, the point always is a point of the spectrum of a completely continuous operator; moreover, it is the only possible limit point for the set of the other points of the spectrum. Thus, the spectrum of a completely continuous operator consists of a finite number of points or of the point and a sequence An �o. All the points An n 2, . . . ) are poles of the resolvent R;. and, hence, eigenvalues of the operator A. To every eigenvalue Jln =ft there corresponds only a finite number of linearly independent eigenvectors and associated vectors. In a finitedimensional space, the eigenvectors, and associated vectors form a basis. In an infinitedimensional space, the picture can be con siderably more complicated for an arbitrary completely continuous operator. There are completely continuous operators which do not have eigenvalues ; their spectrum consists of the single point The Volterra integral operator (see §4, no. 6) can serve as an example of such an oper ators. In connection with this, we call completely continuous opera tors, not having eigenvalues, Volterric. In recent years much effort was applied to the discovery of conditions for which the eigenvectors and associated vectors of a completely continuous operator form a complete system in the space E, that is, a system whose closed linear hull coincides with E. Significant results were obtained for the Hilbert space case (see ch. 2, no. 5). This fact is valid in an arbitrary Banach space : although a completely continuous operator may or may not have eigenvectors, it has without fail proper invariant subspaces, that is, invariant subspaces co inciding neither with E nor with {8}. Numerous examples of completely continuous operators are given by integral operators. If the kernel of an integral operator
Jl0 = 0
( =I,
A.0=0
0
A.0 =0.
1
Ax = f K(t,s)x(s)ds 0 is continuous, then it will generate
a
completely continuous operator
46
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
from C (0, 1 ) into C(O, 1 ). If the kernel satisfies the weaker condition 1
1
I I IK (t, s)l q dt ds < oo 0
0
(q > 1 ) '
then the operator will be completely continuous as an operator from LP (O, 1 ) into LP (O, 1)
G � = 1 ). The condition mentioned is +
not
necessary. Examples exist of completely continuous integral operators acting from LP (0, 1) to LP (0, 1) (l < p < oo), whose kernels are not summable, as functions of two variables, to any power > 1 . The equation x  Ax = y is the analogue in the theory of operators of the Fredholm integral equa tion of the second kind. The basic facts of the Fredholm theory are valid for this equation in the case when the operator A is completely continuous. Consider the adjoint equation g  A 'g = f . If the initial equation has a solution for an arbitrary right hand side in F, then the adjoint equation has a solution for an arbitrary right hand side in E'; in this case the solutions are unique, that is, the homogeneous equations x  Ax = 8
'
g  A' g = 8
have only the trivial solutions x = 8, g = 8. The homogeneous equations have the same finite number of linearly inde pendent solutions, and the dimensions of corresponding radical subspaces coincide. If g 1, . . . , 9m is a maximal sytem of linearly independent solutions of the homogeneous equation g  A' g = 8, then the initial equation x Ax = y has a solution onlyfor those right handsidesfor which gk(y) = O (k = 1 , 2, . . . , m). The three Fredholm theorems are valid for equations of a more general form : Ux  Ax = y , where A is a completely continuous operator and U is a bounded operator
LINEAR OPERATORS
47
having an inverse u 1 • It turns out that these exhaust all linear equations with bounded operators for which the Fredholm theorems are valid.
9. Operators with an everywhere dense domain of definition. Linear equations. Integral operators are basic examples of bounded operators. Differential operators in natural norms, as a rule, are unbounded. Therefore, in the application of the theory of bounded operators to differential equations, these equations are reduced to integral equations. Usually this reduction is done by means of the Green's function or other resolvent kernels. The construction of such kernels is not trivial; therefore, recently the theory of unbounded operators has been developed in a form which allows direct application to the theory of differential equations. Let E be a Banach space and A be a linear operator defined on an everywhere dense set D (A) and acting from D (A) into a Banach space F. For such an operator, the concept of the adjoint operator can be intro duced. If g (y) is a bounded linear functional on F, then the functional g (Ax), defined on D (A), can be bounded or unbounded. If g (Ax) is bounded, then it can be extended in an unique manner to a continuous functional f (x) on the entire space E : as Xn +x, Xn E D (A), we set g(Ax) = limg(Axn )· If these conditions are satisfied, we say that the adjoi nt operator A' is defined on g : f (x) = A'g (x) .
It follows from the preceding that g ED(A') if l g (Ax) l � C ll x ll for xED(A), and for these x A'g (x)
=
g (Ax) .
The adj oint operator plays an essenti al role in the study of the questi on of the solvability of the equation Ax = y
(y E F, x E D (A)) .
(A)
Along with this equation the adjoint equation A 'g = f
(! E E'' g E D ( A '))
i� considered. The following assertions are valid for equations (A) and (A') :
(A ')
48
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
1 . In order that equation (A) be solvable for an everywhere dense set of right hand sides y from F (R (A ) =F) it is necessary and sufficient that the uniqueness theorem be valid for equation (A'), that is, A 'g = () implies g = (). 2. In order that equation (A') be solvable for an everywhere dense set of right hand sides ffrom E', it is necessary that the uniqueness theorem be valid for equation (A). 3. A necessary and sufficient condition for the solvability of equation (A') for every f EE is that the inequality ,
'
(x E D (A) ,
m > 0)
be satisfied. 4. A necessary condition for the solvability of equation (A) for an arbitrary right hand side yEF is that the inequality II A' g jj E' � m lig jjF'
(g E D ( A ) , '
m
> 0)
be satisfied. Properties 13 are valid when E and E are normed linear spaces ; property 4 is valid when F is a Banach space. The last inequality, generally speaking, is not sufficient for the solva bility of the equation (A), but the following assertion is valid : 5. IfD(A ') is everywhere dense in F' and the last inequality is satisfied, then the equation (A) has a "weak solution" for an arbitrary y0 EF in the sense that an element x 0 exists satisfying the identity for arbitrary g E D (A'). If the weak solution x0 belongs to the domain of definition of the operator A , then it is a proper solution. In particular, if the operator A is boun ded and defin ed on the entire space, then the solvability of the equation (A) for arbitrary right hand sides follows from the inequality II A'g iiE' � mllgrll · If the space E is reflexive, then uniqueness of the weak solution follows from the uniqueness theorem for equation (A). 10. Closed unbounded operators. Let E be a Banach space and A a linear operator defined on an everywhere dense set D(A) acting from D (A) into a Banach space F. An operator A 1 is called an extension
49
LINEAR OPERATORS
of the operator A provided that D (A 1) ::;) D (A) an d A 1 x = Ax for xED(A). The restriction of the operator A to the set D c D (A) is the operator A 2 such that D (A2) =D and A2x = Ax (xED). If A is bounded, then it can be extended by continuity to a bounded linear operator A defined on the entire space E by setting Ax = lim Axn n +oo An operator A (possibly unbounded) is called closed if x0E D (A) and y0 = Ax0 follow from Xn +X0 (xn E D (A)) and Axn +Y o· A bounded operator defin ed on the entire space is always closed. The fact that also, conversely, a closed linear operator defined on an entire Banach space is bounded, is very important. If an operator A is not closed, then it allows a closed extension if and only if y = (} follows from Xn +(} (xn E D A)) and Axn +y. If an operator A has a left inverse operator, l:>ounded on the closed set R(A), then it is closed. In particular, only closed operators can have bounded inverse operators. If we consider the operators as acting from D (A) c E into E, then only closed operators can have, for certain .A, resolvents R;. = (A  ).I)  t . The concepts of regular points and points of the spectrum are introduced for a closed operator as for a bounded operator. The spectrum of a closed operator can fill an arbitrary closed set in the complex plane. The investigation of closed operators is sometimes reduced to the study of bounded operators by the following scheme : a new norm ll x ll t = ll x iiE + II Ax iiF
(x E D (A))
is introduced in the domain D (A) of the closed operator A . With this norm, D(A) becomes a Banach space EA. The operator A is bounded as an operator from EA into F. Moreover, an arbitrary operator B, acting from D (A) into F and allowing closure, will be bounded as an operator from EA to F: II Bx ii F � C ll x ll t = C (ll x iiE + II Ax iiF) .
The adjoint operator A' of an operator A with everywhere dense domain of definition is always closed. If the domain D (A') is everywhere dense in F ' , then the initial operator A allows closure. The fol lowing assertions are equivalent for equations (A) and (A') with a dosed operator A :
50
FUNDAMENTAL CONCEPTS OF F UNCTIONAL ANALYSIS
1 . The right hand sides for which equation (A) is solvable form a closed subspace of the space F. 2. The right hand sides for which equation (A') is solvable form a closed subspace of the space E'. 3. If we denote the collection of all solutions of the equation A' g = 0 by N', then equation (A) is solvable for those and only those right hand sides y for which g (y) O for all g EN'. 4. If we denote the collection of all solutions of the equation Az 0 by N, then the equation (A') is solvable for those and only those ffor which f(z) = O for all zEN. In particular, if we succeed in obtaining the lower estimates =
=
JJAxJIF � m JJxJIE (x E D (A)) and I I A' g i i E· � m ll g llr (m > 0) for the closed operator and the operator adjoint to it, then the uniqueness of the solutions of equations (A) and (A'), for arbitrary right hand sides from F and E' respectively, follows. As remarked, problems from the theory of differential equations are one of the stimuli for the study of unbounded operators. Let Q be linear differential operator of order I with sufficiently smooth coefficients defined in a region G of ndimensional space. This operator can be con sidered as an operator acting in LP (G) whose domain D (Q ) LP (G) consists of all functions with partial derivatives of order I continuous in G. Let x (t)ED(Q) and z (t) be a finitary function in G ; that is, an infinitely differentiable function equal to zero in a neighborhood of the boundary of G. The identity c
f Qx (t) z (t) dt = f x (t) Q'z (t) dt G
G
is valid where, Q' is the adjoint differential operator. The identity is obtained by integration by parts ; the boundary terms vanish because of the fin itary nature of the function z (t ) .
(� +: 1}
The fin itary functions form an everywhere dense set in L q = the operator Q', adjoint to Q, is defin ed on it ; therefore the differential operator Q allows the closure Q. The domain of defin ition of the operator Q can now contain functions nondifferentiable in the classic sense (functions with generalized derivatives, and so on).
51
SPACES WITH A BASIS
The solutions of the equation Qx= y belonging to the domain of definition of the closure Q of the differential operator Q in various function spaces are called generalized solutions.
1 1 . Remark on complex spaces. Let E be a complex normed linear
space. Sometimes it is more convenient to introduce the operation of multiplication of a linear functional f(x) by a number A not as indicated in §4, no. 1 but in the following manner : f1 = Af means that
f1 (X) = Aj (X) · The collection of all continuous linear functionals with the operation of multiplication by a number introduced in this way is denoted by E * and is also called the conjugate space of E. All concepts introduced for the space E ' are introduced analogously for the space E*. All facts valid in the space E' are also valid for the space E * ; changes are made only for some formulations : 1 . The operator adjoint to the operator A, considered as an operator from F * to E*, is denoted by A * . Then '
(AA) * = ).A * .
2. For an integral operator A with kernel K(t, s), the adjoint operator
A * has kernel K(s, t). 3 . If the number A belongs to the spectrum of the operator A, then the number ). belongs to the spectrum of A *, and conversely. §
6. Spaces with a basis
I.
Completeness and minimality of a system of elements. A system : ed of elements e 1 , e2, , en , . . . is called complete in a Banach space E
if
•
•
•
the linear hull of this system of elements is dense in E. Obviously, a com plete system of elements can exist only in a separable space. In order for the system { ed to be complete in E, it is necessary and sufficient that th ere
is no linear functional fEE' different from zero and equal to ::cro on all elements ek (k = I , 2, . . . ) (orthogonal to all ek). The system of elements { ek} is called minimal if no element of this �y�tcm belongs to the closed li near hull
of
the remaining elements.
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
52
In order that the system { ek} be minimal, it is necessary and sufficient that a system of linear functionals exist forming with the given system a biorthogonal system; that is, a system { fd E' (k = 1 , 2, . . ) such that J; (e ;) (J ii*) . If the system { ek} is complete and minimal, then the system of functionals {he} is defined in a unique manner. c
.
=
In every separable Banach space there is a complete minimal system. Moreover, we can construct a complete minimal system {ek} such that the functionals .h corresponding to it form a total set ; that is, fk (x) = (xEE, k= 1, 2, ... ) implies that x=O.
0
2. Concept of a basis. A system of elements {ed forms a basis of the space E if every element xEE is representable uniquely in the form of a convergent series
Every basis is a complete minimal system. However, a complete minimal system may not be a basis in the space. For example, the trigonometric system e0(t) = L ez n  1 (t)= sin nt, e 2 n(t)= cos nt(n = 1 , 2, . . . ) is a complete minimal system in the space C [  n, n] but it does not form a basis in it.
Examples of bases 1) In the space L2 [a, b] as well as in an arbitrary separable Hilbert
space H (see ch. II, § 1), every complete orthogonal system of elements forms a basis. Thus the trigonometric system of functions forms a basis in L2 [  n, n] . We can construct nonorthogonal bases in a Hilbert space. For example, if {e;} is a complete orthonormal system in a Hilbert space H, then the system of elements
k
gk = L P;e; i: 1
(k = 1 , 2, . . . )
forms a basis in H if the numbers P ; satisfy the conditions
(n = 1, 2, . . ) . .
*) OtJ
=
0 if i # j and ou
=
1.
SPACES WITH A BASIS
53
The system of functionals forming a biorthogonal system with {g d is given by a system of elements of H:
ek + 1 fk =  ek Pk + 1 Pk 2) In the coordinate spaces c0 and IP (p � 1 ) , the system of unit vectors ek= {O, . . . , 0, 1, 0, . . . } forms a basis. This system does not form a basis in the space c and is not even complete since the element e0 = {1, 1, . . . } 1
1
•
does not belong to the closure of the linear hull of the elements ek(k= 1, 2, . . . ). However, the system e0, et. e2 , forms a basis in the space c. 3) We can construct a basis in the space ofcontinuous functions C[O, 1 J in the following manner: let {r;} (i = 0, 1, 2, . . . ) be a sequence of numbers dense in [0, 1], where r0 =0, r1 = 1, r( # r1 for i#j. We set e0(t) = 1 and e1 (t) = t. Then ek(t) is defined inductively. Let the functions e1(t) be defined for i < k and the segment [0, 1] be divided by the points r2 , , rk 1 into k  1 intervals. Let rk belong to one of these intervals : r.1 < rk < r.2, s 1
• • •
0 � t < t,
1' 1'
t
0,
x�k) (t) =

2;
'
_
2k  1 2k  2 n) , k 1 = < 2 t 2, , . . . , � ( 1 2n+ 1 2n+ 2k 2k  1 < t � n+1 ' n + 1 2 2 for the remaining values of t .
0 k The functions x� ) (t) are arranged in a simple sequence {e;(t)} ( i = I , 2, . . . ) in increasing order of the index and for the same in in ncasing order of k. The system {e1 (t }} is orthogonal and forms a basis i n any space L P [0, 1] ( p � I ). Moreover, it forms a basis in an arbitrary �cparable Orlicz space on [0, I ]. n,
n
54
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
At present there is no solution of the Schauder problem : does every separable space have a basis ? 3. Criteria for bases. Everywhere in this subsection {ei} ( i = 1 , 2, . . . )
denotes a complete minimal system in a Banach space E, and { f;} the system of functionals, forming a biorthogonal system with { e;} . We define a bounded linear operator on E: n Snx = L /; (x) e; . i= 1
The operator Sn is a projector : s; = Sn It projects the entire space onto the ndimensional space Ln spanned by the elements e t . . . . , en . In order for the system {e;} to form a basis, it is necessary and sufficient that the operators sn be uniformly bounded, that is, the inequality n (x E E) II Snx ll = II L f; (x) e; ll � M ll x ll i= 1
be satisfied, where M is a constant. If the system {e;} does not form a basis, then an element x can be found on which II Sn x ll � M for all n = 1 , 2, . . . but for which the series 00
L /; (x) e; n= 1 diverges. If the last series converges for an arbitrary xEE, then the system { e;} is a basis. Moreover, if this series converges weakly for arbitrary xEE, then the system { e;} is a basis. The last assertion sometimes is formulated in the following way: every weak basis is a strong basis. n If the closed linear hull of the elements en , en+ 1 , is denoted by L and the unit sphere in the subspace Ln is denoted by un , then in order for the system {e;} to be a basis, it is necessary and sufficient that a positive constant IX exist such that • • .
e (um r) �
IX ,
where (! is the distance between (Jn and r. If { e ;} is a basis, then the system { .t;} is a basis in its closed linear hull which may not coincide with E'. If E is reflexive, then this hull does coincide with E', and { fi} is a basis in E'. If { f;} is a basis in the conjugate space E', then { e;} is a basis in the space E.
SPACES WITH A BASIS
4.
55
Unconditional bases. The system { e;} is called an unconditional
basis in the space E if it remains basis for an arbitrary rearrangement a
of its elements. The following is an equivalent definition : the basis { e;} is called unconditional if the series 00
L !; (x) f (e;)
i= 1
is absolutely convergent for arbitrary xEE and fEE'. In order for the basis to be unconditional, it is necessary and sufficient that projectors of the form
k
L fn, (x) en,
i= 1
be uniformly bounded for arbitrary finite collections of numbers (nt. . . . , nk) (n; =tni for i=tj). If the unit sphere in the linear hull of the elements of the basis en , , . . . , enk is denoted by sn, , n2 . .. . . nk and the closed linear hull of all the n n remaining elements of the basis is denoted by L ,, .. . , \ then a necessary and sufficient condition for the basis to be unconditional is that a constant fl > exist such that
0
for all finite collections (n 1 , . . , nk)· .
Let U be a bounded linear operator acting in the space E and having a hounded inverse. If the system { e ;} is a basis, then the system { Ue ;} is a basis. If { e;} is an unconditional basis, then { Ue;} is an unconditional basis.
In a Hilbert space H, every orthogonal basis is unconditional. It
can
be shown that an arbitrary unconditional basis in a Hilbert space is rcrresentable in the form { ue;} where {e; } is an orthogonal normalized h:1 �is. Such bases are called Riesz bases. We can characterize them by the followi ng properties : positive numbers m and exist such that
M
m
fo r
l (x, e;) l 2 � ll xll 2 � M L l (x, e;) l 2 L i=l 00
00
i= l
ar bi trary x E H. The syst e m of unit vectors { ed in the spaces c0 and IP (p > I ) forms an u n co nditi ona l basis. The system of Haar functions (see no. 2) forms
56
FUNDAMENTAL CONCEPTS OF FUNCTIONAL ANALYSIS
an unconditional basis in all the spaces LP [0, 1] with p > 1. Unconditional bases do not exist in the spaces C[O, 1] and L [O, 1]. The trigonometric system of functions is a basis in the spaces LP [  n, nJ (p > 1) but is not unconditional. If the system {e;} is an unconditional basis in E, then the system of functionals { .t;} forming a biorthogonal system with {e;} is an uncon ditional basis in E' provided the space E' is separable. 5. Stability of a basis.
Let the system {e ;} form a basis in the space E and let {h;} be some system of elements of E. The question is : for what conditions will the system {e; + h;} also be a basis in E? If { e;} is a basis (unconditional basis) and the elements h; are "sufficiently small" in the sense that 00
11!, 11 11 h;11 < 1 , L: i 1 =
then the system {e; + h ;} forms a basis (unconditional basis) in E. An important corollary follows from this last assertion : if the space E has a basis (unconditional basis) and {
n,
" L.,
c(i)
' •
A basis of polynomials exists, for example, in the space C[O, 1].
CHAPTER II
LINEAR OPERATORS IN HILBERT SPACE
§ 1. Abstract Hilbert space 1 . Concept of a Hilbert space.
Let H be a linear system, with multi
plication by complex numbers, in which to each pair of elements there
(x, y) having the properties : a) (x, y) = (y, x), in particular (x, x) is real ; b) (x1 + x2 , y) = (x1, y) + (x 2 , y) ; c) (A.x, y) = A.(x, y) for an arbitrary complex number A. ; d) (x, x) � O, and (x, x) = O only for x = O. The number (x, y) is called a scalar product. If H is a linear
is assigned a complex number
system
allowing only multiplication by real numbers, then the scalar product is assumed to be real. The following corollaries follow from axioms ad :
I) (x, y1 + Y2) = (x, Jl ) + (x, J2) ; 2) (x, A.y) = A (x, y);
3) The BuniakovskySchwarz inequality :
i (x, y) l ::::; j(x, x) j(y , y) . I n terms of the scalar product in H, a norm
ll x ll = � ran
be introduced, after which H becomes a normed linear space.
I f II is infinitedimensional and complete with respect to the norm i n t roduced, then it is called a (complex or real) Hilbert space. It is obvious
from the definition that every Hilbert space is a Banach space.
2.
Examples of Hilbert spaces.
vcdors
X = {et , e2, . . . , e.}
and y =
As is known, the scalar product of two
{'7t • '12• . . . , '7.}
in ndi mensional Euclid
58
LINEAR OPERATORS IN HILBERT SPACE
ean space is usually found by the formula
(x, y) = L �;1'1; , n
i= 1
and, in ndimensional unitary (complex Euclidean) space, by the formula n
(x, y) = L �;1'/; · i= 1
Analogously, a scalar product is introduced in a number of infinite dimensional spaces after which they become Hilbert spaces. 1 . The complex space /2 becomes a Hilbert space if we set 00
(x , y) = L �;1'/ ; · i= I
2. The space L2 space if we set
( , b) of complexvalued functions becomes a Hilbert a
b
(x, y) = J x (t) y (t) d t. a
3. The complex space L 2 , Q (a, b) of functions which are measurable on the segment [a, b] and have on this segment a modulus whose square
is summable with weight Hilbert space if we set
e(t) (e(t)>O almost everywhere) will be a b
(x , y) = J x (t) y (t)e(t)dt. a
4. The spaces W� (G) of S. L. Sobolev (see ch. 1 , § 2, no. 6) are Hilbert
spaces with respect to the scalar product
Here
(x, y) = J x (t) y (t) dt + J(lat i Dax (t) Day (t))dt. G
5. The space of functions
G
x (t), defined and measurable on the entire
ABSTRACT HILBERT SPACE
axis
(

59
) such that the limit
oo, oo
,
lim
T+ oo
_!___
2T
J lx (t)l 2 dt T
T
< oo
exists, will be a Hilbert space if we set
J (x, y) = lim _!_ x(t ) y(t)dt. 2T T
T+ oo
T
The spaces given in examples 14 are separable ; the space given in example 5 is nonseparable.
x
3. Orthogonality. Projection onto a subspace. Two elements and y of a Hilbert space are called orthogonal, x .ly, if (x, y = O. An element xEH is called orthogonal to the subset G c H, .lG, if (x, y) = O for an arbitrary y E G. Finally, two subsets G and r of the space H are called orthogonal, G .lr, if an arbitrary element xEG is orthogonal to an arbitrary element yET. Let L be a subspace of H. The collection of all elements orthogonal to L forms a subspace M, the socalled orthogonal complement to L. The subspaces L and M have only the single element (} in common. The following is one of the basic properties of a Hilbert space : If L is a (closed) subspace of the space H, then every x E H has a unique representation
x
)
x = y + z,
where yEL and z .lL. The element y is called the projection of x on L. It has the property that in comparison with other elements of L it has the least distance from
x.
Every element of H can be decomposed into a sum of an element of the �ubspace L and an element from its orthogonal complement M. In other words, H is decomposed into the orthogonal sum of L and M : H = L+ M. I n connection with this, we denote M = H· L. A highly useful corollary follows from the preceding : in order for a linear manifold L to be everywhere dense in the space H, it is necessary and sufficient that no element exists which is different from zero and orthogonal to all the elements of the set L.
60
LINEAR OPERATORS IN HILBERT SPACE
4. Linear functionals . It follows from the BuniakovskySchwarz in
equality that the linear functional f (x) = (x, u), for a fixed u e H, is bounded. It can be shown that this exhausts all bounded linear functionals on H; that is, a unique element ueH can be found such that f (x) = (x, u)
for every f (x)eH* ; moreover, II f II H = lluiiH· Thus the conjugate space H* is isometric to the space H itself. The conjugate space is considered here in the sense studied in chapter I, §5, no. 1 2. A Hilbert space is selfconjugate and, hence, reflexive. Every linear functional J, defined on L 2 (a, b), is representable in the form •
b
f (x) =
I x (t) u (t) dt , a
b
where u (t) eL 2 (a, b) and 11/ 11 =
(I 1 u (t)l 2 dty . a
Every linear functional J, defined on /2 , is representable in the form f (x) =
00
L �;c. , i= 1
Remark. It is sometimes convenient to represent linear functionals on a Hilbert space H not by a scalar product in the space H but by a scalar product in some other Hilbert space. Then the conjugate space H* to H will be realized by means of elements of another kind. It is con venient, for example, to represent a linear functional f(x) on the space Wi (G) by a scalar product in L2 (G), that is, in the form f (x) =
I x (t) u (t)dt . G
In this connection, u (t ) will be a generalized function (see ch. VIII). The collection of these functions forms the space W21 (Wi)*. =
61
ABSTRACT HILBERT sf>ACE
5. Weak convergence. In accordance with the general definition of weak convergence (see ch. I, § 4, no. 3), a sequence of elements {xn} c H is called weakly convergent to the element x0 (resp., weakly fundamental) if (xm y)+(x0 , y) (resp., (xn+p• y)  (xn, y)+0) for an arbitrary element yEH. The following properties of weak convergence follow from the reflexivity of Hilbert space : 1) If the sequence {xn} converges weakly to x0 and llxnll + llxo ll , then llxn  xo ll +0, that is, the sequence {xn} converges strongly to x0• 2) The space H is weakly complete; that is, if the sequence {xn} is weakly fundamental, then it converges weakly to some limit. 3) Bounded sets in the space H are weakly compact ; that is, from an arbitrary infinite set of elements of the space H which is bounded in the norm, a weakly convergent sequence can be chosen. 6. Orthonormal systems.
A system e1, e2 . . . , en, . . . of elements of a Hilbert space H is called orthonormal (or orthonormalized) if where D ij is the wellknown symbol which is equal to one for i =j and to zero for i =1j. The trigonometric system
jbr, , 1
is
j 1
rc
cos t ,
vfn sin t , vfn cos 2t , ; sin 2t, . . . 1
1
1
rc
an example of such a system in the real space L 2 (  rc, rc); the system n = 0, + 1, + 2, . . .
is
an example in the complex space L 2 (0, 1 ) . If an arbitrary system of linearly independent elements h1, h 2 , . . . , hn, . . . is given in H, then an orthonormal system can be easily obtained from i t by means of the socalled Schmidt process of orthonormalization. hl ; then we select c2 1 so that h 2  c 2 1 e 1 will be Namely, we set e1 = ll h tl! h 2  Cz t e t orthogonal to e1 and set e2 =  . We further select c3 and 2 e h ll 2  Cz t tll c3 1 so that h 3  c 3 e c 3 1 e1 will be orthogonal to e 2 2 2 and e1 , and set h3  c3 2 e2  c3 1 e1 = , and so on. ll h3  c3 2 e2  c3 1 e . ll t' 1 ·

··
·
62
LINEAR OPERATORS IN HILBERT SPACE
Example. If the system of powers, 1 , t, t 2 , , t n, . . . , is orthonormalized in the space L2 (  1, 1 ), then the system of normalized Legendre poly nomials is obtained. Orthonormalization of this system in the space L2 ,Q (  oo, oo) with weight e (t) =e 12 gives the system of Cebysev Hermite polynomials. If {e;} is an orthonormal system in H, then the numbers c; = (x, e;) are called the Fourier coefficients of the element x with respect to this n system. The linear combination ,L c;e; gives the best approximation i= 1 n of x in comparison with other combinations of the form ,L a;e; ; that is, • • •
i= 1
n n Dn = l!x  L C;e; ll � ll x  L IX;e; ll . i=1 i= 1 n In other words, ,L c;e; is the projection of the element x onto the i= 1
subspace Ln, spanned by the elements e 1, e 2 , enThe formula n J ; = ll x ll 2  L J c;l 2 i=1 is valid for Dn. If the element x belongs to the closed linear hull L of the elements e; (i = 1, 2, ), then X = "' ll xll 2 = ,L J c; J 2 • L... c .e. and i= 1 i= 1 • . .
,
...
00
00
I
I
If x ¢ L, then the element x' = ,L c;e; will be the projection of x onto L, where l! x ll 2 � 1J x ' ll 2 = L J c; \ 2 • 00
i= 1
00
The series ,L c; e; is called the Fourier series of the element x, and i= 1
the last inequality is called the Bessel inequality. We recall that if L coincides with all of H, that is, the linear combi nations of the elements e; are dense in H, then the system { e;} is called complete. A necessary and sufficient condition for completeness is the Parseval equality 1J xll 2 = L J c; l 2 oo
for an arbitrary element x e H.
i= 1
BOUNDED LINEAR OPERATORS IN A HILBERT SPACE
63
A complete orthonormal system { e;} is a basis for the Hilbert space. An orthonormal basis exists in each separable Hilbert space. All separable Hilbert spaces are isometric to the space /2 • § 2. Bounded linear operators in a Hilbert space
1. Bounded linear operators. Adjoint operators. Bilinear forms. For a bounded linear operator A , acting in a Hilbert space H, by definition II A II = sup II A x ll = sup llxli = 1
(x, x) = l
J(Ax , Ax) = sup
x* H x*O
J
(A x , A x) . (x, x)
If y is fixed in the scalar product (A x, y) then a linear functional of x is obtained : f (x) = (A x, y) , where I f (x) l = I( A x, y)l � II A ll ll x ll II Y II . This functional can be represented in the form (A x , y) = (x, u) ,
where UE H. The correspondence y+ u defines a bounded linear operator u=A*y. By definition (A x, y) = (x , A * y) .
The operator A* is called the adjoint operator to A . This definition agrees with the definition in chapter I, § 5, no. 12. A function l(x, y) of two elements x and y of a Hilbert space H is called a bilinear form if l (a txt + a2X2, f3 1 y 1 + f32Yz) = atfJt l (x t , Y 1 ) + a2f1tl (xz, Y 1 ) + + a 1 fJzl (xt , Y2) + azfJ2 l (x2, Y 2) ·
The bilinear form is bounded if l l (x , y)l � c ll xll IIYII 
The least possible value of c in this inequality, or, what is the same, sup ll(x, y)l, is called the norm of the bi li near form.
II X II
=
II y II
=I
64
LINEAR OPERATORS IN HILBERT SPACE
If A is a bounded linear operator, then the form l (x, y) = (Ax, y) is a bounded bilinear form. Conversely, to every bounded bilinear form I (x, y), there corresponds a bounded operator A for which the preceding equality is valid. Furthermore, the norm of the bilinear form is equal to the norm of the operator. The values of the bilinar form I (x, y) are completely defined by the values of the corresponding quadratic form (Ax, x ) In fact .
l (x, y) = [ 1 (x t. x 1 )  l (x2 , x2)] + i [ l (x 3 , x 3)  1 (x4 , x4)] , where x 1 = Hx + y) ; x 2 = � (x  y) ; x 3 = ! (x + i y) ; x4 = ! (x  iy) . The operator A is uniquely defined by its quadratic form (Ax, x) : if (Ax, x) = (Bx, x), then A = B. *) Let the space H be separable and {e;} be an orthonormal basis in H. Then 00
Aek = L a;ke; ; ;� 1
moreover, 00
00
If x = L �jej and y = L 1'/jej , then j� 1 j�1 00
00
" ;:'oj.a . . e. and Ax = " � � j� 1 i � 1
!J
'
The matrix (a;k) is called the matrix of the operator A with respect to the basis {e;}. The matrix (a�) = (dk;) will be the matrix of the adjoint operator A*. It is necessary for the boundedness of the operator A (and this means also for the operator A*) that 00
and
2
L l ak; l < oo k� 1
( i = l, 2, . . . )
.
These conditions are not sufficient for the boundedness of the operator *) The last assertions are valid only in a complex H ilbert space.
\
65
BOUNDED LINEAR OPERATORS IN A HILBERT SPACE
given by the matrix. Examples of sufficient conditions are : 1 . If 00
a � l ; M l k k=l L
00
i = 1, 2, . . . ) � a ( ;i M k l k=l
L
and
'
where M does not depend on i, then the operator A is bounded. 2. If 00
i,
a , l2 ; l < oo k k=
L1
then the operator A is bounded. The number {
00
t is sometimes } a ; l2 l k i,k= 1
L
called the abolute norm of the operator A. This norm does not depend on the choice of the orthogonal basis { e;} in H. Effectively verifiable necessary and sufficient conditions for the boundedness of an operator given in matrix form are not known. In a function space, the most prevalent class of linear operators is the class of integral operators of the form b
Ax = I K(t,s)x(s)ds. a
It suffices for the boundedness of the integral operator in the Hilbert space b) that a number M exist such that
L2 (a,
b
I I K(x,y)l dy � M
b
I I K (x,y)l dx � M. Also, the summability of the square of the kernel K(t, s) with respect to both variables : I I I K(t,sWdtds < oo and
a
a
b
b
a
a
is a sufficient condition for boundedness. 2. Unitary operators. A linear operator U mapping a Hilbert space H onto all of H with preservation of the norm : is
called unitary.
II u X
I I =
X
II '
66
I
LINEAR OPERATORS IN HILBERT SPACE
I
In the coordinate Hilbert space /2 , the operator mappipg the element x onto the element y by means of a fixed permutation of the coordinates of the element x can serve as an example of a unitary operator. In the complex space L2 (a, b), the operator of multiplication by the function eict , where c is a real number, is a unitary operator. The shift (or displacement) operator I
A.x = x (t + s) is a unitary operator in the space L2 (  oo, oo). Indeed
J l x (tW dt = J i x (t + sW dt .  oo
 oo
Analogous unitary operators arise by the consideration of shift oper ators on functions defined on groups with invariant measures or in dynamical systems. Unitary operators have the properties : 1) ( Ux, Uy ) = (x, y ) (x, yE H). 2) The operator u  t , the inverse of the unitary operator, exists, and u 1 = u*
(this property can serve as the definition of a unitary operator). 3) The product of unitary operators is again a unitary operator. Unitary operators form a group. If A. is an eigenvalue of the unitary operator U, that is, an element e =F () exists such that Ue = A.e , then IA.I = 1 . Analytic descriptions of all unitary operators can be given for the space L2 (a, b) (see [39]). A number of transformations used in analysis generate unitary oper ators. The FourierPlancherel transformation, given by the formula g (t) =
J 2n dt
1
J
00
d
 oo
e  •ts _ l •
 is
f (s) ds = Uf (x)
BOUNDED LINEAR OPERATORS IN A HILBERT SPACE
67
. g (t) = ln I j1
or by the simpler formula
oo
e  usf (s) ds ,

oo
in which the integral must be understood as the limit in the mean (with respect to t) of the integral from  N to N as N HXJ , is a particularly important example of these transformations. The operator U is unitary in L2 ( oo , oo ) . The inverse operator is given by the formula
�n I eitsg (s) ds . J 00
u l g (t) = f (t) =
 oo
An isometric operator is a generalization of a unitary operator. Such a linear operator maps a subspace H 1 of a Hilbert space H onto a subspace H2 of the same or another Hilbert space with preservation of the scalar product and, hence, of the norm. In the case where H1 = H2 = H, the isometric operator becomes a unitary operator. 3. Selfadjoint operators. A bounded linear operator coinciding with its adjoint, A = A* , is called selfadjoint. For a selfadjoint operator,
(Ax, y) = (x, Ay) = (Ay, x ) . A bilinear form having the property that l (x, y) = l (y, x ) is called Hermitian. To every bounded hermitian form there corresponds a bounded selfadjoint operator. The quadratic form (Ax, x) corresponding to a selfadjoint operator is real. The numbers m = inf (Ax, x ) and M = sup (Ax, x) JlxJI = l
llxJI = l
respectively are called the lower and upper bounds of the selfadjoint operator. The norm of the operator A is equal to the largest of the numbers lml and I M I : II A II = max ( l m l , IMI ) = sup !(Ax, x)j . •
llxll = I
If the lower bound is nonnegative, that is, (Ax, x) � O
I '
68
LINEAR OPERATORS IN HILBERT SPACE
for arbitrary xEH and A ¥ 0, then the operator is called positive. If a bounded operator is given by a matrix (a 1k), then it will be self adjoint if and only if the matrix corresponding to it is hermitian, that is,
A bounded integral operator in L2 (a, b) with kernel K(t, s) will be selfadjoint if K(t, s) = K(s, t ) Every bounded selfadjoint operator in L2 (a, b) is representable in the form of an integral operator, but by the kernel K(t, s) we must now understand not a usual function but a gener alized function (see ch. VIII.) The eigenvalues of a selfadjoint operator are real ; the eigenvectors corresponding to distinct eigenvalues are mutually orthogonal. Let A be an arbitrary bounded operator. It is representable in the form .
A + A* A  A* = A1 + iA2 , +i A= 2 2i where the operators A 1 and A 2 are selfadjoint. The operators Re A A + A* A  A* and Im A = are called the real and imaginary parts of 2i 2 the operator A. If A is an arbitrary bounded operator, then the operators AA * and A* A are selfadjoint and positive. A + A* is a negative operator, then the operator A is called If Re A = 2 dissipative. 4. Selfadjoint completely continuous operators. If the selfadjoint operator A is completely continuous, then the space H can be decomposed into the orthogonal sum of two subspaces : H = H0 + H ' where Ax0 = 0 for arbitrary x0 EH0 and where an orthonormal basis {x1} exists in the space H' consisting of eigenvectors of the operator A corresponding to nonzero eigenvalues A.1 • Thus, for an arbitrary element xEH x = x0 + L c1x1 = x0 + L (x, x1) x1 i
and
i
Ax = L A.1c1x1 = L; A.1 (x, x1) x1 • I
i
BOUNDED LINEAR OPERATORS IN A HILBERT SPACE
69
In particular, it follows that a selfadjoint completely continuous operator, not vanishing on the entire space, has at least one eigenvalue different from zero. The space H0 consists of eigenvectors of the operator A corresponding to the eigenvalue A. = 0. Selecting in this space an arbitrary orthonormal basis { e;}, we obtain an orthonormal basis { e;} + { e1} for the space H consisting of eigenvectors of the operator A. The eigenvalues and eigenvectors of a selfadjoint completely continuous operator can be obtained by the following process : the form A (x, x) on the unit sphere of the space H attains its largest absolute value at some element x1• It can be shown that Ax1 = A.1x l > where A1 = (Ax1, x1) = + max lA (x, x)l = + II All (the sign + or  coincides with the sign of llxll ; l (Ax1, x1)). Let H1 be the orthogonal complement of x1 in H. The subspace H1 is invariant with respect to the operator A. If the operator A annihilates every element of H1, then the process comes to a stop ; if Ax ¥0, then the form (Ax, x) ¥0, and on the unit sphere of the space H1 it attains its largest absolute value at some element x2 • In this connection, Ax2 = A.2 x 2 where A.2 = + max !(Ax, x) l and x2 E H1 • ll x l l ; l , xeHt
It follows from the construction that 122 1 � I A1 1. The continuation of this process gives a finite or denumerable complete system of eigenvalues and eigenvectors of the operator A in H' . Let A.t, A.i , . . . be the positive eigenvalues of the operator A arranged in decreasing order, and A.! , A2 , be the negative eigenvalues arranged in increasing order (multiple eigenvalues are repeated as many times as their multiplicity). The eigenvalues have the following minimaximal property : let z 1 , . . . , Zn be arbitrary elements of H and M{z 1 , z 2 , . . • , zn) be the maximum of the form (Ax, x) on all elements x satisfying the conditions • • •
Then the smallest value of the function M(z1 , z 2 , • • • , zn ) for all possible systems {z1, z 2 , . . . , zn) of elements of H will be equal to A.: . Analogously, A.; = maxm{z 1 , z2 , . . . , zn), where m (z 1 , z2 , . . . , zn) is the minimum of the zleH
form (Ax, x) on the elements x satisfying the preceding conditions. A selfadjoint completely continuous operator will be positive if and only if all its eigenvalues are nonnegative.
LINEAR OPERATORS IN HILBERT SPACE
70
The properties of selfadjoint completely continuous operators are generalizations of properties of integral operators with symmetric kernels, considered in the theory of integral equations. The equation
X  f.lAX = y is a generalization of the integral equation. If A is a selfadjoint completely continuous operator and the number 1/f.1 does not coincide with any of its eigenvalues, then the formula
A. X = Y + f.l L (y, x ;) X; • ' ; 1  f.lA; gives the solution of the preceeding equation. If 1/p coincides with one of the eigenvalues of the operator A, then a solution exists only under the condition that the element y is orthogonal to all the eigenvectors corresponding to the eigenvalue 1/f.l. In this case, one of the solutions can be obtained by the same formula if terms in it 1 containing A ; =  are discarded. f.1
5. Completely continuous operators. Besides the basic definition of a completely continuous operator, according to which an operator is called completely continuous ifit maps every bounded set into a(relatively) compact set, equivalent definitions exist in a Hilbert space. 1 . A linear operator A is completely continuous if it maps every weakly convergent sequence into a strongly convergent sequence, that is, if Xn�Xo implies that Axn+Ax0 in the norm. 2. A linear operator A is completely continuous if the equality lim (Axm Yn) = (Axo , Yo) n + oo is valid for arbitrary sequences { xn } and { Yn } which are weakly convergent to x0 and Yo ; that is, the form (Ax , y) is a weakly continuous function of x and y. If A is completely continuous, then A* is completely continuous. This assertion is useful : if AA* is completely continuous, then A is completely continuous. A finitedimensional operator in a Hilbert space H is representable
BOUNDED LINEAR OPERATORS IN A HILBERT SPACE
in the form
71
Ax = L (x, xk) Yk = L xk ® Yk , k= 1 k= 1 where xk and Yk (k= 1, 2, . . , n) are fixed elements of H. A representation is possible for completely continuous operators which is analogous to the above representation of a finitedimensional operator. The numbers J.1 > 0, for which nonzero solutions of the system n
n
.
{ Ax* = J.lY , A y = f.lX
exist, are called singular values of the operator A, and the corresponding solutions x, y are called associated fundamental Schmidt elements. The numbers J.1 2 are eigenvalues of the positive selfadjoint completely continuous operators AA* and A* A. Therefore, there exist only a denu merable number of singular values J.l i; moreover J.l i +0 for i oo .* ) The representation +
A = L J.liXi ® Yi i= 1 holds, where x i, Yi are associated fundamental elements corresponding to the singular values J.l i· The series converges in the operator norm. Ex plicitly, the preceding equality has the form 00
00
Ax = L J.li (x, x) yi . i= 1 An analogous representation 00
A * x = L J.liY i ® xi i= 1 holds for the adjoint operator. If A is selfadjoint, then x i = Yi• and the previously examined represen tation is obtained. If the operator A is given by the matrix (a ik), then it suffices for its complete continuity that 2 < oo . a l i k L i, k = 1 00
l
•) There may be only a finite number of singular va lues JJJ. (Editor)
72
LINEAR OPERATORS IN HILBERT SPACE
Analogously, it suffices for the complete continuity of an integral operator in L2 (a, b) that b
b
a
a
J J iK (t, sW dt ds < oo . Neither of these conditions is necessary. An integral operator with a symmetric kernel satisfying the last con dition is called a HilbertSchmidt operator. The question of the completeness of a system of eigenvectors and associated vectors of a completely continuous operator is important in the theory of completely continuous operators (see ch. I, §5, no. VIII.) A basis of eigenvectors exists for a selfadjoint completely continuous operator. However, if a finitedimensional operator is added to a self adjoint completely continuous operator, then the operator obtained may not always have a complete system of eigenvectors and associated vectors. If, for example, the onedimensional operator 1
t
J {1  s) x (s) ds 0
is added to the integral operator with symmetric kernel K (t, s) =
{(t(  1) s
for s � t s  1 ) t for s � t ,
then a Volterra operator
{O � s, t � 1 )
t
J (t  s) x {s) ds 0
is obtained which does not have eigenfunctions. From the available list of criteria for completeness we deduce the following : let A be a completely continuous operator such that the values of the form (Ax, x) for arbitrary xEH are contained in the sector of the complex plane : n
l arg � l �  (e � 1) . 2 e
The system of eigenvectors and associated vectors of the operator A is
BOUNDED LINEAR OPERATORS IN A HILBERT SPACE
complete in the space H if its singular values order have the property lim n111l"rn = 0 '
f.l n
73
arranged in decreasing
in particular, 1j the series
converges. These conditions are not very suitable in that the singular values of the operator A appear in them. For g > 1 in these conditions, the singular values can be replaced by the eigenvalues of the real or imaginary part A + A* A  A* of the operator A the operator or . li 2 The system of eigenvectors and associated vectors is complete when g = 1 , that is, for a dissipative operator, if the operator has finite trace
)
(
However, this assertion becomes invalid if the singular values are replaced by the eigenvalues of the real or imaginary part of the operator A. If both the real and the imaginary parts of the dissipative operator A have a finite trace, then the system of eigenvectors and associated vectors is complete in the space H.
6. Projective operators. The selfadjoint operators having the simplest structure are projective operators. Let L be a (closed) subspace of the space H. The operator which sets in correspondence to every element x its projection y on the subspace L : y = PLx is called the projector onto the subspace L, or, more precisely, the projective operator PL. By definition, PLx =x for an element xEL. A projective operator is selfadjoint ; its square is equal to itself and, hence, it is positive. Conversely, if a bounded linear operator P has the properties P * = P and P 2 =P, then it is the projector of the space H onto its range of values. The norm of a projective operator is equal to 1 . If L is finitedimensional, thenPL is finite dimensional and, consequently, completely continuous. If L is infinitedimensional, then PL is not completely continuous.
74
LINEAR OPERATORS IN HILBERT SPACE
If L 1 and L 2 are orthogonal subspaces, then PL, PL1 = 0, and conversely. In this case the operators PL, and PL1 are called orthogonal. Properties ofprojective operators 1) In order for the sum of two projective operators PL, and PL2 to be
a projective operator, it is necessary and sufficient that these operators be orthogonal. If this condition is satisfied, then 2) In order for the product of two projective operators PL, and PL2
to be a projective operator, it is necessary and sufficient that the operators PL, and PL1 commute. If this condition is satisfied, then
The projective operator P 1 is called a part of the projective operator p2 if p1p2 = p2p1 = p1 . 3) The projective operator P 1 is a part of the projective operator P 2 if and only if the subspace L 1 is a part of the subspace L2 • 4) In order for the projective operator PL1 to be a part of the projective operator PL,, it is necessary and sufficient that the inequality be satisfied for all xEH. 5) The difference PL, PL1 of two projective operators is a projective operator if and only if PL2 is a part of PL,· If this condition is satisfied, then PL, PL1 is the projector onto L 1 L 2 (the orthogonal complement of L 2 with respect to L 1). 6) A series of mutually orthogonal projective operators ·
is always strongly convergent, and its sum is a projective operator P. The subspace L, onto which this operator projects , is called the orthogonal sum of the subspaces Ln onto which the operators Pn project :
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
75
§ 3. Spectral expansion of selfadjoint operators
1. Operations on selfacijoint operators. The sum of two bounded selfadjoint operators is again a selfadjoint operator. Moreover, an arbitrary linear combination of selfadjoint operators with real coefficients is a selfadjoint operator. The sum of positive operators is also a positive operator. A product of bounded selfadjoint operators will be selfadjoint if and only if these operators commute. If, in this connection, the factors are positive, then the product is positive. The set of selfadjoint operators is closed with respect to weak con vergence ; that is, the limit of a weakly convergent (see ch. I, § 4, no. 3) sequence of selfadjoint operators is a selfadjoint operator. In the set of selfadjoint operators, an order relation can be introduced by setting A � B if A  B is a positive operator. In this connection, inequalities between the operators have the basic properties of ordinary inequalities between real numbers. However, for two different self adjoint operators, it is impossible to talk about one always being greater than the other, since it is possible that the form {{A  B)x, x) will be greater than zero for one x and less than zero for another x. In this case, the operators A and B are called noncomparable. Since there exist comparable and noncomparable selfadjoint operators, we say that a partial ordering or a semiordering exists in the set of all selfadjoint operators. The existence of a partial ordering allows the introduction, in the set of selfadjoint operators, of several concepts such as, for example, sets of operators bounded above or below, lower and upper bounds for the bounded set of operators, monotone increase and mono tone decrease of a sequence of operators, and others. The following is an important property of bounded sequences of selfadjoint operators : If {An} is a monotone increasing sequence of mutually commutative selfadjoint operators, bounded above by a selfadjoint operator B which commutes with all the An, then the sequence {An} converges strongly to a selfadjoint operator A � B, and
A = sup A n
• .
To every selfadjoint operator A t h e re co rres ponds the partially ordered
76
LINEAR OPERATORS IN HILBERT SPACE
ring KA of all bounded selfadjoint operators commuting with A. The ring KA , generally speaking, is noncommutative. This ring contains the operator A and an arbitrary polynomial
n 2 + + anA a = a0 + a + A P(A) 1A 2 · · ·
in A with real coefficients. The correspondence between operator poly nomials and polynomials in a real variable is linear and multiplicative, that is, if P (t) = ctQ (t) + f3R (t) , then P (A) = ctQ (A) + f3R (A) , and if then
P (t) = Q (t) R (t) ,
P (A) = Q {A) R {A) .
A more profound fact is the positiveness of this correspondence in the sense that if P{t):?: O on [m, M], where m and M are lower and upper bounds for the operator A, then P{A) � 0. It follows from the positiveness of the correspondence that if a monotone increasing sequence of poly nomials {Pn {t)}, uniformly bounded on the segment [m, MJ by the number K, converges to a function F(t), then the sequence of polynomials {Pn(A)} is also monotone increasing and bounded by the operator Kl and, hence, has a limit B =lim Pn (A). This operator is, naturally, denoted n
by F(A) and called a function of the operator A. The operator F(A) belongs to KA and, moreover, commutes with an arbitrary operator from KAIn particular, we can introduce the function B=JA for a positive operator A. The operator B is positive and B2 = A . It is defined uniquely by these properties. JA can be defined as the limit of the sequence of polynomials Bn defined by the recurrence relation B0 = 0 ,
Bn + 1 = Bn + HA  B; ) .
The operator A 2 is positive for an arbitrary selfadjoint operator A ; therefore, naturally we denote JA 2 =I AI.
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
77
2. Resolution of the identity. The spectral function. Functions corre sponding to characteristic functions of intervals of the real axis are an important class of functions of operators. Since the square of a charac teristic function is equal to itself, the square of the corresponding self adjoint operator will be equal to itself; that is, the operator will be pro jective. We denote, in particular, by E;. the operator corresponding tQ the characteristic function of the halfaxis ( oo , A) (i.e. to the functi � / : equal to zero for t )'; 1 and unity for t < 1). A certain set of projective operators E;. ( oo < A, < oo ) is called a resolution of the identity generated by the operator A and has these properties : 1) E;. � Ell or, what is the same, E;. Eil= E;. for A < 11 ; 2) E;. is continuous from the left with respect to .1, that is E;.  o = lim Ell = E;. ; 
I
�

!l � }. _ Q
3) E;. = 0 for AE( oo , m) and E;. = 1 for A.E(M, oo ) where m and M are the lower and upper bounds of the operator A ; 4) the operator E;. commutes with an arbitrary operator from KA. The operator function E;. is called the spectralfunction of the operator A ; the operator EA = Ep  Ea is called the spectral measure of the interval A = [oc, PJ. This measure has the property of orthogonality : if .1 1 n.d 2 = 0, then EA, EA, = 0. The operator A can be reconstructed from its spectral function or measure. It can be shown that M+ O AdE;. , A= ,
Jm
where the integral on the right is an abstract Stieltjes integral. The abstract Stieltjes integral b
J j (A.) dE;. a
with respect to spectral measure is understood to be the limit with respect to the operator norm of integral sums
78
LINEAR OPERATORS IN HILBERT SPACE
where the Ll k are the subintervals into which the interval [a, b] is par titioned and vk is an arbitrary point inside Llk . From the spectral representation of the operator follow the formulas M+O
f A dE;.x ,
Ax =
m M+O
(Ax, x) =
J ).d(E;.x, x) ,
m M+O
II Ax ll 2 =
J A2 d(E;.x, x) . m
3. Functions of a selfadjoint operator. The spectral representation of an operator allows the introduction of a broader class of functions of the operator, which includes the functions defined previously. We set M+ O
f (A) =
f j (A)dE;. , m
if the last integral exists. In particular, it exists for an arbitrary continuous function. The correspondence between functions of a real variable and functions of the operator has the following properties : 1) If J ( A) = a/1 (A) + bf2 (A) , then f (A) = af1 (A) + b/2 (A) . 2) If f (A) = !1 (A)j2 (A) , then f( A) = !1 (A) f2 (A) . 3) j(A) = [! (A)] * , above the function denotes transition to the complex where the conjugate function.
4) II/ (A) II � max If (A) I . m� A � M
79
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
5) It follows from AB= BA that f(A) B = Bf (A) for an arbitrary
bounded linear operator B. 6) Ifj(A.) �
4. Unbounded selfadjoint operators. If A is an unbounded linear oper ator with an everywhere dense domain D (A) in H, then its adjoint oper ator A* will be defined for those elements y for which the functional (Ax, y) is bounded (see ch. I, § 5, no. 1 0). In a Hilbert space, this means that (Ax, y) = (x, y *) ,
(
where y*E H and A*y= y*. An unbounded operator is called selfadjoint if A = A*. In distinction from the case of bounded operators, this means not only the presence of the identity (Ax, y) = (x, Ay)
(x, y E D (A)) ,
but also the coincidence of the domains D (A) and D (A*) of the oper ators A and A*. Thus in order to test the selfadjointness of an operator, it is necessary to show, for every element y for which the functional (Ax, y) is bounded, that yE D (A) and then test the validity of the pre ceding identity. An unbounded selfadjoint operator is always closed. With some modifications, the basic results of spectral theory stated above for bounded selfadjoint operators remain true for unbounded selfadjoint operators ; in particular, the spectral theorem is true. Precisely : let A be an unbounded selfadjoint operator with domain D (A). Then the operator generates a set of projective operators E;.,  oo < A, < + oo , having the properties : I ) E;. � E'" for A < 11 ; 2) E;. is continuous from the left ; 3) E_ 00 = lim E;. = 0, E+ oo = lim E;. = I; A + 
oo
A + +
oo
4) BE;. = E;.B if B is an arbitrary bounded operator which commutes with A . In this connection, a bounded operator B is called permutable (or commutative) with the unbounded operator A if x E D(A) implies BxE D(A) and A Bx = BAx.
LINEAR OPERATORS IN HILBERT SPACE
80
The element x belongs to D(A) if and only if 00
For these elements x,
00
The integral b
integral
 oo
00
00
 oo
 oo
I AdE.�.x is understood to be the limit of the
proper
00
I A, dE x in the sense of strong convergence ;.
a
as
a+  oo , b+ oo .
A selfadjoint operator A is called semibounded below if(Ax, x) � a(x,x) for all XED(A). In this case
I A.dE.�.x . 00
Ax =
a
An operator which is semibounded above is defined analogously. If the function/ (.1) is finite and measurable with respect to all measures generated by the functions u (A) =(E;.z, z) (z E H ) then an operatorf (A) can be defined. This operator, generally speaking, is not bounded. Its domain D (f(A)) is the collection of elements x for which ,
00
 oo
The set D(f (A)) is dense in H; the operator f(A) is given by the formula
I f(A.)d(E.�.x,y) 00
(f (A) x, y) =
 oo
(x E D (f (A)), y E H) and is selfadjoint (for a real function f(A)).
81
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
If the function j(A) is bounded (  oo < A < oo ), then the operator f(A ) will also be bounded. The resolvent is an important example of a bounded function of an operator. If Ao does not belong to the spectrum of the operator A, then the spectral representation 00
I 1
 dE;.X
 00
A  Ao
is valid for the resolvent R;.0 • Hence, in particular, we have the\f()ll�wirtg bound for the resolvent :
/
where d is the distance from the point Ao to the spectrum of the operator
1
A. In this connection d)'; l im Aol and, hence, IIR;.0 II �
.
l im Aol If the operator A is semibounded below, then the function e A :
a
will be a bounded operator. The collection of all functions of a selfadjoint operator A allows an "extrinsic" description. If the collection of all bounded linear operators commuting with A is denoted by R (A), then the set of functions of A coincides with the collection of all closed operators commuting with an arbitrary operator from R(A).
5. Spectrum of a selfadjoint operator. The spectrum of a selfadjoint
operator A is a closed set on the real axis, consisting of all points of increase of the function E;.. Jumps of the function E ;. correspond to eigenvalues of the operator A ; the operator EH 0  E;. 0 is the projector onto the eigenspace corresponding to the eigenvalue A. The eigenvalues form the discrete or point spectrum of the operator A. If the eigenvectors of the operator A form a complete system in the space H, then we say that the operator has a pure point spectrum .In this case, the spectrum of the operator consists of the set of eigenvalues and limit points of this set. _
82
LINEAR OPERATDRS IN HILBERT SPACE
In the general case, the space H can be decomposed into the orthogonal sum of subspaces H1 and H2 invariant with respect to A where the oper ator A has a pure point spectrum in H1 and does not have eigenvectors in H2 . The spectrum of the operator A i n the subspace H2 is called the continuous spectrum. The continuous spectrum and the point spectrum can intersect. The points of the continuous spectrum, the limit points of the set of eigenvalues, and the eigenvalues with an infinite multiplicity form the limit spectrum of the operator A. The limit spectrum of a selfadjoint operator consists of the single point 0 only in the case when the operator is completely continuous. Example. In the space L 2 [0, I], the integral operator with symmetric kernel K(t, r), having the property that 1 2 IK {t, r)l dr 0
J
exists for almost all tE [0, I], generates a selfadjoint operator, the so called Carleman operator. This operator can be unbounded. The point 0 always is a point of the limit spectrum of the Carleman operator. In order for the point }.0 to be a point of the spectrum of the operator A, it is necessary and sufficient that a sequence of elements Xn E D (A) with ll xn ll = 1 exist such that II Axn  Ao Xnll +0. In orderfor Ao to be a point of the limit spectrum, it is necessary and sufficient that a sequence of elements Xn exist, weakly converging to zero and having the preceding properties. The addition to a selfadjoint operator of a completely continuous operator does not change the limit spectrum of the operator. On the other hand, a completely continuous operator with arbitrarily small norm can be annexed to an arbitrary selfadj oint operator such that its spectrum becomes purely point.
6. Theory of perturbations. The last assertion of the preceding sub section can be applied to the theory of perturbations which studies the change of spectral properties for small changes of the operators. Let a set of selfadjoint operators A (e) depending on the parameter e be given, and let D be the set of x for which the limit lim A (e) x = A0x e+ 0
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
83
exists. (It is assumed that x E D(A (e)) for 0 < e < e0 (x).) If the selfadjoint operator A is the closure of the operator A0, then the relation
E;. = lim E;. (e) E+ 0 is valid for the spectral functions E;. (e) and E;. of the operators A (e) and A and for an arbitrary A not belonging to the point specttum of the operator A. The limit is understood in the strong sense. Uniform convergence Qf E;. (e) to E;. (with respect to the operator norm) under the indicated con ditions, generally speaking, does not hold. Moreover, it may not hold even if it is required that the operators A (e) be bounded and uniformly convergent to the operator A. If the new norm ll x ll1 = ll x ll + II Ax il is introduced in the domain D (A ) of the selfadjoint operator A, then D (A), with this norm, will be a Banach space H1 (see ch. I, §5, no. 10). If all the operators A (e) are defined on D (A) and converge to A uniformly with respect to the norm II x ll 1 : II A (e) x  Ax il � C, ll x ll 1
(x ED (A)) ,
where c,�o as e+0, then the spectral function E;. (e) converges uniformly to the function E;. at an arbitrary point A not belonging to the spectrum of the operator A ; that is lim II E;. (e)  E;. ll = 0 . E+ 0 If A o is an isolated point of the spectrum which is an eigenvalue of finite multiplicity m and Lf is an interval separating it from the remaining part of the spectrum, then under the preceding conditions for a suffi ciently small e the spectrum of the operator A (e) in the interval Lf consists of m eigenvalues (taking into account their multiplicity). These eigenvalues A.k (e) (k = 1 , 2, . . . , m) tend to the point A o as e+0. However, we must keep in mind that although EA (e) converges uniformly to EA• the eigen vectors ek (e), corresponding to the eigenvalues Ak (e), may not have a limit as e+0. If Ao is a simple eigenvalue, then the eigenvectors e(e) of the operators A (e) approach an eigenvector e of the operator A. The operator A(e) is called an analytic function of e if where the operators A 1 and A act from
H1
to
II,
D(A 1) = D(A) =H1 , and
84
LINEAR OPERATORS IN HILBERT SPACE
the series converges with respect to the operator norm. Then E;. (e) is also an analytic function of e in a neighborhood of e = O for every A not be longing to the spectrum of the operator A. For the above considered case of an isolated eigenvalue Ao of multi plicity m, and
ek (e) = ek + eek( 1 ) + e 2 ek( 2 ) + · · · .
Let the operator A have a complete system of eigenvectors {en } with corresponding eigenvalues An . If An is an isolated simple eigenvalue of the operator A, then formulas for the determination of the coefficients of the expansion in powers of e of the eigenvalue An (e) of the operator A (e) = A + A 1 e can be obtained. Here, only formulas or first and second approximations are mentioned : where
(the prime on the summation sign means that the term for m = n is omitted). The expansion
is valid for the eigenvector en (e). In physics , these formulas are called the formulas ofperturbation theory. If the operator A has parts of continuous spectrum, then analogous formulas occur where, besides the sums, integrals occur. In the case of an eigenvalue of multiplicity m, in order to obtain the coefficients of the powers of e, one must find eigenvalues and eigen functions of mdimensional operators.
85
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
Problems of perturbation theory are particular cases of the more general problem of the study of the behavior of the function f(A (e)) as e varies, wherej(A) is a given function. The function E;. (e) is precisely a function of such type (see no. 2). For the expansion in powers of e of such functions it is natural to apply Taylor' s formula, assuming the functions f (A) and A (e) are sufficiently smooth. Then 2 (e (A d )) f (� (e)) df 2 +e + ··· f (A (e)) = f (A ) + e de de e=o e=o There are special formulas for the derivatives of functions of operators with respect to a parameter. Here, only the formula for the first derivative is mentioned. It is valid under the assumption that the operator A has a finite absolute norm (see ch. I, § 2, no. 1). If x = L Ck ek > then k f (Am)  f (Ak) dA df(A (e)) "' "' ek , em em , X = L. L. Ck k m de e = O Am  Ak de e = O where it is assumed that
(
(( )
)
)
Multiplicity of the spectrum of a selfadjoint operator. The spectrum of a selfadjoint operator A is called simple if an element u E H exists such that the closed linear hull of all elements of the form E,1 u, where Ll is an arbitrary interval of the real axis, coincides with H. In this case, u is called a generating element. The formulas 7.
X=
and
f j(A)dE;.u , y = f g(A)dE;./l 00
00
 oo
 oo
00
(x, y) =
f f (A)g(A)du(A) , 
oo
are valid for arbitrary x and yE H, where
86
LINEAR OPERATORS IN HILBERT SPACE
and f(A,) and g (A_) are certain functions with squareintegrable moduli with respect to the measure u(A,). The function u(x) is a nondecreasing function of bounded variation on ( oo , oo ) and is called the spectral function of the operator A. It can be shown that, to an arbitrary function f (A,) ( oo < )., < oo ) with a squareintegrable modulus with respect to the measure u(A_), there corresponds some element x for which 

00
X=
J f (A_) dE;.u .
 oo
Thus, the last formula establishes an isometric correspondence between the space H and the space L 2 ,a of all functions f (A,) on the axis ( oo , oo ) for which 
00
 oo
The formula 00
Ax =
J A,f (A_) dE;,u
 oo
is valid for the operator A and, hence, under the isometric correspondence it maps into the operator A of multiplication by the independent variable )., : Af (A_) = Af (A_) , defined for all the functions f(x)EL 2 , a, for which Af(A_) EL 2 , a. A collection of elements u 1 , u2 un is called a generating basis for the operator A if the closed linear hull of the set of all elements £,1 uk (k = I , 2, , n) coincides with H. The spectrum of the operator A is called nmultiple if the minimal number of elements in a generating basis for the operator A is equal to n. The corresponding basis is called a minimal generating basis. Numerous examples of selfadjoint operators with a finitemultiple spectrum are given by ordinary differential operators (see § 5). If u 1 , ... , un is a minimal generating basis for the operator A, then the ,
.
.
.
•
.
.
,
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
formulas
00
00
x=
f fk (A.) dE;.uk , k l I
y=
 oo
and
87
}\ I gk (A.) dE;.uk  oo
00
n
(x, y) = ;J
�
j 1
!; (A.) gj (A.) da;j (A.) ,
where
U;j (A.)
(E;.U; , uj) ,
are valid. The matrix u (A.) = (a ij (A.)) is Hermitian for every A.(  oo < 2 < oo ) and is continuous from the left, and the difference a (11)  u (A.) for 11 > A. is a nonnegative definite matrix. The space H is isometric to the Hilbert space L 2 , u of the vector functions / (2) = { /1 (A.) , . . . , Jm (A.) } (  oo <2 < oo ) , for which =
i. t I f; (A.)fj (A.) da;j (A.) < 00
II J IIL2,u =
1

oo
00 ,
and the scalar product is introduced by the formula 00
 oo
where the integrals are understood in a particular sense (see [37]). Under the isometric correspondence, the operator A transforms again into the operator of multiplication of all components of the vector function f (A.) by the independent variable 2: 00
 oo
In the general case, for a selfadjoint operator acting in a separable Hilbert space H, the space can be represented in the form of an orthogonal \Lim of subspaces Hk (k = I, 2, . . . ) such that each subspace Hk is invariant with respect to the operator A and the operator A has simple spectrum in it.
In conclusi on, it remains to remark that someti mes it is convenient •
88
LINEAR OPERATORS IN HILBERT SPACE
to use noneigenvectors as generating elements (see [37]). In this case all the formulas are unchanged, but the functions a(A.) need not be of bounded variation on (  oo , oo ) . 8. Generalized eigenvectors. In § 2, no. 5 it was remarked that if A is a completely continuous selfadjoint operator, then its eigenvectors ek (k = 1 , 2, . . . ) form a basis in the space H; that is, for every xEH. The formula
co
J dE;.X
X=
 co
is a generalization of the preceding for the case of an arbitrary self adjoint operator. A more natural generalization would be the formula co
 co
where e;. is an element of H satisfying the equation Ae;. =A.e;. (an eigen vector or 8), and the weight dQ (A.) plays the role of the coefficients ck in the series expansion. However, the simplest noncompletely continuous selfadjoint operators in a Hilbert space, such as the operator of multi plication by x in L 2 (a, b) or the operator of differentiation in L 2 (  oo , oo ), do not have eigenvectors in these spaces. In fact, if the relation xy (x) = A.y (x) is satisfied for some function y(x) E L2 (a, b), then the function y(x) must be equal to zero for x :;d and can be different from zero only for x = ).. But in the space L 2 (a, b) there is no nonzero element having this property. Nevertheless, the operator of multiplication by x has eigen functions, namely delta functions 15 (x  A.) which are generalized functions (see ch. VIII, § I ) and do not belong to L 2 (a, b). The examples mentioned suggest seeking expansions in eigenvectors not belonging to the space H. The difficulty, consisting in having avail able only concepts connected with the space H to construct elements not belonging to it, is overcome in the following manner : a more restricted linear topological or B anach or Hilbert space 11> is constructed !n the
SPECTRAL EXPANSION OF SELFADJOINT OPERATORS
89
initial Hilbert space H. A topology is introduced in cp so that the func tionals (cp, h) (hE H ) are continuous on cP ; then the space H is em bedded in the larger space cP* in which eigenvectors of the operator A are sought. Eigenvectors of the operator A belonging to cP* and not belonging to H are called generalized eigenvectors.*) It turns out that a space cP* can be constructed with respect to tlie>1 space H so that every selfadjoint operator i n H has a complete system of eigenvectors in cP*. In the case of a selfadjoint operator A with a simple spectrum and the generating element u, the expansion with respect to the generalized eigenvectors has the form co
cp =
I g (A)e;. da (2)
 co
for arbitrary cp E cP, where a (2) = (E;. u, u) and the function g (2) is defined by the equality g (2) = (cp, e;.) · The last formulas are analogous to the inversion formulas in the theory of the Fourier transform where, the role of e ;. is played by the ;. ; function e x and a (2) = 2 (see § 2, no. 2). The equation co
co
 co
 co
is a valid analogue of the Parseval equality. For an operator with an arbi trary spectrum, the formulas acquire a more complicated form : co
cp = and
;£1 I
(cp, ey>) ey> day >
 co
co
 co
The spectrum of an operator A is called a Lebesgue spectrum if the func•) The triple o f spaces 4>, H and 4>0 ( 4>
c:
H
c:
4>0) is called equipped Hilbert space.
90
LINEAR OPERATORS IN HILBERT SPACE
tions a (.A.) and .A. are equivalent, that is a (.A.) and A. are absolutely contin uous with respect to one another. In this case da (A.) can be replaced in all formulas by Q (.A.) d.A. where Q (.A.) is a function which is summable on an arbitrary finite interval. § 4. Symmetric operators I . Concept of a symmetric operator, deficiency indices. A linear operator A with an everywhere dense domain D (A) is called symmetric if
(Ax, y) = (x, Ay) for arbitrary x, yED(A). Every selfadjoint operator is symmetric, but the converse does not hold. The domain of definition of the operator A*, adjoint to the symme tric operator A, can be larger than the domain of definition ofthe operator A. On D (A), obviously, Ax= A*x; therefore, the operator A* is an exten sion of the operator A. A symmetric operator always allows closure, and its closure is again a symmetric operator. If a symmetric operator is defined on the entire space, then it is bounded. If the range of a symmetric operator coincides with the entire space, then it is selfadjoint. The point .A.0 is called a point of regular type for the operator A if
II Ax  .A.o x ll > k ll x ll ,
k > 0,
for all xED(A). In other words, this means that the operator A  .A.0I has a bounded left inverse. If, besides this, the operator A is closed, then the range mlkl of the operator A  .A.0 I will be a closed set. If 9R;.0 coincides with the entire space, then the point A.0 will be a regular point of the operator A. The orthogonal complement 91kl of the subspace 9R;.0 is called the deficiency subspace. The dimension n;.0 of the deficiency sub space 91;.0 is called the deficiency index of the operator A at the point .A.0• If a connected set of points of regular type is given for a symmetric operator, then the deficiency index is the same at all the points of this set. For a symmetric operator, all nonreal numbers are points of regular type ; therefore, the deficiency index n + will be the same for all the points
SYMMETRIC OPERATORS
91
of the upper halfplane. Analogously, the deficiency index n_ is the same for all points of the lower halfplane. If there is at least one point of regular type on the real axis, then n = n _ . A closed symmetric operator will be selfadjoint ifand only ifits deficiency indices are equal to zero. The pair of (finite or infinite) numbers (n , n_) shows the degree of deviation of the symmetric operator from a self adjoint operator. One must note that the deficiency subspaces 91;.0 consist of all the solutions of the equation +
+
A * y = AoY . Thus the deficiency index n;.0 coincides with the number of linearly independent solutions of this equation. 2. Selfadjoint extensions ofsymmetric operators. The question arises whether every symmetric operator can be extended to a selfadjoint operator. The answer follows : in order for a symmetric operator to be extendible to a selfadjoint operator, it is necessary and sufficient that the deficiency indices n and n _ of the operator be identical. As was indicated above, this will hold, for example, if there are points of regular type on the real axis. One must note that it is a question here of the extension of the op erator in an initial Hilbert space H. If the deficiency indices of the operator A are not identical, then the space H can be extended to a larger Hilbert space H1 such that the deficiency indices become equal in the larger space. In this larger space, selfadjoint extensions of the operator A will exist. Let A0 be a closed symmetric operator. Every symmetric and, in particular, selfadjoint extension of the operator A0 is a restriction of the operator A �. Therefore, in the construction of such an extension, the question does not arise of how we define it on new elements, but only in what its domain of definition is ; that is, on what elements of D (A6) it is defined. In order to find selfadjoint extensions of the operator A0, it is necessary to find linear subsets in D (A�), containing D (A), on which the operator A 6 generates a selfadjoint operator. It turns out that the set D(A6) has the following structure : +
92
LINEAR OPERATORS IN HILBERT SPACE
where A. is some nonreal number. The sum on the right is direct, that is, an arbitrary element y E D (A*) is representable uniquely in the form
y = x + z;. + z;: , where X E D (A), Z;.E91;. and Z;:E 91;;. If the deficiency indices n + and n_ are equal, then an arbitrary selfadjoint extension A of the operator A0 can be constructed in the following manner : select some linear operator V which isometrically maps the space 91;. onto 91;;. Then the domain D (A) of the operator A will consist of all elements of the form 
y = x + z;. + Vz ;. , where X ED(A) and Z;. E91;.. As already remarked above, the values of the operator A will coincide with the values of the operator A*, that is,
If the deficiency indices are not equal, for example, n + < n _ , then the construction mentioned gives all the maximal symmetric extensions of the operator A0, that is, those symmetric extensions which cannot be extended further with retention of symmetry. The method of constructing selfadjoint extensions described here is due to J. von Neumann. Practically, it is not very effective, since it requires finding solutions of the equation A*y = A:y and constructing of an iso metric operator V. In the following subsection other methods for the construction of self adjoint extensions are considered. 3. Selfadjoint extensions of semibounded operators. A symmetric operator A0 is called semibounded below if for an arbitrary X E D (A0)
(A0x , x) � a (x, x) . All real numbers not exceeding the number a will be points of regular type ; therefore, the deficiency indices of the operator A0 are identical. For the construction of selfadjoint extensions of the operator A0, without loss of generality, we can assume it to be positive definite, that is, a > O. Otherwise, we can consider the operator A0 + kl with sufficiently
SYMMETRIC OPERATORS
large positive
k.
93
If A0 + kl is a selfadjoint extension of this operator,
then A0 + kl kl is a selfadjoint extension of the operator A0. Let the operator A0 be positive definite. In its domain D (A0), a new scalar product can be introduced by the formula
[x, y] = (A0x, y) . The completion of D(A0) with respect to the norm generated by this scalar product will be a Hilbert space H0. It turns out that the elements of the completion are identified naturally with some elements from H, and therefore H0 can be considered as a linear subset of the space H. On the intersection of this subset H0 with the domain of the adjoint operator D (A�), the operator A� is selfadjoint. Thus a selfadjoint extension All of the operator A0 is obtained which is the restriction of the adjoint operator A � to D (A1J =H0 n D (A6). The operator A" is called the Friedrichs or strict extension of the operator A0• The operator A" is positive definite and has the same lower bound as the operator A0 :
(A"x, x) > a (x, x) . The strict extension A" is the simplest. For its construction nothing needs to be known besides the form (A0x, x) generated by the operator A0. In connection with this, the method of Friedrichs for the construction of selfadjoint extensions is one of basic importance in the theory of partial differential equations. For a more detailed description of the domain of definition of the strict extension, it is necessary to study i n more detail the nature of convergence in the norm j(A0x, x) on D(A0) and the structure of the domain of definition of the operator A6 (see § 5 and § 6). The set H0 is the domain of definition of the square root of the operator A Jl :
For an arbitrary positive selfadjoint extension A of the operator A0, the domain D (At) contains the set H0 • The strict extension has the following extremal property : for an arbitrary selfadjoint positive definite extension A of the operator A0
LINEAR OPERATORS IN HILBERT SPACE
94
The operator B = A  1  A ; 1 , which by virtue of the preceding is a bounded positive selfadjoint operator, plays an essential role. On R (A0) the operator B is equal to zero and, hence, we can consider it a's an oper ator acting in the orthogonal complement U of R (A0): and
BU c U .
The subspace U is the deficiency subspace 910, consists of all solutions of the equation A*u = O in H, and is called the null space of the operator A�. The following assertions are valid : The domain of definition of the adjoint operator A � decomposes into a direct sum :
I)
2) The domain of definition of an arbitrary positive definite selfadjoint extension A of the operator A0 decomposes into a direct sum:
D (A) = D (A0) EB (A; 1 + B) U , where B is some bounded selfadjoint positive operator acting in the subspace U. 3) For an arbitrary operator B having the properties described above, the restriction of the operator A� to the set D (A0)EB(A; 1 + B) U is a selfadjoint positive definite operator. Thus, knowledge of the strict extension A IL allows the reduction of the description of an arbitrary positive definite selfadjoint extension to the description of the operator B. The operator B acts in a smaller space than H, the space U. In the theory of boundary value problems in partial differential equations, the subspace U is mapped onetoone onto some space of functions defined on the boundary of the region, and the operator B is connected with the boundary conditions. By means of the operator B, we can describe the structure of the domain of definition of the square root of an arbitrary selfadjoint positive definite extension A of the operator A0 :
The importance of the theory of semibounded symmetric operators is illustrated by the following example. Let a selfadjoint differential expression of order 2m be given in an ndimensional region G of Euclidean
SYMMETRIC OPERATORS
95
space with sufficiently smooth boundary :
Lu = (  lt where
L
Ia! = IPI = m
()( = (ocl , . . . , ocn) , loci = 0( 1 + OCz
Da (aapDPu) ,
lai a + . . . + ocn , D a = a a, n X l . . aXna 
.
and /3, 1/31 and nP are defined analogously. The coefficients aap are assumed to be sufficiently smooth functions of x = (x 1 . . . , xn), and aap = apa · The expression L is called elliptic if the inequality L
lai = IPI= m
n aap��� . . . �:"��� . . . ��" � A L �; m
k= 1
is valid for arbitrary real �1, . . . , �n where 2 > 0 and is independent of X E G. The operator L0, defined by the equality L0u = Lu on the set D(L0) of all finitary functions u (x) (that is, infinitely differentiable functions equal to zero near the boundary of the region G) is symmetric and semi bounded in the space H=L 2 (G). Moreover, constants c > O and k exist such that 2 dx + 1u 2 dx (A0u, u) = (Lu · u + ku 2 ) dx � c u ID"l l L
J G
[f Ia! = m
f
G
G
]
for the operator A0 = L0 + kl. The metric introduced by means of the form (A0u, u) on D(L0) turns out to be equivalent to the metric of the Sobolev space W;'. The space H0 is a subspace W;' of the space W;'. A solution of the equation Al'u = f , 0
where AIL is the strict extension of the operator A0 and / EL 2 (G), is called a generalized solution of the first boundary value problem for the equation Lu + ku = f The theory of extension is illustrated in more detail by the example of the elliptic expression of second order in § 6. 4. Dissipative extensions. A linear operator B with an everywhere dense domain D (B) is called dissipative if for arbitrary x E D (B).
Re (Bx. x) � 0
96
LINEAR O PERATORS IN HILBERT SPACE
If, in the preceding relations, the equality sign holds for all x E D (B), then the operator is called conservative. IfA0 is a symmetric operator, then for the operator B0 = iA0 we have that Re (B0 x , x) = Re (iA 0x, x) = 0 ,
and, hence, the operator B0 is conservative. In a number of problems, the question of the construction of dissipative extensions of the operator B0 arises ; the most interesting are the maximal dissipative extensions, that is, those which cannot be extended further with retention of dissipativeness. The domain of definition of the adjoint operator B� can be decomposed into the direct sum D (B�) = D (B0) EB V+ EB V_ ,
where V± is the collection of all solutions of the equation B�v = + v .
The decomposition is orthogonal with respect to the scalar product [x, y] = (A �x, A �y) + (x, y) (x, y E D (A �) = D (B�)) .
We can indicate the general form of maximal dissipative extensions of the operator B0 which are restrictions of the operator B�. To this end it is sufficient to describe their domain of definition. For an arbitrary maximal dissipative extension B c B� of the operator B0, the domain of definition has the form D (B) = D (B0) EB (l + C) V_ ,
where C is a contractive operator (i.e. 1 \ C II � 1) acting from V_ to V+ . For any such operator C, the operator Bx = B6x on the set indicated is a maximal dissipative extension of the operator B0• § 5. Ordinary differential operators
I . Selfadjoint differential expressions.
An expression of the form
l(y) = qo (x) y
.. + qn (x) y .
with real coefficients q ;(x) (i =O, 1 , . . , n) is called an ordinary linear differential expression of nth order. .
ORDINARY DIFFERENTIAL OPERATORS
97
The expression z * (y)
=
(  1)" (q oy)<"> + (  1)"  1 (q 1 Y)
is called the adjoint differential expression. The expression l(y) is called selfadjoint if l(y) = l * (y). Every selfadjoint differential expression with coefficients differentiable a sufficient number of times is representable in the form (  1)" (Po l"))(n) + (  1)"  1 (p y
l (y) =
· · ·
+
PnY
•
In the sequel it is assumed that the coefficients P; (x) (i =O, 1 , .. . , n) are defined on a finite or infinite interval [a, b] and have continuous deriva tives of orders n  i on [a, b]. Besides this, it is assumed that the function 1 is summable on every finite segment [ct, P] c (a, b). Po (X)
It is convenient to call the expressions defined by the formulas lO
]
=
y'
for k = 1 , 2, . . . , n  1 ,
ky d y[k]  k dx Y [n] =
Y[n +k]
=
d" Y
Po :;; ' ax ank Y Pk d k x"
_
d [n +k 1] (y ) for k = 1 , 2, . . . , n dx
quasiderivatives corresponding to l(y) of the function y. It follows from the definition that /
(y)
=
Y[2n] .
On every segment [a, P] c (a, b) the Lagrange identity /3
J l(y) z dx  J yl (z) dx is valid, where
a
[y, z] =
p
/3
n
'\' L...
k= 1
[y, z]
=
a
{ y[k  1 ] z[2nk]
a
r2n k] [k  1 ]} .
r
z
In the Hilbert space L 2 [a, b] we consider the linear everywhere dense set D� consisting of finitary functions, that is, infinitely differentiable
98
LINEAR OPERATORS IN HILBERT SPACE
functions equal to zero outside some segment [rx , P] (depending on the function) contained entirely in the interval [a, b]. An operator L� is defined on D� by the equality L�y = l(y). It follows from the Lagrange formula that the operator L� will be a symmetric operator. The closure A0 of the operator L0 will also be a symmetric operator. Selfadjoint extensions of the operator A0 are studied in the theory of differential operators. 2. Regular case. The selfadjoint expression l(y) is called regular if 1 the interval (a, b) is finite and the function is summable on the Po (X) entire interval (a, b). If l(y) is regular, then the domain D (A0) consists of all the functions having absolutely continuous quasiderivatives on [a, b] up to the (2n  1 )st order inclusively and a quasiderivative of order 2n belonging to L2 [a, b] and satisfying the boundary conditions lk1 (a) = y[k] (b) = 0 for k = 0, 1 , . . . , 2n  1 . '
The deficiency indices of the operator A0 are equal to 2n. The adjoint operator is given by the equality A6y = l(y), and it is defined on the set D(A6) of all functions having absolutely continuous quasiderivatives up to the (2n  1 )st order inclusively and the quasiderivative y[lnl E L 2 [a, bJ on [a, b ]. Every selfadjoint extension A of the operator A0 satisfies the equality Ay = l(y) on the functions of D (A6) satisfying the system of boundary conditions 2n r, y = I [rxjklk  l ] (a) + P,klk  l ] (b)] = O ( J = 1 , 2, . . . , 2n) , k= l where n I1 [rxjvak, 2nv + l  rxj, 2n  v + l iXkvJ = v= n = I [fJjvfJk, 2nv + 1  /Jpnv+ lf3kvJ (j, k = 1 , 2, . . . , 2n) v=l
·
Conversely, for arbitrary rxik and Pik satisfying the last condition, the operator Ay = l(y) generates a selfadjoint operator on the set of all functions of D(A6) satisfying the system of boundary conditions {riy = 0} . If p0 (x) > 0, then the operator A0 is semibounded below. The strict extension of the operator A0 corresponds to the system of
ORDINARY DIFFERENTIAL OPERATORS
99
boundary conditions
ik1 (a) = 0 and y [k 1 (b) = 0 for
k
= 0, 1 , . . . , n  1 .
The resolvent of an arbitrary selfadjoint extension of the operator A 0 is an integral operator of HilbertSchmidt type (see § 2, no. 5). Conse quently, the resolvent of an arbitrary selfadjoint extension A is a com pletely continuous operator, the spectrum of the operator A is discrete, and the operator A has a complete system of eigenvectors. 1 3. Singular case. If the interval (a, b) is infinite or the function Po (X) is not summable on (a, b), then the expression l(y) is called singular. In this case the picture obtained is considerably more complicated. The domain D (A6) of the operator A6 is obtained the same as in the regular case. The domain of definition of the operator A 0 itself does not always allow description by means of the boundary conditions. It follows directly from the Lagrange identity that D (A 0 ) consists of all the func tions y of D (A6) for which [y, z] I� = 0 for all z E D (A6). The deficiency indices n + and n _ of the operator A 0 are always identical (a consequence of the fact that the coefficients of the expression l(y) are real) and can be equal to an arbitrary integer m between 0 and 2n: 0 � m � 2n . Recall that the deficiency index is equal to the number m of linearly independent solutions in L2 [a, b] of the equation
l(y) = A.y
for nonreal A.. A description of all selfadjoint extensions of the operator cannot always be given in terms of systems of boundary conditions. The conditions dis tinguishing the domain of definition of a selfadjoint extension from the set D(A 6) are given in an implicit form (see [37]). The resolvent of every selfadjoint extention is an integral operator with a Carleman kernel (see § 3. no. 6). If the deficiency index of the operator A 0 is equal to 2n, then the kernel is a HilbertSchmidt kernel. In this case the spectrum of an arbitrary self adjoint extension is discrete. In the general case, the spectrum consists of discrete and continuous parts. The continuous part of the spectrum is the same for all selfadjoint extensions.
100
LINEAR OPERATORS IN HILBERT SPACE
More can be said in the case when the expression l(y) is regular at one 1 of the ends of the interval (a, b). Let a be finite and   be summable on Po (x) every interval (a, c) where a < c
yfkl (a) = 0 , k = 0, 1, . . . , 2n

1,
and [y, zr =O for all zED (A;';). The deficiency index can be an arbitrary integer between n and 2n : n � m � 2n. If the deficiency index is equal to n, then the second condition [y, z]b = O is satisfied for all y, zED (A;';); therefore, the domain D (A6) is described by only the first condition yl kl (a) = 0 (k = 0, 1, . . , 2n  1 ). In this case, an arbitrary selfadjoint extension is also described by means of boundary conditions at the regular end : the domain of the extension consists of all functions from D (A;';) satisfying the conditions .
n
where
rjy = L ctjkyrk  l ] (a) = O (j = 1, 2, . . . , n) , k= 1
Conversely, the above system of boundary conditions selects the domain of definition of a selfadjoint extension of the operator A0 from D(A;';) if the aik satisfy the last system of equalities. If the expression l(y) is singular at both ends of (a, b), then the interval (a, b) can be decomposed by an interior point c into two intervals : (a, c) and (c, b) in each of wpich l(y) will be regular at the end c. If we denote by Ari and A� the operators generated by l(y) on the intervals (c, b) and (a, c ) and by m + and m  the deficiency indices of the operators Ari and A�, then the important formula
is valid for the deficiency index of the operator A0 on the entire interval (a, b). In particular, if the deficiency indices of the operators Ari and A� are equal to n, then the operator A0 will be selfadjoint on (a, b). Con versely, if A0 is selfadjoint on (a, b), then, since m+ � n and m � n, the operators Ari and A� have deficiency index n.
ORDINARY DIFFERENTIAL OPERATORS
101
4. Criteria for selfadjointness of the operator A0 on (  oo , oo ). In this subsection several simple criteria are given, in terms of the coefficients of the expression l(y ) , allowing us to establish the selfadjointness of the operator A0 generated by l(y) on the entire axis  oo < x < oo . As was indicated above, these criteria are simultaneously criteria that on the half axes [0, oo ) and ( oo, OJ the corresponding operators A6 and A� have deficiency index equal to n. If the coefficients of the expression l(y) are constant : 
P o (x) = ao i= 0, Pi (x) = a i, . . . , Pn (x) = an ,
then this expression assumes the form d2n2 Y d2ny l(y) = a o 2t. + a i 2n2 + . . . + any . dx dx
The operator A0, generated on ( oo , oo) by the expression l(y) with constant coefficients, is selfadjoint. Several criteria establish that the operator A0 is selfadjoint if its coefficients are, in a wellknown sense, close to constant coefficients. The operator A0 is selfadjoint on ( oo , oo) in each of the cases described below : 1) the limits lim P o =a0 i= O, lim P i =a i, . , lim Pn = an exist ; 

x  oo
x  oo
1
.
.
x  oo
1 2) the functions , P i, . . . , Pn differ from some numbers , a i , . .,. , an , Po
by functions which are summable on (
3) the functions lim p0 (x) > 0.
(; ) 0
'

oo, oo ) ;
. P i , P2 , . . . , Pn are summable on (
ao

oo, oo
x  oo
) and
All these criteria can be generalized if the following property is used : the deficiency index of the operator A0 is not changed by the addition of a function bounded on ( oo , oo ) to the coefficient Pn (x ) . Thus, in par ticular, the selfadjointness of the operator A0, generated on ( oo , oo ) by the expression 

where q(x) is a function bounded on (

) is implied.
oo , oo ,
1 02
LINEAR OPERATORS IN HILBERT SPACE
Stronger assertions are valid for n =2, that is , for the expression [ (y)
=  y" + q (X) y .
The operator A 0, generated by this expression on (oo, oo) will be selfadjoint if the function q (x) is only bounded below or, more generally, if for sufficiently large lxl q (x) �  kx 2 (k > 0) . ,
The operator A0 is also selfadjoint if q(x) EL 2 (  oo , oo ). Other criteria for selfadjointness and nonselfadjointness of the oper ator A0 generated by the selfadjoint differential expression l (y) are given in [37].
5. Nature of the spectrum of selfadjoint extensions. As was indicated, the spectrum of selfadjoint extensions can be both discrete and con tinuous in the singular case. If we consider the expression l(y) on (0, oo ) , then for the satisfaction of condition 3) of the preceding subsection, the continuous part of the spectrum of every selfadjoint extension of the operator A 0 on [0, oo ) coincides with the entire positive halfaxis A � 0. Points of the discrete spectrum can be found on the negative part as well as on the positive part of the axis. Ifp0(x) > O and the conditions 1) of the preceding subsection are satisfied where a 1 , a2 , . . . , an  l are positive, then only the discrete part of the spec trum can be found in the interval (  oo , an). Only the point ). = an can be a point of condensation of the discrete spectrum on (  oo , an) · The question of the nature of the spectrum is one of the most important in the theory of differential operators. It is of special value in problems of quantum mechanics. It is discussed in chapter VII for the differential operators of quantum mechanics. 6. Expansion in terms of eigenfunctions. In the regular case, a complete orthonormal system of eigenfunctions exists for a selfadjoint extension A in terms of which an arbitrary function from L 2 (a, b) can be expanded in a Fourier series. If the function belongs to the domain of definition of the selfadjoint extension, that is, is sufficiently smooth and satisfies the corresponding boundary conditions, then its Fourier series is uniformly convergent. In the singular case, for a selfadjoint extension, continuous spectrum
ORDINARY DIFFERENTIAL OPERATORS
1 03
can appear, and instead of expansions in series there appear expansions in integrals which are also called expansions in terms of eigenfunctions of the differential operator l (y ). Let u 1 (x, 2), u2 (x, 2), . . . , u 2n (x 1 2) be a system of solutions of the equation l (y) = 2y satisfying the initial conditions
1 if j = k ' 1 [k ] ui (X o 0 if j =/: k ,
){
where x0 is a fixed point of (a, b). For every selfadjoint extension A of the operator A 0 generated by the expression l(y), there exists a matrix function
u (.l.) = (uik (.l.)) (j, k = 1, 2, . . . , 2n)
such that for an arbitrary function f(x) E L2 (a, b) the formula
I iJ 1
f (x) =
 oo
is valid, where the integral converges in the mean square sense. The vector function (
b
·
I
are valid for it where the integral converges in L 2 , The analogue of the Parseval equality holds :
a·
I l f (x) l 2 dx = I i. � 1
00
oo
The multiplicity of the spectrum of the operator A does not exceed 2n . The kernel of the resolvent of the operator A is given by the formula a
K
r
(
)
x, s, f1
00
I
 oo
2 n u k (x, .l.) ui (s, 2) duik ( ) L 2  11 i, k= 1 ,
A
,
where the integral converges in L 2 (a, b) in each of the variables x and s for a fixed value of the other.
104
LINEAR OPERATORS IN HILBERT SPACE
In the case when the expression l(y) is regular at one of the endpoints of the interval (a, b), for example at the end a, and the corresponding operator A0 has deficiency index n, the preceding expansions are simpli fied. Every selfadjoint extension in this case is described by means of a system of n boundary conditions at the end a. In the expansion, not every solution u (x, A.) of the equation l(y) = A.y can be taken, but only those solutions which satisfy the corresponding boundary conditions at the end a. Of them, n solutions will be linearly independent. The matrix u (.A) will be of order n. Thus for the expression of second order
( )
dy _!!_ l(y) = + qy p dx dx on the interval (a, oo) (a >  oo) the expansions assume the form 00
f (x) =
I
 oo
and
00
I
where u(.A) is a numerical nondecreasing function and u(x, 2) is a so lution of the equation l(y) = .Ay satisfying the boundary condition
(pu'  8u)x = a
=
0.
The real coefficient (} corresponds to the given selfadjoint extension. The function u(.A) can be found in the following manner: let u 1 (x, 2) and u 2 (x, 2) be two solutions of the equation l(y) =Ay such that
u 1 (a, 2) = 1 , p (a) u � (a, 2) = 0 , u 2 (a, 2) = 0 , p (a) u ; (a, lc) =  l . Since the deficiency index is equal to unity, then for every nonreal A only one combination of the form u 1 (x, 2) + M(.A) u2 (x, 2) belongs to L2 (a, b). The function u (A.) is found from the function M(A.) : � + .<
u(.A) = lim
lim
1 J Im
� + + 0 e + + 0 1t
6
[M (/c + ie)] dA. .
ORDINARY DIFFERENTIAL OPERATORS
1 05
Examples 1 . Let the differential expression
7.
1 (y)
=  y"
be considered on the interval (0, oo ). Then 1 i

Y1 = e
�2
x
and Y 2 = e

1i

�2
x
will be linearly independent solutions of the equation y" = iy. Of these only y2 EL2 (0, oo). The deficiency index of the corresponding operator A0 is equal to 1. The selfadjoint extensions are defined by the boundary conditions y' (0) = 8y (O) , 
8 u (x, A) = cos JAX + sin JAx JA
where 8 is a real number. Here
will be a solution of the equation l(y) =AY satisfying this condition. Calculation shows that 1 M (!l) = 11 · 0  iyr:.
A+O ) extensions for 8 � 0 have a simple continuous Lebesgue spectrum on (0, 00) . The expansions in terms of eigenfunctions have the form If 8 � 0, then a(A) = O for A < O and u' (A) =
f { 00
f (x) = � cJ> (A) cos JAx + n 0 where
f {
n
( j),.
2
.
Hence, selfadjoint
; sin JAx} A fi+ 8 dA ,
00
cJ> (/c) = f (x) cos J2x + 0
l
2
;� sin JAx} dx .
Formulas for Fourier cosine transformations are obtained when 8 = 0.
106
LINEAR OPERATORS IN HILBERT SPACE
2. A Bessel differential expression has the form 1 (y)
=  y" +
v
2
 4
1
 T y
X
( v � 0) .
If this expression is considered on the interval (0, oo ), then it will be singular on both ends. The corresponding operator A0 will be self adjoint for v� 1 ; it will have deficiency index (1, 1) for O < v < l . Since the general solution of the equation l(y) = y has the form
A.
y
=
A jx Jv (x j).) + B jX Yv (x j).) ,
these facts are easily established by the asymptotic behavior of the Bessel functions as z+0 and z+ oo . All selfadjoint extensions of the operator A0 have discrete spectrum on the interval (0, 1). The deficiency index of A0 is equal to 1 on the interval ( 1 , oo); the continuous spectrum of selfadjoint extensions fills out the positive halfaxis. If we consider the selfadjoint estension corre sponding to the boundary condition y (1) = 0, then the expansion in terms of eigenfunctions has the form
where
I JX Pv (x J).) Yv (JA)  Yv (x JA.)Jv (Ji)} f (x) dx . 1
These formulas are called The Weber inversion formulas. The expansions have the form
f (x) = jx Jv (x J).)
I
where
IJx J. (x J).) j (x) dx , 0
on the interval (0, oo) for v � 1 .
ORDINARY DIFFERENTIAL OPERATORS
107
These formulas yield the Hankel transformations. 8. Inverse SturmLiouville problem. Problems on the restoration of a differential expression of given order by the spectral characteristics of a selfadjoint operator generated by this expression are called inverse problems in the spectral theory of differential operators. Here one variant of the statement of the inverse problem is considered. Let the spectral function u (A.), corresponding to some selfadjoint extension A of the operator A 0 generated by the expression l(y) on (0, oo be known for some differential expression of second order
),
l (y) = 1y' 1 + q (x) y . We are required to find the coefficient q (x) in the expression l ( y) and the form of the boundary condition of the corresponding operator A .
There are a number of methods of solution for this problem, one of which is set forth here. We set 2 u (.:i)   JA for A. > 0 r (A)
=
u (.:i)
n
for A < 0
and find the functions
(F y) = sin J2x sin Jiy d ( ' ) I 00
X'
and
.
 oo
Jc
r
A
o2F(x ) , y x = ) f( , y . ox iJ y The integral equation X
f(x, y) + f! (y, s)K(x, s)ds + K(x, y) = O 0 has a unique solution K(x, y) for every fixed x. The function q (x) is defined by the formula q (x) = 21 dK(x,dx x) . The boundary condition, defining the given selfadjoint extension, has
108
the form
LINEAR OPERATORS IN HILBERT SPACE
y' (O)  Oy (O) = 0 , where 8 = K (O, 0) .
Eigenfunctions u(x, A.) of the equation l(y) = A.y satisfying the boundary conditions u(O, A.) = l , u' (O, A.) = 8 can be found by the formula X
cp(x, A.) = cosjix + J K(x, t) cosjA.t dt. 0
The construction described here can be justified under these con ditions : 1) the integral 0
J ev'fifx da(A.)
 oo
exist for an arbitrary real x, 2) the function
f cos jA.x dr (.A.) 00
a (x) =
1
A.
has continuous derivatives to the 4th order inclusively. If the function q(x) has a continuous derivative, then the indicated conditions are necessary for the solvability of the inverse problem. § 6. Elliptic differential operators of second order
1 . Selfadjoint elliptic differential expressions. Let a differential ex pression of second order be given in selfadjoint form : n
lu = 
Ik 1
i, =
(
)
o ou (x)  a;k (x) + c (x) u (x) , OX ; oxk
where x =(x 1 , , xn) is a point of ndimensional space. It is assumed that the coefficients a;k (x) and c (x) and, hence, the expression lu are defined in some region G and on its boundary. The matrix (a;k (x)) is symmetric. The expression lu is called elliptic if all the eigenvalues of the matrices • . .
ELLIPTIC DIFFERENTIAL OPERATORS OF SECOND ORDER
109
(aik (x)) are bounded below, uniformly on G + F, by a positive constant. This definition agrees with the usual definition (see §4, no. 3). In the sequel it is assumed that the region G is bounded, the boundary r is sufficiently smooth, the coefficients a ik (x) have partial derivatives of first order which are continuous in the closed region G + r, the function c (x) is continuous in G+F and c (x) � O. Green's formula for functions of the class CC 2 ) (G + F) (see ch. I, § 1 , no. 5)
Jzuv dx = J ulv dx  I(�>  u �:) ds G
G
r
a
is obtained by integration by parts. Here   denotes differentiation in the ov direction of the conormal at a point of the boundary r :
where we denote by n the unit outer normal vector to the surface r.
2. Minimal and maximal operators. Lharmonic functions. In the Hilbert space L 2 (G) we consider the set D� consisting of finitary functions on G. The equality L�u = lu on D� defines a linear operator which, by virtue of Green's formula, will be symmetric. The operator L� is positive definite. The closure L0 of the operator L� will also be a symmetric operator which is called the minimal operator generated by the expression lu. The domain of definition of the minimal operator consists of all those functions of the space W} (G) (see ch. I, § 2, no. 6), for which u l r=O and ou  =0. ov r The operator L = L&, the adjoint to L0 in the space L 2 (G), is called the maximal operator. The operator L can be obtained as the closure of the operator L' v defined by the formula L' v = lv on the set of all infinitely differentiable functions v(x) on G + r. The deficiency index of the operator L0 is equal to oo. The deficiency subspace U, orthogonal to the range of the operator L0, consists of all solutions of the equation Lu = O. These solutions are called Lharmonic
1 10
LINEAR OPERATORS IN HILBERT SPACE
functions. A smooth £harmonic function u defines some function q> on the boundary r, the values of which coincide with the boundary values of the function u : cp ( s) =
u (s) ( s E r) .
The operator y, realizing the correspondence u+q>, can be extended by continuity to all the space U. Thus a "generalized" boundary value yu is set in onetoone correspondence to each £harmonic function. For a measure of the extensiveness of the deficiency space U, it can be shown that the collection of boundary values of all £harmonic functions con2(n  1) when n > 2 and tains the space L2 (r) and further LP (r) for p � n for p > 1 when n =2. The collection of boundary values can be character ized completely in terms of generalized functions. In connection with the complicated nature of boundary values of £ harmonic functions, the domains of definition of selfadjoint extensions of the operator L0 , generally speaking, do not allow description by means of boundary conditions in classical form. These domains are described by means of boundary operators not allowing, in the general case, repre sentation within the scope of mathematical analysis. Here the description of selfadjoint extensions corresponding to only three basic boundary value problems of mathematical physics is given.
3. Selfadjoint extensions corresponding to basic boundary value prob lems. The strict extension LIL of the operator L0 is defined by the
formula L/Lu = lu on the set W� (G) of all functions of W� (G), which vanish on the boundary r. 0
The set Wi (G) consisting of all functions of W} (G) vanishing on r is the domain of definition of the square root Lt of the strict extension Lw Thus, D (L/L) and D (L!) are characterized by the same boundary condition but by different conditions of smoothness. The solutions of the equation L/Lu f are called solutions of the first homogeneous boundary value problem for the equation lu f The im portant inequality 0
C1 1\ u \l w? :::;; ii L/L \I L2 :::;; C 2 \\ u /\ w� ·
is valid for these solutions.
ELLIPTIC DIFFERENTIAL OPERATORS OF SECOND ORDER
Ill
The representation
is valid for the domain of definition of the operator L adjoint to L0 (see § 4, no. 3). The functions of D (L0) and L� 1 U belong to Wff (G), and consequently the "nonsmoothness" of the functions of the domain D ( L) is due only to Lharmonic terms from U. The operator L0u =lu, defined on all the functions of Wff (G) satisfying ou the boundary condition  =0, defines another selfadjoint extension of ov the operator L 0 • If c(x) $ 0, then the operator L0 is positive definite on L2 (G); if c(x) = O, then it is positive definite on the subspace of L2 (G) orthogonal to the function identically equal to 1. The domain of definition of the square root of the operator L0 coincides with the space Wl(G). Thus, here the boundary condition is removed in 2 0 0 1 the passage from D(L ) to D((L ) 1 ). The solutions of the equation L0u f are called solutions of the second homogeneous boundary value problem for the equation lu f The inequality
is valid for these solutions. If in Wff (G) the collection of all functions satisfying the boundary conou dition + a (s) w[l =0, where the function a(s) � 0 is sufficiently smooth, is OV
sEF
considered, then the operator La, defined by the equality La u = lu on this collection, will be selfadjoint and will have the same properties as L0 • If the function a (s) is only continuous, or more generally measurable, then the picture becomes considerably more complicated. The domain of definition of the corresponding selfadjoint extension already can contain functions not belonging to Wff (G) for which the concept of a conormal derivative can be undefined not only in the classical sense but also in the sense of Sobolev. In connection with this, the operator of differentiation in the conormal direction is extended. It is defined on functions of W� (G) with the help of the usual generalized derivatives. It is additionally defined on some set of (not smooth) Lharmonic functions. Let cp (s) be the bound
1 12
LINEAR OPERATORS IN HILBERT SPACE
ary value of the £harmonic function u (x) of W� (G) : Then the operator P' (yu) =
o (y 1cp) P'cp = ov [ r
au ovr
is defined. The operator P', defined on such functions q>, is symmetric and positive in L2 (r). Its deficiency index is equal to zero. The closure P of the operator P' is a positive selfadjoint operator. An £harmonic function y lq> corresponds to each function q> ED (P ) The collection Up( G) of all these £harmonic functions already contains all the functions which are needed to define the operator of differentiation in the conormal direction. If w = v + u where v E W](G) and u EUp(G) then we set .
ov Dvw = + P (yu) . ov 
It turns out that the operator L, considered on all those functions of the domain of definition of the operator Dv for which
Dvw(s) + a (s) w (s) = 0 (s E r) ,
generates the selfadjoint operator L". The solutions of the equation L"w f are called solutions of the third boundary value problem for the equation lu =f These solutions may not belong to W� (G) but obviously belong to W� (G). The solutions w (x) have a "strong conormal derivative" in the sense that they can be approximated by smooth functions in Wl (G) such that own I   D vw ov
The construction mentioned in the theory of the third boundary value problem goes through not only for bounded measurable functions a (s) but also for aEL2n  2 (r). The domain of definition of the square root of the operator L" coincides with the space Wi (G). The described extension of the operator of conormal differentiation to
113
HILBERT SCALE OF SPACES
Dv is also justified by the fact that Green's formula ] f(Lw)v dx = J[i2>; k (x)0� 0� + cwv k, OX; oxk dx  f(Dvw)v ds remains valid for an arbitrary function w in the domain of definition of the operator Dv and vEW}.
the operator
G
G
r
All the selfadjoint extensions of the operator L 0 considered have completely continuous resolvents ; therefore, the spectrum of every one of these extensions is discrete and the eigenfunctions form an orthogonal basis in L 2 (G). The resolvent will be a HilbertSchmidt operator if the number of independent variables n � 3. For n > 3 only some power of the resolvent will be a HilbertSchmidt operator. § 7.
Hilbert scale of spaces
1. Hilbert scale and its properties. As already remarked in preceding
paragraphs, for several problems the limitations of one Hilbert space become restricting ; and for the investigation of different aspects of the problem, different Hilbert spaces are introduced (see § 3, no. 8 ; § 6). In connection with this, the concept of a Hilbert scale of spaces was recently introduced. Let H0 be a Hilbert space and J an unbounded selfadjoint positive definite operator in H0 such that
(Jx, x) � (x, x) (xED(J))
We denote by Ha for a � O the domain of definition of the power r of the operator J: (for the definition of r, see § 3, no . 4). The space Ha is a Hilbert space with respect to the scalar product
A scalar product is introduced in the space H 0 for a < 0 by this same formula, and the space obtained by the completion of H0 with respect to the corresponding norm is denoted by H«(a < O).
114
LINEAR OPERATORS IN HILBERT SPACE
The set of Hilbert spaces {Ha} ( oo < a < oo) which is obtained is called a Hilbert scale of spaces. A Hilbert scale of spaces has the following properties : 1) If a < f3, then Hp cHa ; the space Hp is everywhere dense in the space Ha and llx ll a � llxll p . 
2) If a < f3 < y, then the inequality
pa y  {J a a � llx ll p llxll � li x ll r
is valid/or x EHy . 3) The spaces Ha and H a are mutually conjugate with respect to the scalar product in H0. In particular, l (x , Y )o l � ll x ll a IIY II a (x E Ha , y E H_a) ·
(We understand by the functional (x , y)0 the scalar product in H0 if x, yEH 0 and its extension by continuity if xEHa and yEH  a). Let H0 and H1 be two Hilbert spaces with scalar products ( x, y)0 and ( x, y)1 and norms llxll 0 and llxll 1 respectively. It is assumed that H1 cH0, H1 is everywhere dense in H0, and ll x ll o � llxll t
(x E Ht) ·
It turns out that there exists an unbounded selfadjoint positive definite operator J on H0 whose domain of definition is the space H1 and such that The operator J is called a generating operator for the pair (H0, H1) . A Hilbert scale of spaces can be constructed with respect to the operator J which will include the spaces H0 and H1 • The operator J, originally defined on the space H1 and mapping it onto the space H0, can be extended to the spaces Ha. Thus, the operator J can be assumed to be extended to an operator J defined on all the spaces Ha ( oo < a < oo) and mapping Ha onetoone onto Ha t · An arbitrary operator J 1 (l > 0) generates the same Hilbert scale of spaces and can also be extended to an operator defined on the entire scale and mapping Ha onto Ha  l· 2. Example of a Hilbert scale. The spaces W�. As the space H0, we
HILBERT SCALE OF SPACES
115
take the space L2(Rn) where Rn is ndimensional space. We denote by am the Fourier transformation of the function u (x) E L2 (Rn) : a ( �) = s ei (x . �) u (x) dx . Let J be the operator setting in correspondence to the function u (x) the function v(x) whose Fourier transformation has the form I� (�) = D (O = J l + 1�/2 a ( () .
The Hilbert scale constructed with respect to the operator J can be described as follows : the spaces Ha for a� 0 consist of all functions for which /l u ll; = J ( l + / ( / 2t / u ( �W d� < oo . For a < O the spaces Ha are obtained by the completion of L2 (Rn) with respect to the above norm. The scale of spaces obtained is denoted by { W;(Rn)}. Apparently, it is basic for many problems of analysis and the theory of partial differential equations. For positive integers a, the spaces WI(Rn) coincide with the Sobolev spaces (see § 1 , no. 2). The question arises whether a Hilbert scale of spaces exists containing the Sobolev spaces W� (G) defined for a region G of ndimensional space. The answer to this question is not known. However, for every N we can construct a Hilbert scale of spaces H;N ) containing all the spaces W� ( G) for 0 ::::; l::::; N. The construction of such a scale can be carried out by means of an extension to the whole space Rn of functions defined in the region G with preservation of smoothness. For 0 ::::; a ::::; N, the norms in the spaces H�NJ will be equivalent to the norms in the spaces WI(G) : for a = m where m is an integer, // u // wr =
J {/u/2 + I ilfl = G
1
jDPuj2} dx ;
for a = m + A., where m is an integer and 0 < 2 < 1 , 2 2 //u// wr + y = {/u/ + f /Dilu/ } dx +
J G
IP I =
1
LINEAR OPERATORS IN HILBERT SPACE
116
There is no effective description of the norm for negative indices, but it can be defined as a norm in the conjugate space [[ u ll w2· =
sup
l l v l l w•2 = t
J u (x) v (x) dx G
(a > O) .
Thus, for an arbitrary set of bounded indices a, the spaces WHG) have the properties of the spaces of a Hilbert scale. In particular: 1 . The space W/ (G) is contained and is everywhere dense in the space WHG) for a
2. The inequality
y p
p a
[[ u il w1 ::::; C ll u ll h"' [[ u il wt (u E W{ (G))
holds for a
1
0
1
HILBERT SCALE OF SPACES
1 17
operator of order I =2m and a system of boundary conditions (with specific conditions on the coefficients and region), generates a homeo morphism between the subspace Wi (bd.) of the space W{ (G) consisting of all functions of Wi satisfying the boundary conditions and the space L2 (G) : A W� (bd.) L2 .
=
Ordinarily it turns out that in this connection the restriction of the operator A to Wi (bd.) n Wi + •(G) Wi + •(bd.) establishes a homeo morphism between the spaces Wi + •(bd.) and WI( G). Further, the operator A allows extension to the closure wt•(bd.) (O � s � l) of the set wJ (bd.) in the metric of the space wJ• (G) and this effects a homeomor phism between the space wJ  • (bd.) and the space W2 • (bd.) conjugate to Wl(bd.). Finally, the operator A can be extended to all spaces W2s(G) so that it establishes a homeomorphism between the spaces W2 s( G) and w2 l  s(bd.). A nonselfadjoint elliptic operator generates an analogous system of homeomorphisms for specified boundary conditions.
=
(s�O)
4. Theorems about traces. Let the spaces fH�} form a Hilbert scale. We consider functions (  oo < t < oo ) with values in the Hilbert space H1 and having continuous derivatives of lth order in the space H0 (in the sense of the norm of the space H0 (see ch. III, § 1 , no. l )). We introduce the norm
x(t)
1 x 1 ; = I {ll x(t) l � , + � �:� � :J dt 00
 oo
in the set m of all such functions. The space m is completed with respect to this norm, and we are concerned about what can be said regarding the values of the functions obtained and their derivatives of order less than I at an arbitrary point of the real axis, for example at the point t =0. This question can be formulated, in another way, as follows : if the sequence of functions xn (t)em is fundamental in the norm of this space, then what can be said about the convergence of the values of these functions and their derivatives at the point It turns out that the operators setting the elements 1, . . . , 1 I ) in correspondence to the functions x(t)e� are continuous operators
t=O?
xk(O) (k=O,
118
LINEAR OPERATORS IN HILBERT SPACE
2k + l from the space m: to the spaces Hak where rxk 1 . Thus, for an 21 arbitrary function from the completion m: of the space m:, we can talk about the values at a point of the function itself and of its derivatives of order less than l. These values belong to the spaces Hak respectively. Conversely, if a set of elements x0, x1 , . . . , x1_ 1 is given such that k xk EHak ' then a function x(t) e m: can be constructed for which x (O) =xk (k = O, 1 , . . . , l 1). Assertions of the type mentioned are important in the consideration of nonhomogeneous boundary value problems and in other problems where it is required to be able to extend functions on manifolds of smaller dimension to a manifold of greater dimension with the guarantee of specified differentiability properties. We come here to the corresponding result for the scale of spaces W� (see no. 2). Let a set ojfunctions u0 (s) , u1 (s), . . . , u1 _ 1 (s) be given on the boundary r of the region G. In order for this set to be able to serve as values on the boundary for a function from Wi(G) and its normal derivatives, it is neces sary and sufficient that the functions uk (s) belong to the corresponding spaces wi(k+f) (r). Propositions of the type indicated are called theorems about traces. =
CHAPTER III
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
§ 1. Linear equations with a bounded operator 1 . Linear equations offirst order. Cauchy problem.
A linear differential
equation of first order has the form dx dt
=
A (t) x + f (t) ,
where f (t ) is a given function with values in a Banach space E, x = x(t ) is an unknown function with values in E and A (t ) is, for each fixed t,
dx
a linear operator acting in the space E. The derivative  is understood dt to be the limit, in the norm of the space E, of the difference quotient
x (t + At)  x (t) as A t�o. At In this section we consider the case where A (t) is a bounded operator for every t. Under this condition, the properties of the solutions of this linear equation are analogous to the properties of the solutions of a system of linear differential equations, which can be considered as a linear equation in a finitedimensional Banach space. The problem of finding a solution of the equation for 0 � t <
oo ,
satisfying a given initial condition x (O)
=
x0 ,
is called the Cauchy problem for the equation considered. A li near equation is called homogeneous if f(t) = e.
2.
Homogeneous equations with a constant operator.
A unique solution
120
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
of the Cauchy problem exists for the homogeneous equation
dx � = Ax dt with a constant bounded operator A and can be written in the form x (t) = eA1x0 •
The operator eAt is defined by the series 2 t A 2 . . . t nA n . . . + + + eAt = I + tA + ' 2! n! which converges in the operator norm. The bound of the norm of each term of the series gives the inequality The operators eAt, bounded operators :
 oo
ll eAt ll � iiJ A II .
< t < oo , form a oneparameter group of
The estimate of the norm of the operator eAt mentioned above is rough cince it takes into account only the norm of the operator A and does not sonsider the distribution of its spectrum. A more precise estimate is contained in the following assertion : if the real parts of all the points of the spectrum of the operator A are less than the number u, then At � Neat . e II II
Conversely, if this inequality is satisfied, it follows that the real parts of the points of the spectrum of the operator A do not exceed u (ReA � u). In particular, it is necessary for the boundedness of all solutions of the equation on the halfaxis 0 � t < oo that the spectrum of the operator A eli in the closed lefthand halfplane, and sufficient that it lie in the open lefthand halfplane. It is necessary for the boundedness of all solutions on the entire axis oo < t< oo that the spectrum of the operator A lie on the imaginary axis. The fact that this condition is not sufficient is verified by the example of a finitedimensional operator with multiple elementary divisors. Let Ao be an eigenvalue of the operator A to which the eigenvector e0 and the associated vectors e 1 , e 2 , , em _ 1 correspond (see ch. I, § 5, no. 6). 
• • .
121
LINEAR EQUATIONS WITH A BOUNDED OPERATOR
Then the equation has particular solutions of the form
x0 (t)
=
e.l.ot e0 , x 1 (t)
=
e.l.ot (e1 + te0), . . . , Xm  1 (t) =
(
)
m 1 t eo . = e.l.ot em  1 + tem  2 + . . . + ! (m  l )
If the eigenvectors and associated vectors of the operator A form a basis in the space E (see ch. I, § 6), then an arbitrary solution is repre sentable in the form of a series of particular solutions of the form indicated. In particular, if the eigenvectors {en} of the operator A form a basis in E, then the general solution of the equation has the form
x (t)
=
2::Cn e.l."ten .
3. Case of a Hilbert space. Let a homogeneous equation with a constant operator be considered in a Hilbert space H. If the operator A is selfadjoint, then the operator etA is also self adjoint and positive definite. If A = iB, where B is a selfadjoint operator, then the operator eiBt is unitary. In order for all solutions of the homogeneous equation in a Hilbert space to be bounded on the entire axis it is necessary and sufficient that the operator A be similar to an operator iB where B is a selfadjoint operator, that is, A = Q (iB) Q  1 , where the operators Q and Q  1 are bounded. As in the general case, it is sufficient for the boundedness of all solutions on the halfaxis O ::::; t < oo that the spectrum of A lie in the open lefthand halfplane. In a Hilbert space, a criterion can be given generalizing the wellknown Lyapunov theorem : in order for the spectrum of the operator A to lie in the open lefthand halfplane it is necessary and sufficient that a bounded self adjoint positive definite operator W exist such that the operator WA + A * W is negative defin ite. In other words, it is necessary and sufficient that there exists a positive definite form ( Wx, x) for which d(Wx, x)
dt
:::::;
 P (x, x) (P > 0)
for an arbitrary solution x(t) of the differential equation.
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
122
4. Equations of second order. For the equation of second order d2 x �2 + Bx = O dt with a bounded linear operator B in the Banach space E, the Cauchy problem consists of finding a solution satisfying the initial conditions
x (O) = x0 and x' (O) = x 1 • The solution of this problem is given by the formula
x (t) = cos t JBx 0 + �
sin t JB
I3 x 1 ,
J
� sin t jB J are defined by the where the bounded operators cos t B and jB senes .
The series converge in the operator norm. In order for all solutions of the equation of second order to be bounded on the whole axis oo < t < oo , it is necessary and sufficient that the sin t JB operator be uniformly bounded with respect to t . JB In a Hilbert space, it is necessary and sufficient for the boundedness of all solutions on the axis oo < t< oo that the operator B be similar to a positive definite operator. 

5. Homogeneous equation with a variable operator. Now in the equation
dx  = A (t) x , dt let the bounded operator A (t ) on the Banach space E depend continuously on t. The solution of the Cauchy problem for this equation exists and is
LINEAR EQUATIONS WITH A BOUNDED OPERATOR
123
unique. It can be obtained by the method of successive approximations applied to the integral equation t
x (t) = x0 +
J A (r) x (r) dr . 0
Finally, the solution can be written in the form
x (t) = U (t) x0 , where the operator U(t) is the sum of the series t
U (t) = I +
t
t
J A (r) dr + J A (r) J A (r1 ) dr 1 dr +
0 0 which converges in the operator norm. The rough estimate
t max
0
···
II A(r) II
eo ,;;. , ,;;. ,
II U(t) l\ :::::;; is valid for the bounded operator U(t). The operator U(t) can be considered as the solution of the Cauchy problem dU dt = A (t) U , U (O) = I for a differential equation in the space of bounded operators acting on E. A bounded inverse operator V(t) = U  1 (t) exists for every t. This operator is the solution of the Cauchy problem for the operator differential equation dV  =  VA (t) , V (O) = I, dt which is called the adjoint to the preceding. If we consider a more general Cauchy problem for the original equation in which the initial condition is given not at the time t =0 but at an arbitrary time t0 : x (t0) = x0 , then its solution can be written in the form
124
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
The operator U (t, r) = U( t) U 1 (r) is called a resolving operator. It has the properties U(t, s) U (s, r) = U (t, r) and
U (t, t) = l .
In the case where A (t) is constant : A (t ) = A , the resolving operator U(t, r) = eA(t  r). It is assumed in this section that the operator A (t ) (0 � t < oo ) is uniformly bounded : \\A (t)\1 � M. Then the estimate \i x ( t )l l � eMt \lx (0)\1 is valid for the solutions of the original equation. The number : ln \\ x (t) \1 u = hm ''t t + 00
is called the index of exponential growth of the solution. Always u� M. We call the least upper bound of the numbers u, for all solutions of the equation, the leading index u The formula •.
: ln II u ( t)\1 u. = hm t t +oo
is valid for it. The special index, defined by the formula ,.* v
_
1.
In II U(t, r) \1tm r, t t+oo t  r
is an important characteristic of the equation. The inequality \\ x (t) \1 � N,e(a* +s) (t  to) 1\x (to)\\ (t � t0) , where N, depends only on e , is valid for every solution for arbitrary e > 0. The relation u. � u* holds between the leading and special indices. If the operator A (t ) is constant, then the leading and special indices coincide. In the general case they do not coincide. For example, the leading index is equal to 1 and the special index is equal to .J2 for the dx ordinary equation  = (sin In t+cos In t)x. dt The leading and special indices are not changed by a translation of the
LINEAR EQUATIONS WITH A BOUNDED OPERATOR
125
dx argument ; that is, by the passage to the equation   = A (t + a) x. If, in the dt equation, we make the substitution for the unknown function x = Q (t )y, where the operator Q(t) is uniformly bounded on the halfaxis O :::;;; t < oo dQ and has a derivative and an inverse Q  1 (t) which is continuous and dt uniformly bounded on this halfaxis, then the function y satisfies the equation
for which the leading and special indices are the same as for the original equation. An equation is called reducible if, by the above substitution, it can be reduced to an equation with a constant operator. The leading and special indices coincide for a reducible equation. The magnitude of the special index depends essentially on the behavior of the operator function A (t ) at infinity. If the limit A = lim A (t ) exists oo
and the spectrum of the operator A00 lies in the open lefthand halfplane, then the special index is negative. If the operators A (t) (0 :::::;; t < oo ) form a compact set in the space of operators, if the spectra of all limits A00 as t+ oo of the operators*) lie in a halfplane Re A :::::;;  v ( v > 0) and if the derivative A' (t) exists and tends to zero as t + oo, then the special index is also negative. The last condition on A' (t) can be weakened by requiring that, for sufficiently large t, the norm II A' (t )II be less than a sufficiently small number b consis tent with v. Finally, instead of the existence of the derivative we can require that the operator function A (t) for sufficiently large t satisfy a Lipschitz condition I I A (t) A (t1 )11 ::::;; ejt t1 1 with a sufficiently small coefficient e . In a Hilbert space we can give the following criterion for negativeness of the special index. If there exists a Hermitian form ( V(t ) x, y) such that
0 < 1X1 (x, x) :::::;; ( V (t) x , x) :::::;; 1X2 (x , x)
and )
d  ( V (t) x( t), x (t)) :::::;;  P (x (t), x(t)), P > 0 dt
Aoo is called a limit for A (f) as t that lim IIA (ft)  Aool l 0. *
tt
An operator
=
HXl
..rn
if a sequence
ft*OO exists such
126
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
for an arbitrary solution x ( t ) of the homogeneous equation, then the special index is negative. Conversely, we can construct a form ( V(t)x, y) with the properties indicated for every homogeneous equation with a negative special index. The operator V(t) can be obtained, for example, by the formula 00
V (t) =
J U * (r, t) U (r, t) dr . t
6. Equations with a periodic operator. In the equation dx = A ( t) x , dt let the operator function A (t) be periodic with a period w :
A (t + w) = A (t) (0 ::::; t <
oo
).
The resolving operator U(t, 0) = U(t) has the property
U (t + w) = U (t) U (w ) . The operator U(w) is called the operator of monodromy of the equation with a periodic operator. The leading and special indices of an equation with a periodic operator coincide and are equal to the logarithm of the spectral radius of the operator of monodromy (see ch. I, §§ 5, 6) divided by the period : (Ts
ln ru(w) * = (T = ___:___.:._ _
Q)
In particular, in order for the special index to be negative it is necessary and sufficien t that the spectrum of the operator of monodromy lie inside the unit circle. If the spectrum of the operator of monodromy does not enclose zero, then the equation with a periodic operator is reducible. It can be reduced by the substitution x = Q (t) y to an equation with constant coefficients by means of the operator t  In
U( w) w Q (t) = U (t) e .
�
LIN AR EQUATIONS WITH A BOUNDED OPERATOR
127
The logarithm of the operator U(w) can be determined by the Cauchy formula In u (w) = � : ( u (w) Ait 1 In A dA , 
1f
2m

r
where the contour r surrounds the spectrum of the operator U ( w) and does not contain the point A = 0, and where In A is some singlevalue, branch of the logarithm. Estimates can be given for the indices a of exponential growth of solutions in a Hilbert space in terms of bounds of Re (A (t ) x, x). If oc1 (t) (x, x) � Re (A (t) x , x) � oc2 (t) (x, x) , then
�IOC1 (t) dt ro
�Ioc2 (t) dt . ro
�a�
0
0
7. Nonhomogeneous equations. The solution of the Cauchy problem with initial condition x (0) = x0 for a nonhomogeneous equation dx =A (t ) x +f (t ) can be written, using the resolving operator U(t, r ) for dt the corresponding homogeneous equation, in the form t
x(t) = U (t, O) x0 +
J U (t, r)f (r) d1: .
0 The question about the boundedness of solutions on (0, oo ) under the condition ofboundedness off (t ) : sup II / (t) ll < oo
O � t < oo
is important for a nonhomogeneous equation. In order for the solution of the Cauchy problem, with the zero initial condition x (0) = 0, for a nonhomogeneous equation to be bounded on the halfaxis (0, oo ) for every bounded function f(t) it is necessary and sufficient that the leading index of the homogeneous equation be negative. With the satisfaction of this last condition, all the solutions of the non homogeneous equation will be bounded on (0, oo ) . If it is additionally known that the operator function A (t) is bounded
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
128
on the halfaxis, then the necessary condition can be strenghened : for the boundedness on the halfaxis (0, oo) of the solution of the Cauchy problem with the condition x(0)=8 for arbitrary bounded f(t) it is necessary (and, of course, sufficient) that the special index of the homo geneous equation be negative. Examples exist of unbounded operator functions A (t) such that all solutions of the nonhomogeneous equation, for an arbitrary bounded f (t ) , are bounded and the special index is positive.*) Criteria for the boundedness of solutions of the Cauchy problem on the halfaxis ( oo, 0) are obtained from the criteria mentioned by the replacement of the sign of the leading or special index by its opposite. Therefore the question about the boundedness of all solutions of the non homogeneous equation on the entire axis for an arbitrary bounded f(t) is meaningless. A question can be posed about the existence of even one, or of only one, bounded solution for arbitrary bounded/ (t ) For the last question, in the case of a constant operator A (t ) = A, there is a final answer : in order that exactly one bounded solution (on the entire axis) of the nonhomogeneous equation with a constant operator A corresponds to each bounded f (t ) ( oo < t < oo ) it is necessary and sufficient that the spectrum of the operator A not intersect the imaginary axis. A nonlinear differential equation of the form 
.

,
dx = A(t) X + f(t, x) , dt
wheref(t, x) for every t is, generally speaking, a nonlinear operator on x, can be considered as linear with a free member f(t, x (t)) ; then the formula for the solution of the Cauchy problem gives the equation t
J
x (t) = U (t , 0) x 0 + U (t , r) f(r , x (r)) dr . 0 If bounds of growth for the resolving operator U(t, r ) are known, in
particular, if the special index a* is known, then bounds for the solutions of the nonlinear equation are obtained from the resulting integral equa tion. In this way analogies of the Lyapunov theorems about the uniform and asymptotic stability of the solutions of a nonlinear equation are obtained. *)
For unbounded A (t) the leading and special indices are defined the same as in the bounded case.
EQUATION WITlt\A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
129
§ 2. Equation with a constant unbounded operator. Semigroups
1 . Cauchy problem. In this section we consider the differential equation dx  = Ax dt with the linear operator A having an everywhere dense domain of defini tion D (A) in the Banach space E. A function x (t) is called a solution of the equation on the segment [0, T] if it satisfies the conditions : 1) the values of the function belong to the domain of definition D (A) of the operator A for all tE [O, T] ; 2) the strong derivative x'(t) of the function x(t) exists at every point t of the segment [0, T] ; 3) the equation x' (t) = Ax(t) is satisfied for all tE[O, T] . The Cauchy problem on the segment [0, T] is the problem of the dis covery of a solution of the equation on [0, T] satisfying the initial con dition x (O) = x0 E D (A) . If the questions of the existence and uniqueness of the solution of the Cauchy problem and of its continuous dependence on initial data were always solved positively for a linear equation with a bounded operator and, hence, basic attention was given to the behavior of the solutions as t�oo, then the questions enumerated become central for an equation with an unbounded operator. We say that the Cauchy problem is correctly formulated (well posed) on the segment [0, T] if: 1) its unique solution exists for arbitrary x0 E D (A) and 2) this solution depends continuously on initial data in the sense that if x0 , n + 8 (xo , n E D (A)), then for the corresponding solutions xn (t ) it follows that x n (t ) + (} for every t E [0, T]. By virtue of the constancy of the operator A, the correctness on an arbitrary segment [0, T1 ] ( T1 > 0 ), that is, correctness on the entire half axis [0, oo ), follows from the correctness of the Cauchy problem on a single segment [0, T] . Let U(t ) be the operator setting into correspondence the value of the solution x ( t ) of the Cauchy problem x(O)= x0 at time t to every element x0 eD (A). If the Cauchy problem is well posed, then the operator U(t)
130
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
is defined on D(A) and is linear and bounded. Therefore it can be extended by continuity to a bounded linear operator, defined on the entire space E, which we also denote by U(t ) . A family of bounded linear operators U(t ) , depending on the parameter t (0 < t < oo ) , is called a semigroup if
(0 < t1 , t 2 < oo ) .
The operators U(t ) , generated by a well posed Cauchy problem, form a semigroup. Thus, the solution of a well posed Cauchy problem is representable in the form
x (t) = U (t) x0
(x0 eD (A)) ,
where U(t ) is this semigroup of operators. If x0 does not belong to the domain of definition of the operator A, then the function U(t) x0 may not be differentiable and its values may not belong to the domain D (A) of the operator A. We can call the function U(t) x0 a generalized solution of the equation x' = Ax. 2. Uniformly correct Cauchy problem. A correctly formulated Cauchy problem is called uniformly correct if it follows from x0 , n + (} that the solutions xn (t) tend to (} uniformly on every finite segment [0, T] . If the operator A is closed and its resolvent RA (A) (see ch. I, § 5, no. 6) exists for some A., then uniform correctness follows from the existence and uniqueness of a continuously differentiable solution of the Cauchy problem for arbitrary xeD(A). The semigroup U(t) is strongly continuous for a uniformly correct Cauchy problem, that is, the function U(t ) x0 is continuous on (0, oo ) for arbitrary x0 E E. We say that semigroup U(t) belongs to the class ( C0) if it is strong ly continuous and satisfies the condition lim U (t) x0 = X0 0
t + +
for arbitrary x0 EE. The semigroup U(t), generated by a uniformly correct Cauchy prob lem, belongs to the class ( C0 ) . In other words, we can say that all general ized solutions are continuous on [0, oo] in this case.
EQUATION WITH A CONST ANT UNBOUNDED OPERATOR . SEMIGROUPS
131
The limit lim t+
ln II U (t) ll t
OCJ
=w
exists for an arbitrary strongly continuous semigroup U(t). If the semi group belongs to the class ( C0 ), then the estimate II U (t) ll � Mewt
is valid for it. Thus, for a uniformly correct Cauchy problem, the orders of exponen tial growth of all solutions are bounded above. If w = O, then the semigroup is bounded, I I U(t) ll � M, and the Cauchy problem is uniformly correct on [0, oo ) In this ca.s e, an equivalent norm can be introduced in the space E, for example .
ll x ll 1 = sup II U (t) x ll ,
O � t < oo
in which the operators U(t) have a norm not greater than one : II U (t) ll 1 � l .
The semigroup is called contractive in this case. The Cauchy problem for the equation of thermal conductivity: ov o 2 v Jt = ox 2 ' v (O, x) = cp (x)
(  oo < x < oo , O � t < oo)
is one of the simplest examples of a uniformly correct Cauchy problem. Let the space C( oo , oo ), consisting of all continuous bounded functions on the xaxis, serve as the space E. Here the operator A is the second derivative operator with respect to x defined on the set D (A), dense in C ( oo , oo ) consisting of all twice continuously differentiable functions v (x) for which v andv" E C (  oo , oo ). The Cauchy problem is uniformly correct, that is, a unique solution v(t, x) of the thermal conductivity equation, having the property that lim v(t, x) = cp (x) uniformly with respect to x , exists for an arbitrary . t + + 0 function cp E D (A). Furthermore, if cpn (x)ED(A) converge uniformly to cp (x)E D (A), then the solutions vn (t, x ) � v (t, x ) uniformly with respect to x and t on every finite segment [0, T] of variation of t. 
,
132
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
The corresponding semigroup U(t) of bounded operators is given by the integralformula ofPoisson 00
[U (t) cp] (x) =
1 I
2Jnt
( x  s) 2
e  4t cp (s) ds
(t > 0)
 oo
and consists of contractive operators.
3. Generating operator and its resolvent. For the semigroup U(t) the question is raised : for which elements x0 will the function U(t) x0 be differentiable? Differentiability of this function for arbitrary t follows from its differentiability for t = 0. The linear operator U (h) x0  x 0 U' (0) x0 = lim h h+ + 0 is defined on the elements x0 for which U(t) x0 is differentiable at zero. The operator U' (0) is called the generating operator of the semigroup U(t). If the semigroup belongs to the class ( C0), then the domain of defini tion D of the generator U' (0) is everywhere dense; it is closed and commutes with the semigroup on its domain of definition :
U' (O) U (t) x0 = U (t) U' (O) x0 If the Cauchy problem is uniformly correct for the equation x' = Ax, then the operator A allows closure. This gives the generator of the cor responding semigroup U ( t ) :
A = U' (O) . The Cauchy problem is uniformly correct for the equation x' = Ax where A is the generating operator of a semigroup of class ( C0). Thus, if we restrict ourselves to equations with closed operators, then the class of equations for which the Cauchy problem is uniformly correct coincides with the class �f equations for which A is a generator for a semi group of class ( C0 ) . This explains the role which the study of semigroups and their generating operators plays in the theory of differential equations.
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
133
The spectrum of the generato � of a semigroup of the class (C0) lies alway sin some halfplane Re A � w. The class of generating operators can be characterized by the behavior of the resolvents RA (A) of the operators : in order for the operator A to be the generator of a semigroup in ( C0), it is necessary and sufficient that a real OJ and a positive M exist such that
(k = 0, 1 , 2, . . . ) .
for A > OJ
If the operator A in the equation x' = Ax is closed, then the conditions mentioned are necessary and sufficient for the uniform correctness of the Cauchy problem. The estimate I I U(t )II � Mewt for the corresponding semigroup is valid under the satisfaction of the indicated list of conditions. The verification of the necessary and sufficient conditions is difficult since all the powers of the resolvent appear in them. They will be obviously satisfied if
(Jc > OJ) . Such a condition is satisfied, for example, for the equation of thermal conductivity (see example, no. 2). In the presence of the last estimate, the inequality II U ( t) ll � ewt is valid for the semigroup. In particular, if OJ= 0, then I I U(t) II � 1 and the semigroup is contractive. It should be emphasized that the satisfaction of the condition M II RA (.Jc) ll � A  OJ 
(A > OJ)
with M> 1 is not sufficient for the correctness of the Cauchy problem. The semigroup U (t) can be constructed in terms of the resolvent RA(Jc) by the formula
U ( t) x =  lim � 2n:z .... oo
It + iv
f e;.t
It  lv
RA (A) x dA ,
which is valid for xeD(A), t > O and sufficiently large positive
JJ.
The
1 34
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
integral converges uniformly on every interval 0 < B::::;; t ::::;;
1
B
.
The limit
of the integral as t+0 is equal to xf2. The resolvent of the operator A is the Laplace transform (with opposite sign) of the semigroup : 00
RA (Jc) x = 
J0 e  ;.' U (t) x dt .
The integral converges when the real part of Jc is sufficiently large. A uniformly correct Cauchy problem is always the limit of Cauchy problems with bounded operators in the following sense : a sequence of bounded operators A. exists such that the solutions of the problems dx. = A .x., x.(O) = x0eD(A) converge to the solution of the problem dt dx = Ax, x (O) =x0. Moreover, the convergence is uniform on every dt finite interval [0, T]. The operators A. can be constructed by the for mula A. =  nl n2 RA (n) =  nA RA (n). The semigroup U(t ) is the limit of the operatorfunctions e  nARA ( nlt where the convergence is uniform on an arbitrary finite interval [0, T] . 
4. Weakened Cauchy problem. In no. 1 it was required that the solu tion of the equation satisfy the equation for t = 0 also. This requirement must often be weakened. A function x(t), continuous on [0, T], strongly differentiable and satisfying the equation on (0, T], is called a weak solution of the equation x' = Ax on the segment [0, T] . We understand by the weakened Cauchy problem on [0, T] the problem of finding a weak solution satisfying the initial condition x(O) = x0• Here the element x0 does not have to belong to the domain of definition of the operator A . If we leave aside the question of the existence of a solution of the Cauchy problem, then rather general conditions for its uniqueness can be pointed out. The condition
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
135
(this limit is always nonnegative) is sufficient for the uniqueness of the solution of the weakened Cauchy problem. The solution is unique on [0, T h] and can branch for t > T h. If h = 0, then the solution of the weakened Cauchy problem is unique on the entire halfaxis (0, oo ) . On the other hand, a differential equation x' =Ax with an operator A for which IIRA (2)11 < g (2), having a nontrivial solution with the initial condition x (0) =8 exists for every function Q (2) > 0 satisfying the ln Q (2) condition oo (2  oo ). 2 If a real w, a positive M and a /3, 0 < {3 � 1, exist such that for the re solvent of the operator A M IIRA (a + ir) ll � a  w + lr l p +
in the halfplane a > w, then the weakened Cauchy problem has a unique solution on [0, oo) for an arbitrary x0 from the domain of definition of the operator A (x0 ED(A)). The weakened Cauchy problem, in this connec tion, will be correct but not uniformly correct. When the last condition on the resolvent is satisfied, then the solution of the weakened Cauchy problem is given by the formula x(t) = U(t) x0, where U(t) is a strongly continuous semigroup. If x0 ¢;D (A), then the generalized solution U(t ) x0 can be discontinuous at the point t = 0. However, it is always Abel summable to its initial value : 00
lim 2
). + 00
J eM U (t) x0 dt = x0 •
0 If the estimate given above with index /3 > i is valid for the resolvent of the operator A, then we have T
J II U (t) ll dt < oo 0
for the operator U(t). The weakened Cauchy problem in this case will be correct in the mean : T
if x0, n+8, then the integrals J l l xn(t)ll dt tend to zero for the solutions. 0 The derivative of the solution of the weakened Cauchy problem can be discontinuous at the point t 0, but its norm is summable on an arbitrary =
136
finite interval :
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE T
J ll x' (t) ll dt < oo . 0
5. Abstract parabolic equation. Analytic semigroups. An equation x' = Ax, for which the weakened Cauchy problem with arbitrary initial condition x0 E E is correct, is called an abstract parabolic equation. Thus, a unique solution of the weakened Cauchy problem exists for an abstract parabolic equation for arbitrary x0 EE, and this solution continuously depends on initially given data. If the operator A is closed and has a resolvent for sufficiently large positive 2, then the Cauchy problem is uniformly correct for the abstract parabolic equation, and the operator A coincides with the generating operator of the semigroup U(t) of class ( C0) by means of which the solu tion x (t) = U(t) x0 is given. If the Cauchy problem is uniformly correct for the equation x' = Ax, then, in order for the equation to be abstract parabolic, it is sufficient that lim (ln r I I RA (a + ir) ll) = 0
for some real a. Every generalized solution of the abstract parabolic equation is weak and, hence, differentiable for t > O. It follows from the commutivity of the generating operator and the operators of the semigroup that every generalized solution is infinitely differentiable. The operators A k U(t) (t > O, k = O, 1, 2, . . . ) are linear operators, bounded for every t > O. The norms of the operators A k U(t), generally speaking, are not bounded as t+0. An important class of abstract parabolic equations is formed from equations for which all generalized solutions are analytic functions of t and can be analytically extended to some (fixed for a given equation) sector of the complex plane containing the positive real halfaxis. The semigroup U(t) is itself analytically extended to some operatorfunction U(z) which is analytic in the sector. In the sequel such semigroups are called analytic. In order for a semigroup to be analytic it suffices that the estimate ·
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR . SEMIGROUPS
1 37
for the resolvent R;. (A) of the operator A be satisfied in the halj:plane _ Re A > w. The angle of the sector of analyticity can be defined in the following manner : it follows from the indicated estimate for the resolvent that the analogous estimate holds in some sector  cp < arg (A.  w) < cp (cp > i) ; then the semigroup 1t
U (z) is analytic in the sector  t/1 < arg z < t/1, where 1t
t/J = cp   . 2 If the semigroup belongs to the class ( C0), then the indicated estimate for the resolvent is necessary for its analyticity. The estimate Cewt (0 < t < 00 ) ll x o ll ll x' (t) ll � t
holds for the derivatives of all generalized solutions of the equation x' = Ax. Conversely, if the estimate Cewt II A U (t) il �
t
is valid for a semigroup of the class ( C0), then this semigroup has an analytic extension U(z) in some sector containing the positive halfaxis . The inequality k kC k II A U (t) ll �  ewt .
( ) t
holds for the norms of the operators A k U(t). It remains to remark that it follows from the satisfaction of the estimate for the resolvent for some w that it will be satisfied (possibly with another constant M) for an arbitrary w lying to the right of all the points of the spectrum of A. 6. Reverse Cauchy problem. The problem of finding a solution
1 38
LINEAR DIFFERENTIAL EQUATIONS I N A BANACH SPACE
on the interval [0, T] with a given terminal value x ( T) = Xr E D (A) . is called the reverse Cauchy problem. By a replacement of the independent variable, the reverse Cauchy problem for the equation x' = A x can be reduced to the Cauchy problem for the equation x' =  Ax. If the Cauchy problem is well posed for an equation, then the reverse Cauchy problem, generally speaking, is not well posed : for some Xr the solution will, in general, not exist, for others it will cut off, not reaching zero ; the solution will not depend continuously on initially given data. If the direct and reverse Cauchy problems are uniformly correct for an equation, then the operators U(t) are defined and bounded for all t (  oo < t < oo ) and form a group where U (  t ) = u  1 (t). In order for the direct and reverse Cauchy problems with a closed operator A to be uniformly correct, it is necessary and sufficient that constants M and w > 0 exist such that I IR� (.A.) II �
M (I.A.I
_
Y
w
( n = I , 2, 3, . . . )
for all real A with I .A. I > w . The spectrum of the operator A, in this connec tion, lies inside the strip I.A. I < w . We digress from the question of the existence of a solution to the reverse Cauchy problem and consider only the question of the uniqueness of the solution and its continuous dependence on terminal values. The reverse Cauchy problem is called correct (well posed) on [0, T] in the class of bounded solutions if a number b = b (e, M, t0) can be found for every positive M, e and t0 e (O, T) such that the inequality ll x (t0) 11 � b is satisfied for an arbitrary weak solution on [0, T] satisfying the conditions II x (t ) II � M (0 ::::; t ::::; T) and II x (T) II ::::; e. If all the generalized solutions of the equation x' = Ax are analytic in some sector, then the reverse Cauchy problem is well posed in the class of bounded solutions on an arbitrary segment [0, T] . This fact follows from the inequality ll x (to) ll � M ll x (O) II 1  w (to ) llx ( T) II w (to) , where the continuous function w (t), O � w (t) � 1 does not depend on the choice of the generalized solution.
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
139
Equations in a Hilbert space. The differential equation x' =  Bx in a Hilbert space H with an unbounded selfadjoint positive definite operator B is the simplest example of an abstract parabolic equation. The Cauchy problem is uniformly correct for this equation, and the semigroup U(t) corresponding to it can be written in the form 7.
U (t) = e Bt ,
where the function e Bt is defined by means of the spectral decomposition of the operator A (see ch. II, § 3. no. 4) : _ e  Bt 
co
J0
e  u dEA .
The semigroup U(t) will be contractive, II U(t) ll � 1 . All generalized solutions x(t) = e  »t x0 (x0 eH) are solutions of the weakened Cauchy problem and allow analytic extension to the right hand halfplane Re z > 0. The behavior of the weak solutions can be studied in more detail as t+ 0. An arbitrary (nonintegral) power of the operator B is defined by means of the spectral expansion co
If x(O)eD(B a), a: > O, then the inequality
l x' (t) ll �
c
i =a
t
I Bax(O) II
is valid for the solution of the weakened Cauchy problem. In the case a: = � the solutions allow a more precise characterization : in order for the derivative x' (t ) of the solution x (t ) to have a square integrable norm on [0, T], /
T
J0
l x'(t) ll 2 dt < oo ,
it is necessary and sufficient that x(O)e D( Bl).
140
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
It remains to note the useful inequality
t 1 t ll x (t) ll � ll x (O) II T II x (T) I IT ,
which is valid for an arbitrary generalized solution. A linear operator C is called an operator of fractional order with respect to the selfadjoint positive definite operator B if it is defined on D(B) and the operator CB  a is bounded on D (B) for some a:e(O, 1). The greatest lower bound y of the numbers a: is called the order of the operator C with respect to B. In order for the operator C to be an operator of fractional order y with respect to the operator B it is necessary, and if C allows closure it is sufficient, that the inequality
where Ka does not depend on x and b, be satisfied for xeD(B), ex > y and sufficiently small b. If the operator C is an operator of fractional order with respect to a selfadjoint positive definite operator B, then the equation x' =  (B + C)x is abstract parabolic. Generalized solutions will be analytic in some sector containing the positive halfaxis. The behavior of the solutions as t+0 will be the same as for the equation x' =  Bx. In a Hilbert space the generating operators for several important classes of semigroups can be completely described. The description is made in terms connected with the operator itself and not with its resolvent. 1) In order for the operator A to generate a strongly continuous contrac tive semigroup of operators it is necessary and sufficient that it be a max imal dissipative operator with an everywhere dense domain of definition (see ch. II, § 4, no. 4). This criterion can be formulated differently as follows : in order for a closed operator with an everywhere dense domain of definition to be the generating operator of a strongly continuous contractive semigroup of operators it is necessary and sufficient that the conditions
R e (Ax, x) � 0 (x e D (A)) and Re (A * x, x) � 0 (x eD(A * )) be satisfied.
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
141
2) If the operators A and A* have the same everywhere dense domain A + A* of definition and the operator Re A =  is bounded, then the operator 2 A is the generating operator of a semigroup U (t ) of class ( C0) for which the estimate II U (t ) I I � ewt is valid, where w is an upper bound of the operator Re A. 3) In order for the operator A to be the generating operator of a strongly continuous semigroup of isometric transformations ( II U(t) II = 1), it is neces sary and sufficient that it be a maximal dissipative conservative operator with a dense domain of definition. In this case A = iB where B is a maximal symmetric operator. 4) In order for the operator A to be the generating operator of a strongly continuous group of unitary operators, it is necessary and sufficient that A= iB where B is a selfadjoint operator. The group of operators U(t) is representable in terms of the spectral decomposition of the operator B in the form co
U ( t) = eiBt =
J
 oo
ei't dE ;. .
The direct and reverse Cauchy problems are uniformly correct on the entire axis. Generalized solutions U(t) x0 are differentiable only when x0 eD(A). 5) If the condition /Im (Ax, x)/ � [3 /Re (Ax x)/ ,
is satisfied for a maximal dissipative operator, then it is the generating operator for a semigroup allowing analytic extension to the sector 1t
/ arg z/ <   arctan [3. 2 The estimate
is validfor the resolvent ofthe operator in the sector /arg A/ � 1t  arctan f3  e for arbitrary e > 0. The equation x' =Ax is an abstract parabolic equation ; the properties of its solutions are analogous to the properties of the solutions of the equation x'  Bx with a selfadjoint positive definite operator B. =
142
LINEAR DIF'FERENTIAL EQUATIONS lN A BANACH SPACE
Recently, many of the aboveenumerated criteria that an operator be generating for semigroups of one class or another have been extended to certain classes of Banach spaces. 8. Examples of well posed problems for partial differential equations. 1 . Cauchy problem for the diffusion equation. Let b (x) be a continuous bounded function on the entire axis  oo < x < oo . We consider the problem
au a 2 u au  = 2 + b (x)  , at ax ax
u (O, x) = cp (x)
(

00
< X < 00, t > 0) .
a2 The domain of definition of the operator 2 is described in the example ax a2 a the operator 2 + b(x)  is in no. 2. The domain of definition of ax ax assumed to be the same. The Cauchy problem is uniformly correct in the space C( oo, oo ) . 2. Boundary value problems for strongly parabolic systems of equations. Let G be a bounded region in an ndimensional space with a sufficiently smooth boundary r; x = (xu x2 , , xn) a point of the region G ; a a ) L(x ,  an Ndimensional matrix with each element L i i(x,  ) a ax ax linear differential operator of order 2m of the form • • .
(i , j
=
1 , 2, . . . , N) ,
a1 + ·· · + an a where rx = (rxt. . . , r:xn) (rx; � O), lrxl = rx1 + rx2 + · · · + rxm D a = The aX1 . . . a Xn coefficients a\jl (x) are assumed to be real and sufficiently smooth. The a matrix L(x , ) is assumed to be strongly elliptic, that is, ax N I a�j) (xH�' . . . � �n'1il1j > o (  It I .
!Z [
i, j = l lal = 2m
for arbitrary xeG and real �k and 171 with L �� # 0 and L 11? # 0 . For the system of equations
( )
au a   =  L x'  u (x E G ' t > 0) at ax
'
!Zn
.
EQUATION WITH A CONSTA NT UNBOUNDED OPERATOR. SEMIGROUPS
143
one formulates the first boundary value problem of finding a solution u(t, x) = (u 1 (t, x), . . . , uN(t, x)) satisfying the initial condition u (O, x) = cp (x) and the boundary conditions u
au a m  l u=0 = ... = m anl r an r 1
r
'
where n is the direction of the normal to the boundary r. a The strongly elliptic expression L (x, ) generates a linear operator ax in the space L2 (G) of vector functions u(x) whose modulus squared is summable, defined on the smooth functions satisfying the boundary con ditions. The operator allows the closure L, having the properties �
and
Re (Lu, u) � q (u, u) (u e D (L)) Re (L*u, u) � q (u, u) (u e D (L* )) .
Thus, the operator ql L will be a maximal dissipative operator. From this follows the uniform correctness of the first described boundary value problem for the equation u; =  L u in the space L2 (G). The estimate II U (t)ll � e qt is valid for the corresponding semigroup. An analogous problem can be considered in the space LP (G) of func tions for which 1 (p > 1 ) . l l u iiLp = {J l u (x)J P dx} P < oo It turns o ut that for arbitrary p > 1 , the estimate
is valid in some sector larg A. I < cp ( cp >  ) for the resolvent of the operator 2  L generated by a strongly elliptic matrix. Th us, the semigroup generated by a strongly parabolic system for the first boundary value problem in LP (G) is analytic. The solution of the prob lem exists for an arbitrary initial function cp(x)E 1:1,( G) and is an analytic n
144
LINEAR DIFFERENTIAL EQUATiONS IN A BANACH SPACE
function of t. The initial condition is satisfied in the sense that ll u (t, x)  tp (x)ll rp + 0 a&
t + 0 .
The estimate of the derivatives
where C does not depend on cp, is valid for all solutions. For the simpler case of one equation with a scalar function u ( t, x), the concept of a strongly elliptic differential expression coincides with the concept of an elliptic expression (for real coefficients!) (see ch. II, § 6, no. 1 ). The indicate.d estimate is valid in an arbitrary sector Jarg .1.1 � cp with cp < n for the resolvent of the corresponding operator L . Therefore, the solutions are analytic in the entire righthand halfplane. Recently, operators were studied which are generated by a strongly a elliptic expression L(x,  ) on functions satisfying the boundary conax
ditions not of the first boundary value problem but some general boun dary conditions (the socalled Lopatinskii conditions). In the spaces LP (p > 1), the estimates obtained for the resolvents of operators are analogous to those mentioned above. Thus, for the solutions of corre sponding boundary value problems for a strongly parabolic system, the same conclusions are valid as for the first boundary value problem . 3. Cauchy problem for an equation with constant coefficients. We con sider the equation
where
X = (x1 , . . . ,
xn) is a point of the ndimensional space Rn and
a a P (  ,  ) is a linear differential expression with constant coefficients at ax
containing derivatives with respect to t of order not greater than m  1 . . l) 1 = urn the equation can Wtth the su bstltutwn u = u1 0 u1 = u2, . . . , u(mbe reduced to the system '
aul
at
'
1
·
. . a um 1 = Urn , a�m = . = Uz, , at
at
f pk (ax8 )
k= l
U
k.
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
145
The system thus obtained can be treated as the equation u; = Au in the Hilbert space of vector functions L2 (Rn)· The operator A is the closure of the operator defined by the matrix
0 0
1 0
0 1
0 0
0 0 0
0
0
0 0 0
0
1 pk
on sufficiently smooth functions such that Pk(
a
) uk E L2 (Rn)·
ax The question is raised : when will the equation u; = Au be abstract parabolic? It turns out that it is necessary and sufficient that these con ditions be satisfied : a) the equation is correct according to Petrovskil, that is, for all roots of the equation sm = P(s, i�) 
for an arbitrary real vector � = (�1, . . . , � n) , the inequality Im s < C is satisfied, where the constant C does not depend on the choice of the vector � ; b) for arbitrary a > O a b can be found such that llm sl � a 1n 1 Re sl  b . If the stronger inequality
h I Im s l � a 1 1 Re sl  b1
is satisfied for some real b1, a1 > 0 and h � 1 , then the semigroup cor responding to the problem will allow analytic extension to some sector containing the positive real halfaxis. Finally, if h > 1 , then the semigroup is analytic in the righthand halfplane. 4. Schrodinger equation. In the 3dimensional space R3, the equation
at/f i at 
is considered, where
L1
= 
L1t/J + v (x) t/1
is the Laplace operator.
146
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
Under certain conditions on the function v (x ), the opera tor H, obtained by closure of the operator defined by the differential expression  Ll + v (x ) on finitary functions, will be a selfadjoint operator in the Hilbert space L2(R3). The direct and reverse Cauchy problems are uni formly correct on the entire axis. The equation generates a group of unitary operators U(t) = e  iHt (see ch. VII, § 1 , no. 4). 5. Symmetric hyperbolic systems. We consider the system of equations au au  = IA;(x) � + Bu , at ax;
where x = (x 1 , . . . , xn) ; u(t, x) is an mdimensional vector function ; A ; (x) (i= 1 , 2, . . . , n), B (x) are mdimensional square matrices depending smoothly on x ; A; (x) are symmetric matrices. If, for example, it is assumed that A;( x) and B(x) are periodic func a ax.
tions in all variables, and the operator A;  + B is considered on periodic '
differentiable functions, then the real part of its closure will be the a A. bounded operator B + B*  � (case 4, no. 7) . The problem of the ax. '
'
determination of periodic (with respect to spacial coordinates) solutions of the system will be uniformly correct on an arbitrary segment [0, T] in the space L2 (Q), where Q is a period parallelepiped. If the condition * aA ; ( i = 1 , 2 , . . . , n) B + B    � 0 ax.
'
aA ; . . . . . . 1s sat1sfi ed , that 1s, th e matnces B + B* are negatlve d efimte, then ax;
the corresponding operator will be a maximal dissipative operator and the corresponding semigroup will consist of contractive operators. a1 ; Upon satisfaction of the condition B + B* �0 in some region G, ax. a the operator A;  +B will be dissipative on the functions satisfying the ax. '
'
condition
J (L A;n; u, u) dx � 0 r
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR . SEMIGROUPS
1 47
on the boundary r of the region, where n 1 are the components of the normal to r. For example, this condition is satisfied on finitary functions. Recently, the forms of boundary conditions defining maximal dissi pative extensions of the corresponding operator were studied in detail.
9. Equations in a space with a basis. Continual integrals. If a basis { e;} exists in the Banach space E, then solutions of differential equations cah be sought in the form of expansions with respect to the basis. In this subsection, several formulas for solutions of a formal nature will be men tioned. Questions about the convergence of the expansions will not be considered. If a basis { e1 } of eigenvectors of the operator exists in the space E, then the solution of the Cauchy problem can be written in the form x t) L c1e;.'1e1 ,
A x' =Ax, x(O) = x0
( = where the A; are eigenvalues of the operator A and the coefficients c1 can be found from the expansion x0 = L c 1 e1• If we now consider the equation dx = (A + B) x, i
i
dt
then the solution of the Cauchy problem for it can also be sought in the form L a1( t ) e;. An infinite system of differential equations
x(t)=
da . dt
"'
' = ,A.a.' ' + f' b'kak is obtained for the functions a;( t ) , where ( b;k ) is the matrix which is assigned to the operator B in the basis (e;) . The method described of finding solutions in the form of their expan sions with respect to a basis of eigenvectors of the operator is called the Fourier method. Now let an arbitrary basis { e ;} be given in the space E, forming a biorthogonal system with the collection of linear functionals {!;}. Let x1 (t) be the solution of the Cauchy problem
A
dx = Ax ' dt
x1 (0) = e1 •
148
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
The system {x; (t)} is called a fundamental system of solutions with respect to the basis {e;} . The matrixfunction with elements si i(t) /;(x1(t)) is called the fun damental matrix of the equation in the basis { e;} . If x0 is the element having the expansion x0 = I c ;e ; in the basis, then i
the solution of the Cauchy problem with the initial condition x(O) = x0 can be written in the form i, j
The functions s i i(t) satisfy the identity
) ;k (t) sk1 (r) = s ii (t + r) . 2 k If the operator A is such that si i (t) ;;?: O and 2: Si i (t)= 1 , then the j numbers s i i(t) can be treated as the probabilities of the passage of some system from the state e; to the state e1. In this case, the matrix (Si i(t)) des cribes the socalled Markov process with a denumerable number of states. Now let the operator B be such that the functionals /;(x) are eigen vectors for the operator B* with eigenvalues ex; : f; (Bx) = cxJ; (x) . We formulate the problem of finding the fundamental matrix (S ii (t)) for the equation dx  = (A + B) x . dt It turns out that this matrix can be evaluated by the formula n+ l I IXjk.dtk k s ;;, (.d t 1 ) s;,;2 (At2) S;"1 (Atn + 1 ) , Sii (t) = lim I I . . . I e = t 1i = 1 i2 = l in = 1 n+ 1 where A tk > O and I A tk = t ; the limit is taken with respect to par00
00
00
• • •
k= !
titions of the segment [0, t] as max A tk +0. The last limit is often written in the form of a socalled continual integral
Sii (t)
=
t
f0a(t)dt
Je
dJ1;1 .
EQUATION WITH A CONSTANT UNBOUNDED OPERATOR. SEMIGROUPS
1 49
In this connection, we understand the integral
J (a(r)) dJ.lii of the functional 4>, defined on the set of all stepfunctions values in the set of eigenvalues of the operator B*, to be lim max
00
I L . . . I 4> 1 1
Lltk + 0 i! = 1
00
iz =
00
in =
a(r) with
(ai!, . .,;Jr)) S;;1 (A t1 ) . . . S;n1 (A tn + 1 ) ,
where the function !X i t , ··· . d r) is equal to Jl ik on the segment A tk. If we consider the equation
dx = (A + f (B))x, dt then its fundamental matrix is written in the form
Up to now the case of a discrete basis was considered. Sometimes the solutions are expanded in terms of continual bases. Thus, for example, we take the collection of all generalized eigenfunctions of a selfadjoint opera tor in a Hilbert space with a continuous spectrum as such a basis (see ch. II, § 3, no. 8). Then, in all the preceding formulas, the infinite sums are replaced by integrals. In particular, the continual integral is no longer the limit of nmultiple sums, but the limit of nmultiple integrals.
Example. The heat conduction equation : ou ot
The continual basis consists of functions
e1 =
&(x  i) (  oo < i < oo)
(see ch . VIII, § I , n o. 1), the eigenfunctions of the operator of multiplica
1 50
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
tion by the variable. The fundamental system of solutions (see no. 2) is u i (t, x) =
1
2Jnt
00
f
e
 (x  s)2 4t
ii (s  i) ds =
 oo
1
2 Jnt
e
 (x  i)2 4t
The fundamental matrix is 00
sii (t) =
f
ui (t, x) ii (x  i) dx = ui (i) =
 oo
1 2Jnt
e
(i  N 41
•
The continual integral
f cP (ct(r)) d
Jl ii
on the interval [0, T] is defined as follows : a partitioning of the interval [0, T] is made by the points 0 = t0 < t1 < t2 < < tn< tn+l = T. We con sider all continuous functions ct(t) which are linear on each of the sub intervals Atk such that ct(O) = i and ct(T) =j. Let ct (tk) = ctk (k= 1 , 2, . . . , n) ; then the functional cP (ct ( r)) becomes some function of n variables cP (ct(r)) = cP(ct 1 , ct2, . . . , ctn) on the class of functions considered. Then n · · ·
00
�
00
where ct0 = i and ctn+ l =j. The continual integral obtained is called a Wiener integral. The fundamental matrix for the equation 2 ou 8 u at = ax2 + V (x) u can be written in the form
EQUATION WITH A VARIABLE UNBOUNDED OPERATOR
151
§ 3. Equation with a variable unbounded operator
1. Homogeneous equation. When considering the homogeneous equation dx (0 :::;; t :::;; T)  = A (t) x dt with an operator A (t ) which depends on t, we generally assume that for every t this operator is the generating operator for a semigroup having certain properties. Moreover, the dependence of A (t) on t is assumed to be smooth. In order to formulate conditions of smoothness for an unbounded operatorfunction A (t ), it is natural to assume that this function is defined for different t on the same elements of the space. In connection with this, the following assumption is made in this subsection. The domain of definition of the operator A (t) is not dependent on t :
D (A (t)) = D . We can now formulate two sets of conditions which guarantee the correctness of the statement of the Cauchy problem for an equation with a variable operator: I. l ) The operator A (t) is the generating operator of a contractive semi group for every t E [O, T] and, moreover, the inequality
ll RA( t) (A.) fl
:::;;
1 for Re A. ;;:,:: 0 1 + Re A.
is valid for the norm of its resolvent. 2) The bounded operatorfunction A (t) A  1 (s) is strongly continuous ly differentiable with respect to t for arbitrary s. ILl) The operator A (t) is the generating operator of an analytic semi group for every t E [0, T] and the estimate
IIRA (t ) (A.) fl
:::;;
c for Re A. ;;:,:: 0 , 1 + [ A. [
where C does not depend on t, is valid for the norm of its resolvent. 2) The operatorfunction A (t) A  1 (s) satisfies the Holder condition
ff [A (t)  A (r)J A  I (s) lf :::;; C1 J t  r[1 ,
where C1 does not depend on
s,
t and
r
and 0 < y :::;; I .
1 52
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
When the sets of conditions I or II are satisfied, then there exists a resolving operator U(t, s) (t ;,:s), bounded and strongly continuous with respect to the variables t and s, for 0 :!{,_ s :!{,_ t :!{,_ T. In this connection U(s, s) = l. The solution of the Cauchy problem for the equation x' = A (t) x with the initial condition x(O) = x0 E D is unique and is given by the formula x (t) = U(t, 0) x0 • If the initial condition is given for t = s, x (s) x0, then the solution has the form x(t)= U(t, s) x0• The operators U(t, s) have the property =
U (t, r) = U (t, s) U (s, r) Generalized solutions U(t, 0) x0 of the equation x' = A (t) x for arbitrary x0 EE are continuous functions on [0, T]. If the conditions II are satisfied, then every generalized solution is a solution of the weakened Cauchy problem, i.e., is differentiable and satisfies the equation for all c1 tE(O, T]. The estimate ll x' (t) ll :!{,_  ll xo ll holds for derivatives of the weak t solutions. For the conditions I, the resolving operator U (t, s) can be constructed as a "multiplicative integral" n U (t, s) = lim TI UA <,, > (A t) . i= I Here s = t0 < t 1 < · · · < t n = t is a partitioning of the segment [s, t], A ti = ti  ti_ 1 , 1:i are interior points of the segments [ti_ 1 , tJ ; UA <•> is the semigroup generated by the operator A ( r ) The limit is taken for max A t, 0 and exists in the strong sense. The resolving operator U(t, s) can be obtained for the conditions II as a solution of the integral equation .
�
t
U (t, s) = UA (s) (t) +
J U (t, r) [A (r)  A (s)] UA(s) (r  s) d1:. s
If the operator A (t ) allows an analytic extension to a region contain ing the interval (0, T), then, with the satisfaction of the conditions II, all generalized solutions U(t, 0) x0 will be analytic in some region con taining (0, T).
EQUATION WITH A VARIABLE UNBOUNDED OPERATOR
153
Remark. In the conditions I and II, A. can be replaced everywhere by 2  w and the operator A (t) A  1 (s) by the operator (A (t)  wl) x (A (s) wl) l This case reduces to consideration of the replacement x = e"'t y . 
.
2. Case of an operator A (t ) with a variable domain of definition. In the case when A (t) is a differential operator, its domain of definition usually consists of sufficiently smooth functions satisfying certain boundary con ditions. The independence of the domain of definition D (A (t )) of t assum ed in the preceding subsection means, in the applications, the independence of t of the coefficients in the boundary conditions ; therefore, the removal of this condition is of considerable interest. It turns out that some frac tional power A a (t) of the operator A (t) (for the definition of frac tional powers of operators, see no. 4) has, in several cases, a domain of definition consisting of functions not restricted by boundary conditions and, hence, the domain of definition does not depend on t. Condition 2) of the set II can be replaced by the following : 2') For some aE(O, l ), the operators A a (t ) have a domain of definition independent of t, where the operator Aa (t ) A  a (s) satisfies the Holder condition with the index y > 1  a : If conditions 1) of group II and 2') are satisfied, then a resolving operator U (t, s) exists which has the same properties as in the satisfaction of the conditions of II. It remains to remark that the verification of condition 2 ' ) is difficult, since there are no explicit formulas for fractional powers of concrete, for example differential, operators. This condition can be verified for oper ators in a Hilbert space which satisfy conditions 5), § 2, no. 7. Let A (t) be a maximal dissipative operator for every t for which the form (A (t) x, x) has the properties ! Re (A (t) x, x)l ;;:. <5 (x, x) and IIm(A (t) x, x)l :::;; p IRe (A (t) x, x)l , where the constants b >0 and P do not depend on t. Then the operator A (t) satisfies condition 1) of II. The form (A (t ) x, x) can be extended by means of closure to a form A t (x, x ), defined on a wider set of elements than D(A (t )) . It is assumed that the domain D of the form At (x, x) does
154
LINEAR
DIFFERENTIAL EQUATIONS IN A BANACH SPACE
not depend on t and the form satisfies a Holder condition of the form
iAt (x, x)  A, (x, x)i :::;; M i t  rjY (x, x) on D. The domain of definition of the operator A a (t) for O < a < t does not depend on t under these conditions, and Thus, if y ;,:: t, then the operator A (t) satisfies condition 2').
3. Nonhomogeneous equation. The method of variation of para meters can be applied to solve a nonhomogeneous equation dx = A (t) X + j (t) . dt �
Then the solution of the Cauchy problem with the initial condition x ( 0) = x0 can be formally written as t
x (t) = U (t, 0) x0 +
I U (t, s) f (s) ds . 0
If the resolving operator U (t, s) is strongly continuous with respect to t and s and the function f(s) is continuous, then the above integral exists and is a continuous function of t. However, generally speaking, it will not be a differentiable function of t; therefore, we must assume that the above formula gives a generalized solution of the nonhomogeneous equation. If the conditions of set I, no. 1, are satisfied, then the generalized solu tion will be a true solution of the Cauchy problem under the conditions that x0 E D and the function f(t) is continuously differentiable. The last condition can be replaced by the following : f (t )ED for all t E [0, T] and the function A (0) f (t ) is continuous. If the conditions of set II, no. 1 are satisfie d, then the generalized solu tion will be a solution of the weakened Cauchy problem (see § 2, no. 4) for an arbitrary x0 E E and functionsf(t), satisfying the Holder condition
llf (t) f (r) ll :::;; C i t  rio
(0 < b :::;; 1).
Moreover, if x0 ED, then the above formula provides a true solution of the Cauchy problem. In the investigation of a nonhomogeneous equation this fact is useful :
EQUATION WITH A VARIABLE UNBOUNDED OPERATOR
1 55
the operatorfunction A(t) U(t, s) A  1 (0) is bounded and is a strongly continuous function of the variables t and s under conditions I or II. Moreover, if the function A (t) A  1 (0) has a second strongly continuous derivative, then the operatorfunction A 2 (t) U(t, s) A  2 (0) is bounded and is a strongly continuous function of t and s. 4. Fractional powers of operators. For selfadjoint positive operators in a Hilbert space, fractional powers are defined by means of a spectral expansion (see § 2, no. 7). Let A be a closed linear operator with an everywhere dense domain D (A), having a resolvent on the negative halfaxis, and satisfying the condition (A. > 0) . The operator
00
is defined for 0 < 1X < l and x E D (A). The operator I allows a closure which is called a fractional power of a the operator A and is denoted by Aa. Moreover, if the operator A has a bounded inverse operator A  1 , then we can directly define bounded operators which are fractional powers of the operator A  1 : 00
The operator A  a is the inverse of the operator A a. A negative power A a can be defined by the formula 00
for the indices IX > 1 which makes sense for fractional IX contained between 0 and n. If IX tends to an integer k < n, then lim A  a = A  k where the limit is understood in terms of the operator norm.
LINEAR DIFFERENTIAL EQUATIONS IN A BANACH SPACE
156
a < oo ) form a semigroup of bounded operators The operators of class C0). This semigroup is uniformly continuous (with respect to the operator norm) for t > 0. If the operator is the generating operator of a strongly continuous semigroup of operators U(t) for whose norm the estimate (w > O) I U (t) ll :::;:; Me rot
(
Aa(o::;;; B I
is valid, then the estimate
(2 > w) is valid for the resolvent of the operator and, hence, fractional powers can These powers can be expressed by a be defined for the operator . semtgroup :
A= B. 00
A a = ( Bf" = r (a) fra1 U (r) dr 1
(0 < ct. < 1 ) .
0
An important "inequality of moments" holds for fractional powers of operators : if a and f3 have the same sign and I a I < 1 /31 , then
x
where the constant K(a, /3) does not depend on the choice of the element (if /3 > 0, then (O < ct < 1 ) and it The resolvent is defined for 2<0 for the operator can be found by the formula
xED(AP)).
Aa
It follows from this formula that the inequality
(2 > 0) , with the same constant M as for the analogous inequality for the operator is valid for the operator
A,
A a.
EQUATION WITH A VARIABLE UNBOUNDED OPERATOR
157
It follows from the inequality II A.(A + AJt 1 ll ::::; M for all A. > O that the operator A.(A + AJt 1 is uniformly bounded in every sector of the complex plane I arg A. I ::::; cp for cp not greater than some number n  t/1 (0 < t/1 < n). Then the operator A.(A" +AJ) 1 will be uniformly bounded in every sector ! arg A. I ::::; cp for cp < n  rxt/J. In particular, the operator (  A )t is the generat ing operator of an analytic semigroup. If the operator  A is a generat ing operator of class (C0), then the operator (  A )" will be the generating operator of an analytic semigroup for 0 < rx < I . If O < rx, fJ < I , then (A")P = A "P. The following fact is very important in applications. If A and B are two positive selfadjoint operators in a Hilbert space and D(B) :::;) D (A), then the inclusion D (B") :::;) D (A")
holds for 0 < rx < I . More generally, let A and B be two positive selfadjoint operators acting in the Hilbert spaces H and H1 , respectively. If Q is a bounded linear operator from H into H1 such that QD (A) c D (B) and II BQ u ll :::;; M I ! Au l l
(u eD (A)) ,
then QD(A") c D (B") and (u e D (A ")) . li B" Q u i! ::::; M1 II A"u l l 1  ". Q = M" II II We can set M 1 An analogous statement with another constant M1 is valid for the case when A and B are maximal dissipative operators in the spaces H and H1 respectively. For Banach spaces, the last inequality is proven for the replacement, in the right hand side of the inequality, of the norm I I A" u ll by the norm !lAP ull with fJ > rx. The validity of the inequality for identical indices is not established. Fractional powers of operators play an essential role in the investiga tion of nonlinear differential equations.
CHAPTER IV
NONLINEAR OPERATOR EQUATIONS
Introductory remarks In this chapter the equation
is considered, where
A
x = Ax
is an operator (generally speaking, nonlinear)
defined in some Banach space E with range of values in the same space. The operator 1
Ax(t) = J K [t, s, x (s)J ds 0
and, in particular, the operator 1
Ax (t) = J K (t , s)f[s, x (s)J ds 0
can serve as examples of the operator The first of these is usually called a is called a
Hammerstein
operator.
A. Urysohn
operator while the second
The first question occurring in the study of the indicated equation involves the existence of a solution. This question is often formulated
in this form : is there a fixed point for the transformation A? The operator
A can be
defined on a part T of the space E; then we talk
about fixed points of the operator which belong to T. If we are required to find a solution having an additional property, then a subset T0 c T of elements having this property is selected, and a fixed point is sought in T0 • For example, in the problems where non
negative solutions are sought, we set T0 = T n K where K is the corre sponding cone of the nonnegative elements of E (see ch. V).
NONLINEAR OPERATORS AND FUNCTIONALS
1 59
The second question involves the uniqueness of a solution ; that is, the uniqueness (in T0) of a fixed point of the transformation A. Theorems of nonuniqueness are of basic interest for non linear equa tions in many cases, that is, existence theorems for two or more solutions. For example, in various problems of stability theory and the theory of waves, a (trivial) solution of the problem is known in advance and the determination of other (nontrivial) solutions is a fundamental goal. In those cases when the solution is not unique, the problem concerns itself with the number of solutions or with upper and lower bounds for this number. We often consider the equation x = A (x; A.) , where the operator A (x; A.) depends on a numerical parameter A, (in some problems the parameter can be an element of some space). For equations with a parameter, several new problems arise which are connected with a change in the number of solutions for a change of the parameter. Those critical values of the parameter A, for which the solutions branch or merge are of particular interest. The simplest example of an equation with a parameter is given by the problem involving eigenvalues and eigenvectors of a linear operator, that is, the problem concerning solutions of the equation I
x =  Ax ;,
'
where A is a linear operator. Here the trivial solution x = () occurs for all .l. # O. Those values of A, for which other solutions appear are called eigen values. Analogously, ). is called an eigenvalue for a nonlinear operator if the equation Ax = A,x has a solution x # (); x is called an eigenvector of the operator A . §
1. Nonlinear operators and functionals
I . Continuity and boundedness of an operator. Let the operator A be defined on the set T of the Banach space E and let its values belong to the Banach space £1 • The operator A is called continuous at the point x0 e T if II Ax.  Ax0II +O follows from ll x.  x0 11 + 0 (x.e T).
160
NONLINEAR OPERATOR EQUATIONS
An operator A is called weakly continuous at the point x0 if the weak convergence of Ax" to Ax0 follows from the weak convergence of the sequence x" to x0• Sometimes we consider operators A which transform a weakly con vergent sequence of elements into a strongly convergent sequence, and operators A which transform a strongly convergent sequence of elements into a weakly convergent sequence. If the space E1 is the number line, then the operators with values in the space E1 are called functionals. An operator A is bounded on Tif sup II Ax \I < oo .
xeT
In contrast to linear operators, the boundedness of a nonlinear oper ator on some sphere does not imply its continuity. The continuity of the operator A on some set T implies its boundedness on a neighborhood of each point (local boundedness) ; but, an operator A which is continuous at every point of a closed sphere might not be bounded on the entire sphere (ifthe space E is infinitedimensional). The operator defined on the entire space 12 by the equality Ax = (�I•
�� . . . . , ��. . ) (x = (el, �z, . . . , �"' . . . )) . .
can serve as an example of such an operator. This operator is continuous at every point of the space 12 , but it is not bounded on any sphere S(O, r) for r > 1 . If the set T is compact, then a continuous operator A is bounded on the set. An operator A is completely continuous on T if it is continuous and transforms every bounded part of the set Tinto a (relatively) compact set of the space E1 • An operator A satisfies a Lipschitz condition on Tif
An operator A, satisfying a Lipschitz condition, is continuous. 2. Differentiability of a nonlinear operator.
Operators, defined on sub sets of the real line, are called abstract functions. Let x (t ) (a � t � b) be an
NONLINEAR OPERATORS AND FUNCTIONALS
161
abstract function with values in the Banach space E. The derivative x'(t) of the function x(t) is defined as the limit as A t+0 of the difference quotient : x (t + A t)  x (t) 1. X ' ( t) = Im . o At Llt + If the limit is considered with respect to the norm of the space E, then the derivative is strong; if the limit is considered in the sense of weak convergence in the space E, then the derivative is called weak. An operator A, acting from the Banach space E into the Banach space £1 , is differentiable in the sense of Frechet at the point x0 if a bounded linear operator A' (x0) exists, acting from E into E1 , such that where
A (x0 + h)  A (x0) = A' (x0) h + m (x0 ; h) , .hm \\ m (x0 ; h) \\ = 0 . 1\ h \\ [lh[l + 0
The operator A' (x0) is called the Frechet derivative of the operator A at the point x0 • We say that the operator is uniformly differentiable on the sphere T = { II x \I ::::;; a} if I lim \\ m (x0 ; h) \\ = 0 llh [l + 0 \\ h \ 1 uniformly with respect to x e T. A bounded linear operator A' (x0) is called the Gateaux derivative of the operator A at the point x0 if
A (x0 + th)  A (x0) . A' (x0 ) h = hm t t+0 =
'...:_� __

for all h e E. In other words, A' (x0) is called the Gateaux derivative if A' (x0) h is a strong derivative at the point t = O of the function A (x0 + th) of the variable t with values in the space E1 :
d A' (x0) h =  A (x0 + th) dt
t = .()
The Frechet derivative ofthe operator A, if it exists, is also the Gateaux derivative. If the Gateaux derivative A' (x) exists in a neighborhood of the
NONLINEAR OPERATOR EQUATIONS
162
point x0 and is continuous at the point x0 (as an operator on x), then it is the Frechet derivative. We call the expression A' (x0) h the Frechet differential (the Gateaux differential, respectively) of the operator A at the point x0 • If the operator A is completely continuous, then its Frechet derivative A' (x0) is a completely continuous linear operator. If the operator A has a Gateaux derivative A ' (x) on the convex set T, then the equality l (A (x + h)  Ax) = l (A' (x + r h) h) , where T = r (l)e(O, 1), holds for every pair of points x, x + heT and every linear functional l from the space Ei conjugate to E1 • This equality is called the Lagrange formula. An operator A is called asymptotically linear if it is defined on all the elements x with a sufficiently large norm and if a linear operator A' ( oo ) exists such that II A (x)  A' ( oo ) x II 0 . = lim II II II X II X
+ 00
The operator A' ( oo) is called the derivative at infinity of the operator A.
3. Integration of abstract functions. The Riemann integral of an abstract function x(t) (a ::::;; t ::::;; b) with values in a Banach space E is defined as the limit of Riemann sums : b
J x (t) dt = a
lim
max Jtk+ 0
I x (rk) A tk k =
1
(A tk = tk  tk _ 1 ; tk _ 1 ::::;; rk ::::;; tk) ·
If this limit exists for an arbitrary sequence of partitionings of the interval [a, b] and does not depend on the choice of this sequence and the choice of points Tk , then the function x (t ) is called integrable according to Riemann. The integral is called strong if the sums converge to it with respect to the norm of the space E; if the sums are weakly convergent, then the integral is called weak. A strongly continuous abstract function is strongly integrable accord ing to Riemann. It remains to remark that the norm of an abstract func tion, strongly integrable according to Riemann, can be a scalar function which is not integrable according to Riemann.
NONLINEAR OPERATORS AND FUNCTIONALS
163
The usual properties of an integral hold for the integral of an abstract function. In particular, if the abstract function x (t) has a continuous derivative x' (t) on [a, b], then the formula b
J x' (t) dt = x (b)  x (a) a
is valid. The integral representation of the increment of an operator A having a continuous Gateaux derivative : 1
A (x + h)  A (x) =
J A' (x + th) h dt 0
stems from this formula. The Bochner integral is a generalization of the Lebesgue integral to abstract functions. The abstract function x(t) (a � t � b) is called inte grable according to Bochner if it is strongly measurable and l l x(t) l l is a scalar function summable according to Lebesgue. In this connection, an abstract function x(t) with values in the space E is called strongly meas urable if it is a uniform limit of a sequence of finitevalued functions. If the space E is separable, then the concept of strong measurability of an abstract function coincides with the concept of weak measurability. A function x (t) is called weakly measurable if the scalar function f(x(t)) is Lebesgue measurable for every linear functional feE*. The Bochner integral is defined in the following manner : first let x (t) be a finitevalued function k
x (t) = x ; (t e A ;, A ; n A i = O (i # j ) , U A ; = [a, b]) . Then
b
J
(B) x (t) dt = a
1
t1 X; mes A ; . ;
Now let x(t) be an arbitrary function which is integrable according to Bochner, and let xn(t) be a sequence of finitevalued functions converging to x(t) . Then, by definition, b
(B) x (t) dt = !�� (B)
J a
b
J x" (t) dt . a
164
NONLINEAR OPERATOR EQUATIONS
The Bochner integral has many of the usual properties of the Lebesgue integral. In particular,
II (B)
b
b
a
a
f x(t) dt l � f l x (t) l d t.
b
(B)
f x(t) dt = l, (B) f x(t) dt. a
A
If is a bounded linear operator acting from the space E into the space with values in E is integrable according £1 , and the abstract function to Bochner, then the function with values in E1 is also integrable according to Bochner and
x(t) Ax(t)
(B)
b
b
a
a
f Ax(t) dt = A ((B) f x(t) dt) .
A x(t)
If is an unbounded closed linear operator, if the values of the func tion belong to its domain of definition and the function is integrable according to Bochner, then the preceding equality is also valid. In this connection, the integral on the right belongs to the domain of definition of the operator
Ax(t)
A. 4. Urysohn operator in the spaces C and Lr Let the function K(t, s, x) be continuous with respect to the collection of variables (0 � t, s � 1 , l x l �r). Then the Urysohn operator 1
Ax(t) = JK[t, s, x(s)] ds 0 is defined on all the functions of the sphere S(O, r) of the space C, and its values belong to C. The operator A is completely continuous on S((), r). If the continuous derivative K; (t, s, x) exists, then the operator A is differentiable in the sense of Frechet at every interior point x0 (t) of the
NONLINEAR OPERATORS AND FUNCTlONALS
1 65
sphere S(O, r ). Its derivative A ' (x 0) is defined by the formula 1
A' (x0) h (t) =
J K� [t, s , x0 (s)] h (s) ds .
0 For the operator A to be asymptotically linear, it suffices that the func tion K(t, s, x) satisfy the condition I K (t, s , x)  Kro (t, s) x j ::::;;
(x)
The operator A' ( oo) is expressed by the formula 1
A' (oo) h (t) =
J Kro (t, s) h (s) ds . 0
It is necessary for the consideration of the Urysohn operator on the entire space C or in the space LP that the function K( t, s, x) be defined for all values 0 ::::;; t, s ::::;; I,  oo < x < oo . If this function increases faster than an arbitrary power with respect to the variable x (for example, contains exponential nonlinearities), the Urysohn operator will not be defined on any space Lr In connection with this, restrictions are imposed on the growth of the function K(t, s, x) with respect to x. Let IK (t, s , x) j ::::;; R (t, s) (a + b j x j a0) (0 ::::;; t, S ::::;; 1 ,  00 < X < 00) , where cx 0 � 0 and the function R ( t, s ) is summable, with respect to both variables, to some power /30 > 1 : 1
1
J J i R (t, s) jP0 dt ds < oo . If
0 0
CXo
::::;;
f3o

1,
then the Urysohn operator acts, and is completely continuous, in every space
1 66 LP,
NONLINEAR OPERATOR EQUATIONS
where p > 1 and
In some cases it is convenient to consider the Urysohn operator as an operator acting from one space Lp , into another space LP2 • The Urysohn operator acts from Lp , into LP2 and is completely continuous if rxoflo P 1 > 1 , P1 �  1 , 1 < P2 ::::;; Po · flo Here, the condition rx0 ::::;; flo  1 is, naturally, not assumed to be satisfied. If the function K(t, s, x ) contains essentially nonpower nonlinearities, then in several cases the Urysohn operator is completely continuous in some Orlicz space. As in the case of the space C, it is natural to look for the derivative of the Urysohn operator, acting in the space LP, in the form of an integral operator with a kernel K�(t, s, x0 (s)) . However, the differentiability of the Urysohn operator as an operator acting in LP does not follow from the continuous differentiability of the function K(t, s, x) with respect to the variable x. For example, the operator 1 (0 ::::;; t ::::;; 1 ) sin (ex<s>) ds Ax (t) =
J 0
acts and is completely continuous in any Lr However, the operator 1 exo(s) cos exo(s) h ( s) d s 0 is not defined on LP if the function exo(t) is not summable. In order for the integral operator with kernel K� [t, s, x0 (s)] to be the Frechet derivative of a Urysohn operator acting in the space LP (p > 1 ) , it suffices that the function K�(t, s, x) be continuous with respect to x and that the inequality (0 ::::;; t, s ::::;; I,  CX) < X < CX) ) J K; (t, s, x)l ::::;; a + b J x l p  1
J
be satisfied.
NONLINEAR OPERATORS AND FUNCTIONALS
1 67
Operator f Let the function f(t, x) be defined for O � t � l ,  ro < x < oo . It is assumed everywhere in the sequel that the function f(t, x) is continuous with respect to x and measurable with respect to t for every x. The equality fx (t) = f[t, x (t)] defines an operator! Iff(t, x) is continuous in both variables, then the operator f acts in the space C and is continuous and bounded on every sphere. If the operator f acts from the space LP1 into the space LP2, then it is continuous and bounded on every sphere. In order for the operator f to act from LP 1 into LP2 it is necessary and sufficient that the inequality PI P2 lf(t, x) l � a (t) + b l xi be satisfied, where a(t)eLp2 · One must keep in mind that the operator f does not have the property of complete continuity (except for the trivial case when f(t, x) does not depend on x). If the function/ (t, x) and its derivative f� (t, x) are continuous in both variables, then the operator f, considered as an operator in the space C, is differentiable in the sense of Frechet. Its Frechet derivative has the form f ' (x0) h (t) =f� [t, x0 (t)] h (t) . Let f act from LP1 into LP2. The differentiability in the sense of Frechet of the operator f does not follow from the existence of the continuous derivativef�(t, x). The inequality PI 2 1 P , If� (t, x) l � a 1 (t) + b 1 l x i where a 1 (t)e L P I P2 , serves as a sufficient condition for which the operator 5.
PI  P2
of multiplication by f�[t, x0(t)] is the Frechet derivative of the operator f acting from LP 1 into LP2, where p 1 > p2• In some cases the operator f is differentiable at certain points of the space LP 1 and not differentiable at others. 6. Hammerstein operator. If the kernel K(t, s) is continuous, then the linear operator 1
Kx(t) =
f K (t, s) x(s) ds 0
168
NONLINEAR OPERATOR EQUATIONS
acts from an arbitrary space LP and from the space C into every space Lp, and into the space C and is a completely continuous operator. In orderfor the operator to act from Lp , into LP2 and be completely continuous, it suffices that the inequality 1 1
f0 f0
IK(t, s) l ' dt ds
{
< oo
p1 be satisfied, where r = max p 2 , . p1  1 The Hammerstein operator 1 Ax(t) = Kfx (t) =
}
f K (t, s) f [s, x (s)] ds 0
is a particular case of the Urysohn operator considered above. The pos sibility of the representation of the Hammerstein operator in the form of a product Kf allows the search for less restrictive conditions for complete continuity of this operator in the spaces LP . For the complete continuity of the Hammerstein operator in the space LP it suffices that the operator fact from LP into some space LP 1 and that the linear operator K be a completely continuous operator acting from LP1 to LP . If the operator f, considered as an operator from LP into LP •' is differ entiable at the point x0 and the operator K acts from Lp, to LP , then the Hammerstein operator A = Kf is also differentiable at the point x0 ; moreover, 1
A' (x0) h (t) = Kf ' (x0) h (t) =
J K (t, s) J: [s, x0 (s)] h (s) ds . 0
7. Derivatives of higher order. Derivatives of higher order of abstract functions are defined in the usual manner. The situation is more compli cated with the derivatives of higher order of operators. An operator B(x1, x2 ) (x 1 , x 2 eE) with values in the space E 1 is called bilinear if it is a bounded linear operator with respect to each variable. A bilinear operator B (x 1 , x 2 ) is called symmetric if
B (x 1 , Xz) = B(x2 , x 1 ) .
NONLINEAR OPERATORS AND FUNCTIONALS
169
If we set x 1 = x2 = x in the symmetric bilinear operator B(xl > x2), then the operator B2 (x) = B (x, x) is called a quadratic operator. An operator A , acting from the Banach space E into the Banach space E1 , is called twice differentiable in the sense of Frechet at the point x0 if
A (x0 + h)  A (x0) = B 1 (h) + 3_B2 (h ) + w 2 (x0 ; h) , where B 1 = B 1 (x0) is a linear operator with respect to h, B2 = B2 (x0) is a quadratic operator with respect to h and . llwz (x0 ; h)li hm = 0. 2 ll h ll ll h ll > 0 The quadratic operator B2 (x0) is called the second Frechet derivative of the operator A at the point x0 : B2 (x0) = A" (x0) . The expression B2 (x0) (h) is called the second Frechet differential of the operator A at the point x0. We sometimes consider the second successive Frechet derivative of the operator A , the Frechet derivative of the first Frechet derivative. The second successive Frechet derivative is a bilinear operator with respect to the increments h and h 1 • If a function K(t, s, x), which is continuous with respect to the collec tion of variables, is twice continuously differentiable with respect to x, then the Urysohn operator, acting in the space C, has a second Frechet derivative 1
A" (x0) h =
f0 K;2 [t, s, x0 (s)] h2 (s) ds .
The second successive Frechet derivative of the Urysohn operator is given by the formula 1 A" (x0) (h 1 , h2 ) = K;2 [t, s, x0 (s)] h 1 (s) h 2 (s) ds . 0 A quadratic operator B(x0) is called the second Gateaux derivative of the operator A at the point x0 if
f
d2 A (x0 + th) l , = o 2 dt
=
B (x0) (h)
170
NONLINEAR OPERATOR EQUATIONS
for arbitrary heE. The expression B(x0) (h) is called the second Gateaux differential of the operator A at the point x0 • The second successive Gateaux derivative of an operator A is defined as the Gateaux derivative of the first Gateaux derivative. It is a bilinear operator. One must keep in mind that, generally speaking, the existence of the second Gateaux derivative does not follow from the existence of the second Frechet derivative of the operator A. For example, the scalar 1 function f(t)= t3 cos 2 has a second Fn!chet derivative which is equal to t zero at the point t = O, but it does not have a second Gateaux derivative at this point. 8. Potential operators. Differentiable functionals are a particular case of differentiable operators. If
where 1 is a linear functional depending on x and
lw (x ; h)l . lim
llhll > 0

ll hll
�
=
0.
If the functional
X
for ll x ii # O.
ll x ll
1 71
NONLINEAR OPERATORS AND FUNCTIONALS
The norm is a differentiable functional in the spaces
1
cP (x) = ll x ll then
=
1
LP (p > 1 ) : if
P dsy , J i x (s)I ( 0
lx (s)l p  1 sgn x (s) grad cP (x) = ll x ll p  1 :___:___:_:__=,,__:___:_ 
for l x ll #0. The Gateaux gradient of the functional cP(x) is defined by the equality
d cP (x + th) dt
t=
o
= T(x) (h) .
An operator which is the gradient of a functional is called a potential operator. A bounded linear selfadjoint operator A acting in a Hilbert space H can serve as an example of a potential operator. It is the gradient ofthe functional cP (x) =HAx, x). The operator fx (t) = f[t, x (t)] , acting from
LP into LP' (� +; = 1 ) , is another example of a potential ,
operator. It is the gradient of the socalled HammersteinGolombfunctional cP (x) =
LP.
1
t x( )
0
0
J [f f (t, u) du] dt ,
defined on the space Let B be a bounded linear operator acting from the Banach space E into the Banach space £1 . Let cP (y) be a differentiable functional defined on £1 . Then the functional
F (x) = cP (Bx) ,
defined on E, is also differentiable and grad cP (x) = B * grad i/> (Bx) , where B* is the adjoint of B.
172
NONLINEAR OPERATOR EQUATIONS
§ 2. Existence of solutions
1. Method of successive approximations. Let the equation x = Ax be given, where A is some nonlinear operator. The method of successive approximations remains a basic method for the proof of the existence of solutions of this equation, that is, the proof of the existence of fixed points of the operator A . It consists of forming the sequence (n = 1 , 2, . . . )
from some initial element x0 and proving that this sequence converges to some element x*, thus establishing the equality x* = Ax*.
Example (EXISTENCE OF A SOLUTION FOR THE VOLTERRA EQUATION). In a non linear Volterra equation t
x (t) =
J K [t, s, x(s)] ds 0
let the functions K(t, s, x) and K� (t, s, x) be continuous with respect to the collection of variables t, s ;;:,: O,  ro <x< ro , and let
I K (t, s, x)l � cp (x) , where cp (x) is a nondecreasing function. If the differential equation du = cp ( l u l ) dt has a solution on the segment [0, w] satisfying the condition u(O) = O, then the Volterra equation has a solution x* (t), defined on [0, w]. If we set x0 (t)= O, then the successive approximations t
xn(t) =
f K [t, s, Xn l (s)J ds 0
(n = 1 , 2, . . . )
will converge uniformly to some function x* (t) on [0, w] which is a solu tion of the Volterra equation. This is implied by the fact that all xn(t)
EXISTENCE OF SOLUTIONS
173
are in the region  u (t )� x (t ) � u (t ) and satisfy the relation M (ur l xn(t)  Xn  1 (t) l � L n ! ·

(n = 1 , 2, . . . ) ,
where M and L are constants such that and
I K (t, s, O)I � M
(O � t, s � w)
IK (t, s, x 1 )  K(t, s, x 2 )1 � L l x 1  Xz l (0 � t, s � w ,  u (s) � x 1 , x 2 � u (s)) . 2. Principle of contractive mappings. In most cases, the method of successive approximations reduces to the verification of the conditions of the following general principle. PRINCIPLE OF CONTRACTIVE MAPPINGS. Let T be a closed set of the Banach space E. Let the operator A map T into itself and be an operator of contraction, that is, satisfy the Lipschitz condition
with a constant q < 1 . Then the equation x=Ax has a unique solution x* in T which is the limit of the successive approximations xn = Axn_ 1 for an arbitrary initial approximation x0 E T. In the conditions of this principle, the role of T is usually played by either the entire space E or some sphere S (O, r). In some cases, the set T must be constructed in a special way. Example. In the integral equation 1 x(t) =
f0 K [t, s, x (s)] ds +f (t)
let the functions K(t, s, x) and f(t) be continuous and K(t, s, x) satisfy a Lipschitz condition with a constant q < I with respect to the variable x. This integral equation can be considered as an operator equation in the space C of functions which are continuous on [0, I]. The operator defined by the right side of the equation satisfies a Lipschitz condition with a constant q < I . Therefore, by virtue of the principle of contractive
174
NONLINEAR OPERATOR EQUATIONS
mappings, the integral equation has a continuous solution which is a limit of the successive approximations 1
xn (t) =
J K [t, s, xn _ 1 (s)] ds +f (t) 0
(n = I, 2, . . . ) .
It is sometimes convenient to use a corollary of the principle of con tractive mappings : let the operator B also map the closed set T of the space E into itself and commute with the operator A, satisfying the conditions of the principle of contractive mappings, i. e. ,
BAx = ABx
(x E T) .
Then the fixedpoint of the operator A is a fixed point (possibly not unique) of the operator B. In particular, if some iterate Bn of the operator B satisfies the conditions of the principle of contractive mappings on the set T, then the fixed point x* of the operator Bn is also a fixed point of the operator B. In this case x* is the unique fixed point of the operator B. It remains to stress that all the cases when the solution of a nonlinear equation can be obtained as the limit of successive approximations are not exhausted by the principle of contractive mappings. 3. Uniqueness of a solution. Under the conditions of the principle of contractive mappings, the solution of the equation x = Ax is unique in T. However, the uniqueness of the solution does not generally follow from the uniqueness of the solution in T. For example, the equation 1 x(t) = x2 (t) dt
J 0
satisfies the conditions of the principle of contractive mappings in the sphere ll x ll �! of the space C and has a unique solution x0 (t ) = O in it � however this equation has a second continuous solution x 1 (t )= I . One must keep in mind that the uniqueness of the solution of some operator equation in the Banach space E does not imply the uniqueness of the solution of this equation considered in a wider space. Examples of linear integral Volterra equations exist which have, besides unique continuous solutions, nonsummable solutions.
EXISTENC E OF SOLUTIONS
1 75
4. Equations with completely continuous operators. Schauder principle. The principle of contractive mappings imposes on a continuous operator the rigid restriction of strict contraction . If we consider completely continuous operators, then this condition can be weakened considerably. ScHAUDER PRINCIPLE. Let the operator A be completely continuous and map a closed bounded convex set T into itself Then the equation x = Ax has at least one solution (uniqueness of the solution is not guaranteed) in T. If the set T is compact, then it is sufficient that the operator A be continuous. In the application of the Schauder principle to the study of concrete equations, we first construct a space E in which the operator A is com pletely continuous. For the convex set Twe take some sphere of the space E. Furthermore, it is necessary to select the radius and center of the sphere so that the operator A maps the sphere into itself. For example, let the completely continuous operator A have the property
II Ax il � a + b ll x ll''
(x eE , rx, a, b > 0) .
If a number r > O satisfying the condition
a + bra � r exists, then we apply the Schauder principle to the operator A in the sphere S(O, r). Such a number r always exists for ac < 1 and for ac= 1 and b < 1 . If ac > 1 , then r exists under the condition that min ( bsa  s) �  a .
O � s < oo
The Schauder principle only states the existence of a solution and does not give a method for finding it. In the case when the operator A is linear, we can indicate a method for finding solutions. Starting with some x0 e T, we construct the successive approximations Xn = Axn _ 1 (n = 1 , 2, . . . ) . Under the conditions of the Schauder principle, the se quence of elements 1 N 1 2N = N L Xn n=O 
is compact and all of its limit points are solutions of the equation x = Ax. We can formulate an assertion for which both the Schauder principle and the principle of contractive mappings are special cases.
1 76
NONLINEAR OPERATOR EQUATIONS
COMBINED PRINCIPLE. Let the operator, defined in a closed bounded convex set T, allow the representation A = A 1 + A 2 , where A 1 is completely continuous and A 2 satisfies a Lipschitz condition with constant q < 1 . If the condition (x, y e T)
is satisfied, then the equation x =Ax has at least one solution in T. The Schauder principle is proved by topological methods. These same methods (see no. 5) allow it to be strengthened. THE STRENGTHENED SCHAUDER PRINCIPLE. If the completely con tinuous operator A does not have eigenvectors with eigenvalues greater than I, on the boundary r of the closed convex set T containing 0 as an interior point, then a solution of the equation x = Ax exists in T. Thus, it is possible not to require that the boundary r of the region be mapped by the operator A into T. It suffices that there be no vectors on it which the operator A "expands" (Ax = gx, Q > 1 ). Frequently the last condition is verified significantly more easily. For example, if a linear functional f0 (x) exists such that f0 (x0 ) > 0 and f0 (Ax0) �f0 (x0) for every point x0 of the boundary r, then the condition of the strengthened Schauder principle is satisfied. In particular, if a completely continuous operator A is defined on the sphere S(O, r) of a Hilbert space H and has the property that ( ll x \1 = r) , (Ax, x) � (x, x) then the strengthened Schauder principle is valid for it. 5. Use of the theory of completely continuous vector fields. Let a completely continuous operator A be given on the boundary r of the sphere S of the Banach space E. The collection of elements of the form x  Ax (x er) is called a completely continuous vector field on r. A solu tion of the equation x = Ax(xe r) is called a zero of the field. An integer y (r), the socalled degree (rotation) of the vector field (see [23]), is set in correspondence to every completely continuous vector field x  Ax without zeros on r. The degree can be positive, negative or zero. PRINCIPLE OF NoNZERO DEGREE. If A is a completely continuous operator on the sphere S and the degree of the vector field x Ax on the boundary r of the sphere S is different from zero, then a solution of the equation x = Ax exists in S.
1 77
EXISTENCE OF SOLUTIONS
The Schauder principle and the strengthened Schauder principle are special cases of the principle of nonzero degree since the degree is equal to 1 for the conditions of these principles. Two vector fields xA 0 x and x  A 1 x are called homotopic on r if a completely continuous operator A (x ; rx) (xET, O � rx � 1) with respect to both variables exists such that and
(x E T)
A (x ; 0) = A 0 x , A (x ; rx) # x
(x E T, 0 � rx � 1 ) .
An operator A (x; rx) (xET, 0 � rx � 1 ), continuous with respect to both variables, will, in particular, be completely continuous if it is completely continuous for every fixed rx and uniformly continuous with respect to rx relative to x ET. The degrees of homotopic completely continuous vector fields are iden tical. This fact allows the application of the following method for the proof of the existence of a solution of the equation x = Ax with a com pletely continuous operator A. We introduce the parameter A so that the operator A (x ; A.) is completely continuous, A (x ; 1)= Ax and A (x ; A.) # x (xE r), 0 � A, � 1 . If it is now known that the degree of the vector field x  A (x ; 0) on r is different from zero (for example, if A(x; 0) satisfies the conditions of the Schauder principle), then it is immediately implied by the equality of the degrees of the fields x  A (x ; 0) and x  Ax that the initial equation x = Ax has at least one solution on S. This method of the proof of the existence of solutions is called the LeraySchauder method. The evaluation of the degree of a vector field is carried out by the methods of combinatorial topology. The character of the degree is known for some classes of vector fields. For example, if
 x  A ( x) x  Ax # ll x  Ax il II  x  A (  x) ll
c :
on the sphere ll x ll = r (at symmetric points of the sphere, the vectors of the field are not directed identically), then the degree of the field is differ ent from zero (moreover, it is odd). Let Ax0 = x 0 and the equation x = Ax not have solutions different fro m x0 in some neighborhood of the point x0• Then the vector field x  Ax has the same degree on all the spheres llx  x0 ll = r of sufficiently small
1 78
NONLINEAR OPERATOR EQUATIONS
radius r. This common degree y (x0) is called the index of the fixed point x0 of the operator A. If the operator A is differentiable at the point x0, where the linear operator A' (x0) does not have 1 as an eigenvalue, then
Y (x0) = ( 1 )P , where f3 is the sum of the multiplicities of the eigenvalues of the operator A' (x0) which are greater than 1 . We understand by the multiplicity of an eigenvalue of a bounded linear operator B the dimension of the corresponding eigenspace. The multi plicity of every eigenvalue of a completely continuous operator is finite. If 1 is an eigenvalue of the operator A' (x0), then the evaluation of the index y(x0) is complicated ; this evaluation uses derivatives of higher order. Here only partial results are obtained. Let the equation x = Ax have a finite number ofsolutions x1, . . . , xk in the sphere S. Then the degree y (T) of the field x  Ax on r is connected with the indices of the points x 1 , , xk by the equality Y (r) = Y (x l ) + . . . + Y (xk) . 
•
• • •
This property of the degree can be applied in proofs of uniqueness theorems. If the degree y (r) ofthe vector field x  Ax on r is in absolute value equal to 1 and if the index of every possible solution has the same sign, then the solution is unique by virtue of the preceding. Conversely, if the degree y (r) is known and the index y(x0) of the known solution x0 turns out to be different from y (r), then the equation x  A x has at least one more solution besides x0 on S.
Example (EXISTENCE OF A SECOND SOLUTION FOR AN URYSOHN EQUATION). Let the operator A, defined by the right side of the Uryson equation 1
x (t) =
J K [t, s, x (s)] ds ,
0 be completely continuous in LP and differentiable at the origin of this space, where 1 A' (O) h (t) =
J K� ( t, s, O) h (s) ds . 0
EXISTENCE OF SOLUTIONS
1 79
If the operator A satisfies the conditions of the Schauder principle on the sphere S (e, r), then the degree y (r) of the field x  Ax on the sphere ll x ll = r is equal to 1 . Let K(t, s, 0) = 0. Then the equation has a zero solu tion. If 1 is not an eigenvalue of the completely continuous linear operator A' (8) and the sum of the multiplicities of its eigenvalues greater than 1 is odd, then y ( 8) =  1 . Thus, y (r) # y (8) and the Urysohn equation has at least one nonzero solution in S (e, r ). 6. Variational method. The variational method of proof of theorems about the existence of solutions consists of the construction of a solution of the operator equation as an extremal point of some functional. A functional F(x), defined on a Banach space E, is called weakly continuous if it is continuous in the weak topology (J (E, E') in the space E (see ch. I, § 4, no. 3). If the space E is reflexive, then by virtue of the compactness of an arbitrary sphere of E in the weak topology, a weakly continuous functional assumes its least and greatest values on every sphere. The gradient of a smooth weakly continuous functional on a Hilbert space is a completely continuous operator. Let A be_ a potential operator in the Hilbert space H. VARIATIONAL PRINCIPLE. If the operator A is the gradient of a weakly continuous functional F(x) and lim U (x, x)  F (x)] = oo ,
ll x ll + oo
then a point x0 exists in H at which the functional i (x, x)  F(x) assumes its least value and which is a solution of the equation x = Ax. 7. Transformation of equations. In the study of operator equations, we often find ourselves transforming equations into a form which is convenient for the application of one principle or another from which the existence of a solution follows, or for the application of some approx i mation method for finding a solution. Basic forms of the transformation of operator equations are the same as for ordinary equations : a) the addition to both sides of the equation of the same element ; b) the application to both sides of the equation of the same operator ("multiplication by an operator") ; c) replacement of the variable.
1 80
NONLINEAR OPERATOR EQUATIONS
In the first transformation, the equation changes to an equivalent one. If a bounded linear operator B is applied to both sides of the equation, then every solution of the intial equation will be a solution of the new equation. The converse will be true if the inverse operator B 1 exists. Thus, the change to the new equation by a transformation b) can add extraneous solutions if zero is an eigenvalue of the linear operator B. If a replacement of the variable of the form x = Cy is carried out in the equation, where C is some operator, and the solutions y* of the new equation can be found, then in order to obtain a solution x* of the initial equation, it is necessary to verify that y* is in the domain of definition of the operator C, and then x* = Cy*. Moreover, in transformation c), part of the solutions can be lost. This occurs if there are solutions x* which are not representable in the form Cy. The transformation c) happens to be especially useful in that the operator C can act from another space E1 into the space E in which the solution x* is sought. Therefore, it is natural to consider the new equation (with respect to y) in the space £ 1 • It sometimes turns out that the equation is simpler in E1 • In the transformation of equations in infinitedimensional spaces we encounter a specific situation : the transformed equation contains oper ators which are not closed but allow closure. In this connection, it is natural to study the equation with closed operators. In this case, new solutions can appear which are usually called generalized solutions. The basic difficulty frequently is the proof that a generalized solution belongs to the domains of definition of the operators occurring in the equation, before their closure, and, hence, is a true solution. '
8 . Examples. Decomposition ofoperators. 1 . PREPARATION OF AN EQUATION FOR THE APPLICATION OF THE METHOD OF SUCCESSIVE APPROXIMATIONS. In the equation Bx
=
f,
let the operator B be linear, bounded and have a spectrum lying inside the right halfplane Re A. > O of the complex plane. After multiplication of both sides of the equation by the number k and the addition to both sides of the element x it reduces to the form
x = (I

kB) X + kf.
EXISTENCE Of' SOLUTIONS
181
For sufficiently small k, the operator 1kB will have a spectrum lying inside the unit circle and, hence, we apply the method of successive ap proximations to determine the solution of this new equation (which is equivalent to the old equation). It is sometimes convenient to apply an analogous transformation of the equation : we replace multiplication by a number k with multiplication by a suitably chosen operator K. The equa tion x = Ax, with a completely continuous operator A, is transformed into the form x  Bx =Ax Bx, where B is a completely continuous'linear operator. If the number 1 is not an eigenvalue of the operator B, then this equation is equivalent to the equation 2.
EQUATIONS WHICH ARE CLOSE TO LINEAR EQUATIONS.
x = c1  Br 1 (A  B) x . If, on the sphere ll xll =r, the operator (I B)  1 (A  B) does not have eigenvectors corresponding to eigenvalues greater than 1, then, by virtue of th€? strengthened Schauder principle, the equation obtained has at least one solution in the sphere ll x ll � r. There will not be such eigen vectors if the operator A is close to the operator B in the sense that [[ Ax  B x ll � ll x  Bx l l . 3. DECOMPOSITION OF OPERATORS. Let the operator B be linear in the equation x = BCx and allow "decomposition" into two factors : B = B 1 B2 , where B1 and B2 are linear operators. Every solution of the equation is representable in the form x = B 1 y. This replacement reduces the equation to the equivalent equation
A
special form of decomposition of an operator is often convenient : B= BaBt a, where Ba and B 1  a are fractional powers of the operator B. In connection with this, the theory of fractional powers of linear oper ators was, in recent years, amply developed (see ch. III, § 3, no. 4).
1 82
NONLINEAR OPERATOR EQUATIONS
The Hammerstein equation 1
x (t)
= I K (t, s)f [s, x (s)] ds 0
is the simplest example of an equation of the type considered. Let the kernel K(t, s) be sy mmetric, bounded and positive definite. It gen erates a completely continuous positive definite operator B in the Hilbert space £ 2 [0, 1]. If {e; (t)} is a complete orthonormal system of eigen functions of the operator B, and A; are the corresponding eigenvalues, then the operator Bt is defined by the formula 00
Btx (t) = L jA.;c;e; (t) , i= 1
where the c; are the Fourier coefficients of the function x(t) : 1
c; =
f e; (s) x (s) ds . 0
With the replacement x = Bty, the Hammerstein equation reduces to the form We can show that the operator Bt acts from the space £2 [0, 1] into the space M[O, 1]. Therefore, if the function f( t, x) is continuous, then the operator fJPy will be a continuous operator, acting from L 2 [0, 1] into M [0, I]. In this case, the operator B±jBt is completely continuous in L 2 [0, I]. For the operator BtfBt in L 2 , it is convenient to verify the conditions of the strengthened Schauder principle in the form indicated at the end of no. 5. In fact, If the function f(t, x) does not increase too quickly with respect to x, then on a sphere of sufficiently large radius r : I IY II r. Therefore the equation y = BtfBty has, by virtue of the indicated principle, at least one solution =
QUALITATIVE METHODS IN THE BRANCHING OF SOLUTIONS
1 83
y* inside the sphere IIYI I < r. Hence x* =Bty* will be a solution of the Hammerstein equation. Moreover, y*EL2 [0, 1 J and, hence, x*EM[O, 1]. The proof mentioned for the existence of a bounded solution of the Hammerstein equation can be successfully carried out if, for example, the function f(t, x) satisfies the inequality 2 2 xf (t, x) < ax + b (t) Jx l  y + c (t) , 1 where 0 < y < 2, b (t)EL; [o, 1], c (t) EL 1 [0, 1] and a <  . At
In the example considered, the operator Bt acts from L 2 to M. In more general cases it acts from L2 to LP (p > 2) and is completely continuous. Moreover, it can be naturally extended to the operator (B�") * which acts from Lv ' into L2
(; ; = 1) . If the nonlinear operator f acts from LP +
,
into LP '' then the equation
will be an equation in L2 with a completely continuous operator and the reasoning mentioned above will be applicable to it. If y* EL2 is a solution of this equation, then x* =H"y* ELP is a solution of the equation The operator Bt (Bt)* is an extension of the operator B, considered originally on the space L2• Therefore the solutions of the last equation can be considered as generalized solutions of the initial equation. In the case of the Hammerstein equation they turn out to be true solutions. §3. Qualitative methods in the theory of branching of solutions
In this section the equation x = A (x, f.l)
is considered, where f.1 is real. It is assumed everywhere that the operator A is uniformly continuous in f.1 relative to the elements x of an arbitrary bounded set. If, on the basis of one of the principles studied in the preceding section, we succeed in establishing the existence of a solution of the equation for f.1 = f.10, then we succeed in the majority of cases in proving the existence
1 84
NONLINEAR OPERATOR EQUATIONS
of a solution for close values of the parameter p, since the conditions of the applicability of the corresponding principle are not violated for small changes of the operator A (x, p0). For the determination of the magnitude of the segment [Jl o  a, p0 + b] on which these conditions are retained, it is necessary to estimate II A (x, J1)  A (x, p0) 11 , for which a priori estimates for the solutions of the corresponding equations are frequently required. 1 . Extension of solutions, implicit function theorem. If x0 is a solution
of the equation x =A (x, p0), then it is natural to expect that the equation x = A (x, p) will have a solution x(Jl) close to x0 for values of the para meter J1 close to Jlo· In establishing this fact, the general implicit function theorem plays an important role. In the equation F (x, u) = O , let x be an element of the Banach space El> u an element of the Banach space E2 , and F(x, u) an operator with values in the Banach space E3. We understand a solution of this equation to be an operator X (u), defined on some set of elements uEE2 with values in E1 , such that F(X(u), u) =O. An analogue of the usual theorem on the existence of an implicit func tion holds : if F(x0, u0) 0, if the operator F(x, u) is continuous and is continuously differentiable according to Frechet with respect to the variable x for I I x  x0 II �a, II u  u0 II � b and if the linear operator F� (x0, u0) has a bounded inverse, then a solution X(u) of the equation F(x, u) = O is defined in some neighborhood of the point u0 • This solution is unique. The operator X(u) is continuous. If the operator F(x, u) has a Frechet derivative with respect to u of a specified order, then the operator X(u) has a Frechet derivative of the same order. In the application of the implicit function theorem to the equation x = A (x, p), the requirement of the invertibility of the operator F� (x0, u0) is naturally replaced by the requirement that 1 not belong to the spec trum of the operator A� (x0, p0). Under the satisfaction of this condition and the continuity of the operator A�(x, Jl) with respect to (x, p) in a neighborhood of the point (x0, u0), a unique continuous solution x(p) of the equation x = A (x, Jl) exists such that x (J10) = x0. =
2. Branch points.
If the operator A(x, p) is completely continuous
QUALITATIVE METHODS IN THE BRANCHING OF SOLUTIONS
1 85
for every J1 in the equation x = A (x, p,), then topological methods can be applied to the investigation of the behavior of the solution as J1 varies. Let x0 be an isolated solution ofthe equation x = A (x, p,0), having a non zero index. On a sufficiently small sphere S surrounding the point x0, the degree of the field x  A (x, p,0) will be different from zero. Hence the degree of the field x  A (x, p,) will be different from zero on S for J1 close to p,0• It follows from the principle of nonzero degree that at least one solution x (p,) of the equation x = A (x, p,) exists inside S. Reducing the radius ofthe sphere S, we can choose a solution x (p,) such that lim ll x (p,)  x0 ll = 0 . In this sense we can speak of the continuity of the solution x(p,) at the point p,0• The pair (x0, p,0) is called a branch point for the equation x= A (x, J1) if for every e > O we can find a J1 such that jp, p,0j < e and the equation x=A (x, p,) h as at least two solutions lying in the eneighborhood of the point x0 • It is implied by the arguments of the preceding subsection that the pair (x0, p,0) is not a branch point if the operator A� (x, p,) exists and is continuous with respect to (x, p,) in a neighborhood of the point (x0, p,0) and if 1 is not a point of the spectrum ofthe operator A; (x0, p,0). But if we assume only the existence of the operator A� (x0, p,0), not having 1 as a point of the spectrum, and do not assume the existence of the operator A ' (x, p,) in a neighborhood of the point (x0, p,0), then the pair (x0, p,0) may turn out to be a branch point. Let the equation x=A (x, p,) have, inside the sphere S, for every J1 close to p,0 and different from it, only a finite number of solutions, where the operator A� (x, J1) exists at these pointsolutions and 1 is not an eigen value. Then the number of such solutions differs from the degree of the field x  A (x, p,) on the sphere S by an even number (the index of every solution is + 1 , the sum of the indices is equal to the degree ofthe field). As long as the degree of the field x A(x, p,) on S is not changed by changing J1 close to p,0, then the number of solutions of the equation x =A (x, J1 ) as J1 passes through Jlo can be changed only by an even number. This statement is called the principle of the preservation ofparity of the number of solutions. I f the index of the solution x0 is equal to zero, then a solution x(p,) of
186
NONLINEAR OPERATOR EQUATIONS
the equation x = A (x, f.l), generally, may not exist in a neighborhood of the point x0 for f.1 close to f.lo· The solutions can "flow into" x0 for J.l+ f.lo  q and not exist for f.l > f.lo ; then the pair (x0 , f.lo) is called a point of cessation of solutions. The solutions may not exist for f.l < f.lo and "flow out" from the point x0 for J.l+ f.lo + 0; then the pair (x0, f.lo) is called a point of appearance of solutions.
3. Points of bifurcation, linearization principle. The concept of a point of bifurcation is closely related to the concept of a point of branching. Let us assume that A (e, f.1) = 8. Then the equation x = A (x, f.l) has the trivial solution x = 8 for all values of the parameter f.l· The number f.lo is called a point of bifurcation for this equation (or for the operator A (x, f.l)) if to any e > O there corresponds a value ofthe parameter f.1 in the segment if.l  f.loi < B for which the equation has at least one nonzero solution x (J.l ) satisfying the condition llx (J.l) ll < B . Unlike the definition of a point of branching, it is assumed in the definition of a point of bifurcation that one family of solutions is known a priori, defined for all values of the parameter. We speak about a "branch" of solutions from the given family. In the definition of a point of bifurcation, we do not speak about those values of the parameter for which the equation has small nonzero solutions. These values can form a discrete set or even coincide with f.lo· The generality of the concept of a point of bifurcation allows us to obtain general, simple theorems concerning methods of determining these points. At the same time the concept of a point of bifurcation describes reason ably completely the occurrence of the nonzero solutions. For a linear equation x = J.1BX with a completely continuous linear operator B, the points of bifurcation coincide with the characteristic values of the operator B (values inverse to the eigenvalues). If the operator A (x, f.1 ) is continuously differentiable in the sense of Frechet, then by virtue of the implicit function theorem its points of bifurcation can only be those values f.1 for which 1 is a point of the spectrum of the operator A� (e, f.l). Let A� (e, f.l) = J.lB where B is a com pletely continuous linear operator which does not depend on f.l· If 1 is an eigenvalue of the operator J.1B, then f.1 is a characteristic value of the operator B. Thus, in this case, the points of bifurcation are characteristic values of the operator B. The question is raised : is every characteristic value of the operator B a point of bifurcation? In the general case, as examples indicate, the answer is negative.
QUALITATIVE METHODS IN THE BRANCHING OF SOLUTIONS
1 87
We call the principle according to which the determination of points of bifurcation is reduced to the determination of the characteristic values of the linear operator B the linearization principle. The following statement serves as a basis for this. principle : if the completely continuous operator A (x, fl) (A (8, fl) = 8) has a Frechet derivative A�(e, fl) = flB at the point 8, then every oddmultiple (in particular, simple) characteristic value of the operator B is a point of bifurcation of the operator A (x , fl) . If a characteristic value of the operator B has an even multiplicity, then further analysis is required which uses more than just the linear part flB of the operator A (x , fl). Let the completely continuous operator A (x , fl) allow the representation
A (x, fl) = flEX + C (x, fl) + D (x , fl) , where B, as above, is a completely continuous linear operator; the oper ator C(x, fl) consists of terms of k  th order of smallness where k > 1 is an integer, that is, and
II C (xr, fl)  C ( x2 , fl) ll � q (e) llxr  x2 ll 1 (llxr ll � Q, ll x2 ll � Q, q (e) = O (ek )) ;
and the operator D (x , fl) consists of terms of a higher order of smallness IID ( x, fl) ll � L ll xll k + l .
Let fl o be a characteristic value of the operator B of even multiplicity f3 ; let the elements e 1 , . . . , ep form a basis in the eigenmanifold of the operator 1 B corresponding to the eigenvalue � and let the elements g 1 , , gp form fl • . .
o
a basis in the eigenmanifold of the operator B* corresponding to the same eigenvalue (flo is real). The vector field F in a {3dimensional vector space defined by the equality where (i
=
1 , . . . , [3) ,
plays an important role. Let the field not be degenerate (that is, it vanishes only at the point e 1 = e 2 = . . . = ep = 0) and let its degree on the unit sphere be equal to Yc · The following statement hold s :
1 88
NONLINEAR OPERATOR EQUATIONS
If Yc # 1 , then !lo is a point of bifurcation of the operator A (x, 11) . For the application of this statement it is necessary to be able to con struct the field F, to prove that this field is not degenerate, and to be able to calculate the degree Yc· To calculate the degree, it is useful to know that the degree of an even field (F(x)= F( x)) is an even number. For example, let k = 2 and !lo be a characteristic value of the operator B of multiplicity two (/3 = 2). In this case the field F will have two com ponents 1] 1 and 1] 2 each of which will be a quadratic form with respect to e1 and e2 : '1 1 = a u e i + 2a 12e1e2 + a 22eL '12 = hu e i + 2b 12e1e2 + b 22e� . 
If one of these forms is positive or negative definite, then the field F is not degenerate and its degree is equal to zero. If neither of the forms is sign constant, then it suffices for the nondegeneracy of the field that the straight lines e1 = ae2 which one M the forms is annihilated con sist of points on which the second form assumes nonzero values. The slopes a 1 ans a2 of straight lines on which the first quadratic form becomes zero are determined by the quadratic equation a 1 1 +2a1 2 a + a22 a 2 = 0.
Example (POINTS OF BIFURCATION OF THE HAMMERSTEIN EQUATION). In the equation 1
J
x (t) = 11 K (t, s) f [x (s)J ds 0
with a bounded symmetric kernel K(t, s ), let the function f(x) be differ entiable as many times as desired,/(0) = 0 and f'(O) = 1 . The linearization of this equation yields a linear integral equation with the kernel K(t, s ). If the characteristic value /lo of the kernel K(t, s ) has an oddmultiplic ity, then it will be a point of bifurcation for the initial equation. Let /lo have multiplicity 2 and let e 1 (t), e 2 (t) be the eigenfunctions corresponding to it. If{" (O) # O, then the operator C will have the form 1
C (x (t), 11) = 11 K (t, s) x 2 (s) ds .
J 0
QUALITATIVE METHODS IN THE BRANCHING OF SOLUTIONS
1 89
Therefore the components of the vector field F are given by the equal ities 1 1 1
'7 1 = ei '72 = ei
f ei (t) dt + 2 e1 e2 f ei (t) e2 (t) dt + e� f e 1 (t) e� (t) dt ' 0 1
0
1
0
1
f ei (t) e2 (t) dt + 2 e1e2 f e1 (t) e� (t) dt + e� f e� (t) dt . 0
0
0
For example, if e 1 (t) = 1 , e2 (t) = J2 cos2nt, then The first quadratic form is positive definite, the degree of the field F is equal to zero and, hence, llo is a point of bifurcation for the Hammerstein equation. The following question is interesting : for which values ofthe parameter /1, greater or less than /lo. does the equation x = A (x ; 11 ) have small non zero solutions ? Let !lo be a simple characteristic value of the linearized equation x = 11Bx, let e be a corresponding eigenvector, and let g be an eigenvector of the adjoint operator, where ( e, g) = 1 . The vector field F is given in this case by the number K =  (C (e, !10) , g) . ·
The following statements hold : 1) If the order k of smallness of the operator C is an even number, then the equation x = A (x, 11) has small nonzero solutions for 11 < !lo and for 11 > /lo· If the operator A (x, 11) is sufficiently smooth, then the nonzero solution is unique for every 11 (close to f.lo) . 2) If k is odd, then small nonzero solutions exist for 11 > !lo and are absent for 11 < !lo in the case when K < 0; if K > 0, then small nonzero solu tions exist for 11 <11 0 and are absent for 11 > llo · Two nonzero solutions exist for corresponding values of 11 · 4. Examples from mechanics. a) EULER PROBLEM CONCERNING STABILITY FOR BUCKLING OF A BEAM. 1 The deflection y(e) of a beam of u nit length with variable rigidity Q (e) = EJ •
190
NONLINEAR OPERATOR EQUATIONS
under the action of a longitudinal force 11 is given by the solution of the differential equation
for zero boundary conditions y (O) = y ( l ) = O (see figure). The function
is Green's function for the operator y" with zero boundary conditions.
The differential equation of the buckling of the beam reduces to an integral equation with the replacement y" (�) =
Then

cp ( �) .
1
y ( e) =
J K (e, 1J) cp (1J) d1J 0
and the equation for cp (e) assumes the form 1
cp < �) = /lQ ce) K < �. 1J) cp (IJ) a'l
J 0
1
1
_
[f Ki (e, IJ) cp (IJ) dl]r 0
This equation has the zero solution for all values of the parameter 11· For some loads /1, the equation can have nonzero solutions by which the
QUALITATIVE METHODS IN THE BRANCHING OF SOLUTIONS
191
forms of loss of stability are determined. The load /l o is called the critical Euler load if, for some loads close to llo• the equation has small nonzero solutions. In other words, the critical Euler load is a point of bifurcation for the equation of buckling of the beam. The determination of the critical load is one of the important problems of the theory of elasticity. The integral equation obtained can be regarded as an operator equa tion of the form x = A (x, 11) with a completely continuous operator in the space C. Linearization leads to the equation 1 cp (0 = !1Q ( O K (e, 1J) cp (1J) d1J .
J 0
If e(e) is a nonzero solution of this equation for 11 = llo• then the function 1
y ( e) =
J K (e, 1J) e (1J) d1J 0
will be a solution of the equation
y" (e) + !loQ ( e) y (� ) = 0 satisfying nonzero boundary conditions. It is implied that every charac teristic value of the linaerized equation is simple and, hence, is a point of bifurcation. The corresponding values of f.1 give the critical loads. The operator C has the form 1 1 llQ e) K ( � , IJ) x (IJ) d1J Ki ( e , IJ) x (IJ) d1J C (x ( e), 11) = _
[f
; f 0
0
J
for the equation considered and, hence, has a third order of smallness (k = 3). Here 1 1
x =  �Je2(e{J Ki (e, 1J) e (1J) d1JJ ae < o ; 0
0
therefore nonzero solutions appear for 11 > Jlo . This corresponds fully with the physical meaning of the problem : loss of stability occurs then when the load exceeds the critical value.
192
NONLINEAR OPERATOR EQUATIONS
In the literature, another equation for the buckling of a beam is some times encountered : y" (t) + fl(! (t) y (t) [1 + y' 2 (t)]t = 0 for the boundary conditions
y (O) = y (1) = 0 . It corresponds to the case where the curvature is expressed not as a func tion of the arc length e but as a function of the coordinate t. In this connection it is assumed that approximately Q (e)= Q (t) ; and in the boundary conditions, the change of the coordinate of the nonfixed end of the beam, representing a magnitude of the third order of smallness in comparison with the buckling of the beam, is not taken into account. However this magnitude of third order manifests itself in the sign of x. In this case, the expression 1 1 e 2 (t) K/ (t, s) e (s) ds dt > O x=
�f 0
[f
J
0
is obtained for x, and nonzero solutions are obtained for 11 < 11o for the corresponding integral equation which contradicts the physical meaning of the problem. Thus, the disregard of magnitudes of the third order of smallness in equations leads to an improper description of the' problem concerning the forms of the loss of stability of a compressed beam. b) WAVES ON THE SURFACE OF AN IDEAL INCOMPRESSIBLE HEAVY FLUID. The investigation of such waves was reduced by A. I. Nekrasov to the solution of the integral equation 21t K (t, s) sin x (s) x (t) = 11 ds ···
1 0
'
s
+
J sin x (u) du 0
where 11 is a numerical parameter and
K ( t, s)
1 3
= 
L
n= l
sin nt sin ns . n
QUALITATIVE METHODS IN THE BRANCIDNG OF SOLUTIONS
193
This equation can be regarded as an operator equation in the space C on the segment [0, 2n]. It has the zero solution for all values of 11 · Points of bifurcation of this equation correspond to the values of the parameters for which waves arise. The linearized equation has the form 27t
x (t) = !l
J K (t, s) x (s) ds ; 0
its characteristic values are the numbers lln = 3n and the corresponding eigenfunctions are en (t) = sin nt. All the characteristic values are simple ; therefore they will be the only points of bifurcation for the Nekrasov equation. The operator C has the form 27t s C (x (t), !l) =  !1 2 K (t, s) x(s) x ('r:) d1: ds .
J 0
J 0
It is a magnitude of the second order of smallness (k = 2) for small x ; therefore the Nekrasov equation has small nonzero solutions for 11 < lln and 11 > lln• where !ln is an arbitrary point of bifurcation. 5. Equations with potential operators. For the equation
X = 11Ax , where A is a completely continuous operator which is the gradient of a weakly continuous functional in a Hilbert space, the principle of lineariza tion for the determination of the points of bifurcation is strengthened considerably. If A ((}) = (}, the operator A is continuously differentiable, and its derivative A'((}) = B is a completely continuous selfadjoint operator, then every char acteristic value of the operator B, independently of its multiplicity, is a point of bifurcation of the nonlinear equation x= 11Ax. As an example we can again consider the Hammerstein equation with a bounded symmetric positive definite kernel : 1 x (t) = 11 K (t, s)f[s, x (s)] ds , f (s, 0) = 0 , f ' (s, 0) = 1 .
J 0
194
NONLINEAR OPERATOR EQUATIONS
As in § 2, no. 7, we can transform it into the form y = !lBtfBty .
The operator BtfBt is the gradient of the functional 1 Bh
cl> (x) =
J ds f f (s , u) du .
0 0 If the operator f is differentiable, then the Frechet derivative of the operator BtfBt at the point () is a linear integral operator B with kernel K(t, s). All characteristic values of this operator are points of bifurcation for the equation y= !lB tfBty . The inverse replacement B+y=x indicates that the points of bifurcation of the last equation coincide with the points of bifurcation of the initial Hammerstein equation. 6. Appearance of large solutions. In no. 2 a general pattern was de scribed of the change of solutions for a change of the value of the parame ter. This pattern is relative to the case when solutions in some sphere were considered. In a more general case, the norms of the solutions can increase indefinitely for a change of the values of the parameter. We may encoun ter such a case when the solutions with large norms appear for values of the parameter greater than some critical number. Here one theorem is mentioned which describes the appearance of solutions with large norms. Let the operator A (x, 11) be asymptotically linear, where A;, ( oo , 11) = f.lB. Let !lo be a characteristic value of oddmultiplicity of the completely con tinuous linear operator B. Then, for any e, R> 0 a 11 can be found which satisfies the inequality 1!1  !lol < e andfor which the equation x = A(x, 11) has at least one solution whose norm is greater then R. 7. Equation of branching. We assume that 1 is an eigenvalue of the derivative A�(x0, !lo) of the completely continuous and continuously differentiable operator A (x, 11) . For the sake of simplicity we restrict ourselves to the case where the invariant subspace E0 corresponding to 0 this eigenvalue consists only of eigenvectors. We denote by E the invariant subspace of the operator A� (x0, !lo) which is complementary to E0 • We represent every element xEE in the form
QUALITATIVE METHODS IN THE BRANCHING OF SOLUTIONS
1 95
0 Let P and Q be the projectors onto E0 and E defined by the equalities Px = u, Qx = v. The equation x = A (x, fl) can be rewritten in the form of the system
y = PA (x 0 + y + z, fl)  Px0 , z = QA (x0 + y + z, fl)  Qx 0 , where y = P (x  x0) , z = Q(x  x0) . If y and fl  fl o are sufficiently small, then the second equation has a unique small solution z = R(y, fl) . There fore the question of the solvability and the construction of a solution of the equation x = A (x, fl) is equivalent to the question of the solvability of the equation y = PA (x0 + y + R (y, fl), fl)  Px o . The last equation is an equation in a finitedimensional space. It is called the equation of br({nching. Analytical and topological methods can be applied for its investigation. 8. Construction ofsolutions in the form of a series. Let x0 be a solution of the equation x = A(x, flo) · Let the operator A (x, fl) be analytic in a neighborhood of the point (x0 , flo) in the sense that it is representable in the form of a Taylor series
A (x, fl) = x0 + L (fl  fl oY Cii (x  x0) , i+j;:; l where the C ii(h) (i�O, j � O) are operators havingjth order of smallness with respect to h ; in particular, the C;0 (h) = Cw are some fixed elements of E. As above, the linear operator C0 1 = A�(x0, fl o) plays a special role. Let the operator A (x, fl) be completely continuous. Then the operator C01 = A� (x0, flo) is also completely continuous. If 1 is not an eigenvalue of the operator Co lt then the equation x = A (x, f.l) has a unique solution x (fl) for values of fl close to fl o · This solution, as it turns out, is representable by a series To determine the elements x1 , x2, . . . , this series is substituted in the equation, then the right hand side is developed in a series in powers of (f.l  flo) and the coefficients of the identical powers of ( Jl  fl o) are equa
196
NONLINEAR OPERATOR EQUATIONS
ted. As a result we arrive at the system of equations
x 1 = Co 1 (x l) + C10 , X2 = Co 1 ( x 2) + C1 1 (x 1) + Co2 (x l ) + C2o . •
0
•
•
•
0
•
0
•
•
0
•
0
•
•
•
0
0
•
The linear equations which are written out can be solved successively. The series for x (11 ) is convergent for 111  11o I sufficiently small. Majorizing numerical series are usually constructed to estimate the radius of con vergence. Now let 1 be an eigenvalue of the linear operator C0 1 . In this case, the question concerning the number of solutions of the equation x = A (x, 11) for values 11 close to 11o becomes more complicated. Such solu tions can sometimes be found in the form of the series 2 1 x (11) = Xo + (11  11ol X1 + (11  11 ol X2 + · · · �
�
with respect to fractional powers (k is a natural number) of the increm�nt 11  11 o · To determine the elements X t. x2 , , we again substitute the series for x (11) in the equation and compare the coefficients of identical fractional powers of 11  110 • For example, the equations • • •
x 1 = C01 ( x t ) , X2 = Col ( x2 ) + Co 2 (x l) + C 1 o , •
•
•
•
•
•
•
0
0
0
•
•
•
0
are obtained in the case k = 2. The first equation is a homogeneous linear equation. Its solution has the form where e1, , e. is a basis for the subspace E0 of eigenvectors correspond ing to the eigenvalue 1 and 1)(1, . . . , il(s are arbitrary numbers. To determine the numbers il( t . , 1)(_, conditions for the solvability of the second equation are used. These conditions can be written in the form • . •
• • •
(i = 1, 2, . .
.
,s
),
where /1, ,f. is a complete system of eigenvectors (linear functionals) of the operator Cri'1 , adjoint to Co t. corresponding to the eigenvalue 1 . • . .
QUALITATIVE METHODS IN THE BRANCIDNG OF SOLUTIONS
1 97
The conditions of solvability are represented by a system of s non linear equations with s unknowns. If it can be solved, then the element can be found. Simultaneously, we can state that the second equation can be solved (with respect to x 2 ) . Its solution is again defined to within s arbitrary constants : The coefficients /31 , . . . , /32 are determined from the conditions of the solvability of the third equation, and so on. The determination of the elements x 1 , x2 , . . . becomes more difficult if it is impossible to determine the coefficients oc 1 , . . . , oc. from the conditions of the solvability ofthe second equation. Here it is necessary to draw upon the conditions of solvability of the following equations. If we do not succeed in constructing the solution in the form of a series in powers of 11  J10, (11  Jlo}" , then we try to construct the solution in the form of a series in powers of (11  Jlo}", and so on.
CHAPTER V
OPERATORS IN SPACES WITH A CONE
§ 1. Cones in linear spaces
Cone in a linear system. A convex set K of elements of a real linear system is called a cone if this set contains, together with each element x(x # 0), all the elements of the form tx for t ;;::: 0 and does not contain the element  x *). 1.
Examples
1 . The collection of all nonnegative functions space
C(O, 1)
forms a cone in this space.**) .
x(t) (tE [O, 1])
of the
Analogously, the sets of all nonnegative functions of the space
LP (O, 1 ), spaces .
the space
M(O, 1 ),
and the Orlicz spaces form cones in these
2. The set of positive operators forms a cone in the space of bounded linear selfadjoint operators acting in a Hilbert space (see ch. II, § 2, no. 3). 3. The sets of elements with nonnegative c oordinates will be cones in the coordinate spaces
lP , m, c.
4. In function spaces, it is sometimes necessary to study cones which are narrower than the cone consisting of all nonnegative functions. These cones are determined by a system of additional inequalities. Exam ples are the cone of nonnegative nondecreasing functions :
and the cone of nonnegative convex upwards functions :
*) If the last condition is not satisfied, **) For the definition of the spaces, see
then the set is called a wedge. ch. I, § 2, n o . 5.
'
CONES IN LINEAR SPACES
1 99
The cone K in the linear system E is called generating if an arbitrary element x E E is representable in the form of the difference of two elements of the cone : x = x1  x2 (x1 , x 2 E K). The cone of nonnegative functions of the space C[O, 1 J is generating. We can represent every function x (t)E C(O, 1) in the form of a difference of nonnegative functions x+ (t) and x_ (t) :
x (t) = x+ (t)  x_ (t) , where
x (t) , { X+ o t 0, x (t) = { _  � C t) ,
if x (t) � 0 , if x(t) < 0 , if x (t) � 0 , if x(t) < O .
All the cones considered'in examples 13 are generating ; however, not every cone has this property. For example, the cone of nonnegative non decreasing functions (example 4) is not generating in the space C(O, 1 ) since only functions of bounded variation can be represented in the form of a difference of nondecreasing functions. 2. Partially ordered spaces. The real linear system E is called a linear partially ordered space if a relation x ::s:;y is defined for some pairs of elements x, y E E and if the sign :::::; has the usual properties of the sign of inequality. We mean the following properties : 1) it follows from x::s:;y that tx ::s:; ty for t � O and ty :::;:; tx for t < O, 2) it follows from x ::s:;y and y ::s:; x that x =y, 3) it follows from x1 ::s:;y1 and x 2 ::s:;y2 that x 1 + x2 :::::: y1 +y2 , 4) it follows from x::s:;y and y ::s:; z that x ::s:; z. The relation :::::; is usually called inequality, and we say that x is less than or equal to y if x::s:;y. It remains to remark that the sign :::::; establishes a total ordering rela tion in the space of real numbers in the sense that two arbitrary numbers a and b can be united with this sign (either a :::::; b or a � b) ; the sign :::::; , generally speaking, does not have this property in a linear space. The origin of the term "partial ordering" is connected with this situation. The collection K of all elements x of a partially ordered space for which O ::s:; x forms a cone in this space. Conversely, if' t h e cone K is given in a
200
OPERATORS IN SPACES WITH A CONE
linear system E, then a partial ordering can be introduced in this system by setting x::s:;y if y  x E K. Thus the consideration of linear partially ordered spaces is equivalent to the consideration of linear systems with a cone. If K is the cone of nonnegative functions in the space c (or Lp), then the partial ordering relation acquires a simple meaning : x::s:;y if x (t) ::s:;y(t ) for all (or almost all) values of t. 3. Vector lattices, minihedral cones. The partially ordered linear space E is called a vector lattice if the following property is satisfied : 5) for any two elements x, yEE there exists an element zEE such that x ::s:; z, y :::;:; z, and z ::s:; C for every element C having the same property
The element z is called the least upper bound or the supremum of the elements x and y and is denoted by z= sup(x, y). The existence of an infimum for any pair (x, y), that is, an element u which has the following properties : u :::::; x, u :::::; y and if v :::::; x, v :::::; y, then v :::::; u, follows from the existence of a supremum for arbitrary elements x, y in a vector lattice. In this connection we write u= inf(x, y). The cone formed by the elements x;;;;. e of a vector lattice is called minihedral. If we denote the collection of elements of the form x + u (u E K) by Kx, then the minihedralness of the cone means that for any x, y E E an element z can be indicated such that Kx n Ky = Kz. In this connection z = sup (x, y). Every element of a vector lattice allows the representation x = x +  x_, where X + = Sup (x, e) and is called the positive part of X and Where x_ = sup ( x, e) and is called the negative part of x. The element lxl = x + + x_ is called the absolute value of the element x. The relation 
 I X I :::::; X :::::; I X I
is valid. The cones of nonnegative functions are minihedral in the spaces C, L P , M. The cones of nonnegative sequences are also minihedral in the spaces lP, m, c. Not every cone is minihedral. The usual circular cones of threedimensional Euclidean space are the simplest examples of non minihedral cones.
201
CONES IN LINEAR SPACES
Kspaces. Let M c E be some subset of the partially ordered space E. If all the elements of M are less than or equal (in the sense ofthe opera tion of comparison :::;:; ) to some element zEE, then z is called an upper bound of the set M and the set M itself is called bounded above. An upper bound z of the set M is called the least upper bound of M (we write Z=Sup M) if the relation C � z is satisfied for every other upper bound ' of the set M. The definitions of boundedness below, lower bound and greatest lower bound of a set of elements from E are introduced anal ogously. If the space E is a vector lattice, then a least upper bound exists for every finite set of elements x 1 , x 2 , x. and can be defined by means of the recurrence relation 4.
• • •
,
The vector lattice E is called a Kspace if each of its bounded above nonempty subsets has a least upper bound. In a Kspace, every bounded below nonempty set of elements has a greatest lower bound. The cone of elements x � (} in a Kspace is sometimes called strongly minihedral. The spaces LP , M, lP, m are Kspaces ; the vector lattices C and c are not Kspaces. In Kspaces, convergence with respect to the ordering can be intro duced which is called (a)convergence. For convenience of description, two new "noncharacteristic elements" oo and  oo are adjoined to the Kspace with respect to which we assume that  oo :::;:; x :::;:; oo for all elements x E E. Then we write sup M = oo for a nonbounded above set M c E and inf M =  oo for a nonbounded below set. If we introduce oo into the number of elements of the set M in addition to the characteristic elements x E E, then we assume that sup M= oo and if we introduce  oo , then we assume that inf M =  oo . Let x.(n = 1 , 2, . . . ) be an arbitrary sequence of elements of the Kspace E. The upper and lower limits of this sequence are the elements defined by the relations lim x. = inf [sup (x., Xn + I • . . . )] , lim x. = sup [inf(x x. +
n + oo
n + oo
n
n
•.
1 , • • •
)] .
202
OPERATORS IN SPACES WITH A CONE
These elements can be finite or equal to
ro
or
ro .
If lim xn = lim xn, n + oo then the sequence xn is called (a)convergent, and the common value of its upper and lower limits is called simply the (o)limit and is denoted by (o) lim xn. n + (o)convergence is a basic form of convergence in a Kspace ; (o) convergent sequences have several usual properties of convergent sequen ces. For example, if (o)lim xn = x and (o)lim yn = Y and both limits are finite, then the sequence xn + Yn is (o)convergent and 
oo
(o) lim (xn + Yn) = (o) lim Xn + (o) lim Yn . n + n + Moreover, in order for (o) lim xn =x it is necessary and sufficient that n + (o) lim I x"  xl = e. n + The property of completeness with respect to (o)convergence is an important property of a Kspace : in order for the sequence xn to have a finite ( o)limit it is necessary and sufficient that oo
oo
oo
oo
(o) lim [ sup lxk  xml ] n + oo k,m';3; n
=
e.
The sequence xn is called (t)convergent to the element x if from an arbitrary subsequence xn · we can choose a particular subsequence xn · such that (o)lim xn = x. If a sequence is (a)convergent, then it is (t) k convergent. In the Kspace M, (a)convergence of a sequence of elements coincides with convergence almost everywhere, and (t)convergence in M means convergence in measure. (a)convergence and (t )convergence coincide in the Kspaces Ip(p � 1 ). '
'k
;k
5. Cones in a Banach space. If the system E is a Banach space, then a cone in the Banach space E is any cone of the system E which is a closed set in the space E. If the cone K in the Banach space E is generating, then a constant M exists such that for every XEE we have the representation x = x 1  x2 (x1 , x2 E K) in which ll x1 11 :::::: M II x ll and ll x2 ll :::::: M I I x ll . A solid cone is a special case of a generating cone. A cone is solid if it contains at least one interior element. The cone of nonnegative functions of the space C[O, 1 J can serve as an example of a solid cone. Functions
•
CONES IN LINEAR SPACES
203
with a positive minimum are interior elements of this cone. The cone from example 2, no. 1 and the cone of nonnegative sequences of the space m of bounded sequences are also solid. The cones of nonnegative functions of the spaces Lp[O, 1 J (p � 1) and the nonnegative sequences of the coordinate spaces lv(P � 1) do not have the property of solidity. Thus, not every generating cone is solid. Every generating cone is solid in a finitedimensional space. The cone K is called normal if there exists a b > 0 such that the inequal ity II e t + e2 1 1 � b is satisfied for all et , e2 E K, II e1 II = II e 2 11 = 1 . If the cone K is normal, then l ift + fz l l �
b
2
max { llft ll , l lfz ll }
for arbitrary elementsft /2 E K. The set of normal cones is very important. The cones of nonnegative functions are normal in the spaces C, Lv (P � 1), lv m ; 1 can be taken as the constant b which appears in the definition of a normal cone in these examples. The cone of positive linear operators can serve as an example of a normal cone in the space of self adjoint operators. Not all cones are normal. For example, the cone of nonnegative functions in the space c( l ) [0, 1 J of continuously differentiable functions x(t) does not have the property of normality (see ch. I, § 2, no. 5). If the cone K is normal, then the partial ordering established by this cone in E has the following property : a new norm 1 1 . . . li t can be intro duced in E which is equivalent to the originally given norm and such that the inequality ll x ll t :::::; IIY I I t follows from the relation e :::;:; x :::;:; y . The con verse statement is valid. By virtue of the equivalence of the norms I I · . · 11 1 and I I · . . 11 , this property of a normal cone can be formulated in the form of the following criterion: the cone K is normal if and only if the inequality 8 :::::; x ::s:;y implies the scalar inequality I xll :::::; M IIY II where M is a constant. Several further criteria for normality of a cone exist. One of these is connected with the concept of a u0norm. Let u0 be some fixed nonzero element of the cone K. The element xEEis called u0measurable if ,
for some nonnegative tt and t2 • Let Euo be the set of all u0measurable elements, oc (x) the lower bound of the numbers 1 1 , f3(x) the lower bound of the numbers 1 2 for X E Eu a ·
204
OPERATORS IN SPACES WITH A CONE
This set is a linear space. If we now set l l x l luo
=
max {oc (x), f3 (x)}
for all elements x of E" o ' then the set Euo becomes a normed space, and the nu mber ll xll u o is called the u0norm of the element x. For the normality of the cone K it is necessary and sufficient that the inequality
where the constant M does not depend either on x or on u0, be satisfied. It follows from the last inequality that the normality of the cone K guarantees the completeness of the space Eu with respect to the u0norm. 6. Regular cones. The cone KcE is called regular if the partial order ing generated by it has the property that the convergence of a sequence x (n = 1 , 2, ) with respect to the norm of the space E follows from the relations n
and
. . .
(n = 1 , 2, . . . ) ,
where u is some element of the space E. In other words, the cone K is regular if every monotone (with respect to the cone) and bounded (also with respect to the cone) sequence of elements of the space converges with respect to the norm. The cone of nonnegative functions of the space Lp(P � 1) serves as an example of a regular cone ; the cone of nonnegative functions of the space C is an example of a cone which is not regular. The cones of non negative sequences in the spaces lp(P � 1) and numerical sequences con vergent to zero in the space c0 are regular. The cone of nonnegative sequences of the space m of all bounded sequences does not have the property of regularity. The fact that every regular cone is normal is important. Analogously, we can raise the question of convergence in the norm of every monotone (in the sense of the partial ordering generated by the cone K) sequence of elements of the Banach space E bounded with respect to the norm. It is clear that such convergence occurs far from always. For example, in the space C[O, 1] the monotone, with respect to the cone of
205
CONES IN LINEAR SPACES
nonnegative functions, sequence xn(t) = 1  t " is bounded with respect to the norm and at the same time is not convergent. In this connection one more class of cones is it;�troduced. A cone K is called completely regular if the partial ordering generated by it is such that for an arbitrary sequence xn the convergence of the sequence xn with respect to the norm of the space E follows from the relations x 1 :::::; x 2 :::::; · · · :::::; xn :::::; and from the scalar inequality ll xnll :::::; C(n = 1, 2, . . . ) where C is a constant. The cones of nonnegative functions in the spaces Lp (P � 1) and the cone of nonnegative sequences in lP(p � 1) can serve as examples of completely regular cones. The cone of nonnegative sequences in the space c0 is, as indicated above, regular, but it is not completely regular : the sequence X1
= ( I , 0, 0, . . . ), X 2 = (1, 1 , 0, 0, . . . ) , . . . , Xn = (1, 1 , . . . , 1 , 0, 0, . . . ), . . . • • •
of elements of the space c0 is monotone with respect to this cone, bounded with respect to the norm, but it does not converge. This example shows that not every regular cone is completely regular. However every com pletely regular cone is a regular cone. A regular cone K is completely regular if it satisfies one of the following conditions : 1) the cone K is solid ; 2) the space E is weakly complete. A functional f (x), not necessarily linear, is called positive if f (x) � 0 for xEK ; the functional is called strictly positive iff (x) > 0 for xEK, x =f (} The functional f(x) is called monotone with respect to the cone K if the relation (} :::::; x :::::; y implies the inequalityf(x) ::s:;f(y) . The positive functional f(x) is called strictly increasing if for any hn E K (n = 1 , 2, . . . ) with llhn ll �eo > 0 (n = 1, 2, . . . ) it follows that lim f (h 1 + h 2 + · · · + hn) = 00
•
If a monotone strictly increasing functional can be defined on the cone K, then the cone K is regular. If on the cone K a strictly increasing functional f (x) that is bounded on every sphere can be defined, then the cone K is completely regular. A n example of a stri c tly in creasing functional which is monotone on
206
OPERATORS IN SPACES WITH A CONE
the cone of nonnegative functions of the space LP [0, 1] (p � 1) is the p th power of the norm of an element : 1 P dt . f (x) = lx (t) I 
J 0
A result of a negative character is valid : a solid minihedral cone i n an infinitedimensional space can not be regular. 7. Theorems on the realization of partially ordered spaces. Let K be a normal cone in a separable Banach space E. Then there exists a oneto one linear and continuous mapping of the space E into a subspace of the space C(O, 1) for which the elements of K and only they map into non negative functions. If E is not separable, then an analogous statement with the replace ment of C(O, 1) by the space C(Q) of continuous functions defined on some compactum Q is valid. In the case when a solid normal minihedral cone K occurs in the Banach space E, there exists a linear onetoone and continuous mapping of the space E onto all of the space C(Q), where Q is a compactum, for which K maps onto the set KQ of all nonnegative functions on Q. The following is a special case of the last statement : let E be an ndimensional space and K c E a minihedral solid cone ; then a basis (e�> e 2 , , en) exists in E such that the set of all vectors x = �1e 1 + � 2 e2 + · · + �nen with nonnegative coordinates � i � 0 (i = 1 , 2, . . , n) coincides with K. If a minihedral cone K is given in a separable Banach space E and ll x + Yll = ll x ll + IIYII for x, y E K, then there exists a linear onetoone and continuous mapping of the space E onto the s pace L(O, 1 ), for which K maps onto the set of all almost everywhere nonnegative functions of L(O, 1 ). If E is not separable, then L(O, 1) is replaced by the space L(Q) on some compactum with a measure. • • .
·
.
§ 2. Positive linear functionals
1 . Positive functionals. The positive linear functionals are the most important class of positive functionals (i.e. functionals such that f(x) � O for x � 8).
POSITIVE LINEAR FUNCTIONALS
207
If the cone K in a Banach space is generating, then every positive linear functional is continuous. In the sequel, we understand a positive linear functional to be a continuous functional. Positive linear functionals exist for every cone K. Moreover, for every xEK(x # 8), a positive continuous linear functional f can be indicated such that f(x) > O. If the space E is separable, then a continuous linear functional f(x) can be constructed such that f(x) > O for all xE K(x ¥ 8). If the cone K is solid and u0 is an arbitrary interior element of it, then f(u0) > 0 for every positive functional[ At least one positive functional / can be found such that f(v0) = 0 for an element v0 E K which is not an interior element of the cone K. On an element not belonging to the cone K, at least one of the positive functionals assumes a negative value. Thus, every cone K is characterized uniquely by a set K' of positive linear con tinuous functionals. The set K' E' of positive linear functionals cannot be a cone : it still does not follow from /EK' and f¥ 8 that f¢K' (that is, the set K' is, generally speaking, a wedge). However, K' is a cone if K is a generating cone. The cone K' in this case is called a conjugate cone. Starting from the general form of the linear functionals in concrete function spaces, it is not difficult, as a rule, to establish the general form of the positive linear functionals in these spaces. Let K be the cone of nonnegative functions. Then the general form of a positive linear functional in the space LP (Q) (p � 1) (where Q is some set) is defined by the formula c
f (x) =
J x (t) cp (t) dt , n
where cp ( t) is a nonnegative function from the conjugate space Lq (Q)
(q=pP l·) , L00
=
M. The general form of a positive linear functional in
the space C(O, 1) is described by the formula 1
f (x) =
J x (t) dg (t) , 0
where g (t) is a nondecreasing function.
208
OPERATORS IN SPACES WITH A CONE
Thus, for the cone K of nonnegative functions, the cone K' coincides with the cone of nonnegative functions of the space Lq (the case E = Lp) and with the cone of nondecreasing functions (the case E = C ) . Analogously, the general form of a positive linear functional f in the space !P (p � l) with the cone K of nonnegative sequences is given by the equality f (x) = L X;Y; i= 1 Y = (Yt• Jl, J3, . . . )) , 00
where y = (y1 Jl, y 3, . . . )E lq is an arbitrary nonnegative sequence. The following statement is important : in order for the conjugate cone K' to be generating, it is necessary and sufficient that the cone K be normal. Thus, if the cone K is normal, then for every fE E' there exists a represen Tation f /1  (2 in the form of a difference of positive functionals/1 and /2 . This representation is not unique : ,
f = C ft + g)
(!2 + g )
for arbitrary gEK'. If the cone K is solid and minihedral, then a minimal representation j = f? f� ( f� , j� E K ') exists having the property that !?
POSITIVE LINEAR FUNCTIONALS
209
Another theorem about extension can be formulated for vector lattices. Let E0 be a linear manifold in the vector lattice E, having the property that for any xEE an x' cE can be found majorizing the absolute value of the element x :
l x l :::; x ' . Then every positive linear functional defined on E0 (the cone K0 c E0 consists of xEEo n K) can be extended to a positive linear functional defined on the entire space E. 3. Uniformly positive functionals . The positive linear functional f(x) is called uniformly positive if a positive number a exists such that
f (x) ?o a ll x ll The functional
f (x) =
(x E K) .
J x (t) dt n
will be uniformly positive in the space L(Q) of functions summable on some set Q with the cone K of functions which are nonnegative on Q. In the spaces LP where p > 1 , there are no uniformly positive (on the cone K of nonnegative functions) linear functionals. There are also no uniformly positive linear functionals in the space C. If a uniformly positive functional exists, then the cone is normal and even completely regular. The converse statement is not true. In order for a uniformly positive functional to exist, it is necessary and sufficient that the cone K be included in another cone K1 such that every nonzero element x0 EK be con tained in K1 together with its spherical neighborhood of radius q I x0 I I , where the number q does not depend on the choice of x0 E K. The cone K(F), constructed with respect to a closed bounded convex set F not containing zero in the following way : K( F) consists of all elements x E E which allow the representation x = tz where t ?o 0 and z E F, always has the last property. 4. Bounded functionals on a cone. An additive and homogeneous functional f(x), defined only on elements of the cone K of the Banach
210
OPERATORS IN SPACES WITH A CONE
space E, is called bounded if 11! 11 + =
X E K, [I x ll � l
sup
l f (x) l < oo .
If an additive and homogeneous functional is defined on a solid cone and is nonnegative on it, then it is bounded. A bounded, additive and homogeneous functional, defined on a generat ing (in particular, on a solid) cone, is uniquely extendible to a continuous linear functional defined on the entire space E by the equality f(x) = f(x1) f(x2) where x = x1  x2 and x1, x2 E K. If the cone is not generating, then the indicated extension of the functional to the linear hull of the cone gives an unbounded functional. It turns out that under the condi tion of normality of the cone, every bounded, additive and homogeneous functionalf(x) on K is majorized on the cone K by some positive contin uous linear functional, that is, a (x)EK' exists such that l f (x)l :::;; (x)
(x E K) .
If a norm is introduced in E such that llxll :::;; IIYII for e :::;; x :::;; y, then the functional can be chosen so that II II = Il l+ I I for nonnegative f and II I I :::;; 2 11/+ II in the general case. §
3. Positive linear operators
1 . Concept of a positive operator. Let E be a Banach space with cone K. A linear operator A is called positive if it maps the cone K into itself: AK K. A positive linear operator has the property of monotonicity : for arbitrary elements x, yE E, x :::;; y implies Ax :::;; Ay. In the case of finitedimensional spaces with a cone consisting of vec tors with nonnegative components, the positive linear operators are defined by matrices with nonnegative elements. The linear integral operators c
Acp (t) =
J K (t, s) cp (s) ds n
with nonnegative kernels K(t, s) (t, sEQ ; Q is a closed bounded set of a finitedimensional space) are the most important examples of positive linear operators acting in different spaces of functions. If the kernel
POSITIVE LINEAR OPERATORS
211
K(t, s) satisfies conditions such that the operator acts in a corresponding space, then this operator is a positive linear operator for the cone K of nonnegative functions. If the cone K is solid and an n can be found for every nonzero q> of K such that A " r:p is an interior element of the cone, then the operator A is called strongly positive. If the integral operator acts in the space C, and if some iterate of the \ kernel K( t, s) : KN (t, s) =
J ... J K (t, s1 ) K (s1 , s2) . . . K (sN _ 1 , s) ds 1 . . . dsN  l • n
n
is positive, then this operator will be strongly positive in the space C. A positive linear operator A is called u0bounded below (u0 is a fixed element of K) if for every nonzero q>EK a natural number n and a posi tive number rx = rx(q>) can be found such that rxu0 :::; A " r:p. Analogously, operators are defined to be u0bounded above. It turns out that if the operator A is u0bounded above and below, then the relation
rx (r:p) u0 :::; A " r:p :::; fJ (r:p) u0 is satisfied for every q>EK for some n and rx(r:p), f3(r:p) > 0. The operators, satisfying the last relation, are called u0positive. Let the integral operator act in LP(Q) (p "?3 1). If for this kernel K(t, s) the inequality K (t, s) "";3 m > 0 is satisfied, then the operator will be u0bounded below if we choose as u0 the function u0 (t)= I . In this connection the operator cannot have the property of u0boundedness above. 2. Affirmative eigenvalues. An eigenvalue .A0 t= 0 of the positive oper ator A is called affirmative if it has a corresponding eigenvector e0 in the cone K. This element is called a positive eigenvector of the operator A. An affirmative eigenvalue is always positive. An affirmative eigenvalue has an important property : if the cone K is generating and the operator A is u0positive, then an affirmative eigenvalue is simple and greater than the moduli of the remaining eigenvalues. This statement, generally speaking, loses force if the condition of u0positiveness of the operator A is dropped. I f the element u0 itself is a
212
OPERATORS I N SPACES WITH A CONE
positive eigenvector of the operator A, then it suffices to require u0boundedness above of the operator A instead of u0positiveness. Several theorems on the existence of affirmative eigenvalues can be formulated for completely continuous operators. Let the closure of the linear hull of the cone K be all of the space E. If a positive linear completely continuous operator A has eigenvalues different from zero, then it has an affirmative eigenvalue A.0 not less than the modulus ofany other eigenvalue. The number .A0 is always an affirmative eigenvalue for the operator A' acting in the conjugate space E' with the cone K'. In practice it is convenient to use the following statement : for the positive linear completely continuous operator A, let an element u exist such that u = v  w(v, wEK),  p¢K and A Pu� (l(u((I( > O) for some natural number p ; then the operator A has an affirmative eigenvalue .A0 where A.0 � f)()(. The number .A0 is also an affimative eigenvalue of the operator A' . The preceding results obtain further development if the cone is solid. Let A be a completely continuous linear operator, strongly positive with respect to the solid cone K. Then: 1) The operator A has one and only one (normalized) eigenvector x0 inside K: 2) The adjoint operator A' has one and only one normalized eigenvector
l/1 in K ':
moreover l/1 is a strictly positive functional. 3) The eigenvalue A.0 corresponding to these elements is simple and exceeds the modulus of every other eigenvalue of the operator A. Conversely, if a completely continuous operator has properties 1), 2) , and 3) , then it is strongly positive with respect to K. Theorems on the existence of affirmative eigenvalues can be illustrated with a Fredholm integral equation b
f K (t, s) cp (s) ds
=
A.cp (t)
a
with a nonnegative kernel K(t, s) continuous on the square a < t, s < b. If a system of points sl > s2, ,sP of (a, b) exists such that K(s1 , s2) K ( s2 , s3) K (sp  1 > sP) K (sP , s1 ) > 0 , • • •
• • .
213
POSI TIVE LINEAR OPERATORS
then the equation has a positive eigenvalue .A0 not less in modulus than every other of its eigenvalues. At least one nonnegative solution (eigenfunction) of the integral equation corresponds to this number .A0• If for every continuous nonnegative function cp (s) not identically equal to zero an iterate KN (t, s) can be found such that \ b
J KN (t, s) cp (s) ds
>
0
a
then the Fredholm equation has a unique positive eigenfunction. The transposed equation b
J K (s, t) l/t (s) ds
=
A.l/t (t)
a
has a unique positive solution corresponding to the same positive eigen value. The eigenvalue .A0 is in absolute value greater than all the remain ing eigenvalues of the integral equation. Now let the kernel K(t, s) in the integral equation be a nonnegative function measurable on the square a < t, s < b satisfying the condition b
b
a
a
JJ
2 JK (t, s) l dt ds < oo .
If the inequality K (s1 , s2) K (s2 , s3) . . . K (sP , s 1 )
>
0
is satisfied for some p ?: 2 on a set of points (s 1, s2 , • . • , sP) of positive meas ure in the correspondingpdimensional cube, then in this case the integral equation has at least one eigenvalue such that a positive eigenvalue occurs among the eigenvalues having the largest absolute value. At least one non negative eigenfunction of L 2 corresponds to this positive eigenvalue. 3. Positive operators on a minihedral cone.
Let K be a minihedral solid cone and A a positive completely continuous linear operator having a fixed vector v inside K: Av = v . Then the eigenvalues of the operator A, equal in absolute value to one, are roots of an integral power of one. The sets of fixed vectors of the oper
214
OPERATORS IN SPACES WITH
A
CONE
ators A and A' have bases vi> v2, , v, and l/1 1 , l/1 2 , , l/tro respectively, having the properties : 1) The systems v 1 , v2 , , v, and l/1 1 , l/1 2 , , l/t, are biorthogonal : l/t; (vi) = b ii (i, j= 1, 2, . . . , r). 2) For every pair it=j(i,j= 1, 2, . . . , r) • • •
• • •
• • •
• • •
3) Linear combinations I c;v; (or L c;l/t;) are nonnegative if, and only if all the coefficients are nonnegative. In the linear manifold M 1 of all eigenvectors and associated vectors of the operator A, corresponding to all eigenvalues equal in absolute value to one, we can choose a basis which always has property 3). The operator A allows the expansion A = U1 + A1 where the operator U1 maps all the space E onto M1 and permutes the elements of the basis ; the operator A1 has spectral radius < 1 (see ch. I, § 5, no. 6). In the finitedimensional case the statements mentioned are applicable to the socalled stochastic matrices, i.e., to the matrices (a;k) with non negative elements satisfying conditions n
I a ik = 1 k= 1
(i = 1 , 2, . . . , n) .
For the integral equation b
J K (t, s) cp (s) ds = A.cp (t) a
with a continuous nonnegative kernel, satisfying the condition b
J K (t, s) ds = 1
(a � t � b),
a
the following statements are obtained : a) All eigenvalues in absolute value equal to one are natural roots of unity. b) The set of eigenfunctions, corresponding to the value 1 = 1, has a basis consisting of nonnegative functions cp 1 (s), cp2 (s), . . . , cp,(s) and having the following properties :
215
POSITIVE LINEAR OPERATORS
1 . At least one point exists in (a, b) for every function
l/t; (s) l/ti (s) = 0
(a ::;::; s ::;::; b ; i # j ; i, j
=
1 , 2, . . . , r) .
4. Nonhomogeneous linear equation. For the nonhomogeneous equation A
r, then the non homogeneous equation has a unique solution
5. Invariant functionals and eigenvectors of conjugate operators. A continuous linear functional f(x) is called invariant with respect to the bounded operator A if f(x) = f(Ax) . In other words, an invariant functional is a fixed vector for the con jugate operator : A'f = f . The following statement is very important : if the collection of bounded positive operators {Ah}, commuting with one another, has a common fixed element inside the solid cone K, then a positive functional F(x) exists which is invariant with respect to all the operators Ah.
Example Let G be a commutative group, E be the space of functions x(g) bounded on G, and the operators Ah(h e G) be defined by the equality
216
OPERATORS IN SPACES WITH A CONE
A hx (g) = x (g + h). The function u (g)= 1 is an interior element of the cone of all nonnegative functions of E, which is fixed for the transformations Ah . There exists an invariant functional F (x (g + h)) = F (x (g))
(h E G) .
If a topology is introduced in the group G, then for the defined conditions the functional F(rx) can be represented in the form of an integral, that is an invariant integral exists on the group (see ch. VI, § 2, no. 1). The following statement is more general than the existence theorem for an invariant functional : a positive functional q>, which is a common eigen vector of all the adjoint operators:
exists for every collection {Ah} of pairwise commutative bounded linear operators mapping the interior of a solid cone K into itself. 6. Inconsistent inequalities. If y x rf:K, then we write x :j;; y. Let the positive operator A be u0bounded below where Then
Ax :{; .Ax
for arbitrary nonzero xEK and .A < .A0, and it follows from that Let the operator A be u0bounded above where then
Ax � A.x
for arbitrary nonzero xEK and .A > .A0• Let the operator A be u0positive where
NONLINEAR OPERATORS
217
then the elements A.0x and Ax are incomparable for arbitrary nonzero xeK(x:;6 cu0) : The theorem formulated here is applicable to the comparison of the eigenvalues of two operators. Let A 1 and A 2 be two linear operators, A 1 x � A 2 x for xeK, A 1 is u0bounded below where A 1 u0 � A.0u0• Then every affirmative eigenvalue A 2 of the operator A 2 is less than A 0 • § 4. Nonlinear operators 1 . Basic concepts. Positiveness and monotonicity for nonlinear operators are defined the same as for linear operators. An operator A is positive if AKc K and monotone if Ax � Ay follows from x�y. Unlike linear operators, in the case considered, monotonicity does not follow from the positiveness of the operator. The operator A is strongly differentiable with respect to the cone K at the point x0 if A (x0 + h)  Ax0 = A' (x 0 ) h + w (x0, h) for all h e K, where A' (x0) is a linear operator and lim
h e K, Jihll + 0
The linear operator A' (x0) is called the strong derivative with respect w(x0, h) to the cone of the operator A at the point x0 • In the case where h ll ll weakly tends to e for ll hii .....,. O (h e K), we speak of the weak derivative with respect to the cone. It turns out that the strong derivative A' (x0) of a completely continuous operator with respect to the cone K transforms every bounded set Tc K into a compact set, and the strong derivative with respect to a generating cone K of a completely continuous operator is a completely continuous operator . Along with the derivative with respect to a cone, the derivative at in finity plays an important role in the investigation of nonlinear operators. An operator A is called strongly differentiable at infinity with respect to
218
OPERATORS IN SPACES WITH A CONE
the cone K if a continuous linear operator A' ( oo) exists for which x_ll )_ ( oo I I A_x_A_'_:_ _ ' sup = 0. lim [[ x[[ R + ao ll x ll ;;>R, x e K In this connection A' ( oo) is called a strong asymptotic derivative with respect to the cone K. Analogously, the concept of weak differentiability at infinity is developed. 2. Existence of positive solutions. Here the equation x = Ax with a positive operator A is considered. The solutions of this equation will be fixed elements of the operator. Let the positive continuous operator A have a strong asymptotic derivative A' ( oo ) with respect to a cone and let the spectral radius of the operator A' ( oo ) be less than one. It suffices that one of the following conditions be satisfiedfor the existence of at least onefixed point in the cone: a) the operator A is completely continuous, b) the operator A is monotone and the cone K is completely regular, c) the space E is reflexive and the operator A is weakly continuous. For a completely continuous operator the condition of existence of A' ( oo) can be replaced by the following : an R exists such that for all e > 0, Ax ;;j:: (l + e)x (xeK, ll x ll � R). The collection of elements x for which x0 � x � u0 is called the conical interval <x0, u0) . It suffices for the existence, for an operator monotone on the interval <x0, u0), of at least one fixed point that the operator transform <x0, u0) into itself and that one of the following conditions be satisfied: a) the cone K is strongly minihedral, b) the cone K is regular, the operator A is continuous, c) the cone K is normal, the operator A is completely continuous, d) the cone K is normal, the space E is reflexive, the operator A is weakly continuous. In the satisfaction of conditions b)d), the fixed point of the operator A can be obtained as the limit of the sequence xn = Axn _ 1 (n = l, 2, . . . ). If it is additionally known that a unique fixed point of the operator A lies in (x0, u0), then the successive approximations Yn = AYn  1 (n 1, 2, . . . ) =
NONLINEAR OPERATORS
219
'
converge with respect to the norm to the solution for any y0 e <x0 , u 0 ) in cases b)c). 3. Existence of a nonzero positive solution. When Ae = e, then not unfrequently the question arises of the existence in a cone of a second (different from e) fixed point for a positive A. In several cases the answer to this question can be obtained. We say that the positive operator A (Ae = e) is a contraction of the cone on the part from r to R(O < r < R) if and
Ax :$ x (x e K, Jl x[[ � r, X '# e) Ax i (1 + e) x
(x e K, [[x[[ � R) .
for all e > O. The operator A (Ae =e) is called an expansion of the cone on the part of the cone from r to R(O < r < R) if
Ax i (1 + e) x (x e K, [[x[[ � r, X '# e) for all e > O and
Ax :$ x (x e K, [[x[[ � R) . Let the positive completely continuous operator A be an operator of contraction (expansion) on some part of the cone K; then the operator A has at least one nonzero fixed point on K. The verification of the conditions of contraction or expansion is facilitated if two linear operators A _ and A + exist for which (x e K) . The conditions of contraction will automatically hold if
(x e K, [[ x[[ � r) , (x e K, [[x[[ � R) . Analogously we can verify the conditions for expansion. In this connnec tion it suffices to construct the operator A _ only on the elements of small norm, and the operator A + on elements with a large norm. In the con struction of the operator A _ we often use the derivative A'(e) with respect to the cone and in the construction of A + , the derivative A' ( oo ) at infinity with respect to the cone (see [25]).
220
OPERATORS IN SPACES WITH A CONE
4. Concave operators. Let u0 be a fixed nonzero element of K. The operator A is called u0concave on K if it is positive and monotone and positive numbers 11. and f3 exist for arbitrary nonzero xEK such that and 1J = 1J (x, a, b) > O can be found such that A (tx) � (1 + 1J) t Ax
( a � t � b)
for arbitrary xEK with the condition x � yu 0 (y > 0) and for every seg ment [a, b] c(O, 1 ). The set of those A for which the equation
Ax = AX with a completely continuous u 0concave operator has a nonzero solution in the cone K form some interval (a, f3). The equation cannot have more than one solution different from e in the cone K for every A E (a, {3). If A1 > A2 (A1, A2 E(a, /3)), then for the corresponding solutions of the equation x 1 and x2 in K the inequality x 1 � x2 is valid. If Ae e and the strong derivative A ' (e) with respect to the cone is a completely continuous operator, then the upper bound f3 is a positive eigenvalue of the operator A ' (e). If, in this connection, the operator A' (e) is u0positive, then f3 coincides with the unique positive eigenvalue of the operator A' (e). If the operator A has a strong asymptotic derivative A' ( oo ) with respect to the cone and is a completely continuous u0positive operator, then 11. is a eigenvalue of the operator A ' ( oo). The Uryson integral operator =
1
Ax (t) =
J K (t, s, x (s)) ds , 0
in which the function K (t, s, u) is continuous and K(t, s, u) � O for u � O can serve as an example of a nonlinear positive operator in the space C(O, 1). This operator is monotone if the function K(t, s, u) does not decrease as u increases. Moreover, if K (t, s, 0) =0 and for u2 > u 1 the inequality
NONLINEAR OPERATORS
221
is satisfied for every t for almost all s, then the integral operator will be u0concave. In this connection we take as u0 the function identically equal to 1 . 5. Convergence of successive approximations. Let the equation x = Ax with the u0concave operator A on the normal cone K have a unique non zero solution x* in K. Then the successive approximations xn = Axn _ 1 (n = 1, 2, . . . ) converge to x* whatever the nonzero initial approximation x0 e K is. Moreover, the successive approximations will converge in the u0norm, which, as was indicated in § 1 , no. 5, is stronger than the initial norm of the space E. The condition of u0concavity can be weakened, requiring only the satisfaction of the inequality A (tx) � tAx (O � t � 1) for A (tx). Then the convergence of the successive approximations will occur for arbitrary nonzero initial approximations of K if the cone K is regular or if the operator A is completely continuous.
CHAPTER
VI
COMMUTATIVE NORMED RINGS (BANACH ALGEBRAS)
§ 1. Basic concepts
Commutative normed rings.*) A complex Banach space R, with elements x, y, . , on which there is defined an associative and commutative multiplication xy which is commutative with multiplication by complex 1.
.
.
numbers, distributive with respect to addition, and continuous in each
factor, is called a
commutative normed ring (Banach Algebra) .**)
In the general theory of commutative normed rings, we can restrict ourselves to the consideration of rings with an element
e
such that
ex = x
for every
xe R.
identity element, that is,
an
If the ring does not have an
identity element, then one can be formally adjoined to the ring ; i . e . , we consider the collection of elements of the form adj oined identity element and norm
ll A.e + x ll = IA.I + ll x ll .
x
A.e + x,
where
e
is an
is an arbitrary element of R, with the
In every normed ring with an identity element, we can change the norm to an equivalent norm so that the relations satisfied for the new norm.* **)
ll xy l l � ll x ii · IIYII , II e ll = 1
A set K of elements of the ring R is called a
system of generators
are for
this ring if the smallest closed subring, with an identity element, contain ing K is R. The identity element is not included among the generators. 2.
Examples of normed rings.
1 . Let *)
C (0, 1)
be the space of all complex functions, defined and con
For definitions of a ring, group, algebra,
and other algebraic
definitions,
see any standard text on higher algebra.
**)
From point of view of the terminology of modern algebra, the term "Banach algebra" is more precise, but here the term "normed ring" introduced originally in the works of I. M. Gel'fand, is retained.
***)
Thus the multiplication, which was assumed to be continuous in each factor
separately, is actually continuous in
x
and y simultaneously [Editor].
BASIC CONCEPTS
223
tinuous on the segment [0, 1], equipped with the norm //x/1 = max /x (t)/ . C is a normed ring (with the identity element x (t)= 1) with respect to the usual multiplication. 2. Let c< n> (o, 1) be the space of all complex functions on the same segment [0, 1] which possess a continuous nth order derivative equipped with the norm n max k J x( ) (t)/ \ o "' r "' r . = /x/1 1 k! /...; k=O ·

c
1/x l/ · 1/y/1 . 3. Let W(O, 2n) be the space of all complex functions x (8), continuous on the circle 0 � 8 � 2n and expandable in an absolutely convergent Fourier series im8 e x (e) = " c ' m m =L, oo with the norm 1/ x/1 =
1:
l cm l · The space Wforms a normed ring (with the usual multiplication) where again 1/xy/1 � 1/x/1 · 1/y/1 . We often call the ring W the Wiener ring. 4. Let A be the space of all functions of a complex variable (, defined and continuous on the disk /(/ � 1 and analytic inside this disk, equipped with the norm 1/x/1 = max Jx(OJ . A is a normed ring with the usual 1{1 "' 1 multiplication. 5. Let L 1 (  oo, oo) be the space of all absolutely summable measurable functions on the real line  oo < t < oo with the norm  oo
1/x/1 =
I /x (t)/ dt .
 oo
L 1 forms a normed ring if as multiplication we take convolution : (X * y) (t) =
I x (t  r) y (r) dt .
 ao
More over, 1/x • y/1 � 1/x/1 1/y/1 . ·
I n L1
there is no identity element with
224
COMMUTATIVE NORMED RINGS
respect to the multiplication introduced. We denote by V the normed ring obtained by means of formal adjunction of an identity element to L 1 . 6. Let v< b> be the linear space of all complex functions f(t) of bounded variation on oo < t < oo, satisfying the condition f ( oo )=0 and con tinuous from the right, with the norm 

(!) . 11/ 11 = ( Var oo, oo )
v< b)
is a commutative normed ring if multiplication is defined as convolu tion : 00
 oo
The function B
{
0 � f t < 0, = (t) 1 If t � O
serves as the identity in v< b> , and llell =1. The ring V of example 5 is isomorphically and isometrically (see ch. 1, § 2, no. 4) embeddable in the ring v< b> . This isomorphism is attained if we set the element A.e + x in correspondence with the function
J x(r)dr .  oo Let ct (t) be a positive function, defined and continuous for all real
A.e(t) + 7.
t
values of t and satisfying the condition
for all t1 and t 2 • x (t) for which
L
is the normed space of all measurable functions 00
l l x ll = L
J lx (t)l ct (t) dt < oo .
 oo
forms a commutative normed ring with convolution 00
(x * y) ( t) =
J x (t  r) y ('r) dt
 oo
as multiplication. In this ring there is no identity element. The ring
BASIC CONCEPTS
225
obtained by the formal adjunction of an identity element to L<�> is denoted by v <�>. The ring V from example 5 is a ring v <�> where ct (t)= 1 . 8 . Let L<;.> be the collection of all measurable complex functions x (t) (t � O) satisfying the condition
ll x ll
=
f l x (t) l ct (t) dt < oo , 00
0
where IX (t) is a function as in example 7 but considered only for t � 0. L<;.> forms a normed ring with the multiplication defined by the formula t
(x * y) (t) =
J x (t  r) y (r) dt . 0
Adjoining to L<;.> the formal identity element, we obtain the ring which is denoted by V��>. The rings of examples 7 and 8 are called Wiener rings with weight. 3. Normed fields. An element ye R is called the inverse of an element xE R if xy = e An element of the ring, having an inverse, is called invert ible. If ll x  e ll < 1 , then the element x is invertible and its inverse can be represented in the form of a Neumann series : .
y=
00
(e xt . L = k O 
If every nonzero element of the ring has an inverse, then the ring is called a normed field. Every normed field is isomorphic and isometric to the field of complex numbers. 4. Maximal ideals and multiplicative functionals. A set I of elements of the ring R with an identity element is called an ideal if IR eland II c I where {0} =/= I=/= R (IR is the collection of elements of the form ab where ael and b E R, and I I is the collection of elements of the form a  b where ael and bel). Every ideal is a linear manifold of the space R. If the ideal I is closed, then it forms a subspace. The factor space R/I is itself a normed ring with the natural definitions of the operations and the norm of the coset X defined by II X II = inf ll x ll xoX
226
COMMUTATIVE NORMED RINGS
(see ch. I , § 4, no. 5). This factor space is called the factor ring of the ring R with respect to the ideal I. The concept of a maximal ideal is a central concept in the theory of commutative normed rings. An ideal is called maximal if it is not properly contained in any other ideal. Every maximal ideal is closed. An element XE R has an inverse if and only if it does not belong to any of the maximal ideals. The factor ring, with respect to a maximal ideal, is isomorphic and isometric to the field of complex numbers. Among the continuous linear functionals on R the multiplicative func tionals, that is, those nonzero functionals such that M(xy) =M(x) M (y), play an important role. There is a close connection between maximal ideals and multiplicative functionals : every maximal ideal is a hyperplane M(x) = O for some multiplicative functional, and conversely. This last statement facilitates the description of the maximal ideals of a given ring, reducing this problem to the description of multiplicative functionals. In this way, in the case of the rings C, c< n>, W, A (examples 1, 2, 3, 4), a multiplicative functional represents the evaluation of the functions of the rings at a fixed point of the corresponding domains of definition (segment, circle, or disk). Thus a function x (t ) which is not zero at any point does not belong to any of the maximal ideals of the corre1
sponding ring, and therefore its inverse  also belongs to the ring. This X (t) statement is trivial in the case of the rings C, c< n> and A . However, in application to the ring W, it reduces to the proof of the wellknown Wiener theorem : if the function x (8) is expandable in an absolutely convergent Fourier series and does not vanish, then
1
is expandable
x (e) in an absolutely convergent Fourier series. This idea, applied in the proof of the Wiener theorem, is applicable in many other cases. L 1 forms a maximal ideal in the ring V (example 5). All remaining maximal ideals in V can be described in the following manner. Let s be a real number and M. the collection of elements A.e +xe V for which 00
A. +
J eist X (t) dt
=
0.
 oo
M. forms a maximal ideal in V. Other maximal ideals, besides L 1 and M. (  oo < s < oo ), do not exist in the ring V.
227
BASIC CONCEPTS
An element A.e+ x(t)E V is invertible if and only if
 oo
is different from zero for arbitrary s. The maximal ideals in the rings V and V�a> can be described in an analogous manner (examples 7 and 8). Thus, in the case of the ring V , besides L< a > , maximal ideals Ms of the above indicated type still occur but s can now be an arbitrary complex number for which . In o: (t) . In o: (  t) . hm � Im s � hm t t t > + oo t> + oo 


In the case of the ring
v�a > ,
s is an arbitrary complex number for which
In o: (t) lm s � lim . t t > + oo

The situation is more complicated for the ring v (example 6). How ever, in this case all maximal ideals have also been described (see [9]).
5. Maximal ideal space. The set 9Jl of all multiplicative functionals on R is closed in the weak* topology a (R', R). Every multiplicative func tional has norm 1 , hence 9Jl is a weak* closed subset of the unit sphere of the conjugate space which is compact in this topology (see ch. I, § 4, no. 4). Thus 9Jl is a compact set in the weak* topology a (R', R). This set is called the maximal ideal space. It often is useful to set the elements of the ring R in correspondence with continuous functions on the set 9Jl by setting x (M) =M(x) for xe R, Me 9Jl. The correspondence x + x (M) is a normnonincreasing homomorphism of R into the ring C (Wl) of all continuous functions on 9Jl equipped with the usual norm (i.e., the norm 1 \ x l\ sup jx(M)I). In =
M E IDl
the case when the correspondence is an isomorphism, R is called a ring offunctions. We always have max jx (M)I = lim y/�x·� � 1 \ xll .
M e iDl
n  oo
The intersection of all maximal ideals is called the radical of the ring. x (M)=O for elements in the radical. The elements x for which z/ llx" ll +0
228
COMMUTATIVE NORMED RINGS
are called generalized nilpotent elements. The radical of the ring coincides with the collection of all generalized nilpotent elements. The set of values of the function x(M ) on the space Wlcoincides with the spectrum of the element x, i.e., with the set of those complex numbers A for which x  Ae does not have an inverse in R. The ring R is, by definition, a direct sum of its ideals /1 and 12 if every element x E R can be expanded (in an unique manner) in a sum x = x 1 + x2 , where x1 E /1 , x 2 E 12 • The ring R is a direct sum of its ideals if and only if the space 9Jl is not connected, i.e. is representable in the form of the union of two nonempty disjoint closed sets. 6. Ring boundary of the space Wl. The smallest closed subset r c9Jl on which all the functions ! x (M) ! (x ER) attain a maximum is called the ring boundary of the space 9Jl (G. E. Silov). Such a set exists and is unique. For example, in the case of the ring A (example 4), the maximal ideal space 9Jl can be set into onetoone correspondence with the disk I ' I � 1 ; the ring boundary r consists of points lying on the boundary !( I = 1 of this disk. This and similar examples justify the name of ring boundary. If the ring R is isometrically imbedded in some larger ring R 1 , then every continuous linear functional can be extended to a continuous linear functional over R 1 • In this connection, the property of the functional being multiplicative is not retained. This means that not every maximal ideal of the ring R is the intersection of this ring with some maximal ideal in the larger ring R1 • As for maximal ideals corresponding to points on the boundary r, they all extend to maximal ideals in an arbitrary larger ring R 1 . In this connection, if the norm in the initial ring is the maximum modulus on Wl, then a ring R 1 R exists such that only max imal ideals corresponding to points of r extend to the maximal ideals of this ring. The ring C(r) of all continuous functions on r (in example 4, the ring of all continuous functions on the circle I (I = 1) is such a ring R 1 • ::::>
7. Analytic functions on a ring. A function x ;. with values in a given . normed ring R is called analytic if it is defined in some region of the com plex variable A and the ratio X;. +h  X;.
h
229
BASIC CONCEPTS
x� E R h
converges in the norm to some limit as '> 0. For example, the 1 is analytic in the complement of the spectrum of the function element For analytic functions with values in a ring, a considerable part of the elementary theory of ordinary analytic functions is extendible  in par ticular, the Cauchy theorem, the Cauchy integral formula, and the Liouville theorem. If/ (0 is an entire function, then in the ring for every element an element can be defined by means of the Taylor series. This element M for all M 9Jl. has the property that/ (M) The Cauchy integral formula allows us to make this result more pre cise. Namely, if
e)(xA. x.
R,
f (x)
x,
(x) = f (x( )) E i s some element of a ring and the function f(O i s analytic x in a neighborhood of the spectrum of the element x, then the formula � =f(x) = 2nz J (x  A.e)  1f (A.) dA. , where i s an arbitrary rectifiable contour enclosing the spectrum of x and of analyticity. of the function f(O, defines an element lyERjor ying in thewhichdomain y(M) f(x(M)) Y
y
In this statement, if it is formulated for the ring W, is contained the gen eralization of the abovementioned Wiener theorem, attributed to P . Levy. Thus, to the element of a normed ring, we can apply any function analytic in a neighborhood of the spectrum of In spite of the fact that this class of functions is narrow, already in the case of the ring W it, generally speaking, is impossible to extend it. The Cauchy integral formula effects a homomorphic mapping /(0+ of the ring of functions analytic in a neighborhood of the spec trum of the element into the ring The theory of functions of elements of a ring has many applications. For example, theorems of the WienerLevy type play an important role in the construction of the theory of regular and singular integral equations. We can also apply a function / (( 1 , . . . (n) of several independent vari ables ( 1 , , (n to the collection t > , xn of elements of a normed ring if this function is analytic in a neighborhood of the joint spectrum of the elements x 1 , . . . , xn, i.e., in a neighborhood of the set (in ndimensional com plex space)
x
f(O
f(x)
x
x.
f(O R.
,
. • •
x
• • •
230
COMMUTATIVE NORMED RINGS
8. Invariant subspaces of R' . Let R' be the conjugate space of the normed ring R. A subspace Pc R' is called invariant if from the relation fEP it follows that, for any x0 ER , the functionalfxo (x) f(x0x) also belongs to
P.
If I is an ideal in R, then its orthogonal complement (see ch. I, § 4, no. 5) will be an invariant subspace of R' which is closed in the weak* topology a (R' , R). Conversely, every weak* closed invariant subspace of R' is the orthogonal complement of some ideal of the ring R. The invariant subspace orthogonal to a maximal ideal M c R is one dimensional and consists of multiples of the functional M(x) ; the con verse is also true. Since every ideal is contained in a maximal ideal, every nonzero weak* closed invariant subspace P c R' contains a one dimensional invariant subspace. In particular, let W' be the conjugate space of the Wiener ring W (example 3). The space W' can be identified with the space of all bounded sequences { .. , f_ 1 , f0 , /1 , . . . } such that if .
iko x (B) = L ck e E W, 00
 oo
then
The statement formulated above in terms of invariant subspaces in a conjugate space reduces then, in application to the space W', to the follow ing theorem : every weak* closed subspace of the space of sequences, in
with respect to displacements, contains a subsequence of the form variant ik
{e 6o} . Every weak* closed invariant subspace P c W', containing only one kO i o sequence {e } , is onedimensional and consists of multiples of this sequence.
9. Rings with involution. A normed ring is called a ring with involution if an operation X4X* is introduced in it which has the properties: 1) (x*)* = x, 2) (..h+ .uy )* h* + .uy *, 3) (xy)* =y *x *. The last property is written in a form in which it is extendible to non commutative rings. A most important example of a noncommutative =
GROUP RINGS. HARMONIC ANALYSIS
23 1
ring with involution is the ring of bounded operators on a Hilbert space ; in this case, the operation * represents passage to the adjoint operator. If the involution has the property
4) 1/x*x [[ = IJ x* il " ll x iJ ,
then the ring is isomorphic and isometric to a subring of the ring of operators on a Hilbert space.
A commutative normed ring with involution, having property 4), is iso morphic and isometric to the ring of all continuous functions on its maximal ideal space.
If this theorem is applied to the ring of operators commuting with every operator which commutes with a given selfadj oint or normal oper ator, then an analogue of the spectral expansion of the operator is obtained. The involution x*+x is called symmetric if the element e +x*x is invertible for arbitrary x ER. A linear functional J, defined on a ring R with involution, is called positive if f(x*x) �O for all x E R. Every po sitive functional on a commutative normed ring R with a symmetric involution is uniquely extendible to a positive linear func tional on the space C(Wl) of all complex functions continuous on the space 9Jl of maximal ideals of the ring R, and therefore it is representable in the form
f(x) =
J x (M) dfl (M) , m
where f.l (M) is a positive measure on Wl. If, for every nonzero element x0 ER, a positive functional fo exists such thatf0 (x0 x� ) # 0, then the ring R does not have a (nontrivial) radical. § 2. Group rings. Harmonic analysis
1. Group rings. Let G be a group with a finite number of elements and let R be the linear system consisting of formal linear combinations of elements of the group :
(x E R) , where the x9 are arbitrary complex nu mbers. I n the system R, an opera tion of multiplication of elements can be introduced in a natural way :
232
COMM UTATIVE NORMED RINGS
if x = l: x9g and y=L y9g, then g g g'
g' g"
g"
If we denote the product g'g" =g, then g" = (g')  1 g and the formula for multiplication assumes the form x y = L (L X9• Y(g') 19) g . g g' With the operation of multiplication thus introduced, the system R be comes a ring (algebraic) which is called the group ring of the finite group G. In the sequel, commutative groups are considered, and additive nota tion is used for the group operation. The formula for multiplication is written in the form xy = L (L x g' Ygg') g g g' ·
The group ring of a commutative group is commutative. The formula for multiplication can be written in coordinate form : (xy )9 = L X9• Y9  9• • g'
The group ring becomes a finitedimensional commutative normed ring if we introduce the norm Moreover,
ll xy [[ < [[ x ll · [[ y ll .
The group ring of a discrete commutative group G with an infinite number of elements is constructed analogously. Here it is more conve nient to consider elements of the group ring R to be complexvalued functions x (g) of the elements of the group. The multiplication in R is introduced by the formula (xy) (g) = L x (g ') y (g  g ' ) , g' and the norm by the formula [[ x [[ = L [ x (g) [ . g
GROUP RINGS. HARMONIC ANALYSIS
233
The ring R consists of all functions x(g) for which [ [x [ [ < oo . It is an infinitedimensional commutative normed ring. For the consideration of continuous groups, in the preceding formulas, sums are replaced by integrals. Furthermore, it is assumed that an invariant integral exists on the group, that is, an integral having the prop erty
J x (g + g0) dg = J x (g) dg
for arbitrary g0 E G .
The ring R consists of the space L 1 (G) of summable functions x (g) with the norm
[ [ x [[ =
J l x (g)! dg
and with the operation of multiplication in the form of the convolution
(xy) (g) =
J x (g') y (g  g') dg' .
The group ring of a discrete group has an identity element. It is the function l, �f g = O , e (g) = 0, If g # 0 .
{
The ring L 1 (G) of a nondiscrete continuous group does not contain an identity element, since the function e(g) is equivalent to zero in L1 (G). Therefore, in this case the ring V(G) obtained by means of a formal adjoining of an identity element to the ring L 1 (G) is called the group ring. The formally adjoined identity element can be treated as the <5function on the group G (see ch. VIII). The ring V(  oo, oo) (example 5), corresponding to the group G of all real numbers, is a special case of a group ring. We can introduce an involution in group rings. By definition, x* (g) = x(g) and e* =e for the identity element of the ring. It turns out that group rings do not have (nontrivial) radicals, that is, their radicals consist only of the zero element. Thus a group ring is isomorphic to the ring of functions on the space 9J1 of maximal ideals. 2. The characters of a discrete group and maximal ideals of a group ring. A function x (g ) # O having the following two properties is called a
234
COMMUTATIVE NORMED RINGS
character of the discrete group G: x (g + h) = x (g ) x ( h ) and x* (g) = x (  g ) = x (g) . It follows from the definition that l x(g) l = The function x (g )= 1 is a trivial character. We say that the group G has a sufficient number of characters if a character Xo can be found such that Xo (g 0) :;<: 1 for every g 0 E G. From the characters of the group, we can construct multiplicative functionals on the group ring. If the group is discrete, then the multi plicative functionals are constructed according to the formula
1.
M (x) = L x (g) x (g ) . g
It turns out that this construction exhausts all the multiplicative func tionals on the group ring. Thus, between multiplicative functionals and characters, and hence, between maximal ideals and characters, a oneto one correspondence is established. A maximal ideal M consists of all functions orthogonal to some character :
L x (g) x (g) = o . g
The character is reconstructed from the maximal ideal by the formula where
ego = eo (g  Oo) =
{1
0
if g = Oo , if g # go .
The space 9Jl of maximal ideals is compact in the weak topology, and therefore the space X of characters of the group will also be compact in the corresponding topology. This topology is defined by means of a fundamental system of neighborhoods of the elements Xo : where { x 1 , . . . , xn} is an arbitrary finite set of elements of the group ring and e > O. The character x belongs to U(x0, x 1 , . . . , xm e) if
1 , 2,
. . . , n) . ILxk (g) [x (g )  Xo (g)] l < e (k = g The topology indicated coincides with the topology introduced by L. S. Pontryagin by means of a fundamental system of neighborhoods
GROUP RINGS. HARMONIC ANALYSIS
U(x0 ; F, e) consisting of characters
235
x for which
i x (g )  Xo (g) i < e for g E F. Here F runs through the arbitrary finite subsets of the group G. In the set of characters x, the operation of multiplication is introduced in the natural way : the product x1 (g) x2 (g) of two characters of the group will again be a character of the group. The operation of multiplica tion is continuous in the topology of the space X. Thus, the space of characters of a discrete group is a compact com
mutative topological group. The group of integers {m} is a simple example of a discrete group. The group ring consists of sequences x = {e m} with the law of multiplica tion
(xy)m =
L Cm,dmm'
m' =  oo
and the norm This ring is isomorphic and isometric to the Wiener ring W (example 3). The functions eimB are the characters of the group. For () differing by an integral multiple of 2n, the characters coincide ; therefore we can assume that 0 � () < 2n, and the characters are representable as points on the unit circle. Thus the unit circle is the group of characters of the group of integers. Multiplicative functionals are given by the formula
M (x) =
00
m L cmei B
m =  oo
and coincide , as noted above, with the values of the functions of the Wiener ring at the point e. 3. Compact groups. Principle of duality. An invariant integral (see [38]) exists on a compact group ; therefore the group ring V(G) can be constructed for it. The elements of this ring have the form A.e + x (g ), where e is a formally adjoined identity element and x(g) EL1 (G). A trivial maximal ideal M00 = L 1 (G) occurs in the ring V(G), and the cor responding multiplicative functional is M00 (A.e + x) = A.. The characters of the group are defined in the same way as for a dis crete group, with the additional requirement of conti nuity. A character
236
COMMUTATIVE NORMED RINGS
defines a multiplicative functional according to the formula
M (A.e + x) = A. +
J x (g) x (g) dg .
It turns out that this construction enables us to obtain all nontrivial maximal ideals. The maximal ideal space, after rejection of the one point M00, becomes a discrete set. Thus, the group of characters of a compact
group is a discrete group. A compact group has a sufficient set of characters which form a com plete orthogonal system offunctions on the group G. If we construct the group of characters G' for the group of characters X, it will be a compact group. It turns out that G' is isomorphic to the initial group G. The isomorphism is given by the equality
where CfJ x is an element of the group G'. The topology of the group of characters G' coincides with the topology of the initial group. Thus, the groups G and G' can be identified. If we consider a discrete group G, its group of characters X, and the group of characters G' of the group X, then the groups G and G' are also isomorphic. These last statements constitute the content of the principle of duality of L. S. Pontryagin.
4. Locally compact groups.
A group G is called locally compact if it has a compact neighborhood of zero. An invariant integral also exists on a locally compact group. The elements of the group ring V(G ) have the same form as in the case of a compact group. The maximal ideal space, after removal from it of the trivial maximal ideal M00 = L1 (G), ceases to be compact. Thus the group of characters of a locally compact group itself is a locally compact group. A locally compact group has a sufficient number of characters, and . the Pontryagin principle of duality is also valid for it. The difference from the case of a compact group is that the characters do not belong to the ring L 1 (G), and we must consider them as elements of a conjugate space. The group of all real numbers  oo < t< oo is the simplest example of a locally compact group. The group ring is the ring V (  oo, oo ) (example 5). The functions e ist are the characters of the group for an arbitrary real s .
237
GROUP RINGS. HARMONIC ANALYSIS
The group of characters is also the group of real numbers. The representa tion of an element of the group ring in the form of a function on the space of maximal ideals X
(Ms)
00
=
J
 oo
X
(t) eist dt
corresponds to the passage from a function in L1 ( oo , oo) to its Fourier transform. An analogue of the Fourier transform can be constructed on an arbitrary locally compact group by means of its characters. 
5. Fourier transforms. The passage from a function on the group G to a function on its group of characters X according to the formula Tx (x)
=
I x (g) x (g) dg G
is called a Fourier transform. A uniqueness theorem holds : if the Fourier transforms of two functions coincide, then the functions themselves coincide for almost all g EG. The operator T, considered on the intersection of the spaces L1 (G) and L2 (G), allows closure to an operator isometrically mapping the space L2 (G) onto the space L2 (X). The inverse operator on L1 (X) n L2 (X) is given by the formula
Positive definite functions, that is, functions cp (g ) such that the inequality n
I
n
I1 cp (g k  g ,) �k� l � o k = l 1=
is satisfied for an arbitrary finite set of elements g � > . . . , g n of the group G and arbitrary complex numbers � � > . . . , �m form an important class of functions on the group G. A positive linear functional on the group ring can be constructed from a positive definite function on the group by the formula
f"' (.A.e + x) = .A.cp (O) +
I x (g) cp (  g ) dg .
238
COMMUTATIVE NORMED RINGS
Every positive functional / on V(G) is representable in the form f(2e + x) = AQ + f"' (2 e + x) , where cp is uniquely determined by the functionaljand is a positive definite function on G, and 12 > 0 . The theorem on the representation of a positive definite function on a group follows from the theorem on the representation of a positive functional (see § 1 , no. 8) : a continuous function on a locally compact commutative group G is positive definite if and only if it is the Fourier transform of a nonnegative measure (defined uniquely with respect to cp) on the group of characters: cp (g ) =
J i (g) d
"' (o) 2kk .' n ( n + 2) . . (n + 2k  2) =
where
L1
is the Laplace operator.
.
 2, the convergent integral c/·.z",
� f z;.z"
defines a generalized function z'z" on K. This function is homogeneous : the equality
(z;.z",
is valid for an arbitrary function