Documentos de Académico
Documentos de Profesional
Documentos de Cultura
Functional Analysis q
Linear Functional Analysis
Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
Hong Kong Barcelona
Budapest
Encyclopaedia of
Mathematical Sciences
Volume 19
Editor-in-Chief:
R.V. Gamkrelidze
Institute,
Fontanka
27,
Author
Yu.1. Lyubich, Department of Mathematics, State University
Stony Brook, Stony Brook, NY 11794-3651, USA
of New York at
Translator
I. Tweddle, Department
Tower, 26 Richmond
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Chapter 1. Classical Concrete Problems
..........................
4
6
6
6
7
10
10
11
16
19
19
21
21
27
28
29
34
35
36
39
45
47
51
53
53
56
58
Contents
and Methods
60
61
65
69
70
70
74 f)
76 77
78
82
............................
85
85
85
88
93
95
97
100
101
107
107
109
111
112
113
115
117
119
119
127
132
138
140
144
147
153
163
165
165
174
178
181
Q1. Infinite-Dimensional
Linear Algebra ..........................
..................................
1.1. Bases and Dimension
.................
1.2. Homomorphisms
and Linear Functionals
1.3. The Algebraic Theory of the Index .......................
1.4. Systems of Linear Equations ............................
1.5. Algebraic Operators ...................................
1.6. General Principles of Summation of Series ................
1.7. Commutative Algebra
.................................
4 2. Convex Analysis ...........................................
2.1. Convex Sets ..........................................
...................................
2.2. Convex Functionals
2.3. Seminorms and Norms .................................
............................
2.4. The Hahn-Banach Theorem
................................
2.5. Separating Hyperplanes
........................
2.6. Non-negative Linear Functionals
2.7. Ordered Linear Spaces .................................
9 3. Linear Topology
...........................................
3.1. Linear Topological Spaces ..............................
3.2. Continuous Linear Functionals ..........................
3.3. Complete Systems and Topological Bases .................
3.4. Extreme Points of Compact Convex Sets ..................
3.5. Integration of Vector-Functions and Measures .............
........................................
3.6. w*-Topologies
.....................................
3.7. Theory of Duality
...........................
3.8. Continuous Homomorphisms
3.9. Linearisation of Mappings ..............................
0 4. Theory of Operators ........................................
4.1. Compact Operators
...................................
4.2. The Fixed Point Principle
..............................
...............
4.3. Actions and Representations of Semigroups
.........
4.4. The Spectrum and Resolvent of a Linear Operator
/
Contents
on the Bibliography
................................
197
206
207
211
215
216
220
222
227
232
242
242
244
252
255
261
Bibliography
.................................................
261
Author Index
.................................................
273
Subject Index
.................................................
277
<
Preface
Chapter 1
Classical Concrete Problems
$1. Elementary Analysis
The overwhelming majority of functional relations encountered in mathematics and its applications are non-linear
nthe large but linear
nthe small For
example the linear theory of elasticity is founded on the assumption that stresses
depend linearly on deformations when the latter are small. The method of
mathematical analysis, which is firmly established in the works of Newton and
Leibniz, consists of two basic procedures:
local linearisation of the functions being studied (differentiation) and reconstruction of the global relation according to its local structure (integration .
Functional analysis was conceived in the womb of classical mathematical
analysis but it took shape as an independent discipline only at the end of the
19th and the beginning of the 20th centuries.
1.1. Differentiation.
If we speak about classical analysis using modern language, we can say that one of its fundamental objects is a dlfirentiable real or
complex function of n real variables, i.e. a mapping f: G --f @ (G a domain in
Euclidean space W), having the property that at each point x E G the increment
f(x + h) -f(x) as a function of the displacement vector h can be approximated
by a linear function to within order o( lhl) as Ihl+ 0:
f(x + 4 -f(x)
The uniquely determined
= Mx),~) + 4lW.
(1)
df of tke function
linear part
5 1. Elementary Analysis
that it belongs to the class2 C(G). The class C(G) is defined as the intersection
of all the P(G). The enumerated classes can also be considered on closed
domains, for example on the interval [a, b] c R (and even on arbitrary subsets
x of W).
If the function f belongs to C(G) (I c co) then at any point x E G we have
TaylorS formula
f(X + dx) =f(x)
+ mtl i
i,+..TinE,
dxilarnfaxin dxfl
1
xi
+ o(ldxl Y
i.e. to within order o(ldxI the function f is a polynomial of rth degree in dx.
The inner sum in Taylor
formula is the mth differential of the function f at the
point x.
For a function f E C G) we can formally write down its Taylor series by
passing in Taylors formula to I = co (and of course rejecting the error term). If
the Taylor series of a function f E Cm(G) converges to it in some neighbourhood
of the point x E G, then f is said to be analytic at the point x. A function which
is analytic at all points of the domain G is said to be analytic on G. A function
which is analytic on the whole space is said to be an entire function.
All that has been said above carries over immediately to complex functions of
n complex variables. In the 18th century the latter situation had already begun
to separate out little by little into a distinct branch of mathematical analysis
(complex analysis) in view of the fact that complex analytic functions of one or
several complex variables have very distinctive properties. In particular, from
the definition of an analytic function (complex or real) one easily obtains the
uniqueness theorem: an analytic function which is flat at some point (i.e. such that
all its derivatives are equal to zero at this point) is identically zero.
The standard example of an infinitely differentiable non-zero function which
is flat (at x = 0) is
f(x)
= {z
I: f ;;
In what follows, either it will be stated or it will be clear from the context
whether the functions under consideration and their independent variables are
real or complex.
1.2. Solution of Non-linear Equations. This is one of the first applications of
linearisation. Let us consider the equation f(x) = 0, where f is a real function
which is smooth on a certain interval [a, b] c R. The existence of a root is
guaranteed if, for example, f(u) < 0 and f(b) > 0. The uniqueness of the root is
guaranteed if the derivative f x) > 0 for all x. Assuming these conditions are
satisfied, let us denote the root by X. If some approximation
x,, to the root x is
known then, using (l), we obtain
f(x0)
+f xo)@
- x0) = 0,
*For convenience we also put Co(G) = C(G), where C(G) is the class of continuous functions on G.
from which
XSC1~Xo-flo.
f(xo)
In exactly the same way the next approximation x2 is formed from x1 (if a < x1 <
b) and so on. This iterative process is called Newton
process or the method of
tangents, since formula (2) corresponds to approximating
the graph of the function f by the tangent to it at the point (x,,f(x,)).
Theorem 1. Let f be a function
the conditions:
If x)1 < M.
lf
I% --xl
2/J
d--q,
5M
(m=O,1,2
process is defined on
,... ).
x m+l =fhA
Cm= 0, 1, Z...).
5 1. Elementary Analysis
-f(x*)l
G 41x1 - x21
(5)
A sufficient condition
and
If l s 4.
Theorem 2 is widely used in applications.
Example.
equation
x-ssinx=[
(6)
with unknown x and parameters E, c (0 < E c 1; -co < c < co). Since the
function f(x) = E sin x + [ satisfies the conditions of the previous theorem,
equation (6) has a unique real root X, which can be found by means of the
iterations
x,+i = E sin x, + c (m = 0, 1,2, . . .) with any x0.
Theorem 2 extends without change in statement (or proof) to mappings of the
semiaxis or of any interval into itself4. This is important because many of its
applications are of a local character.
Example (Implicit function theorem). Let f (x1, x2) be a function of class C2 in
a neighbourhood
df
of the point (0,O) E lRz such that f (0,O) = 0. Then if ax
# 0,
2 (0.0)
there exists on some neighbourhood lx1 1 < q a unique function x2 = &xi) which
satisfies the equation f (x1, x2) = 0 and the condition 4(O) = 0.
In fact the equation under consideration can be written in the form
x2
= 1x1 + P(X,,
x2),
(7)
10
Vfb) = 0,
(8)
i.e. x is a statihury (critical) point for $ This necessary condition for an extremum
is one of the first basic results of differential calculus. Its discovery is due to
Fermat (1643). Further development in this direction came about on the creation
of variational calculus (Johann Bernoulli, Euler, Lagrange).
In variational calculus we consider functions of variables which are of a more
complicated nature than points in R. For example, in the classical brachistochrone
problem it is required to find the curve joining two given points in a vertical
plane such that a heavy particle moving along it takes the least possible time.
Here the time of motion is a function on the set of smooth curves joining the
given points. Functions of such a type are given the name functionals. More
abstractly a functional is defined as a real or complex function on any set X (i.e.
as a mapping of X --) R or X + C).
In variational calculus necessary conditions of the type (8) are established for
a wide range of problems on the extremum of functionals. Condition (8) in
coordinates represents a system of n equations with n unknowns xi, . . . , x,:
af 0
ig=
5 1. Elementary Analysis
11
dimensional as in general are the great majority of function spaces5 (with the
exception of the case where functions on a finite set are being considered).
The simplest class of functionals on a linear space is the class of linear functionals, i.e. additive, homogeneous functionals:
Nfl
+fd
= t4fl)
+ bKc2),
Naff) = ati(f)
a linear functional is a
A homomorphism
A of
a homomorphism
of a
on E (with domain of
@f)(x) = g1 4(x)g.
k
f H g is a linear operator
12
elements. Cauchy (1823) gave this idea a precise meaning when applied to
continuous functions. Then Riemann (1854) introduced the class of all functions
for which Cauchy definition of the integral is meaningful. For continuous
functions Stieltjes (1894) introduced integration with respect to a eight function
which allowed the connection to be made between ure integration and discrete
summation. These achievements of the 19th century were superseded in the 20th
century by the works of Lebesgue and his successors, as a result of which the
general theory of measure and integration was created. This was accompanied
by an important extension of the collection of function spaces: there appeared
the space L X, ,u) of Lebesgue integrable functions on any set (space) X endowed
with some measure p, the space L X, p) of measurable functions the square of
whose modulus is in L1(X, p) and, following this, the entire family Lp(X, p)
(1 < p < co) of spaces of measurable functions for which the pth power of the
modulus belongs to L X, p).
A measure p on X is a countably additive function on some a-algebra of
subsets of the set X which takes values in [0, co]. If p(X) < co the measure is
said to be finite. The sets appearing in the associated a-algebra are said to be
measurable. A real function f on X is said to be measurable9 if for any half open
interval [a, b) c [w its inverse image f- [a, b) is a measurable subset of X lo.
Given a measurable function f 2 0, i.e. f(x) > 0 for all x E X, then its Lebesgue
integral with respect to the measure p is defined as
f dP = sup kzo WW %
~,+A
where the supremum is taken over all sequences 0 = a, < a, < a, < . . . , a, + co.
This integral is a non-negative number or + co. In the first case the function f
is said to be Lebesgue integrable or summable.
For a real measurable function whose sign is unrestricted we put f+ = max(f, 0),
f- = - min(f, 0) and say that f is summable if both functions f+ , f- are summable, or equivalently IfI is summable. Further, by definition,
0 1. Elementary Analysis
13
then f(x) # 0 only on a set of measure zero or, as it is usually expressed, f(x) = 0
Equality almost everywhere is an equivalence relation on the
set of all measurable functions; moreover equivalent functions fi, fi are simultaneously summable or not summable and in the first case
almost eoerywhere.
This allows us to identify them in what follows. From now on L(X, 11)and along
with it all the L!(X, FL)are in fact converted into linear spaces which consist of
the classes under the stated equivalence relation.
The space L(X, p) occupies a special place in the family Lp(X, p) (p 2 1)
because of the fact that on it there is a natural scalar product
(this definition is meaningful by virtue of the inequality labI < t(lal + lb12)).
All the properties required in linear algebra of a scalar product are fulfilled here.
Therefore in L2(X, ,u) we have a Euclidean geometry (in general, it is inlinitedimensional). In particular we can speak of orthogonal functions fi, f2: (fi ,f2) = 0.
A function f is said to be normalised if (f,f) = 1. A system of functions { fi}y is
said to be orthogonal if (fi,fk) = 0 (i # k) and orthonormal if (fi,fk) = 8ik.
It is particularly important that we can define in L(X, ,u) the distance d(f, ,fi)
between any two elements by introducing, just as in Euclidean geometry, the
norm [IfI/ = ,/(f,f) and setting d(f,,f,)
= [Ifi -f211. The triangle inequality for
the distance follows from the triangle inequality for the norm
Theorem.
then
14
Theorem.
G llfill
+ Ilfill.
A norm generates a metric (just as in the examples given above) and, consequently, a topology. A linear space provided with a norm is said to be normed. A
normed linear space which is complete with respect to the corresponding metric
is called a Banuch space. Banach not only initiated the use of this important class
of spaces but he also developed its theory deeply. l1
A Banach space in which the norm is generated by a scalar product (i.e.
)If II = J( f, f)) is called a Hilbert spuce12 (without the completeness requirement
it is called a pre-Hilbert space). The space L2(X, ZJ)is a Hilbert space and all the
spaces Lp(X, p) (p 2 1) are Banach spaces. The most interesting and furthest
developed aspects of functional analysis are connected with Hilbert spaces.
Let us return to the question of integration of a function. Let Z be any interval
on the real axis and d(x) (x E I) a non-decreasing function. For any bounded
interval (a, /3) c Z we shall call the increment A(8 - 0) - n(a + 0) its I-measure;
the jumps of the function L at the boundary points of the interval Z will be called
the A-measures of these points. By means of this a certain measure Z.L~can be
uniquely defined first of all on the a-algebra of Bore1 sets13 and then also on the
a-algebra of sets of the form B u N where B is a Bore1 set and p:(N) = 0 (,u* is
the outer measure defined by a measure p, i.e. p*(A) is the inlimum of the measures
of all measurable sets which contain A). The latter is called the Lebesgue a-algebra
for a given function I and the measure pLIis called the Lebesgue-Stieltjes measure
defined by the function 1. The integral of a function f E L Z, pl) is usually written
in the form
ii Even finite-dimensional normed spaces (all of which are complete) are objects which are highly
interesting and also useful for application (for example, in numerical analysis).
r21n honour of Hilbert who introduced (in the setting of L*) methods of Euclidean geometry into
analysis.
i3 This is the smallest a-algebra containing the compact sets Q c 1. If a measure on a locally compact
opological space X is defined on all the Bore1 sets and is finite on all the compact sets then it is
called a Bore! measure.
4 1. Elementary Analysis
15
If f is continuous then this integral is just the usual Stieltjes integral. Consequently for any fE J!,(I, PJ the integral (9) is called its Lebesgue-Stieltjes integral.
The space Lp(Z, pl) is often denoted by Lf;(Z) or more concisely Lf; (provided the
domain of,definition appears in the description of the function A). In the case
J,(x) = x the mesure pI is called Lebesgue measure. For this measure we write
briefly Lp(Z) or even simply Lp if the interval Z is already fixed.
The theory of Lebesgue also provides a suitable context for the indefinite
integral, characterising functions of the form
xf(t) dt + const.
(x E I, a E I)
s
(f is locally summable, i.e. s L on some neighbourhood
of each point) by
means of the property of absolute continuity: the image of any set of measure zero
has measure zero. In this situation F x) =f(x) almost everywhere. Any nondecreasing function I is differentiable almost everywhere and 1 is locally summable, which allows us to extract from I its absolutely continuous part, the
integral of L
We can introduce analogously the Lebesgue measure and integral on the circle
(however this reduces to the case of the interval C-n, n]).
For any two measures pl, pLzon spaces X,, X, we form naturally their direct
product pL1 x pL, on Xl x X,. For this, iffE L X, x X,, pI x p2), we have
F(x) =
(summability
theorem of Fubini.
on L X, ~4)is linear14. In general, integ ration is a powerful tool for the construction of functionals
both linear and non-linear. For example, the simplest form
of functionals arising in variational problems is
ICYI =
sa
(x,Y,Y dx
(Y E C a,bl),
(10)
14There is in addition the important property of positiuity, namely $(f) 2 0 if f 2 0. For a linear
functional positivity is equivalent to monotonicity: if fI Cfi then $(f,) C $(fi).
15Also operators.
16
where f is a given, sufficiently smooth (not lower than C2) function of three
variables. The functional
ction for a conservative mechanical system with s
degrees of freedom has form
12
ICd =
m, 4) dt,
(11)
s t1
where t .is time, q = q(t) (ti < t < t2) runs through the set of curves of class C2
in s-dimensional configuration space, q(t) = (ql (t), . . . , qS(t)) and L is the Lagrange
function, i.e. L = T - V where T is kinetic energy and V is potential energy.
Action has an analogous structure in other physical theories where it appears as
the integral of a certain
agrangian.
In many situations the integral representation of a functional is an inevitable
consequence of its intrinsic properties. For example as F. Riesz showed (1911) any
positive linear functional on C[a, b] has form
tiff)
= S b) dw4,
(1
where Iz is some non-decreasing function on [a, b]. It is essentially uniquely
defined, i.e. to within an additive constant on the set of points of continuity. In
other words, the measure p1 is uniquely defined.
Example.
If n(x) = 0(x < x0), J.(x) = 1(x > x,,) then $(f) =f(x,)
= S,,[f].
* (fySy +f,dy
s II
(oariation)
dx,
(12)
variation
(13)
This is a second order differential equation with respect to the extremal function
y (under the condition of regularity f,,,, # 0).
17
Analogously from the variational principle and the expression (11) for action
in mechanics we can deduce the equations of motion in generalised coordinates,
which are called the Lagrange equations:
d 8L --...--=
aL
aqk
dt 84,
Here L = T - V where
with certain coefficients ~~,~~(q). We can resolve this system with respect to
dj (1 < j < s) since the matrix (pkj) is non-singular. The equilibrium positions of
the conservative mechanical system under consideration coincide with the stationary points of the potential energy:
-=
av 0
aqk
Here Akj = -
a2v
aqkaqj
pkj(O)dj
Equations
i
j=l
Akjqj
0.
(15)
small oscilla-
qj=$ CjlQt
(1 <jGs)
to the form
16The assumption about the stability of the equilibrium position q = 0 which is contained in this
sentence is valid if the potential energy V(q) has a strict local minimum at the point q = 0 (LagrangeDirichlet theorem).
Chapter
18
1. Classical
(j, +
o:Ql
Concrete
Problems
= 0
a%
@=a
,a%
ax'
u=
-.L
J>P
(16)
Equation (16) arises systematically in the description of oscillations of onedimensional systems. Moreover in certain situations it can be exact (for example,
in electrodynamics). For - co < x, t < co its general solution has the form
u(x, t) =f(x
(17)
where f, g are arbitrary functions (of class C . Thus u(x, t) turns out to be the
superposition of two truvelling wuoes moving with speed a along the x-axis in
opposite directions. Given the initial conditions
u(x, 0) = $w, Utb, 0) = w
(i.e. the initial configuration
(18)
he
n-dimensional
Au = $I $.
wave equation
has form e
at2
= a2Au, where
of small transverse
A is the Laplace
vibrations
operator
of a uniform
or Laplacian:
membrane.
19
+ 1
2u
*+lzt
_ W)
x
s at
&*
For a finite string (0 < x < 1) boundary conditions must be given in addition,
for example
u(0, t) = 0, u(l, t) = 0.
(19)
If moreover 4 E C2, $ E C1 and 4, #, II/ vanish at the ends of the interval [0,1]
(this is all necessary in order that u E C2), then the solution is found as before by
D lembert
formula where 4, $, which are initially given for 0 < x < I, are then
extended to the whole axis as odd periodic functions (with period 21).
We give one further classical partial differential equation - the heat equation
for a bar?
au
at=
a2u
(20)
(u = u(x, t) is the temperature at the point x at time t > 0). In order to extract a
unique solution it is necessary to give the initial condition u(x, 0) = d(x) and, if
the bar is finite, boundary conditions at each end such as (19).
-c-c
X
T
u2T
-oJ,
where w = const. The boundary conditions (19) require that X(0) = 0, X(I) = 0.
I*
The n-dimensional
heat equation
au
1Daniel Bernoulli and Euler were in possession of this method long before Fourier.
20
T,(t) = A, cos y
t + B, sin 7
t.
The series
formally satislies the equation, since it is linear, and clearly also the boundary
conditions. The initial conditions (18) immediately
reduce to the problem of
expansion in trigonometric series:
4(x) = nzl A, sin yx,
I)(X) =
l B,?
sin y x.
(21)
At one time a bitter controversy arose over this problem between Euler and
D lembert, which was only resolved in the 19th century after the work of
Fourier and after N.I. Lobachevskij (1835) and Dirichlet (1837) gave a more
general definition of the idea of a function and established the first tests for the
convergence of Fourier series. In the present context the Fourier series (21) of the
function 4 is defined by the Euler-Fourier2 formulae for the Fourier coefficients
An=;
((x)sinyxdx,
s0
(22)
on the
I&X) sin F
dx.
s0
With this the procedure for solving the problem on the vibrations of a linite string
is formally completed.
B. = &
A.cosyt+B,sinyt
with coeflicients defined by formulae (22), (23), converges uniformly on the strip
O<x<l,
- 00 < t < co to a function u(x, t) of class C2 which satisfies the
equation for vibrations of a string, the boundary conditions ~(0, t) = 0, ~(1, t) = 0
and the initial conditions u(x, 0) = 4(x), u,(x, 0) = Ii/(x).
The formulae are first encountered not in the work of these authors but in that of Clairaut (1754).
3If the expansion (21) for 4 converges uniformly on [O, fJ then formulae (22) necessarily hold.
21
T
lCT
-m.
If 0 < x < 1 then under the boundary conditions (19) we obtain the same spectrum as in the previous problem and the same X,,(x), but now
T,(t) = e-rwo2~,
which gives
u(x, t) = g
&?(nn~l)2t
sin Ex
l
n=1
In the problems considered above, the solution u(x, t) is obtained in the form
of a series in the eigenfunctions
sin yx
d2
where
a, = i
b. = i j:
22
2x
: f (x)eeinx dx
s x
(c-, = ;E, for real f ). Moreover the partial sums of the series (25) may be written
in the form
S,(x) = , & c,einx.
m.
= &
Q,,(x - 5)f(5)
of bounded oariation
d&
(26)
einx
InlCm
sin(m
*lx.
(27)
sin X
2
lim
= D,,,(x)f(x)
m-e0 s -z
of zero then
dx = 0.
sWe note that in formula (26) there appears the important convolution
functions on the circle:
(B *f)(x) = &
; dx - Of(t) de.
s I
For the convolution of functions the corresponding Fourier coefkients are multiplied
(Conoolution
Theorem).
together
23
lim
m h(x)e dx = 0.
i4-Q-J I -00
In turn, the principle of localisation follows from property 2): the convergence
and the sum of the Fourier series of a function f~ L at a point x depends only
on the germ of the function at this point.
The insufficiency of continuity alone even for the convergence of the Fourier
series at each point was first shown in 1876 by Du Bois-Reymond in an example,
after which Lebesgue (1905) constructed a continuous function whose Fourier
series converges everywhere but not uniformly. This line of development, which
compromised pointwise convergence as a tool for reconstructing a function from
its Fourier series, was brought to a definite conclusion when A.N. Kolmogorov
(1926) produced an example of a function in L whose Fourier series diverges
everywhere. In another direction, N.N. Luzin raised the question of the convergence almost everywhere of the Fourier series of any function f~ L This
question was resolved positively by Carleson (1966) and shortly afterwards Hunt
extended Carleson result to all Lp (p > 1).
A natural method of reconstructing a function from its Fourier series is
provided by Cesriro summation
Sk(X)
j&
j.
Mx)
4(x
&
Of63
dt,
(28)
2
(k + 1) sin G
(29)
1 _ &!-)
einx;
points,
24
1) &fin F,(x) dx = 1;
2) h-n,+, F&c) = 0 uniformly
on [ -
T,
rr] outside of any neighbourhood
of zero;
As compared with the Dirichlet kernel, 3) is the most significant property and
it leads to quite different behaviour of the FejCr averages from that of the partial
sums of the Fourier series.
Corollary
f c,einx
II= -03
for the correspondence between functions and their Fourier
analogous notation in the real form).
T,(x) = 1
Inl<m
c,,,einx.
(30)
The theory of Fourier series for functions of class L2 turns out to be particularly satisfactory. The reason for this is the Euclidean geometry and completeness of the space L Moreover the L2-theory is most important for many
applications, which is explained in particular by the physical significance of the
integral of the square of the modulus. For example, the total energy of a string
(in the approximation
for small vibrations) is equal to
E=f
pu;+Izu:)dx;
I0
in quantum mechanics the integral of the square of the modulus of the wave
function Ii/(x, t) over any volume Q is the probability of finding particles in this
volume and so on.
In the Lz-theory it is necessary to consider for f E L2 the quantity
&if)
E.(f)=&
from which follows immediately
and, consequently,
; If(x) - $,,(412dxs n
is easily obtained:
; I If(
dx - ,I); lnk12,
Bessel inequality
(31)
5 2. The Fourier
Method
polynomial
and Related
Questions
25
(30)
(32)
(here Pythagoras
theorem can be applied because of the orthogonality of the
function j i S,,,to all einx (I nl < m) and therefore to all trigonometric polynomials of degree <m, in particular, S,,, - T,). The inequality (32) shows that the mth
partial sum of the Fourier series of the function f (and only it) gives the best least
squares approximation
to f by means of trigonometric
polynomials of degree
<m (the error of this approximation
is equal to ,/E,(f)).
If f is continuous we
can assume that T,(x) converges uniformly to f as m + co and moreover in L2.
The last assertion is correct for any f~ L2, so that f can be approximated in L2
to any degree of accuracy by continuous functions. But then from (32) E,(f) + 0
as m + co; thus we have
Theorem. For any f E L2 its Fourier series converges to f in L2, i.e. in mean
square.
In this case Parseual
$,
IC 2= ;I;
If(xN dx.
5%
On the other hand the Riesz-Fischer theorem on the completeness of the space
L2 implies the convergence in L2 of any trigonometric series
f c,einx
n= -m
(33)
But then (33) will be the Fourier series of its sum. Thus we have a one-to-one
correspondence 9 between the function space L2( -7c, rr) and the space of
scalar (complex) sequences (c,)? which satisfy condition (34). The latter space,
which is usually denoted by 12, is L2(Z, p) where p is the measure on Z such
that9 p( (n}) = 1 for all integers n. Parseval equality says that 9 is an isometry.
Moreover 9 is an isomorphism of the linear spaces L2( --a, rr) and 12. It is
lthough
Parseval wrote down this equality away back in 1799, it was only proved considerably
later.
*The notation
1 is also used for the space of one-sided numerical
(complex)
sequences, the squares
of the moduli of whose terms form convergent
series (we also retain it when we restrict to real
numbers). The spaces Ii (p 2 1) are defined similarly.
91f r({n))
= pL,, where (p(.) is a given (weight) sequence of positive numbers, then we use the notation
1, (similarly
Ii (p 2 1)).
26
appropriate
identity
Chapter
1. Classical
Concrete
Problems
where not only f but also g belong to L the d, being the Fourier coefficients of
g. This means geometrically that (9f, 9g) = (f, g).
The problem of absolute convergence occupies a significant place in the theory
of trigonometric series. In the case of a series of the form (25 absolute convergence everywhere is clearly equivalent to c. lc,I < a. In the case of series (25) it
is equivalent to
but this is no longer obvious (in one direction) and constitutes the Denjoy-Luzin
theorem: if the series
zl (a, cos nx + b. sin nx)
x .
as 6 + 0. We say thatfbelongs to the Lipschitz C~USS with exponent a (0 < u< 1)
and write f~ Lip a if o(6) = O(6 . S.N. Bernshtein proved that if f~ Lip a for
some a > f then its Fourier series is absolutely convergent
while counterexamples are to be found in Lip 3.
The functions with absolutely convergent Fourier series form not only a linear
space but also an algebra, for if
fib) = n= ?-53
and f = fr
c;l)einx,fz(x)
f2, then
lo Or Hiilder class.
I1 A more general condition
is PO o(S)S-qz
f
-al
C$!)einx
9000923
27
C =
I= -co
c!$j2)
(n = 0, * 1, &- 2, . . .),
Zf f E W and
with the
G llfill
* Ilf2ll;
e and
IMI = 1.
1111
llll
llllll
IllIIIIIIII11
Ill1
IIIlllll
II11
lllll
Illlllll
III
#f&j&+
MAT0003040368%
28
and the Reisz-Fischer theorem. In the case of a complete system (and only in this
case) the functions f~ L2(X, p) are uniquely determined by their Fourier coefficients. In order that the series converge in L X, p) to f, it is necessary and
sufficient that Parseval equality should hold:
All three systems are complete in the corresponding Lz. In other words, in each
of these spaces the system of powers is complete. As a result of Weierstrass
theorem on the uniform approximation
of continuous functions by means of
polynomials, the system of powers is complete on a finite interval for any Q. On
an infinite interval whether it is complete depends on the weight. For example,
if p(x) = e + xt then the system is not complete since
m
xke- lxl sin(2n In 1x1)dx = 0
(k=0,1,2,...)
s -G3
(any function which has arbitrary close mean-square approximations by means
of polynomials for the given weight is orthogonal to sin(2a In /xl)). This example
of Stieltjes is covered by the following test (M.G. Krejn, 1945).
141t is possible to carry it out so that the highest coefficients are positive; this is assumed in the sequel.
Theorem.
+)(--co
29
complete in Li.
* In o(x)
dx > - 00, then the system of powers is not
s -,1
2) P(-4
= p(x);
dx = -co,
ODda(x)
(x)
<1 2
sJ-T-2
and put
S.N. Bernshtein
approximation
Levin developed
the uniform case
2.6. The Power Moment Problem. Let us denote by Z any one of the three
intervals (- co, co), [0, co), [0, 11. Let a(x) (x E I) be a non-decreasing function
and as in Section 2.5 suppose
fI
xk da(x)
(k=0,1,2,...)
30
Chapter
1. Classical
Concrete
Problems
are called the power moments of the function G (measure p,). If we interpret the
measure ,LQ,as a mass distributed on Z then m, is the mass of the whole od ml
is the static moment, m2 is the moment of inertia relative to the point x = 0. The
problem is to give an intrinsic characterisation of the class of moment sequences
(mr)2. For any Z an obvious necessary condition is the positivity of the Hankel
matrix A4 =
(t?lk+j)kqlj=o:
$ m,+j&<j
This condition is also sufficient in the case Z = (- 00, 00) (theorem of Hamburger).
In the case Z = [0, co) necessary and sufficient conditions are provided by positivity of the matrices M and M, = (mk+j+l)&=e; in the case Z = [0, l] we have
an analogous (but more cumbersome) criterion, although the following theorem
is of greater interest.
Theorem of Hausdorff. In order that (mk) be a moment sequence on [0, l] it is
necessary and suflicient that the inequalities
A ,z
t (-1,
j=O
(n,k=0,1,2
> 0
mk+j
,...)
hold.
As before the necessity is trivial. The proof of the sufficiency can be carried
out on the basis of the uniform approximation
of a function f~ C[O, 1) by means
of the Bernshtein polynomials
Bnkf)=kio00
;f
g Xk(l - x) !
This tool was proposed at one time by S.N. Bernshtein for the proof of
Weierstrass theorem. At the same time we obtain from Weierstrass theorem
the definiteness of the moment problem on an interval, i.e. the uniqueness of the
corresponding measure. As the example of Stieltjes shows, definiteness fails in
general on the semiaxis or the whole axis. Carleman (1922) found the following
suflicient condition for definiteness of the moment problem on [0, co):
-
= co.
co.
The sequence mZk = T(k + i) along with mzk+i = 0 (k = 0, 1, 2, . . .), which cor-
31
responds to the weight p(x) = iemX2, provides an example where this condiJn
tion is satisfied.
The investigation of the moment problem is closely connected with the properties of orthogonal polynomials. Let us consider the moment problem on the
whole axis for the sequence (mk);. We introduce on the linear space of real
polynomials the linear functional
8(R) = 2 m,r,
R(x) = i r,xk .
k=O
(
1
If the given moment problem is solvable and C(X) is any solution of it, then
k=O
8(R)
m R(x) da(x).
s -Co
(36)
Thus, all solutions generate one and only one linear functional and one and only
one scalar product (P, Q) = @(PQ) (it is important for this construction that the
polynomials form not only a linear space but also an algebra). Positivity of this
scalar product follows from positivity of the functional 8, while for its nondegeneracy it is necessary and sufhcient that the set of points of increase of the
function 0 be infinite, which in turn is equivalent to the rank of the matrix A4
being infinite - this will be assumed from now on. It follows from what has been
said that the system of orthogonal polynomials (s(x)): is uniquely defined by
the moments (mk)$ even in the indefinite case. We shall assume further that
m, = 1 and then P,(x) E 1. The theory set out below was established in 1922
in the works of Hellinger, Nevanlinna and M. Riesz and was subsequently
developed further by H. Weyl(1935).
Emma. &f the series xk Ipk(z)12 converges even at one point z (im z # 0) then
it converges uniformly on each bounded region.
ipktz)12)-
is called the WeyZ radius at the point z. The lemma means that on the set im z # 0
the Weyl alternative holds: either r(z) = 0 for all z or r(z) > 0 for all z.
Let us introduce the system of polynomials
Q,(Z)=@x~+~(x))
associated with (Pk)$. If we
(k=0,1,2,...),
put
w,(z) =
O3 da(x)
__
s -,x-z
(im z # 01,
32
Chapter
Q&J
1. Classical
Concrete
-wAzP&)
Problems
+ cd4
where the ck(z) are the Fourier coefficients of the function (x - z)-. Bessel
inequality gives
Consequently
the point w,(z) belongs to the set W(z) defined by the inequality
This is a disk of radius r(z) for r(z) > 0 and a point for r(z) = 0. Correspondingly
W(z) is called the Weyl disk or point.
Theorem 1. The moment problem is definite in the point case and indefinite in
the disk case.
In fact, if in the point case there are two solutions a,(x), a2(x) then w,,(z) =
w,,(z) for im 2 # 0, i.e.
J~*$g$J~m$!g,
from which it follows that r~i(x) essentially coincides with a2(x) (see Section 1.5).
In the disk case for each point c E W(z) there exists a solution a(x) of the
moment problem such that w,,(z) = {. A solution a(x) is called z-extremal if the
point w,(z) lies on the boundary of the disk, i.e. equality is attained in (38) or in
other words, Parseval equality holds for (x - z)- Therefore, if the system of
powers is complete in L$ then a(x) is z-extremal for all z (im z # 0). The converse
is true even in a strengthened form.
Theorem 2. Zf the solution o(x) is z-extremal
system of powers is complete in Lz.
Since this solution is z-extremal for all z, we can call it simply extremal. The
unique solution in the point case is referred to in the same way. Analogous to
Theorem 2 we have the following theorem.
Theorem 3. Zf the moment problem has a unique solution o(x) then the system
of powers is complete in La.
The description of all solutions of the moment problem in the indefinite case
is of considerable interest. Some further concepts are required for its presentation.
We say that a function U(Z) which is analytic on the half-plane im z > 0 belongs
to the class of Neualinna (in short - the class N) if im U(Z) 2 0. For example, any
function w,,(z) is of this type.
Theorem 4. In order that the function u(z) (im z > 0) be in the class N it is
necessary and suflicient that it permits an integral representation
u(z) = a2 + /I +
where z(x) is a non-decreasing
Corollary.
OD1+x2
~
-Q, x-2
bounded function,
dW,
33
* dr(x)
-a,
X
s
lim
y+ +a0
lyw(iy)l
< 00.
in each bounded
al 1(4 = z T Qk(0)Qk(zh
a12(4 = 1 + z T Pk(0)QkW~
a,,(z) = - 1 + z 7 QkWk(z)9
az2(4 = z T SWJ&9.
all(z)
( azl(4
a12(4
az2b9 >
an entire matrix-function
of the complex variable15 z, which is real on the real
axis and identically unimodular, i.e. det N(z) E 1.
Theorem 5. The formula
m do(x)
-=
.-,x-z
- al&M4
a2 1(zM4
- all(z)
- az2(4
la&l
maxi.k
I4
= o
.
34
xpk(x) = a,P,W
(k=0,1,2,...)
+ bkec+l(x) + BkPk-16)
(39)
Here
P-,(x) z 0,
pk
(xpk(x)>
ak =
pk-l(x))
conditions
(xpk(x)9
b/c =
pk(x))y
=
(xpk-l(x)v
pk(x))
fxp,tx),
=
bk-,
pk+l(x))
>
o,16
(b-, = 0).
P,(x) = 1, PI(x) = 7.
under
Q,,(x) = 0,
Ql(4 = ;.
The ignite
matrix
is called the Jacobian matrix of the given moment sequence (mk); (or of the
given a(x)). We remark that J uniquely determines the moments since from the
expansion
Xk
ckjpj(x)
j=O
there follows mk = $0 (k = 0, 1, 2, . . .). It can be shown that for any real matrix
of the form (40) with positive off-diagonal elements the sequence of polynomials
(Pk(x))g defined by it is orthogonal with respect to some a(x) (- co < x < co).
We shall consider the matrix J as a linear operator on the (infinite-dimensional)
linear space of sequences (vk)z of complex numbers. The recurrence formula (39)
(with the replacement of x by an arbitrary z E C) shows that the sequence
II(z) = (4(z)); is an eigenvector for J with corresponding eigenvalue z. Clearly,
apart from multiples, it is unique in this property, i.e. all eigenspaces of the
operator J are one-dimensional. The Weyl alternative implies that either II(z) $ 1*
for all z (im z # 0) (the point case) or II(z) E 1* for all such z (the disk case)l. The
criterion for definiteness of the moment problem is thereby reformulated in terms
of the spectral theory of Jacobian matrices.
l6The latter follows from the positivity of the leading coefficients.
utting this differently we can say that equation (39) (with x replaced by z (im z # 0)) either has
a unique solution up to normalisation in I2 or all its solutions belong to I2 and moreover for all z.
the same case results.
35
2.8. The Trigonometric Moment Problem. Suppose there is given on the circumference of the unit circle a measure p which is generated by a non-decreasing
function a(x) (-a < x < K). Then the numbers
..=d=,=l~nd~=~~~e
o(x)
(n=O,fl,f2
)...)
are called the power moments of the measure p or the trigonometric moments of
B. Here again the problem is to give an intrinsic characterisation of
the class of moment sequences (c,)?~. Analogous to the theorem of Hamburger
we have
the function
(n + 1 - Im()c,e 2 0
36
Chapter
1. Classical
Concrete
Problems
and
for all (A,, . . . , A,) E R and all (rl, . . . , <J E C (n = 1,2, 3, . . .). In the terminology of ,probability theory we can say that the characteristic functions of the
probability laws are exactly the continuous positive functions satisfying the
normalisation condition c(0) = 1.
2.9. The Fourier Integral. This continuous analogue of Fourier series appeared
(1822) in the solution of the problem of heat conduction in an infinite rod. In this
case the boundary conditions drop out and only the initial condition U(X, 0) =
4(x) remains. The formal solution obtained by separation of variables has the
form (for K = 1)
E(A)e-azreiAx dl,
(43)
--m
m qi(x)epiAx dx.
s
Here it is required that 4 E L although, after eliminating
heuristic transformations from (43) and (44), the formula
(44
where
E(A) =
-(x-Y)%40&)
dy
:
I
(44*)
i
j
E(l)e dA,
is clear from
combining
relation
m eiixPy
I --m
which
is the typical
form
dl
= 2716(x
situation.
- y),
the orthogonality
1
1
summablelg.
31
&
E(A)?
dl
s
I N
(from which 4 may be determined almost everywhere by differentiation). Consequently 4 is uniquely determined by its Fourier transform.
The L*-theory of the Fourier integral is contained in the following theorem of
Pluncherel. Let Q belong to L*( - co, co). Then the Fourier-Plancherel
transform*
E(I) = Jili
&
:N &)eeiax
dx
&
-1 E(IZ)e dl
f
Parseual equality
38
Chapter
1. Classical
Concrete
Problems
i.e. @ belongs to the linear space H* of functions f(2) which are analytic in the
half-plane im 1 c 0 and satisfy the condition
If(a - @)I* da -z co.
classes HP.
The theorems stated below on functions of class H* are due to Wiener and
Paley.
Theorem 1. Any function f E HZ can be represented in the form
f(A) = &
1:
&c)e-iAx
dx
(45)
s
s
* IWtMI
-m
dl < o.
1 +I2
and, conversely, for any function + E L*( - co, co), + 2 0 such that
m Iln WI dl < co
,
--coTTF
t
I
I
I
b d(x)emiAx dx.
sa
By means of this identity E(I) is extended to the whole complex plane as an entire
function of exponential type p < max(lal, lbl) and in addition, by the theorem
of Plancherel, it belongs to L* on the real axis.
Theorem 2. If f (A) is an entire function of exponential type p which belongs to
L* on the real axis, then it can be represented in the form
9 2. The Fourier
Method
and Related
39
Questions
2.10. The Laplace Transform. We shall say that the function f(t) (t > 0) belongs
to class A if it is Lebesgue measurable and there exists k = k(f) such that
f(t) = O(ekt) almost everywhere. The class A is a linear space. If f~ A then the
analytic function
f(A) =
03f(t)e-A dt
s0
is defined in the half-plane re 1 > k(f). The class 2 of all such functions is also
linear. The homomorphism
f H f is called the Lupluce transform and under it f
is called the preimage for its image f The preimage is uniquely determined for
a given image, and moreover the Riemunn-Mellin
inversion formula holds:
(a > k(f)).
This formula opens up a wide range of possibilities for calculations using the
theory of residues. For example, let f(A) be continued to the whole plane as a
meromorphic function with poles {&} T. To make matters precise let us assume
k(f) > 0 and put A0 = 0. If f is bounded on some sequence of left semicircles
with diameters [a - /?.i, bl + j&i] (j?. + co), then
[ 1
df (7) dt = lim
c
Res ?eA
-+cop~-rrlia, A=&
(46)
the limit being uniform on each finite interval. The residues in (46) have the form
of degree < n, - 1 and nk is the order of the
Consequently,
40
under certain additional requirements also f itself) turns out to be the sum with
brackets of the series of quasipolynomials &Pk(t)e@
(here the & are indexed
according to increasing 11, - u I); moreover convergence is uniform on each finite
interval.
Formula (46) is called the Expansion Theorem.
We note further that if the function f~ n is continuous then
f(t)
lim
LZj!(!!!+tfln)(F),
n-tm
the limit being uniform on each finite interval (a, b) (a > 0). The specific character
of this formula of Widder, which goes back for its idea to Laplace
ethod of
larger numbers lies in the fact that use is made in it of the value of the Laplace
transform only for large I > 0.
A most important property of the Laplace transform is that it transforms the
operation of differentiation in n under the condition f(0) = 0 into the operation
of multiplication
by 1 in ;?. 22 The more general relation
.m =m, -f(O)
is valid for all absolutely continuous f such that f E n (in this case f~ /i
automatically).
If f, g are two functions of class n then the function
s0
t - 4d4
dt
(47)
with constant coefficients and right hand side f~ A. Let y be the solution under
he Fourier transform has a similar property but in the class 15
here is an analogous result for the Fourier transform on L - co, co) and this is not accidental.
The Fourier convolution is converted into the Laplace convolution if!(t) = 0 and g(t) = 0 for t < 0.
The Laplace convolution has a causal character in the sense that the value of (f t g) (t) depends only
on the values of the functions f and g in the region z < t. From the physical point of view the
fundamental difference (within a great formal similarity) of the Laplace transform from the Fourier
transform lies in the causality of the corresponding convolution.
41
the initial conditions y (O) = 0 (0 < k < m - 1). It is easy to see that y E A. From
(47) we have
Q(t) = CF.".
[ 1
$
The quasipolynomial
Q(t s0
r)f(z) dT.
Example 2. Now let the right hand side in equation (47) be one-periodic:
+ 1) = f(t). Let us assume that the characteristic polynomial does not have
roots of the form in (where n is an integer). Then there is a unique one-periodic
solution, namely
f(t
y(t) =
f
&einz
la=-co
(48)
(the c, are the Fourier coefficients of the function f ). We shall obtain this solution
by applying the Laplace transform:
D(A)j(A)
+ A(n) =&).
(49)
Here A(n) is a polynomial of degree < m - 1 whose coefficients are linear combinations of the initial values y(O), . . . , y - (O). From the periodicity of f it
follows that
f(A) = $$
(h(l)
= I (l
- t)e dt)
fl(4
A(4
= D(A)(e - 1) - o(n)
42
Since the left hand side is regular at the root A, of the polynomial
D(A),
in
Q(t - z)f(z) dz - c
s0
= $
(t - ~)a-Y(T) dz
s0
of the
theorem
f(t) = ; ( r(l
(51)
a(t)
= O(eTk .
(52)
be a> k)
transform.
(53)
2 4 I = O(Nk).
IPI=1
When a, E 1 the series (53) delines the Riemunn c-function
which plays an
43
extremely important role in analytic number theory, where it and certain other
Dirichlet series find application.
If the function Q is non-decreasing and satisfies the condition (52) for every
k > 0 and S(A) (A > 0) is its Laplace-Stieltjes transform then the function S( - A)
is absolutely monotonic, i.e. it is infinitely differentiable and all its derivatives
(including itself) are non-negative.
Theorem of Bernshtein.
= SW ext da(t),
0
Dirichlet introduced these series for the proof of his theorem (1837) that there are infinitely many
primes in any arithmetic progression (CI+ nd), , where a, d are relatively prime natural numbers.
44
Chapter
1. Classical
Concrete
Problems
to the domain
(7) dz
s0
(55)
sm
F(l)e dl
be C < 0)
cc
to the domain Pa, it has limiting values Y(t f i0) and
df(t)e-
dt.
s0
f(t + s) do(s) = 0
(-m<t<m)
fa
where o is a complex function of bounded variation (o # constant on neighbourhoods of the points a, b). In the case o(s) = s the equation is equivalent to
the periodicity (with period b - a) of the function f and its mean over the period
being zero. Naturally we expect to have for p.i.m. a theory similar to that of
Fourier series. In fact from the equation for f we may formally calculate the
aplace transform:
@&I
>
45
where
b
A(n) =
eeAs do(s),
dt.
sD
The function f is meromorphic and the roots A,, I,, A,, . . . of the characteristic
equation A(n) = 0 are its poles. Formally applying the Expansion Theorem (in
differentiated form) we arrive at the generalised Fourier series:
where
P,(t) = e-
Res [&)e ]
1=1,
dt /D:Js)e-
ds = O(e-
of determining
tor
46
YW
- htY(4
= 0,
Y(b) + h,Y@) = 0
(h, , hz E Iwu {co}; if, for example, h, = co then the corresponding condition has
the form y(a) = 0). In 1836 Sturm and Liouville independently constructed a
systematic theory of this problem (with q(x) = 0), which generalises the theory
of Fourier series of the form (21) at least for smooth functions. In the 20th century
the Sturm-Liouville
problem and its generalisations became an essential attribute of the mathematical apparatus of quantum mechanics. The point is that in
quantum mechanics physical quantities are described by linear operators?, their
possible values are the eigenvalues of the corresponding operators and they are
realised in the states of the quantum mechanical system, these being defined by
the corresponding eigenfunctions. In particular -y + q(x)y is the operator of
total energy (Hamiltoni~n~~):
(-y) corresponds to kinetic energy, q(x)y to
potential energy. The eigenvalues of an energy operator are the energy levels of
the system.
We shall consider the operator L on ~!,:(a, b), (a(x) = c r(c)&), assuming for
simplicity that its domain of definition D(L) consists of those functions of class
C2 [a, b] which satisfy the boundary conditions. A basic property of the operator
L is contained in the identity
bxY,l~
(Yl, Y2 E w-49
f(x) = zl
c,Y,(x)
(G = (5 Y.) = ~bfWd4~C4d~).
a
series converges to
in Lg(u, b),
41
The proof of this theorem was outlined in essence in the works of Sturm and
Liouville but it was not until 1896 that a rigorous version was given by V.A.
Steklov.
One of the methods of proof is based on the variational theory of eigenvalues
developed by Fischer and Courant. In this theory the equation L[y] = Ay is
regarded as an Euler-Lagrange equation for the problem of finding the minimum
of the quadratic functional (L[y], y) under the constraint condition (y, y) = 1. It
turns out that the desired minimum is the least eigenvalue ;1, and is attained at a
corresponding eigenfunction y, . If we already have the eigenvalues il, < * * * < 1.
and corresponding eigenfunctions y, , . . . , y, then
1n+l =
min{Wyl,
Y) : (Y, Y)
(f~W4
follows from the construction
it the expansion in Lz(a, b):
(56)
L[f
1 =
kzl
tL[f
1, Ykbk
ktl
ckAkYk.
The uniform convergence of the series (56) follows from this as a result of the
bound
IfWl G constIILCflll
(requiring a separate proof).
2.12. The SchrSdinger Operator on the Semiaxis. Let us consider the operator
L[y] = -y + q(x)y (0 < y < co) with boundary condition y(0) - hy(0) = 0
(h E 88u (co}). We assume the potential q(x) to be real and (for simplicity)
continuous. On L(0, co) the set of eigenvalues of such an operator can turn out
to be finite or even empty. For example, with q(x) = 0, h = 0 we require non-
48
f(x)= CLtit.7u y.
Let us introduce
the non-decreasing
function
Then
where
4,(4 =
Parseval
bf(xM(x, 4 dx.
f0
n general it is mixed. For example, in the spectrum of the hydrogen atom there is a countable
system of discrete energy levels Em< 0 corresponding to the states of an electron which is connected
with the nucleus and a continuous spectrum E > 0 corresponding to the possibility of unbounded
escape from the nucleus.
soThe first work of H. Weyl on the Schrodinger operator (so called much later) on the semiaxis
appeared in 1909-1910 well in advance of the foundation of quantum mechanics. The results set out
below are essentially contained in these works.
49
lf(412
dx
CQ iFbtn)12
s -Co
dPb(&
= , m
-(x)+(x,
A) dx
-+ s 0
exists in Lz( - co, co), the Weyl inversion formula31
(58)
f(x) = lim
N+x
-N
FW(x,
oalf(xN2 dx =
4 dA4
(5g*)
equality
OIf 412WI
s --a,
is valid.
([a] denotes the integral part) for 1 > 0 while &,(A) = 0 for I < 0. Letting b + 00
we obtain the unique limit function:
p(l) = ;Jn
After the substitution
(A > 0);
= t
where
31Although the inversion formula is obtained heuristically from (57) on letting b -) co, its rigorous
derivation requires a deeper argument.
50
pxdx
G(p)
=plis:f(x)cos
(both defined in L(O, co)). Plancherel
theorem has now been obtained for the
case of an even function32.
In general the spectral function p(A) is not unique. The situation here is
analogous to the power moment problem.
We shall consider $(x, [) for complex c (im c # 0).
Lemma. Zf $(x, co) c L O, co) fir some co (im co # 0) then tj(x, c) E L2(0, co)
for all ( (im [ # 0).
The quantity
-
>
is called the Weyl radius at the point [. The lemma implies that on the set im c # 0
we have the Weyl alternatiue: either r(c) = 0 for all c or r(c) > 0 for all [. The
Weyl disk or point is defined by the inequality
a, IwJ/(x, CC-)+ 0(x, 01 dx < - s,
f 0
where 0(x, c) is the solution of the equation L[y] = cy for the initial conditions
0(0, [) = -sin ~1,0 0,[) = cos a. The non-emptiness of the set of values of w
defined by the inequality (59) is established by passage to the limit from the finite
interval. Now the inequality
b
Iw$(x,
(60)
s 0
is satisfied if the solution &(x, [) = w$(x, c) + 0(x, LJ is such that the Wronskian
of the pair 4,,,, &, equals zero at the point b. Such solutions clearly exist and
therefore (60) defines a certain disk C,. Obviously C,, c Cb2 if b, > b,. Thus
C, = fib C, is non-empty and this is also the disk or point of (59). This argument
provides the reason for calling C, the limit disk (or point). Its principal result is
Theorem 2. Zf im c # 0 the equation L[y]
L2(0, co).
In the disk case, all solutions of the equation L[y] = cy belong to L O, co)
while in the point case only multiples of 4(x, [) belong to L2(0, co). Incidentally,
it is clear from this that the realisation of one or other case of the Weyl alternative
does not depend on the boundary condition at zero, i.e. it depends only on the
potential. For example, if the potential is bounded from below or if it only admits
321t is appropriate to note that even Fourier deduced his integral formulae heuristically by passage
to the limit from series.
51
a bound q(x) > - cx (c a positive constant) then the point case occurs (HartmanWintner, 1949).
Theorem 3. The spectral function
the disk case.
The
The
of the
y(0) -
operator
(1953).
functions
33As a result of (58) and (58*) the spectral function provides a measure on the spectrum with respect
to which the system of generalised eigenfunctions is orthonormal in the sense that
52
the mean
f(t) dt
exists.
This allows us to introduce on the space AP the scalar product (f, g) = m(c)
(AP is not only a linear space but also an algebra).
Let us consider the continuous family of exponentials: cl(t) = ei (A E R). It is
orthonormal: (e,, e,) = &,. The following definition therefore arises naturally:
CA(f) = (f, e,) is called a Fourier coeflicient of the a.p.f. f. Following on from this
we introduce formally the Fourier series
f(t) - C cl(f )e
In fact this series contains no more than countably many non-zero terms, since
Bessel
inequality
ktl I%,(f )I2 G (f9f)
holds for any collection (1,) . . . , A,}. The set of those I for which cl(f)
called the Bohr spectrum of the function f.
# 0 is
n the terminology of physics this is the superposition of harmonic oscillations with arbitrary
frequencies. In particular, the coordinate functions describing small oscillations of conservative
mechanical systems are of this type.
8 3. Theory of Approximation
53
equality
= (f,f),
which means that the Fourier series of an a.p.f. f converges to f in the sense of
and introduce the distance d(f,,f,) =
+
co,
where
S, is the nth partial sum of the
1% -fill then 4f, 4) -, 0 as n
Fourier series of the function f for an arbitrarily chosen enumeration of the
non-zero terms. It is also possible to sum the Fourier series of an a.p.f. f to f
using a method of FejCr type (Bochner, 1927).
Bohr mean-square: if we put llfll = ,/(f,j)
= -sgn R(x,)
(1 < k 6 n + 1).
t On the space C(X) of continuous functions on any compact set X the natural metric (corresponding
to uniform convergence and therefore called uniform) is defined as d(fr&) = jlfr -frl[, where
llfll = maxlf(x)l. Thus the deviation d(f, P) is the distance from P to J
.e.long before the appearance of Weierstrass theorem.
3The proof of this theorem was developed gradually and a complete version did not appear until
the work of Bore1 (1905).
54
Corollary.
The polynomial
of best approximation
for given
f, n is unique.
x = tan i.
T,(x) = 1 (n = 0),
T,(x) = &
cos(n arccos x)
(n 2 1);
function
W(0) = 2 a cos(b )
k=O
where 0 < a < 1 and b is an odd integer > 1. This is a continuous 2n-periodic
function which does not have a derivative at any point if ab > 1. It is given by
its Fourier series which converges uniformly to it. It can be shown that the partial
sums of this series are the best uniform approximations
for W by means of
trigonometric
polynomials. This situation is encountered extremely rarely, in
contrast to best approximations
in mean square where it is always the case.
For any complex 2aperiodic continuous function f we shall denote by E,(f)
the deviation from f of trigonometric
polynomials4 of degree <n. From the
formula for the partial sum of a Fourier series we obtain the bound
max
0=0<2n
If (4 - U4l
G @, + W,(f
),
(61)
4The same notation is also used in an algebraic situation and also in subsequent generalisations
which cover both situations.
5 3. Theory
55
of Approximation
where
L, =
= PnWl dx
s --z
is the Lebesgue constant (the L-norm of the Dirichlet series). Weierstrass
theorem shows that E,(f) + 0 as n + co. At the same time L, -+ co but grows
slowly. In fact as Lebesgue showed (1910),
$1,
n + O(1).
Inequality (61) can therefore be useful as a bound for the rate of convergence of
a Fourier series. This is the case for example with smooth functions since we have
Theorem of Jackson. Let f be a function of class C on the circle and let q(6)
be the modulus of continuity of its rth derivative (r 2 0) (o,(6) E o(6), the modulus
of continuity off ). Then
E,-,(f)
< BP,
n-
ly f ( E Lip a then
= O(ner- .
(64
A bound of type (63) was also obtained by Jackson in the algebraic situation.
In the class of functions which are analytic on the interval [a, b] we have the
bound of S.N. Bernshtein
T
1n-n l&(f)
< 1,
(65)
n-ccl
and moreover if f is an entire function
hm YE,(f)
= 0.
(66)
+oD
To S.N. Bernshtein is due a series of deep results, the main point of which lies
in the fact that the rate of convergence of the deviation E,(f) to zero determines
the class of smoothness of the function f. We formulate in particular those
theorems which provide converses to the bounds mentioned above.
Theorem 1. Zf (64) is satisfied for a function f E C[a, b] then f l C a, b] and
f ( belongs to the Lipschitz class with exponent a on each interval [aI, b,] c (a, b).
Theorem 2. Zf (65) is satisfied for a function f~ C[a, b] then f is analytic on
[a, b] and in the case of (66) it is an entire function.
inequality
56
Chapter
1. Classical
Concrete
Problems
1, z, . . .) Z
(z E @I;
(-co<x<co)
substitution
(x > 0).
The inequality
E det(u,(xi))&o
# 0
for all pairwise distinct x,,, . . . , x, provides a criterion under which (u& is a
Chebyshev system. Clearly any Chebyshev system is linearly independent.
Let X be any compact space. Given a system (u& c C(X) and f E C(X) there
clearly exists a linear combination
Calruk with least deviation from f in the
uniform metric.
Tonelli-Haar-Kolmogorov
Theorem. Let (u& c C(X) be a linearly independent
system. For any f E C(X) there is a unique linear combination P = 1 I,+, with least
deviation from f in the uniform metric if and only if (ut)$ is a Chebyshev system.
Interestingly, it is far from true that Chebyshev systems (with n > 0) exist on
each compact space (Mairhuber, 1956; Sieklucki, 1958).
Theorem. If a real Chebyshev system (with n > 0) exists on the compact space
X then X can be embedded homeomorphically in the circle.
51
5 3. Theory of Approximation
1961).
min re F(x){f(x)
- P(x)}
xeMp
F = I&,
> 0
alruk, where
uo,
1)" =
u,
n-l
c
ankuk
k=O
where for n 2 1 these have least deviation from zero in the L-metric. It can be
shown that the function v, has exactly n roots xnl < *** < x,, in (a, b).
Let f E C[a, b]. We put P. = ~~=o I&u&, where the A&are defined by the system
of equations
Pntxn+l,i)
=f(-%+l.i)
(1
i G
n +
1).
58
Chapter
1. Classical
Concrete
Problems
a&&. Further
the least
Is a
f(x)
v,+Ax) dx .
1 sin(n + 1) arccos x
= ji
J(1 - x2) -*
from Markov
theorem.
We note that the Chebyshev polynomials
on (- 1, 1) with weight J(l - x2).
Problem.
sa
t.&(x) da(x) =
Ck
&i. a&u&(X)2 0
(67)
the inequality
should hold. It can be shown that if uo(x) > 0 (a < x < b) then this condition is
sufficient for the solvability of the moment problem under consideration. Suppose that it is satisfied and let v = n < CO. The Chebyshev-Markov
problem
consists of finding for a given continuous function Q(x) 2 0 (a < x < b) and a
given point 5 E (a, b) the greatest value of the integral
0 3. Theory
59
of Approximation
uo . . . u,
> 0,
x0..
. x, >
(
if a G x0 < * * * < x, < b (0 < m < n). We assume further that the function Q
associated with this system does not disturb its arkovicity
even if in the
following sense:
e-0
C-0
Q(x)
d4x)
>
dq(4.
sa Q(x)
sLI
Thus the function a,(x) is the desired solution of both extremal problems and
moreover this solution does not depend on 8.
The piecewise-constant solutions of the moment problem naturally lead to
quadrature formulae of Gaussian type. In fact if x1 < . * * < x, form the set of jump
points of this solution with pl, , . . , p, the corresponding jumps, then the approximate formula
ab.fCx) Wx)
x j$ Pjftxj)
is exact for all linear combinations f = CzCo &a, and solutions d of the given
moment problem. The function d can be assigned a priori, then the moment
sequence is calculated and the nodes Xi and weights pi (1 G j < r) are already
5The first work
of P.L. Chebyshev
(1874)
was devoted
= 1.
60
defined by it. In particular in the Gaussian case, i.e. U&X) = xk (0 < k < 2r - 1;
a < x < b), we have the roots x1, . . . , x, of the orthogonal polynomial P,(x) as
nodes, while the weights (the Christoff
coeficients) are calculated from the
formula6
Qr(xj)
~_ _ Res-QAx)
pi = Pi(Xj)
(1 Gj < I),
XCXj P,(x)
Ip(x)l < L
(a < x G b)
(69)
sa
uk(x)P(x)
dx
(0
ck
<
k <
n)
(70)
with respect to the given real System (n& c L(a, b). We call this statement of
the question the L-problem of moments. The criterion for its solvability is closely
connected with least deviations from zero. It is immediately clear that if a solution
to the problem (69)-(70) exists, then for any collection of numbers (at& satisfying
the condition
n
c
k=O
akCk
l,
61t is appropriate to note here that in any L:(I) the roots of the orthogonal polynomials are simple
and lie inside the interval I.
0 3. Theory of Approximation
61
N.I. Akhiezer and M.G. Krejn (1938) obtained the following result.
Theorem. Suppose the system (u& has the property that for each non-zero linear
combination 1 akuk its set of zeros has measure zero. The L-problem (69)-(70) has
a solution if L 2 1. This solution is essentially unique if and only if L = 1,in which
case it has the form
~(4
= L w
OM,
0bIQ(x)1 dx = min
I
on the set of functions
of the form
of
3.5. Interpolation and Quadrature Processes. The exact solution of the best
approximation
problem lends itself relatively rarely to an effective description
while an approximate solution can require complicated calculations.
The more traditional method of approximating
functions, namely interpolation, is effective and simple from the computational
point of view, although the
approximations provided by it can turn out to be very coarse.
Interpolation of a function consists of constructing those linear combinations
of the given system (u& (e.g. polynomials) which take the same values as f at
the given points (interpolating points). A more refined procedure also requires
equality of derivatives up to a given order (with corresponding assumptions
about their existence) so that, for example, from this point of view Taylor
formula is an interpolation formula. We restrict ourselves below to interpolation
with simple interpolating points, i.e. without regard to derivatives.
If (u& (v < co) is a Markov system on the space X then for any interpolating
points Xi E X and for any numerical values yj (0 < j < n < v + 1) there is a unique
linear combination
P = &, il,u, such that P(x,) = yj (0 <j < n). It can be
expressed in the following form:
where C& is the linear combination of the functions uc, . . . , u, which satisfies the
conditions @Jxj) = ~5,~(in this sense the systems of functions (CD& and interpolating points (Xi) are biorthogonal).
i.e. up to values on a set of measure zero.
he solution $ is not unique but the signs of any two solutions coincide almost everywhere.
62
=
w(xk)(x
xk)
where
If this polynomial
formulag:
is defined. If N.i.s. converges even only pointwise on some set containing all the
interpolating points, then its sum f is a solution of the interpolation problem
(j=O,1,2
,... ).
= Yj
The convergence of N.i.s. with some or other interpolating
points in the
complex plane (in particular, with the interpolating
points xj = j lo) has been
studied in considerable detail. It is interesting that this tool can be used for the
proof of the famous theorem of Lindemann that the numbers CI# 0 and e cannot
be simultaneously algebraic (for a = 1 this yields the transcendental nature of
the number e and for a = zi that of the number rr) (A.O. Gel ond, 1929).
Let (uk)g be a Markov system and for each n = 0, 1, 2, . . . let interpolating
points xnj (0 < j < n) be given. The sequence of linear combinations
ftxj)
similar (but of course more complicated) formula resolving the problem of interpolation
multiple interpolation points was constructed by Hermite.
lo In this case N.i.s. converges or diverges simultaneously with the Dirichlet series
mt-l?%
ck
k=,,
with
pn
i
k=O
63
nkUk
(n = 1,2, . ..)
ankuk
k=O
which have least deviation from zero in the L metric, then, under the conditions
of Markov
theorem, the interpolation process for any f E C[a, b] converges to
f in the L metric. In particular, the sequence of interpolating
polynomials
constructed on the roots of the Chebyshev polynomials of the second kind, i.e.
on the interpolating points
=
XJ
COS
f$
converges.
For uniform convergence of an interpolation process it is necessary to impose
stringent restrictions on the function f. For example, if f is an entire function the
sequence of interpolating polynomials converges uniformly to f on any interval
[a, b] for any distribution of the interpolating
points. This follows easily from
the formula for the remainder term
f(x)
2j - n
= n
(O<j<n)
process of Lagrange
type
64
formulae
s X
(n = 0, 1,2, . ..)
(73)
for the integral which is regarded as a functional off E C(X). The coefficients
pni are found, for example, by requiring that the formula is exact for f = uk
(0 < k < ,n), where (uJ2 c C(X) is a Markov system. Under this condition
formulae (73) can be written in the form
s X
fdcL=
P,&
(n = 0, 1, 2, . . .),
s X
where (PJF is the interpolation process for f corresponding to the given system
of nodes. Therefore if (P,): converges to f in the L-metric, the quadrature
process for f also converges:
S[
b
Pj
44
1
2
O(Xj)(X
Xj)
do(x).
Ilf - call G (1 + ~ %,
where
hf. = max i
xeX
k=O
I&(x)1.
The interpolation
65
n-ml
= 0.
(74)
The analogous condition for convergence of the quadrature process has the form
lim p.E. = 0
-Yo
I ==a)
,
I
(75)
where f runs through the class of all functions which are continuous on the entire
plane, analytic outside of some compact set Q/ c M, bounded in modulus by
unity and tend to zero at infinity.
he rather simpler notion of analytic capacity (denoted by y(M)) is defined by the same formula
(75) but on a larger class of functions f, namely those which are defined and analytic outside of some
compact set Q, c M, bounded in modulus by unity and tend to zero at infinity. Clearly a(M) d y(M)
and it can be shown that if M is open then a(M) = y(M). T e analytic capacity of a disk is equal to
its radius.
66
c@\X)
= @\int
X)
(76)
where (necessarily)
1
cn = 2% s ,jJ=l+e
t,k being the inverse function
of 4 and
(n = 0, LZ
E
> 0 suficiently
. ..X
(77)
small.
We can say that this is a conformally invariant form of the Taylor expansion.
Formulae (77) result from the fact that the systems (F,(ll/(C))), and (c--l);
are biorthogonal on the circle IQ = 1 + E. Now suppose that we are given any
system of analytic functions of the form
P(Z) = $i
+ *=$+1 F
0 3. Theory
of Approximation
67
But then we can associate with each function f(z) which is analytic on the disk
IzI < R (R > r) the series
2C
n=O
Z)
(78)
with coefficients
1
fKM0
dl.
cn = 2ni s ,jcp
The problem of convergence arises once again in this context, also the moment
problem etc.
In particular, if we are given a sequence of interpolating points {z.}; lying in
the disk IzI < I, then on putting
II-1
P,(Z) = kgo (z - G-l
we will have
n-1
cp,(d = kGo(z - Zk)
(n = 0, 1,2, . ..)
= W
(n = 0, 1,2, . . .),
moment
(79)
where
The uniqueness of the solution of problem (79) is clear since the set of interpolating points has a limit point in the disk IzI < r < R. The question of conditions
for solvability of this problem is complicated even in the case where the set of
limit points for the interpolating points is finite.
In a more subtle version of the problem of type (79) we require that the function
f belongs to the class of analytic functions on the disk lzl < I and all limit points
of the set of interpolating points lie on the boundary IzI = r. Such a solution (if
68
it exists) need not be unique but if, for example, we require in addition that the
function f is bounded (and for notational simplicity we take r = 1) then the
solution is unique if and only if
O(1 - kl)
= mJ
(80)
(theorem of Blaschke). The lack of uniqueness when (80) fails is shown by the
straightforward construction of the Blaschke product:
m
>
j,k=O
2 w.o z z)
:)n=O
which formally solves problem (79). The investigation of convergence of the L.i.s.
requires powerful analytic tools although the problem is simplified substantially
for a particular choice of interpolating points and class ofinterpolating functions.
r3A more general version of the Nevanlinna-Pick problem consists of interpolating in a given disk
or half-plane by means of analytic functions with values from another disk or half-plane.
r41t is possible to construct the function o in such a way that all the roots z, are simple and there
are no other roots.
0 3. Theory
69
of Approximation
f(z)
=& s:n4(4e-iZx
dx,
(81)
Substituting
(n = 0, 1, 2, . . .).
x)l< Z&r
(n=0,1,2
)... ),
(84)
OCXGl
where I > 0 and M. > 0 is a sequence which does not increase too rapidly, then
4 cannot be flat if it is non-zero. For example, with M, = n! it follows from (84)
that 4 can be extended analytically to the disk IzI < r-l and therefore if 4 is flat
then 4 = 0.
Let (M,), be any sequence of positive numbers. We say that the function
4 E C [0, l] belongs to the class C(M,) if there exists I > 0 for which the bounds
(84) are satisfied. The class C(n!) coincides with the class of analytic functions on
[0, 11. The class C(M,,) is said to be quasianalytic if it does not contain flat
functions 4 # 0, i.e. if the corresponding uniqueness theorem holds in C(M,). The
problem of quasianalyticity, which was formulated by Hadamard in 1912, con-
70
sists of characterising those sequences (A&), for which the class C(M,) is quasianalytic. Its complete solution was obtained by Carleman (1926) and in rather
different terms by Ostrowski (1930).
Carleman
Criterion.
Put
p.=infvM,
(n=1,2,3
,... ).
fan
Ostrowski
Criterion.
be quasianalytic
> 0).
be quasianalytic
Put
GK
(x 2 0; n = 0, 1,2, . ..).
belongs to L2(0, co) and & (O) = 0 (n = 0, 1,2, . . .). Then its Fourier-Plancherel
transform E(1) belongs to the class H2 on the lower half-plane and satisfies the
bounds
Ml
(n= 1,2,...),
IWI G Inl
from which we obtain that on the real axis 1&(A)1 < (T(A))- and, consequently,
4, ln IEaGUl dA = _
00;
-03 1 +A2
of
by the expression
71
+ * -* + p,(x)y
FiCYl z ab 12 d4Y*Yx))
I(
of the form
drib = 0
(87)
The coefficients pi(x), an(x) are assumed to be at least continuous and the pi
are functions of bounded variation. If the measures 7i are concentrated at the
ends of the interval, then conditions (87) become boundary conditions; in the
general case they may be called distributed conditions.
The general solution of the equation Icy] =f(f~ C[a, b]) may be written in
the form
y =
YO +
i
j=l
AjUj,
= -&[Yo]
(89)
Let us assume that under conditions (87) the homogeneous equation I[u] = 0
has only the trivial solution. This is equivalent to the homogeneous system
i
Fi[Uj]Bj
= 0
(1 < i < n)
j=l
having only the trivial solution. But then the system (89) has a unique solution
for any y,, i.e. the equation L[y] =f has a unique solution for any right hand
side fe C[a, b]. For the explicit construction of the inverse operator I,- we have
to find a particular solution by the method of variation of parameters. As a result
it turns out that the operator L-r is integral,
with a certain kernel G. This kernel is called the Green function of the operator
L and has the following structure
G(x, 67 =
(91)
or this, in turn, it is necessary (but not sutlkient) that the linear functionals Fi (1 C i Q n) on
CR- a, b] are linearly independent. A necessary and sufficient condition is the linear independence
of the restrictions of the F, to the space of solutions of the equation l[u] = 0.
72
4C% 01 = 0
(x + 5)
Vi),CW, 01 = 0
(1 < i 6 n).
(92)
The description which has just been given can be substantially reduced if we
invoke the terminology of the d-function: for each fixed r the Green function
G(x, 5) is the solution of the differential equation
LCG(x,01 = 6(x - 0
which satisfies the conditions (92). From this point of view formula (90) is obvious
since formally
b
1
Gb, 51f(O dt
[S a
IS
b
1 sa
=
W - tlf(O
dt =fW
1s
1*cz-J =(-l)
(93
and certain boundary conditions. These are connected with the resulting integration by parts of Lagrange formula
sa
where Y(x) = (y(x), y x), . . . , y - (x)), Z(x) is defined similarly and B is a bilinear
form of the vectors Y 2. It is easy to see that the form B is non-degenerate.
Subjecting the function y to the boundary conditions (87), wz extract an ndimensional subspace R of the 2n-dimensional space zf vectors Y = (Y(a), Y(b)).
Let us consider the s_ubspace R of those vectors 2 = (Z(a), Z(b)) which are
orthogonal to every YE R in the sense that
ere for simplicity the concept of the adjoint operator is interpreted in the real sense, which is the
natural one if the functions pk, a&, q are real.
13
0 4. Integral Equations
~i[Z]
(1 GiGn),
(87*)
L[y]z
dx =
s a
s (I
Moreover, if z E C a, b] and the identity
yL*[z]
dx.
(94)
L[y]z
dx =
yw dx
sa
s a
is satisfied for ally E D(L) and some w E C[a, b], then z E D(L*) and w = L* [z].
The differential operator L is said to be self adjoint if L* = L. Such an operator
necessarily has even order n = 2m and the form
l[y] = yt2m) + (nl(x)y Putting y = L- [f]
+ *.* + n,(x)y.
b
G*(x,
ss0 (I
where G* is the Green
51fCW3
dx
dt
W,
ss a
WWf(5)
dx
&,
LI
function is symmetric:
G(x, 5) = W, 4.
It is now clear that the systems (vi )); and (vi
;in (91) are bases for the space
of solutions of the equation 1* [u] = 0, the udjoint equation for I[u] = 0.
Example. The Green
function of the operator
interval [a, b] with boundary conditions
YW
- hlY(4
= 0,
l[y]
y b) + Jw@) = 0
W, 67=
= y - q(x)y
UlW2(5)
(x
0,
~2w41(5)
(x
09
on the
14
Chapter
1. Classical
Concrete
Problems
where (ur , u2} is a basis for the space of solutions of the equation Z[u] = 0, chosen
so that u1 satisfies the condition at the point a, u2 satisfies the condition at the
point b and the Wronskian of this pair of solutions is equal to 1.
We note that by inverting the operator L the problem of finding eigenvalues
and eigenvectors,
L[w] = Iw,
(95)
is reduced to the homogeneous
w(x)
G&v
h~(0
dt
s II
with parameter 1 (generally complex). If A is not an eigenvalue then the inhomogeneous equation
LEYI = LY + f
can be solved for any J Moreover
it is equivalent
Y(X) =
G(x,
BY(~) dt + g(x),
sa
where g = L- [S]. Thus the types of linear integral equations which have been
introduced are naturally associated with linear differential operators and can be
used for the study of the latter (in particular, this applies to the Sturm-Liouville
problem). The realisation of this approach requires an independent development
of the theory of linear integral equations which was carried out in the works of
Fredholm, Hilbert and Schmidt; these played a prominent role in the creation
of functional analysis.
A
4.2. Fredholm and Volterra Equations. The definition of these classes does not
require linearity; indeed non-linear integral equations also form an interesting
and important subject.
Suppose we are given an interval Z c R, a function g(x) on Z and a function
K(x, 5, y) on Z x Z x IR. An equation of the form
W, 5, y(5)) d5 = g(x)
b E 1)
sI
is called a Fredholm integral equation of the 1st kind. By a Fredholm integral
equation of the 2nd kind we mean an equation of the form
(x E I).
W, t> ~(0) dt + ~$4
sI
Now let us put Z, = I n (- co, x] (x E I). If in the above we change the interval
of integration to Z, we obtain Volterra integral equations of the 1st and 2nd kinds
in the respective cases.
A general theory of integral equations of the 1st kind is lacking even in the
linear case, although certain particular equations of this type have been successY(X) =
15
Y(X) =
K(x,
sD
t,
y(5))
&
(a
g(x)
<
<
b)
- VY,)W
GM
b I(YI(~) - YAOI d5
s ll
TY,) G Mb - 44y,,
~2).
Therefore the solution not only exists and is unique but it is also the uniform
limit on [a, b] of the iterative process
Y,+,(X) =
b K(x, 5, v.(O)
s0
& + g(x)
(n=O,
1,2,...)
x K(x, t, ~(0)
Ia
solution
dt + g(x)
on the
(a s x < b)
interval
[a, a + 1) where
1=
Y(X)),
Y(O) = Yo
(96)
76
condition
equation
(96) is equivalent
xf(5,
I
~(0)
&
to the Volterra
integral
+ ~0.
integral equation
Y(X) =
b W, Oy(t) dt +f(x)
(97)
Ia
on the interval a d x G b, assuming that the kernel K is continuous on the square
a < x, < < b andfe C[a, b]. We also look for a solution in C[a, b] (all functions
considered can be complex).
The equation (97) is the continuous analogue of the system of algebraic linear
equations
Yi= iil
jYj+ fi
The Fredholm alternative is the central result of the theory: either the equation
(97) has a solution for any f~ C[a, b] or the homogeneous equation
b
u(x)
K(x,
MO
dt
(a < x < b)
WV)
drl
(99)
I II
has a non-trivial solution. In addition to this we can assert that
1) the subspace of solutions of equation (99) is finite-dimensional,
2) a necessary and sufficient condition for the solvability of equation (97) with
fixed f is that f be orthogonal to all solutions of the adjoint homogeneous
equation
b
u(x)
Kh
w*)
I (I
of the spaces of solutions of equations (99) and (99*) are
3) the dimensions
equal.3
Thus the basic facts of the theory of finite systems of algebraic linear equations
hold for Fredholm integral equations of the 2nd kind. When applied to the
problem of eigenfunctions,
b
w(x)
sa
W.-G
t%W
d5,
3If they are equal to m then 2) reduces to m orthogonality conditions (J 0,) = 0 (1 Q k d m) where
(ok)7 is a basis of the space of solutions of equation (99*).
77
W, OY(X)Y(~) dx dt
ssI I
just as a (Hermitian) quadratic form is associated with any Hermitian matrix.
The variational method applied to the functional Q under the condition
s
leads to the following fundamental
Theorem.
, Iv(x)12 dx = 1
result of Hilbert-Schmidt
theory.
WG 5)W & = 0
sI
with a Hilbert-Schmidt
kernel has only the trivial solution in L Z). Then the
eigenvalues form an infinite sequence tending to infinity and a complete orthonormal system (4 ; can be constructed from its eigenfunctions.
A most important consequence of this result is the theorem on the complete
system of eigenfunctions in the Strum-Liouville
problem and, in general, for any
linear self-adjoint differential operator on a finite interval.
Let us denote by 12,the eigenvalue corresponding to the eigenfunction 4 Then
we have the expansion of the kernel
= 0)
78
(101)
!!!!!
n=1
'
where the c, are the Fourier coefficients of the function y with respect to the
system (4Jp, i.e.
cn
=sYC&W
- dx.
I
The identity (101) is just the theorem on the reduction of the Hermitian
functional Q to its principal axes.
We note that the relation
quadratic
follows from (100) on applying Parseval equality. If all the 1, > 0 (which is
equivalent to positivity of the functional Q) and if further the kernel K is
continuous and the interval Zis finite, then the expansion (100) converges uniformly
on the square Z x Z (theorem of Mercer). In particular
K(x,
x)
f
n=l
y,
I
4.5. Equations with Difference
Kernels.
5 4. Integral Equations
19
a,
A(I) =
s
-CO
(-co<1<00)
a(t)e- dt
In fact, using the language of Fourier transforms, we can express equation (102)
algebraically:
Y(I) = A@)Y@)
+ G(1).
If A&) = 1 and the equation is solvable we obtain G(1,) = 0, which cannot hold
for all g E ~5.Conversely, if A@) # 1 for all 1, then by Wiener theorem
(1 - A(I))-
where B is the Fourier transform
equation (102) is satisfied by
= 1 + B(1),
of a certain function
bEL
Consequently
y=g+b*g
80
(1 - A(n))- = [l + B+(L)][l
where
B+(A)
=sm
b, (t)e dt,
+ B-(A)]
(--oo<~<oo),
B-(A) =
m b- (t)e- dt
s0
are functions which can be extended analytically to the upper and lower halfplanes respectively (b, E L(0, co)), If v > 0 the factor 1 + B-(A) can always be
chosen so that it has no roots in the lower half-plane. Moreover
(1 + B-(A))- = 1 +
om/I?-(t)e- dt
(im 3, < 0),
I
where b- E L(0, co). The factor 1 + B+(A) will have exactly v roots (counted
according to multiplicity)
in the upper half-plane (by the principle of the argument); their locations can be chosen arbitrarily, after which the factorisation will
be uniquely determined (so that for v = 0 it is unique). More specifically
1 + B+(A) = R(1)(1 + c(A)),
where R(A) is a rational function which has the prescribed roots in the half-plane
im A> 0, a pole of order v at 1= - i and no other roots or poles, and
(1 + c(A))- = 1 +
m j?+ (t)e dt
s0
(im A 2 0)
ln(1 - A(A)) =
m l(t)e- dt
(-co<~<oo)
s -UZ
= exp[ -I:a
l(t)e-
t]
exp[ -J:
l(t)eei dt]
81
4 4. Integral Equations
(--oo<x<a3),
where
W
= -
m 4x - 5)~(5) d5
(x < 0)
s0
and h(x) = 0 (x 2 0). Converting to Fourier transforms and using the factorisation we obtain
w
1 + B-(n)
The left hand side
L(0, co) extended
Fourier transform
and /I(x) = 0 if x
function from L(
Consequently
= (1 + B+(Q)G(L)
+ (1 + B, (I))H(L).
O3f(x)e- x dx.
s0
(105)
The arguments outlined can be reversed, i.e. (105) is in fact the solution (it is
unique because of the previous part).
For v > 0 each factorisation produces a certain solution from formula (105).
Thus we obtain an infinite set of solutions parameterised by divisors of roots with
total multiplicity
v which lie in the upper half-plane. The space N of solutions of
the homogeneous equation
u(x) =
(106*)
m a(< - xb(O 4
(x > 0)
s0
has index (-v) > 0 and so the space N of its solutions is lvl-dimensional.
If
equation (103) is solvable in this case, S must be orthogonal to all solutions of
equation (106*) (and, conversely, under this condition equation (103) is solvable).
Thus a necessary condition for solvability of equation (103) for all g is that v 2 0.
e draw attention to their similarity with separation of variables.
sA suitable construction of a basis in N shows that each function u E N is absolutely continuous,
u E I, and u(co) = 0.
9 More precisely, transposed, but we will not make use of this term, relying upon the context.
82
Chapter
1. Classical
Concrete
Problems
This situation is related to the Fredholm case but in general there is the significant
difference that dim N # dim N. The difference dim N - dim N is equal to the
index v in all cases, i.e.
dim N =
v (v>O)
0 (v<O)
dim N =
0
I -v
(v 2 0)
(v<O)
+ w3
(r E n
(107)
(r
u(Oge(C)
E n
(108)
~440
a(0
E r)
now arises, the solution of which (unique if #J co) = 0) is given by the integral of
Cuuchy type
at infinity.
83
where gi, ge are the exponentials Of 4i, &. They are continuous on Gi, G, analytic
on Gi, G, respectively and have no roots. Such a factorisation is defined up to a
common constant factor in gi and ge,
If v > 0 we can nullify the index by dividing a(<) by [, after which we can
apply the previous factorisation. This produces the desired factorisation of a([);
moreover g,(z) = G((z(-) as lzl+ co which leaves the possibility of multiplying
the solution obtained, (gi, ge), by an arbitrary polynomial of degree <v. All
solutions are realised by this process, so that the space of solutions turns out to
be (v + 1)-dimensional I.
Just as in the Wiener-Hopf method, using the factorisation we can effectively
solve the inhomogeneous problem by eparation of variables. It turns out to be
solvable for all b(c) if v > - 1. A criterion for solvability when v < - 1 is the
orthogonality condition
Wh(O
4 = 0,
fr
where hi is the interior part of the general solution of the associated homogeneous
problem
(v 3 0)
ov + 1
(v < 0)
dim N =
0
I -v-
(v> -1)
1 (v< -1)
1.
(109)
(in addition
the condition
84
assumption
where the integral is to be interpreted as a principal value, i.e. as the limit of the
integral along the arc T,(c) = {r: r E r, Ir - cl > E} as E + 0. Substitution in (107)
leads to a singular equation of the form
(110)
where cc(c)= (1 + a([))/2, /3(c) = (1 - a([))/2, r(c) = b(c) and consequently
43 + P(C) = 1.
Conversely, if an equation of the form (110) is given and cr([) + /l(c) # 0
everywhere on r, then (by inversion of the above calculations) it reduces to the
Riemann-Hilbert
problem with coefficients
which allows us to define the index of equation (110) and to obtain the same facts
that are available for the Riemann-Hilbert
problem. These constitute the theory
of Noether:
1) finite-dimensionality
of the space N of solutions of the homogeneous
equation;
2) a criterion for solvability of the inhomogeneous equation in terms of
orthogonality of the right hand side to the space N of solutions of the adjoint
homogeneous equation;
3) the equality dim N - dim N = v (the difference from (109) is explained by
the fact that f,(m) = 0).
The theory of Noether remains valid for more general singular equations
eak singularity:
K, 4 = our: - V),
O<K<l.
Chapter 2
Foundations and Methods
$1. Infinite-Dimensional
Linear Algebra
All the ituations discussed in this paragraph will be concerned with a prescribed basic linear space E # 0 over an arbitrary (within the limits of this section)
field K (or simultaneously with several such spaces). The elements of the space
E, i.e. oectors or points, will usually be denoted by lower case Roman letters, while
lower case Greek letters will be used for scalars, i.e., elements of the ground field.
The identity of the field will be denoted by the figure 1.
The standard function spaces (C, Lp etc.) lie outside of the range of traditional
linear algebra, since it is restricted to the finite-dimensional
case. Consequently
in the context of functional analysis this restriction is unacceptable, the reason
being that the spaces considered below are generally infinite-dimensional.
Infinite-dimensional
linear algebra comes in at the foundation of functional
analysis as a suitable
lgebraic model for the theory of linear topological spaces,
which occupies a central place in functional analysis.
1.1. Bases and Dimension. By an algebraic basis, or Hamel basis, of a space
E we mean a maximal (with respect to inclusion) linearly independent subset.
Using Zorn lemma (or, equivalently, transtinite induction) we can easily prove
Theorem 1. Any linearly independent subset
basis.
Corollary
of a space
E can be extended to a
Moreover the extending vectors2 form a basis of a subspace M such that the
whole space E is the direct sum of L and M (L i M = E, i.e. M is a complement
for L).
Let a basis (ei}is, of the space E be given (I is an index set). In the unique
representation of a vector x E E with respect to the given basis the set of indices
of coefficients which are different from zero is finite. It is called the support of the
vector x and is denoted by supp x.
We have
x
ie&
ti%
his word can be omitted in the purely algebraic context of the present paragraph.
Zlf they exist.
86
(the & are called the coordinates of the vector x with respect to the given basis).
Any other basis {Uj}j,J can be expressed in the form
uj
1
isI
&jet,
and b =
(Bjk),
C
je.l
bjk
(k
E 1)
then
jTJ
Q&k
hk
(i,
k E 0
zI
fijiail
sjl
09
l E 4.
(1)
If x has coordinates (&)ier with respect to the initial basis and (~?i)~,~with respect
to the new basis, then
5i
jTJ
QijVj-
f?i
(i E I)
hl
of the set Z into the set of finite subsets A of the set .Z. The preimage of each A is
finite. Therefore card Z G card J. Similary, card .Z G card I. 0
The power (cardinality) of any basis is called the algebraic dimension of the
space E and is denoted by dim E. As a common index set for all bases we can
take the least ordinal of power dim E. Let us denote this by n.
We note that, after identifying .Z with I, we can write formulae (1) in the form
ub = e, bu = e, where e = (~5,~)is the identity matrix. The invertible matrices
whose columns have finite support form a full linear group GL(n, K).
On any set S a function rp: S + K is said to have finite support if its support
supp cp = {s: q(s) # 0} is hnite3. The functions of finite support form a subspace
Q+,(S, K) of the space @(S, K) of all functions (it is a proper subspace if the set
S is infinite). In @,-JS,K) we have the canonical basis {~~}sos: d,(t) = &(t E S). The
rp on a topological space is said to have compact support if its support supp cp=
{s: q(S) # 0) is compact. This coincides with the present notion of we give S the discrete topology.
3A function
0 1. Infinite-Dimensional
representation
Linear Algebra
a7
of any rp E @e(S, K) is
cp =
(Pws = c (PM4.
SOSUPPCP
SES
Since, clearly, dim @e(S, K) = card S, any cardinal is the dimension of some
linear space.
The dimensions of isomorphic spaces (E, x E,) are obviously equal and conversely isoinorphism of the spaces follows from equality of their dimensions (a
l-l correspondence between the bases can be extended by linearity to an
isomorphism of the spaces).
Example. If B is a basis in E, then E x Q&3, K). Hence it follows in particular
that, if K is infinite or E is infinite-dimensional,
then4
card E = max(dim
E, card K).
We can use this formula to calculate the cardinal v = dim @(S, K) for infinite S
(if S is finite then v = card s). We have
card @(S, K) = max(v, card K).
But v 2 card K, for even if S = N E (0, 1,2, . . .} the functions cp,(s) = ~1(a E K)
are linearly independent. Thus
dim @(S, K) = card @(S, K) = (card K)cardS.
We note that dim @(S, K) > card S, i.e.
dim @(S, K) > dim @e(S, K).
(2)
An interesting quantity associated with a linear space E (dim E > 1) is the least
power for a set of proper subspaces covering E. It turns out to be
countable if E is infinite-dimensional
equal to card K if E is finite-dimensional
and K is infinite,
and K is infinite,
then clearly
88
is equal to the sum of their dimensions. In the general case the dimension of a
sum of subspaces does not exceed the sum of their dimensions.
The intersection of all subspaces which contain a given non-empty set X c E
(i.e. the smallest such subspace) coincides with the linear hull Lin X, i.e. with the
set of all linear combinations of vectors in X. The dimension dim(Lin X) is called
the rank of the set X and is denoted by rk X. It is equal to the power of any
maximal linearly independent subset Y c X (each such Y is a basis in Lin X).
One of the fundamental constructions of linear algebra associates with each
subspace L c E the factor space E/L, consisting of the classes of the following
equivalence relation (congruence modulo L):
x=y(modL)ox-yEL.
The class of the vector x is denoted by [xl. The operations on the classes are
defined naturally by
dxl
= Caxl,
which are valid because the right hand sides of these identities are independent
of the choice of representatives of the classes on the left hand sides.
Any complement M of the subspace L is isomorphic to E/L (the natural mapping x H [x] serves as the isomorphism). All complements of a given subspace
L are therefore isomorphic and consequently they all have the same dimension.
It is called the codimension of the subspace L and is denoted by codim L. If
codim L = 1 (and only in this case) L is a maximal (with respect to inclusion)
subspace different from E and is called a hyperplane.
A subset of the form A = x + L (x E E, L a subspace) is called an afme
manifold. In addition we put dim A = dim L, codim A = codim L. If dim A = 1
it is called a line; if codim A = 1 it is a hyperplane.
1.2. Homomorphisms
and Linear Functionals. For any linear spaces E,, E,
the linear space of homomorphisms6
h: E, + E, is denoted by Hom(E,, E,). In
particular Hom(E, K) is the space of linear functionals on E. It is called the
algebraic conjugate of E and is denoted by E.
If a basis B = {e,}iel is given in E then to each f E E there corresponds
a function f(ei) on B and this correspondence is an isomorphism
(so that
E x @(B, K)) since
f(x)
iTI
f(ei)W),
where the &(x) are the coordinates of the vector x with respect to the basis B.
We stress that the coordinate functionals ti do not form a basis in EX if E is
infinite-dimensional,
since the infinituples (f(ei))i,, can be chosen arbitrarily and
6According to the standard definition, (h, + h,)x = h, x + h,x, (c&)x = a(hx). Further, if h: E, -B E,,
g: E, + E, are homomorphisms, then their product (superposition or composition) gh: E, + E, is
also a homomorphism.
n other words @,JS, K)* x @(S, K) for any S.
0 1. Infinite-Dimensional
Linear
Algebra
89
not just with finite support. However the set {gi)is, is linearly independent
because of the biorthogonality relation: &(er) = 6,. As a result of (2), dim EX >
dim E if (and only if) E is infinite-dimensional.*
If L is a subspace of E, the restriction & of any f E EX is a linear functional
on L. Conversely, any f0 E L# can be extended to an f E EX by choosing any
complement M for the subspace L and setting f(x) = fe(u) where x = u + u
(u E L, u E M). Hence it follows in particular that if e E E and e # 0 there exists
f E E such that f(e) # 0 (it is possible to make f(e) = 1). The same result shows
that the linear functionals on E separate points: if x, y E E and x # y there exists
fc E such that f(x) # f(y).
When applied to L = Lin{e}, the extension method described above shows
that, if a subspace V does not contain the vector e, there exists f E E such that
f(e) # 0 and flY = 0. Any subspace is therefore the set of solutions of a system
of linear equations h(x) = 0 ({J) c E).
The subspace
L = {f: f E E, f(x) = 0 (x E L))
of linear functionals which are orthogonal to the subspace L is called the
annihilator9 of L. In the language of projective geometry the mapping L I-+ L * is
a correlation. It is an involution in the following sense. For any subspace N c E
we define its annihilator in E to be the subspace
N, = {x: x E E, f(x) = 0 (f E N)).
Then (L 1)1 = L. We note the further obvious implications
L1cLz*L+L:,
Nl = Nz *Wd.
= WA1
= (card K)d E.
his definition
remains meaningful
for any non-empty
of E.
lo In general (Nl)
# N although
N c (NJ .
90
Chapter
2. Foundations
and Methods
h)(f) =fM
(xEE,fEE+).
It is injectiue since, if f(x) = 0 for all f E E#, then x = 0. Identifying
E we obtain the inclusion
reflexive if E# = E.
Im u with
If E is infinite-dimensional
IfN=OthenN,=EandN =E .
@#f)(x) = f&4
(xEEl,fEE;).
Coker hX x (Ker h)
Im h# z (Im h)#.
Q 1. Infinite-Dimensional
Substitution
the identities
Linear
91
Algebra
of formulae
The homomorphism
h# #: Ef + E2## is the natural extension of the homomorphism h from E, to Ef # .
The linear spaces over a given field K are the objects of a category in which
the homomorphisms
serve as morphisms. The following contravariant
adjoint
jiinctor acts on this category: E H E#, h H h# ((gh)# = h#g#, l# = 1). The
more general contravariant
functor E H Hom(E, E,) (E, is a fixed space) transforms the morphisms by means of the commutative diagram
PI,,,=1
x = u + u (u E L, u E M) of an
P (Px = u) for which P2 = P,
onto L along M and 1 - P is
in L, (uj} is a basis in M and ti
to the basis (q} U {vi} in E
Px = c &Ui.
I
Associated with any decomposition
of a space E into the direct sum of
subspaces L, is a family of projections P, which are defined by the conditions
Im P, = L,,
Ker PO = 1. L,.
the mapping
h H h is linear:
(h, + h,)*
92
Chapter
2. Foundations
and Methods
zero and
x = 1 P,x.
d
under consideration
i.e.
n Ker P, = 0.
0
subspaces.
the operator
A.
5 1. Infinite-Dimensional
Linear Algebra
93
h: E, + E, is said
If h is a Noetherian
h) - dim(Coker
h).
A Noetherian homomorphism
with zero index is called a Fredholm homomorphism. In this case surjectivity is equivalent to injectivity, i.e. the Fredholm
alternative holds: either the inhomogeneous equation hx = y is solvable for all y
or the homogeneous equation hx = 0 has a non-trivial solution.
It follows from formulae (3) and (4) that the homomorphism
h# is Noetherian
if and only if the homomorphism
h is Noetherian and in the Noetherian case
cork h# = def h 2
def h# = cork h,
so that ind h# = -ind
h. It is appropriate
homomorphisms
is Noetherian
and its
operator.
Proof.
We take a complement of the subspace L = Im A in the whole space:
L i M = E. If x = u,, + u (uO E L, u E M) then x = u + (1 - A)u, where u =
94
Chapter
2. Foundations
and Methods
If the operator
T: E ---* E is Noetherian
R, (left and right regularisors) such that
R,T=
1 -A,
R, and
TR, = 1 - B,
any operator
+ 1/R,),
and
R,(T + V) = (1 + R,V)(l
are Fredholm
- (1 + R,V)- )
R, = ind T.
and
95
(i E I),
(5)
for solvability
(consistency)
holds for any system of scalars (jli)iel with finite support. Conversely, if (6) is
satisfied the system (5) is equivalent to a subsystem corresponding to a maximal
linearly independent subset F, c F.
If we introduce the homomorphism
h,: E + @(I, K) by means of the formulae
(hFx)(i) = fi(x) (i E I) we can write the system (5) in the form of an equation
h,x = y, where y(i) = vi (i E I).
Lemma. Zffi,...,fm
are linearly independent linear functionals, the system of
equations h(x) = vi (1 < i d m) is solvable for any right hand sides vI, . . . , q,,,.
Proof. Because the set M = { 1, . . . , m} is finite we can identify @(M, K)# with
@(M, K). The conjugate homomorphism
h; maps any collection (Ai) to cE1 A&.
By the condition hz is injective. Consequently h, is surjective. 0
Corollary. If the system (5) has finite rank (i.e. rk F is finite) then for its
consistency it is necessary and sufficient that condition (6) be satisfied.
Example. Let (ei}i E1 be a basis in E and { ri}i EI the set of coordinate functionals.
Because of their linear independence the system of equations ti(x) = qi(i E I)
satisfies condition (6) but it is solvable only for those (qi)ior of finite support.
In the general case, (6) is necessary and sufficient for the consistency of each
finite subsystem of system (5), i.e. for the reduction to the case of linearly
independent equations.
If the system (5) is consistent and x0 is some particular solution of it, then the
general solution is described by the formula x = x0 + z, where z is the general
solution of the corresponding homogeneous system:
fi(z) = 0
(i E I).
(50)
96
Chapter
2. Foundations
and Methods
X0
codim NF = rk h,.
In the case where the system consists of one equation f(x) = 0 (f # 0) we have
codim NF = 1, i.e. NF is a hyperplane (NF = Ker f). Conversely, for any hyperplane H c E there exists a linear functional f # 0 such that Ker f = H. In fact,
since dim(E/H) = 1, we can identify E/H with the held K and then f: E + E/H
is the natural homomorphism.
Finally, if Ker fi = Ker f2 then fi and f2 are
proportional. Thus the set of hyperlanes in the space E can be identified with
the projective space PE consisting of the classes under proportionality
on
E#\ (0).
In contrast to the very concise representation of the system (5) in the form of
the equation h,x = y we can represent it in scalar form
jJ$J
ccijC;j
Vi
ti
E 03
(7)
where (c;i)jsJ are the coordinates of the vector x E E with respect to some basis
Iej>jeJ and aij = J(ej). In addition only solutions of finite support are sought.
However we can dispense with this restriction if the rows in (7) have finite support.
From the geometric point of view a system of the form
(i E I)
(5#)
is obtained, where xi is the vector with coordinates (IX~~)~,,and f is an unknown
linear functional. The condition
ftxi)
Vi
T &xi = 0 =s T liqi = 0
(A E Go(Z, K))
(6
is sufficient for solvability of the system (5#) since, if the system (xJisl is already
linearly independent, it can be extended to a basis, after which the functional f
is defined by the values vi on the vectors xi and is extended arbitrarily on the
other basis vectors. With this we have established
Theorem of Toeplitz. For consistency of the system of equations (7) when in
each equation the set of non-zero coefficients is finite13, it is necessary (triuial) and
sufficient that all its finite subsystems be consistent.
I3 We shall sometimes
a Toeplitz
system.
$1. Infinite-Dimensional
Linear
Algebra
97
polynomial
1. Any projection
is 1 - II.
= 0}
on
98
Chapter
2. Foundations
and Methods
1 0
...
/z 1
...
/I
...
...
...
...
2,
Let A be an algebraic
Cp(n)
(n
operator
s , ,
j=l
2, are pairwise distinct. Then {S} 1 is the set of all eigenvalues of the
whereA,,...,
operator A, their orders are equal to the corresponding rj and the space E decomposes into the direct sum of the maximal root subspaces:
E = I+ &) i ... i Vi &).
(8)
In particular, if all the roots of the minimal polynomial are simple (and only
in this case), the operator A is diagonalisable, i.e. E decomposes into the direct
sum of the eigenspaces.
A decomposition of the identity operator 1 into the sum of root projections PI,
. ..) P, is defined by the decomposition (8). These projections have the following
explicit representation:
Pk = cp,(A), where (pk is the interpolating
polynomial
corresponding
to the conditions:
cp,(Aj) = djk;
The root projections
case
q# Aj) = 0
(1 Q i < rj - 1; 1 <j
< s).
5 1. Infinite-Dimensional
99
Linear Algebra
R,(A) = y,
where 8 is a polynomial in two variables (deg, 8 < deg cp- 1).
The operator p(A) is naturally defined for any rational function p whose poles
are different ,from A,, . . . , A,. Moreover the Lagrange-Sylvester formula holds:
fi
(A
Aj),
j=l
where vi < rj + l(1 < j d s). Let us consider the linear equation cp(A)x = 0 with
unknown x E E. The subspace of its solutions N = Ker q(A) is invariant under
A and the operator Al, is algebraic. Its minimal polynomial coincides with cp.
By the spectral theorem
N = w (n,) i ... i w&)
(these subspaces, which are root subspaces for A, are contained in N) and, in more
detail, Al, has a Jordan structure.
We will discuss two classical examples from this point of view.
Example 1. The difirentiation
operator D = d/dt on the complex function
space Cm(R) has all I E C for its eigenvalues (of infinite order). The root subspace
w,(A) consists of all functions of the form z(t)e , where rc is a polynomial of
degree <k - 1.
If follows from the spectral theorem that the general solution of the linear
differential equation with constant coefficients
cp(D)x = 0
p(l) = fi (A - ij)
j=l
>
7rj(t)e
j=l
cw,
= 5,+1
(n = 0, 1,2, . ..)
loo
q(A) = fi (A - ljp;
j=l
cp(O)z 0
>
)... ),
(x = (Lk E L)
(9)
(i.e. the equation (T IL)#o - 0 = - &,, where &-, is regarded as a linear functional
on L). Further, we shall say that L admits summation if a generalised sum exists
on it. We denote the class of such subspaces by z.
On the subspace C of those x for which the series to + <I + t2 + * * * converges
the ordinary sum is clearly a generalised sum. Many classical methods of summation on their natural domains of definition (in particular, the Cesaro method)
satisfy condition (9).
A method of summation on L 3 C is said to be regular if it coincides on C
with the ordinary sum. Such a method always exists if L E Z.
We denote by N the subspace of sequences of the form (z(n)); where II runs
through all polynomials. This is the maximal root subspace of the operator T for
A= 1.
Theorem 1. Let L be an invariant subspace. Then L E Z ifand only if L
n N = 0.
Naturally,
Zorn
if and only
0 1. Infinite-Dimensional
Linear Algebra
101
Let us now consider the invariant subspace Q of all sequences of the form
where all the lj are different from 1 and 0 and the 7rj are
t, = Cpl
polynomials. Since Q n N = 0, we have Q E L: The operator T - 1 on Q is
surjective and so the generalised sum on Q is unique. It can be described in a
straightforward manner. To this end we note that if x E Q the power series
c. e <,,[ converges on some disk Ill < r,. Its ordinary sum f,(c) (the generating
function of the sequence x) is rational and its poles are ;Z: (1 < j < m).
?Tj(n);ljn,
102
by 2I = K i E with multiplication
(LX+ x)(/I + y) = ap(a, /I E K, x, y E E)16. The
subspace E c 8 is the unique maximal ideal.
In any algebra $Ieach element a generates the ideal (a) = { y: y = ax (x E 2l)}.
The set of non-invertible
elements therefore coincides with the union of all
maximal ideals. Consequently an algebra 2I is a field if and only if its unique
maximal ideal is the zero ideal. If I # 3 is an ideal in 2I then the factor algebra
2l/I is a field if and only if the ideal Z is maximal.
The radical of an algebra Iu is the set Rad 2Iof those elements a E $Isuch that
e - ax is invertible for all x E $I.
Theorem. The radical of an algebra 9I coincides with the intersection
maximal ideals.
of all its
1. Rad 2I is an ideal.
If the element
invertible:
a E 3 is nilpotent,
(e
a)-
1
k=O
uk.
But if a is nilpotent so also is ax for any x E 9I. Therefore all nilpotent elements
are contained in the radical of the algebra. They form an ideal called the nil
radical. The nil radical coincides with the intersection of all simple ideals (an ideal
I # 9I is said to be simple if there are no divisors of zero in the factor algebra
%/I). If E is finite-dimensional
its nil radical coincides with its radical.
For each maximal ideal M in an algebra 2I the field +3/M is an extension of
the field K, for the latter is canonically embedded in 5X/M by means of the chain
of natural homomorphisms1 7:
We call an algebra spectral if U/M = K for all M, i.e. the codimensions of all
maximal ideals are equal to one. This condition is a fundamental axiom of the
I6 In general if 21L,is an algebra without identity the construction
= K i I, with multiplication
(a + x)(B + y) = a/3 + (ay + fix + xy) is called the unitisation of 9llo.
I7 We stress that a homomorphism of algebras with identities must by definition map identity to
identity.
0 1. Infinite-Dimensional
Linear Algebra
103
general spectral theory set out below, the principal prototypes of which are to be
found in the spectral theory of operators, in the theory of commutative Banach
algebras and in algebraic geometry. In what follows the algebra 2I is assumed
to be spectral.
The spectrum of an element a E 2I is the set spec a of those 1 E K for which the
element a - le is not invertible * . The remaining I E K are said to be regular,
the set of regular values is denoted by reg a and it is called the resolvent set for
a since it is the domain of definition of the resolvent R,(u) = (a - le)-.
Example. If a = ue then spec a = {a}, R, = (a - I)- . In particular
{0}, spec e = (1).
We note two fundamental relationships for the resolvent:
uR, = e + AR,;
spec 0 =
R, - R, = (A - ,u)RAR,.
Proof. We have:
spec a = {A: 3M E 2X s.t. a - i2e E M} = {A: 3M E !UI s.t. ii(M) = A}. 0
Corollary
Non-invertibility
to be quusinilpotent
Corollary
algebra.
is quasinilpotent.
coincides
with
the radical
of the
I8 For this and the definitions which follow immediately it is not necessary that the algebra be spectral
or even commutative.
104
In particular, if a is idempotent,
the identity then spec a = (0, l}.
Corollary 4 (Spectral Mapping
spec e(a) = B(spec a) holds.
0 the identity
This 0(a) is invertible if and only if e(n) # 0 (A E spec a). Consequently we have
the functional calculus p H p(a) for all rational functions p which are regular on
spec a and spec p(a) = p(spec a).
For a maximal ideal M the natural homomorphism
X~M:9 -P %/M = K is a
character, i.e. a linear functional on % which is multiplicative in the sense that
XM(ab) = XM(a)XM(b)(XM # 0 and in fact X,,Je) = 1). The mapping M H X~ of the
space of maximal ideals to the set of characters is bijective.
Let us consider some specific examples of spectral algebras.
Example 1. An algebra % over the field [F2 = (0, l> is said to be Boolean if all
its elements are idempotent, i.e. the identity x2 = x is satisfied in 9I. Any Boolean
algebra 2l is spectral since B/M is a Boolean field for any maximal ideal M and
then M/M = IF, since the identity element is the unique invertible idempotent.
We now observe that any Boolean algebra %is semisimple since, if a E Rad 2l,
we have spec a = (0) so that a - e is invertible and then from the identity
a(a - e) = 0 it follows that a = 0.
The image $I of the algebra 3 under the canonical functional representation
is isomorphic to 2L However the functions G E % take only the values 0 and 1
and can therefore be identified with the corresponding subsets of the space 93 of
maximal ideals. Moreover a sum transforms into a symmetric difference while a
product becomes an intersection. Thus we obtain the result that any Boolean
algebra is isomorphic to a certain algebra of sets (theorem of Stone).
Example 2. Let K = [F, be any finite field (card K = q = pm, where p is prime
and m > 1). The identity x4 = x holds in K. If it holds in a field E 1 K then E = K
since the equation x4 = x cannot have more than q roots in E. Therefore if this
identity holds in an algebra 9I over K then 3 is spectral. Also it is semisimple
19These concepts are defined just as in the case of a linear operator. An algebraic operator on the
space E is an algebraic element of the algebra L(E) and its spectrum in L(E) coincides with the
spectrum in the previous sense if the minimal polynomial of the operator factorises into linear factors
over the field K.
4 1. Infinite-Dimensional
Linear Algebra
105
to some
(10)
(11)
has a root.
f,(t) = 0
(fv E KC-U, v E NJ
in the unknown t = (ti)iot (N is some index set), it is necessary (trivial) and suflicient
that each of its finite subsystems be consistent.
In fact, the ideal J c K [X] which is generated by the system (f,), EN is different
from K[X]. In the contrary case 1 E J, i.e. there exists a collection f,,, . . . , f,,
such that
Ill
forsomeg,,...,
g,,, in K[X];
m) is not consistent,
106
Chapter
Condition
2. Foundations
and Methods
a)
(a
E K,
{X,),,
K in
(%>h
1.
A proof is required only in the case card S > card K. But then there exists
R c S such that card R = card K and, since R is infinite, the algebra @(R, K) is
not spectral. But @(R, K) x @(S, K)/J, where J is the ideal of functions which
vanish on R. Moreover a factor algebra of any spectral algebra is also spectral.
This follows from the general result: if h: ?I + 2 is a surjective algebra homomorphism then the preimages of the maximal ideals of the algebra f? are all
maximal ideals of the algebra 9I which contain Ker h; further a/hx f!/M.
We remark incidentally that if the field K is finite, K = IF,, then the algebra
@(S, K) is spectral since the identity (pq = cpis satisfied in it.
Corollary 2. Zf S = N = (0, 1,2, . . .} and K is any infinite field then for any
non-standard maximal ideal M c @(S, K) the field E = @(S, K)/M is infinitedimensional over K.
8 2. Convex Analysis
107
In the Wiener algebra W all maximal ideals are standard. Thus if a function
f E W has no zero then it is invertible in W.
9 2. Convex Analysis
From now on the ground field K will always be R or @ and where possible we
will not distinguish these cases. As before the underlying space will be denoted
by E.
2.1. Convex Sets. A set X c E is said to be conuex if for each pair of points
x1, x2 E X the mterval
1,7,20 )
( g1 =, =
is said to be a convex combination of the points x1, . . . , x,.
k=l
(12)
The analogous assertion over R is false. For example, the Banach algebra C over R is not spectral.
See the proof of the Gel and-Mazur theorem in Section 4.17.
108
Chapter
2. Foundations
and Methods
k
k=l
PkXk
($/k~+
is called an absolutely conuex combination of the points x1, x2, . . . , x,. In order
that a set X be absolutely convex it is necessary and sufficient that for each finite
collection (xk)l of points of X all of its absolutely convex combinations belong
to x.
The intersection of any family of absolutely convex sets is absolutely convex.
If X is absolutely convex then ai X is also absolutely convex.
The absolutely convex hull of an arbitrary set X c E is defined (in two equivalent ways) similarly to the convex hull.
0 2. Convex Analysis
109
is called a non-negative combination of the points x1, . . . , x,. In order that a set
X be a wedge it is necessary and sufficient that for each finite collection (x& of
points of X all its non-negative combinations belong to X.
The intersection of any family of wedges is a wedge.
For any set X c E the set of all non-negative combinations of all possible
collections (xk)l of points of X is the smallest wedge which contains X, i.e. it
coincides with the intersection of all wedges containing X.
If K is a wedge, the intersection &,=r CK, if non-empty, is the largest subspace
contained in K. It is called the edge of the wedge. A wedge is said to be a cone if
its edge is zero or empty. The closed ray {y: y = ;Ix (2 2 0)} and the open ray
{y: y = Izx (A > 0)} defined by a non-zero vector x are cones. In the first case
the edge is zero, in the second it is empty.
For any wedge K # E its algebraic interior ai K is a cone (its edge is empty).
A wedge K is said to be generating if
,& cK = E*
Convex sets of one type or another arise naturally in connection with linear
inequalities. Let F = (f;:)iE, be some system of real linear functionals on E. Then
the set of solutions of the system of linear inequalities
fitx) 2 Vi (i E 1)~
(13)
where the ri are arbitrary real numbers, is convex and the set of solutions of the
corresponding homogeneous system
fi(x) > 0 (i E 1)
(13,)
is a wedge. Its edge is the space of solutions of the homogeneous system of linear
equations A(x) = 0 (i E I). The wedge of solutions of system (13) is therefore a
cone if and only if the system F is total. For the strict analogue of system (13,),
i(x) > 0 (i E I), the set of solutions is a cone (it may be empty).
2.2. Convex Functionals.
A real functional
(~9 Y E E)
110
(CI E R).
111
Let X be a convex absorbent set. Since in this case I,X c I,X (0 < il, < A,),
it is natural to introduce for X its gauge (or Minkowski functional):
q&)
= inf{l:
A > 0, x E AX}
(x E E).
3) I
= HP(x).
1. If g is a sublinear
functional
seminorm.
Example 2. If f is a linear functional then IfI is a seminorm.
The gauge of any absolutely convex absorbent set is a seminorm. The mapping
p H Dp is a bijection of the set of seminorms onto the set of absolutely convex,
absorbent, algebraically open sets. This property allows us to introduce seminorms in a purely geometric way.
Associated with each seminorm p is the subspace Ker p = {x: p(x) = 0}, its
kernel. Seminorms with zero kernel are norms. In other words, norms are those
seminorms whose unit spheres contain no rays.
It follows from the inequality
IP(X) - P(Y)1 G p(x - Y)
6qx(x) = 0 (x # 0) if and only if the ray defined by the vector x is contained in the set X.
(14)
112
is a seminorm. We will see in what follows that this is a universal method for
constructing seminorms. In order that the seminorm (14) be a norm, it is necessary and sufficient that the family F be total. In the general case
Kerp = (I Kerfi.
I
2.4. The Hahn-Banach Theorem. Let 4 be a sublinear functional on E and
f a real linear functional on a subspace L c E which satisfies the condition
f(x) < q(x) (x E L). Then f can be extended to the whole space E in such a way
that the majorant 4 is preserved, i.e. there exists a real linear functional f on E
such that f , = f and f(x) < q(x) (x E E).
The proof is carried out with the help of Zorn lemma. The basic step consists
of extending by one dimension, i.e. to a subsFace 1 = L + Lin, {e} where e 4 L.
It is sufficient to choose f(e) so that f(x) + c$(e) < 4(x + ae) for all c1E R, x E L.
But it is easy to see that
inf
xsL,po
Any value of f(e) lying between these bounds satisfies the required condition.
This theorem is one of the basic principles of linear functional analysis.
Corollary 1. Let p be a seminorm on E and f any linear functional on a subspace
L c E which satisfies the condition If(x)1 < p(x) (x E L). Then there exists a linear
functional f on E such that f L = f and If(x)1 < p(x) (x E E).
ProoJ If E is a real space we can extend fin such a way that f(x) < p(x) (x E E).
Substituting -x for x here, we obtain the required result. In the complex case it
is sufficient to extend the real linear functional g = re f so that g x)< p(x) (x E E)
and then to put f(x) = g(x) - @(ix). 0
A linear functional f on E is said to be subordinate to a seminorm p if there
exists c 2 0 such that If(x)1 < cp(x) (x E E). The greatest lower bound of such c
is called the p-norm of the functional f and is denoted by Ilf 11,.Thus
$2. Convex
Analysis
113
Any such functional is said to be p-supporting for the vector x0. It can be
constructed by defining the linear functional f-Jax,,) = ap(x,) on the one-dimensional subspace Lin{x,} and then extending f. to E without change of p-norm.
Let us now denote by Sz the set of linear functionals which are subordinate
to p and such that Ilfll, = 1. Then, as a result of Corollary 3,
p(x) = max If(x)1 = max re j(x).
JEq
feS$
Thus any seminorm is the maximum of some family of linear functionals (necessarily subordinate to it).
The following more general assertion is established similarly to Corollary 3:
if q is a non-negative sublinear functional on E, then for each vector x0 there
exists a q-supporting real linear functional f:
f(xo) = 4(x0),
f(x) G 4(x)
(x E E).
> a},
f by (-f)
and a by (-a)
the hyperplane
altered
but H,
and H-
are
114
Chapter
2. Foundations
and Methods
y E Y we have f(x - y + zo) < 1 and hence f(x) < f(y) and c1E ~up,~~f(x) <
inf,,, rf(y). Since X is algebraically open, the upper bound c1is not attained.
If we now take the hyperplane H = {v: f(v) = a} then X c H- and Y c H,.
Cl
We shall say that a hyperplane
(moreover X and Y can even intersect). Here we can also restrict attention
convex sets. From the theorem on the separating hyperplane follows
to
HnaiX=IZI(i.e.aiXcH-oraiXcH+).
Proof. Let G = {v: f(v) = a} be a hyperplane which almost separates X and
A: f(x) < CI(x E X), f(y) 2 a (y E A). If A = x0 + L, where L is a subspace, we
have f(z) > u - f (xe) (z E L) and so f(z) = 0 (z E L). Consequently, the hypercontains the manifold A. Moreover f(x) < f(xo)
plane H = {Y: f(y) = f(x,))
(x E X) and hence f(x) < f(x,,) (x E ai X), i.e. ai X c H- . 0
Corollary 3. Any convex algebraically open set X # E is the intersection of some
family of algebraically open half-spaces.
In fact, for each point u $ X there exists a hyperplane H(v) such that v E H(v)
and X c H-(v). Since moreover v $ H-(v) we have
X = f-) H-(v).
0+x
ix H-(v),
of the
115
4. Zf X is a convex algebraically
a convex set X if
a K-non-
Jb) =
cpdp
fS
on it. Since 1 E L n ai K, the functional J can be extended to a K-non-negative
linear functional Jon B(S, R). Now for any set M c S we can put &kf) = J(xM),
where xM is its characteristic function. With this we have constructed a finitely
additive extension of the measure p to the algebra of all subsets of the set S.
Correspondingly,
*As above, this term carries over to functionals. We note that if q is a non-negative sublinear
functional then a functional which is q-supporting at the point x,, (4(x,) # 0) is a support functional
for the set {x: q(x) < 4(x,)}.
91t is a cone if the wedge K is generating.
his wedge must therefore be algebraically solid.
116
where the integral can be interpreted as before, i.e. as the least upper bound of
Lebesgue integral sums. Naturally, this integral is only finitely additive.
From what has been said it is also clear that the integral with respect to a
finite, finitely additive measure gives the general form of a non-negative linear
functional on B(S, R).
For the space C(S) of continuous functions on a compact space S we have the
theorem of F. Riesz on the general form of a non-negative linear functional:
f(d=scp
dp,
X
by the
Ci
(i E I),
where the ci are given scalars (the moments of the required functional f with
respect to the system U). It turns out that for solvability of this abstract moment
problem it is necessary and sufficient that the implication
& aiui E K * gI aici 2 0
should hold for systems of scalars (cli)isl with finite support. We only need to
establish the sufficiency. Without loss of generality we can assume that the system
U is linearly independent. Then f is uniquely defined on L = Lin U by the given
equations and from the condition we see that it is (K n L)-non-negative; finally
we can extend it to a K-non-negative linear functional on the wholel of E.
i i A Bore1 measure is said to be regular if its value on each Bore1 set M coincides with the supremum
of its values on the compact sets Q c M and with the inllmum of its values on the open sets G = M.
In what follows, all measures on locally compact topological spaces are assumed to be regular Bore1
measures (or this property is to be proved).
he problem of constructing a linear functional f which is subordinate to a given seminorm p and
satisfies a system of equations f(uJ = ci (i E I) can be solved similarly with the help of the HahnBanach theorem. A criterion for solvability is that the inequality
I;, aicil < Cp(z,
a,,,>
(C is a positive constant)
should hold for all (c&, , oftinite support. Moreover the solution will satisfy the condition 11f I&,,< C.
117
f(v)=
moment
z rp(x)wd
s -X
on C( -71, rr) which satisfies the equations f(e,) = c, (e,(x) = einx, n = 0, +_1,
+2, . . .). A criterion for solvability is that, for each non-negative trigonometric
polynomial
T(x) = En a,einx, the combination
Cna,c, must be non-negative.
But, by the well-known theorem of Fejtr and F. Riesz, the condition T(x) 2 0 is
equivalent to 7(x) = ( V(x)12 where V is a trigonometric polynomial.
The Riesz-Herglotz criterion (positivity of the matrix (c.-,)~,=,)
follows easily
from what has been said.
2.7. Ordered Linear Spaces. This class of spaces was introduced and first
studied by L.V. Kantorovich (1935). A linear space E is said to be ordered if there
is given on it a partial ordering which is compatible with its linear structure:
1) x<y5x+z<y+z;
2) x<y*axGay
(a>O).
xdy-sy-XEK.
The elements of the cone K are said to be non-negative in accordance with
their characterisation by the inequality y 2 0.
Example. The linear space @(S, [w)of all real functions on any set S is ordered
pointwise: cp < +-+cp(s) < e(s) (s E S). This ordering is defined by the cone of
non-negative functions.
Any subspace L of an ordered linear space E is automatically
ordered, the
corresponding cone being K n L where K is the cone in E.
We now describe the standard method of ordering the space of homomorphisms Hom(E, El), where E, is also an ordered linear space and Q c E, is the
corresponding cone. We assume that the cone K in E is generating. Then the
homomorphisms
h: E + E 1 which map K into Q form a cone in Hom(E, E 1) and
the corresponding order is defined by this. In particular, h > 0 means that
x 2 0 => hx 2 0 and, because of the linearity of the mapping h, this is equivalent
to its monotonicity: x < y Z+ hx < hy.
118
Taking E, = Iw(Q = W,) we obtain the natural ordering on the space of linear
functionals on E. It is defined by the cone K#.
We note further here that the definitions of subadditivity, sublinearity and
convexity used in Section 2.2 for real functionals carry over word for word to
mappings of any linear space into any ordered linear space.
Of special interest among ordered linear spaces are lattices, in which (by
definition) for each pair of vectors x, y there exists a supremum sup& y), their
least upper bound 3:
sup(x, y) = min(z: z > x, z > y}.
We note immediately
that14
(2 > 0).
i3The existence of an infmum inf(x, y), the greatest lower bound (also part of the definition of a
lattice), is automatically guaranteed here: inf(x, y) = -sup(-x,
-y).
i41n particular the relation sup(x, y) + inf(x, y) = x + y follows from this.
isIn fact these words are often omitted.
l6 This follows from the subadditivity of the mappings x I+ x+, x H x-. We note also that lax1 =
Ial (xl for all a E R, x E E.
6 3. Linear Topology
The sufliciency is trivial. For the proof of the necessity we put g(x) =
~UPo,,,J-(Y) (x > 0). This functional is additive and positive-homogeneous
on the cone K. It is uniquely extended by linearity to K - K and then to the
whole of E. By construction, g(x) 3 0 and g(x) > f(x) (x > 0). It remains to put
h(x) = g(x) - f(x)- 0
The constructed functional g is sup(f, 0). At the same time we have also proved
Theorem 2. The space E of locally bounded linear functionals
E is a lattice.
on a vector lattice
120
w= u LYV
IWe
is a balanced neighbourhood of zero contained in U.
In any LTS all open sets are algebraically open.
We note further that any neighbourhood ofzero in an LTS is an absorbent set.
We give several simple example of LTSs.
Exainple
lim d(ex, 0) = 0
E-0
(15)
then E, equipped with this metric, is an LTS. The function v(x) = d(x, 0) is a
pseudonorm, i.e. 1) V(X) > 0 (x # 0), v(0) = 0,2) v(ax) < V(X)(Ial < l), 3) v(x + y) <
v(x) + v(y), and it is affinely continuous at zero, i.e. continuous in each direction3:
lim e.+OV(EX) = v(0) (=O). Conversely, any pseudonorm which is affinely continuous at zero determines an invariant metric d(x, y) = v(x - y) which satisfies the
conditions (15), i.e. finally, it generates a linear topology. In this construction are
included not only normed spaces but also certain others of detinite interest.
Example 1. In the space (S, p) of measurable functions on a set S with finite
measure ,u the pseudonorm
vb) =
Ids)l
dp
s 1 + I4al
.e.such that d(x + z, y + z) = d(x, y) (in what follows it is simply called invariant).
3 Continuity in each direction at a given point is clearly a concept which does not require a topology
on E.
0 3. Linear Topology
121
Example 2. The space LP(S, p) (0 < p < 1)4 is defined by the condition
and it is topologised
by this pseudonorm.
Birkhoff-Kakutani
Theorem. If an LTS satisfies the first axiom of countability
then it is metrizable and moreover the metric which defines its topology can be
chosen to be invariant and to satisfy conditions (15).
Remark. The authors mentioned in fact resolved the problem of metrizability
for
topological groups. It was shown additionally that on any (Hausdoti) topological group the bounded uniformly continuous functions separate the points.
Hence it follows that the underlying space of a topological group is completely
regular. In particular any LTS is completely regular. The weaker property of
regularity is obtained from simple considerations: if G is a topological group,
X c G is a closed set, x # X and U is a symmetric neighbourhood of zero such
that U + U is contained in (G\X) - x, then the neighbourhood x + U of the
point x does not intersect the neighbourhood X + U of the set X.
From the Birkhoff-Kakutani
theorem follows
equivalent
122
Remark 2. In the class of all LTSs we can define the property of completeness
by requiring the convergence of every fundamental net (xi) (i.e. a net (xi) such
that limi,k(x, - x,J = 0). For metrizable LTSs this definition reduces to the
previous one.
Remark 3. We can complete any LTS in the standard way. Moreover the
completion of a metrizable LTS will be a Frechet space, the completion of a
normed space will be a Banach space and the completion of a pre-Hilbert space
will be a Hilbert space.
Frtchet spaces have the important property of barrelledness. A barrelled space
is an LTS in which each barrel contains a neighbourhood of zero, where barrel
means an absorbent absolutely convex closed set. The barrelledness of a FrCchet
space follows from Baire theorem that a complete metric space is of the second
category, i.e. it cannot be the countable union of nowhere dense sets. In fact if
an LTS E is of second category then it is barrelled. For the proof we consider an
arbitrary barrel T c E. Since
E=&T
PI=1
and all the nT are closed, for some m the set mT is not nowhere dense, i.e. it has
an interior point; then T also has this property. But if x E int T, then -x E int T
and consequently 0 E int T, i.e. a neighbourhood of zero is contained in T.
We can characterise barrelledness in terms of seminorms. In any LTS continuity of a seminorm p is equivalent to its continuity at zero and for this in turn it
is necessary and sufficient that the p-unit ball DP should contain a neighbourhood
of zero*. For an LTS to be barrelled it is necessary and sufficient that each lower
his also applies to any subadditive functional 4 with q(0) = 0.
sThe last criterion is correct not only for seminorms but also for any non-negative sublinear
functionals. This can also be expressed in the form of a boundedness requirement on some neighbourhood of zero.
123
3. Linear Topology
(i = 0, 1, 2, . . .),
and if the latter converges to (r, then we shall say that the initial series is summable
A. The question arises of the regularity of such
methods of summation. It turns out that for regularity it is necessary and
sufficient that the following conditions be satisfied:
to 0 by means of the matrix
1) lim 1, = 0;
i+cc
2) lim
i+m
i
k=O
il, = 1;
i
k=O
We. shall prove only the necessity of 3). Let us consider the Banach space c of
124
convergent sequences s = (s,); with norm llsll = sup. Is,/. Under the condition,
the sequence of continuous seminorms 1Oi(S)l (i = 0, 1,2, . . .) converges pointwise
and consequently it is pointwise bounded in c. By the Banach-Steinhaus theorem
(C = constant; i = 0, 1, 2, . . .).
It just remains for us to put sk = sgn 1, (0 < k < i) for arbitrary
i.
uk
(0
dp
<
<
4,
where (a,); is a complete Markov system. If the process converges for all f E C(X)
then, by the Banach-Steinhaus theorem,
(C = constant; n = 0, 1, 2, . . .).
1; Pnjf(%jIl G cllf II
Since for each n there exists in C(X) a function f satisfying the conditions
f(xnj) = sgn pnj (0 < j < n), llfll < 1, it follows that
It is easy to show that this necessary condition for convergence is also sufficient.
Returning to general problems, let us assume that a total family {pi}isr of
seminorms is given on the linear space E: thus for any x # 0 there exists i such
that pi(X) # 0. Then we can introduce on E a linear topology with base of
neighbourhoods of zero
Llit,i2,...,im;c
X: max
pi,(X)
< &
(E > 0)9.
lSk<m
1
P
0 3. Linear Topology
I/il,i2,...,imie
=
I
(E > 0)
1 <k<m
on E.
1. The topology
Example 2. Let S be any Hausdorff topological space and let cpE C(S). For
each compact set Q c S we put po((p) = maxo 1~1. The topology of uniform
convergence on compact sets is defined by means of this family of seminorms on
C(S). In this topology C(S) is complete whenever S is locally compact.
We can define similarly on C(S) a topology of uniform convergence on any
family of compact sets {Qi}ie, which covers all of S. The topology of pointwise
convergence is included in this scheme.
The topology of uniform convergence on any set S can be defined in the space
B(S) by means of the norm I(cpII = sup, Icp1,with respect to which B(S) is a Banach
space.
Theorem.
seminorms.
is defined by a certain
total family
of
Thus, for example, if S is o-compact (i.e. it is the union of countably many compact
sets) then C(S) is a locally convex Frechet space (but not a Banach space if S is
not compact).
Let us pass to the consideration of subspaces in an LTS E. They are all LTSs
under the induced topology (LCSs if E is of this type). Particularly important are
closed subspaces. For example, in the case of Frechet (or Banach) spaces only
closed subspaces are Frichet (or Banach) spaces, since subspaces which are not
closed are incomplete. Any closed subspace of a complete LTS is complete.
If L is a closed subspace in any LTS E then the factor space E/L with its
standard topology is an LTS. If in addition E is an LCS then E/L is also an LCS
and if E is barrelled then E/L is also barrelled. If E is metrizable and d is an
This is the weakest amongst those for which all functionalsfi are continuous. Any linear functional
on E which is continuous under this topology is a linear combination of the functionals A (i E I).
r1 This is the strongest topology under which the natural homomorphism j: E + E/L is continuous,
i.e. the open subsets of E/L are precisely those sets whose preimages are open in E. We note that in
addition the homomorphism j is open. If the closedness of L is dropped then E/L fails to be Hausdorff.
126
invariant
metric
(16)
X.Y
127
3
\/
WHO
in which j is the natural homomorphism,
that f is continuous. It just remains to
take account of the uniqueness of the linear topology on K.
Continuity of a linear functional on an LTS is equivalent to its continuity at
zero and this in turn is equivalent to the boundedness of the functional on
some neighbourhood
of zero. In a normed space we can choose the open unit
of zero. We also consider the
ball D = (x: llxll < l} as such a neighbourhood
closed unit ball fi = {x: llxll < l} and the unit sphere S = o\D = (xi llxll = l}.
Boundedness of a linear functional S on any one of these three sets D, D, S implies
its boundedness on the other two and thereby is equivalent to its continuity. But
if f is bounded on S (and only in this case) then it is subordinate to the basic
norm II * II, and moreover its II * II-norm (subsequently called simply the norm of
f) is equal to
llfll = ,w&
x IfbN*
This least upper bound is not altered on replacing the unit sphere by either the
open or the closed unit ball.
In view of what has been stated, linear functionals on a normed space, which
are subordinate to the basic norm, are said to be bounded but in fact this property
is equivalent to continuity. As a consequence of a corollary of the Hahn-Banach
theorem the basic norm can itself be represented in the form
128
Example.
f(x) =
linear functional:
If the sequence (la,l); does not attain its least upper bound then If(x)1 also does
not attain its least upper bound on the unit sphere llxll = 1.
In any LCS the continuity of a linear functional is equivalent to its subordination to some continuous seminorm and is also equivalent to its subordination
to some finite collection of basic seminorms.
For LCSs the Hahn-Banach theorem guarantees that it is possible to extend
any continuous linear functional defined on a subspace to a continuous linear
functional on the whole space. In the case of a normed space there is an extension
which preserves the norm.
The geometric form of the Hahn-Banach theorem in any real LTS guarantees
that if an aftine manifold A does not intersect the interior of a convex set X with
non-empty interior then it is contained in a closed hyperplane which does not
intersect int X. In fact int X is a non-empty convex, open, and consequently
algebraically open, set. We take a hyperplane H 3 A which does not intersect
int X. It is closed since it is not everywhere dense.
The topological variant of the theorem on the separating hyperplane is established similarly: for any non-empty open convex set X and any non-empty convex
set Y which does not intersect it there is a closed separating hyperplane.
A set X in an LTS is called a convex body if it is convex, closed and has
non-empty interior. At each boundary point of a convex body X # E there exists
a support hyperplane; all such hyperplanes are closed and X is the intersection
of all the closed half-spaces of the form f(x) < c1which contain it, where the pairs
f, a correspond to all possible support hyperplanes.
In an LCS there is a further variant of the theorem on the separating hyperplane: for any non-empty closed convex set X and any non-empty compact
convex set Q which is disjoint from X there exists a closed separating hyperplane.
This situation is reduced to the previous one by constructing disjoint convex
neighbourhoods for X and Q. The hyperplane which we want turns out to be
even strictly separating:
0 3. Linear
Topology
129
130
is also normed (as was mentioned above) and even a Banach space (as is easily
verified). We can now consider the second conjugate E** and the canonical
homomorphism v: E + E ** defined by the formula
w
u-1 = f(x)
(XE E,feE*).
Since I(vx)(f)l G llfll llxll and equality is attained at a support functional f for
x, then IIvx II = [[XII, i.e. v is an isometry. Thus we can always consider a normed
space E as a subspace of the Banach space E** and identify the completion of
the space E with its closure E c E ** (E is a Banach space if and only if it is closed
in E**).
A Banach space E is said to be refexive14 if E = E** (i.e. if the canonical
isometry v is surjective). In contrast to the algebraic situation there exist inlinitedimensional reflexive Banach spaces (for example, all Hilbert spaces are of this
type (see below)). If E is finite-dimensional
then it is reflexive since in this case
all linear functionals are continuous.
Reflexivity of the space E* is equivalent to the reflexivity of the space E.
James (1964) established the following remarkable criterion: for the reflexivity
of a Banach space E it is necessary and sufficient that each continuous linear
functional on E should attain its norm on the unit sphere Is. It is appropriate to
note here that for any Banach space the set of continuous linear functionals which
attain their norms is dense in E* (Bishop-Phelps, 1961).
James (1951) constructed an example of a Banach space E for which
dim(E**/E)
= 1, and moreover E** is isometric to E (but of course not
canonically).
A Banach space is said to be uniformly convex if whenever 0 < E < 2 there
exists 6 > 0 such that from the conditions ~/XII = 11yJ[ = 1, [Ix - yll > E there
follows 11(x+ y)/2II < 1 - 6. For example, the spaces Lp(S, ,u) for 1 < p < COare
of this type (Clarkson, 1936). Any Hilbert space is uniformly convex16. All
uniformly convex spaces are reflexive (D.P. Mil an, 1938). This classical result
follows easily from James criterion.
A Banach space is said to be uniformly smooth if for any E > 0 there exists
6 > 0 such that from the conditions llxll = 1, llyll < 6 there follows [Ix + yll +
[Ix - yll < 2 + EII yll. Uniform smoothness (convexity) of the space E is equivalent to uniform convexity (smoothness) of the space E* (V.L. Shmul an, 1940).
Thus all uniformly smoooth Banach spaces are reflexive.
All spaces Lp(S, p) (1 < p < cc) are uniformly smooth. Any Hilbert space is
uniformly smooth.
The unit sphere of a Banach space is said to be smooth at the point x if there
is a unique support functionalf, at this point. A Banach space is said to be smooth
I4 More precisely, topologically
reflexive.
15The necessity is obvious: for any f~ E* there exists a support functional < E E** (r(f) = jlfll,
llrll = 1) and, if E** = E, then r = x E E, T(f) = f(x), llrll = Ilxll.
16As a consequence of the parallelogram
identity:
I/x + yJIz + /Ix - yllz = 2( llxllz + Ijyllz). Inckhtally, this identity for a Banach space implies that it is a Hilbert space (Jordan-von Neumann, 1935).
5 3. Linear Topology
131
if its unit sphere is smooth at each point. In this case there is defined the support
mapping x H f, from the unit sphere of the space E into the unit sphere of the
space E*.
Smoothness of the unit sphere at a point x is equivalent to differentiability
the norm at this point in the sense of GAteaux, i.e. the existence of the limit
of
where o is a uniquely defined real or complex measure. Its proof makes use of
the extension of a functional to B(S) by means of the Hahn-Banach theorem,
which allows the possibility of applying the functional to characteristic functions
of sets8.
ere we have to extend the concept of measure to allow any real or complex expression: w =
wr - o, + q - tw,, where the y (1 < k < 4) are measures in the restricted sense, i.e. they are
non-negative. Summability with respect tow is defined as summability with respect to [WI = C& q.
In the situation under consideration it is possible (but not obligatory) to use the term generalised
measure (in the real case a convenient term is charge, which is associated with potential theory).
*A similar approach was noted in Section 2.6 where we were concerned with non-negative linear
functionals on C(S). Here it is appropriate to point out that non-negativity implies continuity.
132
then it can be proved that IIf II = ~~o~~.The space M(S) of measures on S (with the
properties described above) turns out to be a Banach space which is isometrically
isomorphic to the space C(S)*. In what follows these two spaces are identified.
If S is a locally compact space then continuous linear functionals on C(S) have
a form similar to the previous one but with measures of compact support (a
measure is said to have compact support if it is equal to zero on sets which are
disjoint from some compact set).
In the space Lp(S, p) (1 G p < co) continuous linear functionals are identified
with the elements of the space Lq(S, p), where q is related to p by Young relation
p-l + q-1 = 1:
f(cp) =
ss
cpe 4
of F. Riesz. If E is a Hilbert
f(x) = (X>Y)
of a
(x E 4 Y E E);
8 3. Linear Topology
133
complete if its linear hull is dense, i.e. its closed linear hull coincides with E. For
example any Hamel basis is such a system but a complete system can be considerably thinner.
a4
=s*
-*
cp(t)e dt,
we arrive at the entire functions of class B, with roots at the points d,. Consequently, in order that the given system of exponentials be complete, it is necessary
and sutlicient that {A,} be a uniqueness set for the interpolation problem Il/(&,) =
0 (n E Z) in the class B, The trivial situation of this type is where the set {A,>
has a limit point. Therefore we may consider that I, + co. In this direction we
obtain, for example, the following result (A.I. Khejlits, 1969): let 1, = n + 0(n),
where 0(n) = O(1); then the system (e n is complete in L2( - R, rr) if and only if
- fw)
(the prime signifies the omission of the term with n = 0; convergence of the series
is assumed).
Example 2. Let us consider a non-empty compact set S c C whose plane
Lebesgue measure is zero. The Hartogs-Rosenthal
theorem on the denseness in
19Further, any system of the form (tnk)~&, where 0 G n, < n2 < ... and ck,I n; = DC),is complete
in C[O, l] (theorem of Mirntz).
134
C(S) of the set R(S) of rational functions with poles outside of S can be established
in the following way. Suppose that R(S) is annihilated by the linear functional
on C(S) defined by the complex measure p which is concentrated on S. Then
pj=J-$+o
(z$S).
We take any rectangle X for which p(aX) = 0 and 1,(8X n S) = 0 (A, is linear
Lebesgue measure). Then
X = 5 tiei.
i=O
For example, in lp (1 < p < co) the system (6i)~ is a Schauder basis. Topological
bases in C[O, l] and Lp(O, 1) (1 < p < co) were described by Schauder in the
1920s. Later on much effort was expended on the construction of Schauder bases
in specific spaces2 and also in the proof of the existence of such a basis in any
separable Banach space; however the last conjecture was disproved by Enflo
(1973). Soon after this counterexamples were constructed in subspaces of the
spaces Lp (Davy (p > 2), Shankovskij (p < 2)).
Clearly, any Schauder basis is a complete linearly independent system22. The
coefficients 5i in the expansion of a vector x are called its coordinates (just as in
the algebraic situation). They are clearly linear functionals of x. The image E^of
the homomorphism
x H (&(x))go mapping E into the space of sequences is
called the coordinate space of the given basis.
Example. Let E be a separable Hilbert space and let (e&z be a complete
orthonormal system (it can be obtained by the orthogonalisation
process from
any countable dense set X c E). This orthonormal basis is a Schauder basis in E.
The unique expansion of a vector is its Fourier series
X =
(X,
e&i,
i=O
the coordinate
between E and l2 is
r For example, the question of the existence of a Schauder basis in the space of functions which are
analytic on the unit disk and continuous up to the boundary remained open for a long time. A positive
solution was obtained by S.V. Bochkarev (1974).
*i It can be shown that in the class of Frechet spaces a necessary condition for the existence of a
Schauder basis is that E* be total. Thus, for example, there is no Schauder basis in Y(0, 1) for
O<p<l.
**It also has the stronger property of w-linear independence: c;=,, <,ei = 0 + & = 0 (i = 0, 1,2, . . .).
Q 3. Linear
Topology
135
equality
All this is proved in exactly the same way as for classical Fourier series in
L2( - R, n). As a corollary we obtain the result that all separable Hilbert spaces
are isometrically isomorphic to one another.
In the example considered the coordinate functionals are continuous and the
situation is exactly the same for any Schauder basis (eJ$ in a Banach space E
(see Section 3.8). Moreover the systems (ei); c E and (C)g c E* are biorthogonal: <i(ek) = Si,. Therefore the system (ei): is minimal, i.e. ek 4 Lin{ei}iZk
(k = 0, 1,2, . . .). Conversely for any minimal system of vectors there exists a
biorthogonal system of continuous linear functionals. We note that not every
complete minimal system (even one with a total biorthogonal
system) is a
Schauder basis. In contrast to the Schauder basis, such a system exists in any
separable Banach space (A.I. Markushevich, 1943).
A topological basis (ei): is said to be unconditional if all convergent series
I;,, tiei converge unconditionally,
i.e. their convergence and sum do not depend on the order of the terms. For example, an orthonormal basis in a separable
Hilbert space H is such a basis (but there also exists in H a conditional topological
basis (K.I. Babenko, 1947)). There are no unconditional
bases23 in C[O, l]
(Karlin, 1948).
The system (eiaJ)ne z in L2( - rr, a) is an orthonormal basis for AR= n and it
remains an unconditional topological basis for I, sufficiently close to n.
Theorem of M.I. Kadets.
the condition
(e n ,,z is an unconditional
topo-
to construct
a reflexive
separable
Banach
space having
no unconditional
basis
136
(Dvoretzki-Rogers,
x = C &ei
iel
(the set of coordinates 5i different from zero is no more than countable). Completeness and linear independence are clear in this case also. The coordinate space
is now a subspace E^c @(I, K) which is invariant with respect to rearrangments
(it is symmetric) and moreover each e E E^has no more than countable support.
Example. Let E be any Hilbert space and let B = (ei)isI be a complete orthonormal system (its existence can be shown with the help of Zorn lemma).
This orthonormal basis is an unconditional
topological basis. The unique expansion of any vector x is its Fourier series
X =
C
iEI
(X9
e&i,
In particular,
(i # k).
37
5 3. Linear Topology
( XSS
>
YSSM
d(ui, Vi) < 8 + E. Then the system (Vi)iel will be [S - 2(8 + s)]-distal.
quently, dimt L < dimt M. 0
Theorem 2. Zf 8(L, M) = 0 -K 4 and the subspaces
dimt(E/L) = dimt(E/M).
Conse-
138
Chapter
2. Foundations
and Methods
d( - l uiz7
L, - d( -, ui29
z,,
llzll
6-v
2(1 + E) + r/
is a complete metric on the set of subspaces of a Banach space (I. Ts. Gokhberg
- A.S. Markus, 1959), whereas the original gap does not satisfy the triangle
inequality. Moreover they are topologically equivalent because of the inequality
1 < 8(L, M)/B(L, M) < 2.
3.4. Extreme Points of Compact Convex Sets. For any convex set X in a linear
space E a point x E X is called an extreme (or extremal) point if it does not belong
to any interval [x,, xZ] c X whose end points are both different from x. This is
equivalent to the assertion that if x = (u + u)/2 where u, u E X then u = u. The
set of extreme points of a convex set X is denoted by extr X. Its role is contained
in the following fundamental result.
Krejn-Mil an
Theorem. Let E be an LTS and suppose that E* is total. Then
any compact covex set Q c E coincides with the closed convex hull of the set extr Q.
In this connection we note that the set of extreme points of a compact convex
set need not be closed and hence it can fail to be compact.
Example. Let us consider in R3 any circle C and any interval [x1, x2] not
lying in the plane of the circle but having a common point c with it different from
x1, x2. For the compact set Q = Co(C u [x1, xJ) the set of extreme points is
(C\{C)) {x19 x2).
Nevertheless, if E is finite-dimensional
any compact convex set Q c E coincides with the convex hull of its set of extreme points. This assertion had already
been proved by Minkowski (by induction on the dimension).
In any LTS the closure of a convex set is convex, from which it is clear that
the closure of the convex hull of any set X is the smallest closed convex set
containing X. We call this the closed convex hulP of the set X. The convex hull
of a set which is actually closed may fail to be closed.
Example. The convex hull of the set of points (<, r2) (r > 0) in R2 consists of
zero and the points (5, q) for which q > r2 and < > 0. Its closure contains in
addition the semiaxis (0, q) (q > 0).
25The
closed absolutely
convex
hull is defined
similarly.
5 3. Linear Topology
139
In a finite-dimensional
space the convex hull of any compact set is closed and
consequently it is compact (its boundedness is obvious). In an infinite-dimensional space (even if it is a Hilbert space) the convex hull of a compact set may fail
to be closed. In certain infinite-dimensional
spaces there exist compact sets whose
closed convex hulls are not compact. This is essentially the effect of incompleteness of such a space. In any complete LCS the closed convex hull of a compact
set is compact. For a locally convex FrCchet space a similar assertion holds in
the weak topology: the weak 26 closure of the convex hull of a weakly compact
set is weakly compact (Krejn-Shmul an theorem).
If the closed convex hull of a compact set Q in an LTS E with total E*
is compact then all of its extreme points are contained in Q (D.P. Mil an,
1947).
We now give a proof of the Krejn-Mil an
theorem. First let us assume that
E is an LCS (without loss of generality we can also assume that it is real). We
put K = Co(extr Q). If K # Q then any point z E Q\K can be separated from K
by a closed hyperplane {x: f(x) = a} such that f(x) < a (x E K), f(z) > a. We put
B= maxXGaf(x). Clearly /I > a and so the hyperplane H = { y: f(y) = /I} does
not intersect K or, consequently, extr Q. This situation leads to a contradiction
since we are able to find an extreme point of the compact set Q in H. For this
purpose let us call a closed alline manifold A extreme if A n Q # 0 and whenever
an interior point of an interval [a, b] c Q belongs to A then the entire interval
is contained in A. For example, the hyperplane H is extreme.
Using Zorn lemma it is easy to show that there is a minimal extreme manifold
M contained in H. It is a singleton set since otherwise there exists a continuous
linear functional g which is not constant on M and then the smaller extreme
manifold {y: y E M, g(y) = r}, where y = max xGMnp g(x), is contained in M. But
any singleton extreme manifold must simply consist of an extreme point of the
compact set Q.
Now if we only know that E* is total, from what has been proved the theorem
is true for the weak topology, i.e. Q = e where e is the weak closure of the set
C = Co(extr Q). Denoting the strong closure by c we have C c c c e. But c
is compact, consequently weakly compact and so also weakly closed. Therefore
c = e, i.e. Q = c 0
It is not possible to dispense with the condition that E* be total: there exists
an LTS in which a certain non-empty compact convex set has no extreme point
(Roberts, 1977).
The condition of compactness in the Krejn-Mil an
theorem is also essential;
for example, it is not possible to replace it by closedness and boundedness.
Example. If S is a connected compact space the closed unit ball 1)cp11< 1 in the
Banach space of real continuous functions on S has no extreme points other than
the constants f 1.
26It is also the closure in the strong (original) topology since the space is locally convex.
is called the integral of x(s) with respect to the measure p. Thus, in general,
J E (E*) #, however the most interesting case is when J(f) = f(F) where X E E. If
such an x exists it is unique; it can be identified with J and we put correspondingly
y=ss44dp.
Theorem. Suppose that ,u(S) < CO and x(s) is an integrable vector-function.
If
the closed convex hull X of the set of values x(S) is compact then the integral-of
x(s) with respect to the measure ,u is a vector X E E (further, X E X if p(S) = 1).
Proof. We may assume that p(S) = 1 and that E is a real space. For any
collection {S,}; c E* we consider the continuous mapping Fx = (J(x)); from X
into R Its image is a compact convex set. Moreover (.T(f;:)); E Im F since
otherwise there exists a vector (ai); such that
dp
>I-;g(x)
a g(x(s))
!J
s&(s))
S
141
(P E L(S, PI, A = 9,
~(4 &
sA
while the function x(s) satisties the previous conditions.
In the most developed situation we have to integrate a function x(s) with values
in a Banach space E which is defined and weakly continuous on a locally compact
topological space S provided with a generalised measure o. In this situation the
norm of the function is measurable since 11x(s)11= suplf(x(s))l (f E E*, II f I[ = 1)
and ail the f@(s)) are continuous. If
~(4 =
then x(s) is integrable
XC
s S
is a vector in E; moreover llX\l < v(x). In fact for any E > 0 we can choose a
compact set Q, c S such that
Il~(~)IIW-4 < E.
s s\Q.
The set x(QE) is weakly compact. By the Krejn-Shmul an
closed convex hull is also weakly compact. Consequently
x, E
x(s) dw E E.
s Q.
Further
IIX - X,IlE*+ =
sup
f(x(s))
do < E.
142
llxll =
ss
IWII &
on the a-algebra associated with the measure p is countably additive and in this
sense it is a vector measure. For any vector measure m(A) its ouriution is the
measure
ImA
sA
IIx(s)II G.
The total variation (or norm) of a vector measure m is the quantity IIml1 = Iml (S).
A vector measure m is said to be absolutely continuous if l(rnll < co and
m(A) =
~(4 dImI,
sA
where x E L(S, Iml; E). We say that a real Banach space E has the RudonNikodjm property (RN) if every vector measure with finite norm is absolutely
continuous3. The RN property has a purely geometric characterisation which
is closely connected with the Krejn-Mil an
theorem.
r9Any summable function is the limit almost everywhere of some sequence of simple functions.
Therefore its set of values is separable (except possibly for values taken on a set of measure zero).
30The classical Radon-Nikodjm
theorem asserts that if there are two measures p, v (on a common
o-algebra) and p(A) = 0 + v(A) = 0 then
v(A) =
sA
P dp,
deriuatiue).
143
s dp
IS
belongs to Q and the Krejn-Mil an
theorem implies that any point X E Q is the
centre of mass of some normalised measure p5 on S. A significant refinement of
this result is due to Choquet (1956): if the compact set Q is metrizable then pz
(X E Q) can be chosen to be concentrated33 on extr Q and not on its closure S (in
this situation extr Q is a G, set). Such a pz is called a Choquet measure for the
point X. It can be regarded as a solution of a singular abstract moment problem.
This way of looking at it embraces in a single form all those classical moment
problems in which the required measure is finite (the general case needs an
additional limiting process). For example, in the power moment problem on the
whole axis we have a convex set Q of positive (in the sense of the corresponding
Hankel matrix) sequences (m& (m. = l), its extreme points are {t"}$
(t is a real
parameter) and the integral representation
al
m, =
tk da(t)
(k=0,1,2,...)
s0
corresponds to Choquet theorem.
The converse is false. However, as we see, strongly exposed points play the same role as extreme
points in a series of cases. We note here that any weakly compact convex set in a Banach space is
the closed convex hull of its set of strongly exposed points (Amir-Lindenstrauss, 1968). In a uniformly
convex Banach space every point of the unit sphere is a strongly exposed point of the closed unit ball.
32Thus, for example, if S is a connected compact space with card S > 1 then C(S) is not an RN-space.
33We say that a measure v on a set X is concentrated on a subset M c X if v(X\M) = 0. If X is a
topological space the smallest closed subset on which the measure v is concentrated is called its
support and is denoted by supp v.
144
Chapter
2. Foundations
and Methods
(17)
f&(c)) = kzotk(f
)(c - cdk
(fEE*),
where
<k(f)
= &
f(W)
s I<-[ol=p(C-
4
ro)k+l
Since x(c) is weakly continuous on the circle I[ - [,, 1 = p, we have tk( f) = f (x,),
wherex,EE(k=0,1,2
,... ).
Moreover /lxkII < M(p)pek, where M(p) = sup16-501=P[lx(c) 11.Consequently,
the power series (17) converges absolutely on the disk Ic - co 1 < p for any p < 6,
i.e. finally, on the disk I[ - co1 < 6. Its sum s(c) - x(c) since f(s(c)) = f(x([)) for
all f E E*. 0
The further development of the theory of analytic vector-functions with values
in a Banach space presents no problem. Concerning this we note only the
Cauchy-Hadamard
formula for the radius of convergence of a power series with
vector coefficients ck (k = 0, 1,2, . . .):
The sum of the power series is analytic inside the disk of radius r.
3.6. w*-Topologies.
Let us consider an arbitrary linear space E and its algebraic conjugate space E. Since E c @(E, K) the topology of pointwise convergence is defined on ES. It is called the w*-topology
on E#. The LCS EX is
w*-closed in @(E, K) and consequently it is w*-complete.
If E is an LTS, its conjugate space E* is contained in E and so the w*topology is also defined on E*. This is the weakest topology in which all the linear
34A vector-function
which
is analytic
on the whole
145
functionals of the form R(f) = f(x) (X E E, f E E*) are continuous (these are the
canonical images of the elements x E E in the space (I?*)#). Each w*-continuous
linear functional E* has the stated form.
A fundamental result of a general character concerning the w*-topology is the
Banach-Alaoglu Theorem. Let E be an LTS and U c E a neighbourhood of
zero. Then the set U consisting of those linear functionals on E which satisfy the
condition35 supXScr If(x)1 G 1 (and so are continuous) is w*-compact.
Proof We put A = {I: 111< l> and consider the set of functions @(U, A) with
its product topology. This is compact by Tikhonov
theorem. The mapping
Rf = f Iu of the set U c E* provided with the w*-topology into @(U, A) is a
homeomorphism
between U and RU But, as is easily verified, RU is closed.
Consequently it is compact. 0
theorem follows
146
isometric
2 of the Banach-Alaoglu
theorem.
Sufficiency. The image of the closed ball in E** is w*-compact and consequently w*-closed. On the other hand it is w*-dense in the closed ball of the space
E**. These balls therefore coincide and then we also have E = E**. 0
Corollary. A necessary and sufficient condition for each closed bounded convex
set in a Banach space E to be weakly compact is that E be reflexive.
8 3. Linear Topology
4x0,
147
Y)).3
YEY
If E is a reflexive Banach space then for any non-empty closed convex set Y c E
and any point x0 E E a best approximation
y, E Yexists. In fact it is sufficient to
look for y, in the intersection Yeof the set Y with the ball d(x,, y) < d(x,, Y) + 1.
Now Ye is weakly compact and the functional d(x,, y) = [lx0 - yll is weakly
lower semicontinuous. Therefore d(x,, y) attains its inlimum on Y,. The condition of reflexivity is necessary because of James criterion.
We note in passing that the theorem on the uniqueness of the best approximation for all x0 E E and all convex Y c E holds if and only if E is strictly convex.
Thus in a uniformly convex Banach space E any non-empty closed convex set
Y c E (in particular, any closed subspace) is a Chebysheu set in the sense that a
best approximation
y, E Y exists and is unique for each point x0 E E. N.V.
Efimov and S.B. Stechkin have investigated the general problem of the convexity
of Chebyshev sets (1961).
Example. Let us consider the problem of the best approximation
in a Hilbert
space E38. Let L c E be a closed subspace, let x0 E E and let y, be the best
approximation
in L for x,; put z0 = x0 - y,. Then for any y E L the function
CPM = lb0 + ~~11~= llzol12 + 22 reh
Y) + ~zll~l12
(r E w
- llYol12).
3.7. Theory of Duality. Let E and F be two linear spaces and let (x, y)
(x E E, y E F) be a bilinear functional such that
Vx # 0 3y with (x, y) # 0 and Vy # 0 3x with (x, y) # 0.
Then we say that there is a duality between E and F or that (E, F) is a dual pair
of spaces.
Example 1. Let E be any linear space, let F = E
put (x, f) = f(x). Clearly this is a dual pair.
Example 2. If E is an LTS
(E, E*) is a dual pair with (x,
If (E, F) is a dual pair there
is total on E. Thus F defines
148
by o(E, F). The space E*, the conjugate of E in this topology, coincides with F.
Now we can consider the w*-topology on F = E*. Clearly it coincides with the
topology o(F, E) for the dual pair (F, E) and consequently F* = E. In this way
a complete symmetry between the linear topology space? E and E* is obtained.
This symmetry also appears in the following property of reciprocity of polars.
Theorem. Let (E, F) be a dual pair, X c E (X # 0) and Y = X0, the polar of
X in F. Then Y, the polar of Y in E, coincides with the a(E, F)-closed absolutely
convex hull 2 of the set X. In particular, if X is absolutely convex and o(E, F)closed then X = Y (i.e. X0 = X).
Proof. Since 1(x, y) 1 < 1 (x E X, y E Y) we have X c Y and hence 8 c Y.
Let x0 I$ 2. By the theorem on the separating hyperplane, there exists y E F such
that 1(x, y)( < 1 (x E 2) and 1(x0, y)l > 1. But then y E Y and consequently
x0 4 Y. It follows that 2 = Y. IJ
Corollary 1. Let (E, F) be a dual pair, L a subspace of E and N = L, the
annihilator of L in F. Then the annihilator NL of N in E coincides with the
a(E, F)-closure of L. In particular, if L is a(E, F)-closed then L = N1, i.e.
(Ll), = L.
If (E, F) is a dual pair we say about a locally convex topology on E for which
E* = F that it is consistent with the given duality. The weakest among such
topologies is a(E, F). It turns out that the class of topologies which are consistent
with the given duality can be described and, moreover, among these topologies
there is a finest.
Mackey-Arens
Theorem. Let (E, F) be a dual pair. In order that a locally convex
topology on E be consistent with the given duality it is necessary and sufficient that
it be a topology of uniform convergence on some family of absolutely convex subsets
of the space F which are compact in the topology a(F, E).
Necessity. The polar U of any neighbourhood
of zero U c E is compact in
the topology o(F, E) by the Banach-Alaoglu
theorem, since by assumption
F = E* and o(F, E) is the w*-topology. The initial topology T on E coincides
391n this connection, if E is an LTS to start with, its initial topology is to be replaced by the weak
topology; moreover E* is preserved and it is provided with the w*-topology.
f
0 3. Linear
Topology
149
with the topology of uniform convergence on the family {UO} where U runs
through the family T of absolutely convex (open) r-neighbourhoods
of zero in
E. In fact, if U E T and
XXEE,
sup I(x,y)I < 1 ,
YEW
1
I
then U c I$, since 1(x, y) I < 1 (x E U, y E U), U is compact and the function
cp,(y) = 1(x, y)l on U is continuous and, consequently, attains its maximum.
On the other hand if WE T we can find U E T such that its a(E, E*)-closure 0
is contained in W. Thus 0 = Uoo I> VU and so VU c W.
l,=
This topology is called the Mackey topology and is denoted by z(E, F). An LCS
E whose topology coincides with 7(E, E*) is called a Mackey space. Any barrelled
(and also any metrizable)
Remark. If (E, F) is a dual pair, the closure of any convex set X c E is the
same for all topologies which are consistent with the given duality since, as a
result of the theorem on the separating hyperplane, the operation of closure on
the class of convex sets is determined by the set of continuous linear functionals.
Taking this as his starting point, D.A. Rajkov (1962) constructed a purely
algebraic theory of duality and obtained an algebraic variant of the KrejnMil an theorem.
Many important questions in the theory of duality for LTSs are connected
with the consideration of bounded sets.
Theorem of Mackey. Zf (E, F) is a dual pair the same sets are bounded for all
topologies on E which are consistent with the given duality.
Proof. By the Mackey-Arens theorem it is sufficient to show that each o(E, F)bounded set X c E is z(E, F)-bounded. Moreover we need consider only the
case where X is countable: X = {x.}?. If it is not r(E, F)-bounded there exists
{a,,}? c R, such that lim,,, a, = 0 but {a,,x y does not tend to zero in the
topology z(E, F). Then we can find E > 0 and an absolutely convex o(F, E)compact set Q c F such that maxy.o I( a,x,, y)l > Efor an infinite set of suffices
Y (we can suppose that this holds for all n). There is therefore a sequence
150
~Y,>Y
= Qsuchthat<x,,Y.> + cc. At the same time, sup 1(x,, y) ( < co for each
y E F because X is o(E, F)-bounded. This situation contradicts the BanachSteinhaus theorem. 0
Corollary.
ness.
Q 3. Linear
Topology
151
In the proof of this theorem we use the following criterion for a@*, E)continuity of a linear functional on E* which goes back to Banach.
Theorem. Let E be a complete LCS. Zf a linear functional on E* is w*-continuous on the polars of the neighbourhoods of zero in E, then it is w*-continuous
(and
consequently it is generated by an element x E E).
Proof. Let us consider the linear space L of all linear functionals on E* which
satisfy the condition of the theorem. Clearly L I> E and (L, E*) is a dual pair.
We can introduce on L the topology p of uniform convergence on the polars of
the neighbourhoods of zero in E since the functionals in L are bounded on these
sets. This locally convex topology induces the initial topology on E and, since E
is complete, it is therefore closed in L.
By the Banach-Alaoglu theorem the polars of the neighbourhoods of zero in
E are compact in the topology o(E*, E). On each such set P the topology a(E*, L)
is on the one hand stronger than a(E*, E), since L 3 E, while on the other hand
it is weaker, since it is, by construction, the weakest topology under which all
the glP (g E L) are continuous. Consequently p is a topology of uniform convergence on a certain family of absolutely convex o(E*, L)-compact sets. By the
Mackey-Arens theorem it is consistent with the duality of the pair (L, E*), i.e.
L* = E*. But then if f E L* and f IE = 0 we have f = 0. Thus L = E. 0
of C [0, l] is defined
lIdIn = max
ObkSn
43 For example,
the Banach
space C(X),
by the countable
max
OSx+l
where
Iv l
family
of norms
(n=0,1,2
X is an infinite
compact
,... ).
set, is such a space.
152
neighbourhood
of zero U c E. It suffices to establish
the absence of this for some E > 0 we can construct
and {f y c U such that If,(z)1 > E for all m, lf,(z Ifm(xn)l < +E for m > n.
This leads to a contradiction on taking a weak limit
w*-limit point f of {f ?. 0
continuity at zero. In
inductively ix.}? c X
x,)1 c 3.s for m < n and
point x of {x.}? and a
For the proof we associate with an arbitrary LCS E the space E^ of linear
functionals f on E* for which all the sets f- O) n U are w*-closed (U runs
through the neighbourhoods of zero in E) or, equivalently, whose restrictions
f IuO are w*-continuous. We embed E in ,!? in the canonical way and introduce
on i the topology of uniform convergence on the polars U The preceding
theorem follows from the following assertion.
Lemma. For any LCS E the space .@is its completion, i.e. 1) E is complete, 2)
the topology of E is induced by the topology of i?, 3) E is dense in E^.
Proof The completeness of E^follows from the w*-compactness of each polar
U and the completeness of the space of continuous functions C(U . The
coincidence of the induced and the initial topologies on E is trivial. Finally the
topology on E^ is consistent for the dual pair (fi, E*) by the Mackey-Arens
theorem, i.e. @)* = E*. Moreover if E is not dense in E^there exists f E (I?)* such
that f # 0, f IE = 0, which is a contradiction since f E E*. 0
Remark. The criterion for completeness of an LCS can be reformulated as
follows: E is complete o each linear functional on E* which is w*-continuous
on the polars of the neighbourhoods of zero is w*-continuous.
An LCS E is said to be fully complete (or to be a Ptdk space) if each subspace
of E*, whose intersections with the polars of the neighbourhoods of zero in E
153
are w*-closed, is w*-closed. It follows from the previous result that a fully
complete space is complete. The converse is false, however the following holds.
Banach-Krejn-Shmul an
Theorem.
space is fully
complete.
154
UCT
451f E is a normed space, the norm on U(E) satisfies the usual properties: I/h,h, 11< llhl 11Ilh,II,
11111
= 1. Thus the continuous linear operators on E form a Banach algebra f!(E) which is noncommutative if dim E > 1. We note here that in any Banach algebra the operation of multiplication
is continuous.
461n the case of E* it coincides with the w*-topology.
set X in an LTS E is said to be precompoct if it is relatively compact in the completion of the
space E (a set is said to be relatively compact if its closure is compact). The property of precompactness
is weaker than relative compactness (in a complete space they are clearly equivalent), but it is
convenient to work with it in view of Hausdorff
criterion: X is precompact o for any neighbourhood
of zero U there exists a finite U-net, i.e. a covering of the set X by neighbourhoods of the form x, + U
(1 < i < m).
It is also called the principle of uniform boundedness.
4QIf E, is a normed space uniform boundedness on some neighbourhood of zero is equivalent to
uniform continuity. If both spaces E,, E, are normed then uniform continuity of a family F is
equivalent to boundedness in norm: sup,,, jlhll < co.
5 3. Linear
155
Topology
For any LTSs E,, E, and any h E Homc(E,, E,) the conjugate5 homomorphism h*: Ef + ET is defined by the formula @*f)(x) = f&x) (x E E,,fe
Ef).
It is important to know for what topologies on the dual space h* will be
continuous.
Lemma. Let ET and E,* be provided respectively
where hS2, c 52,. Then h* is continuous.
Ilh*II = IIW
We note in passing the duality relations
of the homomorphism
h E Homc(E,, E,)
annihilators considered below are relative
necessarily dual pairs).
With no restrictions on the LTSs E, and
Ker h* = (Im h)
and the inclusions
Im h* c (Ker h),
Im h c (Ker h*)l
on51 ET). If Ez is total we have
Im h = (Ker h*)l.
ore
precisely, the topological
conjugate.
Clearly h* = h#),; and moreover
note the following
formulae:
1) (hi + h,)* = h: + hz; 2) (ah)* = ah*; 3) (hg)*
h E Homc(E,,
E,) has a continuous
inverse then h* is invertible
and (h*)- =
not be continuous.
If E, is a reflexive
Banach space the w*-topology
on ET coincides with the
normed space ET and the weak closure of any subspace of ET coincides with
Im h* c ET. We also
= g*h*; 4) l* = 1. If
(h-l)* but (h*)- may
weak topology
of the
its strong closure.
156
Chapter
2. Foundations
and Methods
Lemma 1. Let E,, E, be LCSs and let h E Hom(E,, E,). A necessary and
sufficient condition for closedness of h is that the space A,, of those f E Ez for which
h#f E E: be dense in Ef under the w*-topology5.
Proof. .For any LTSs E,, E, the conjugate space of E, x E, is naturally
identified with E: x Ef in accordance with the formula cp(x, y) = g(x) + f(y)
(g E E:, f E Ez). Moreover for any h E Hom(E,, E,) the annihilator of its graph
I, is the subspace Ih of functionals of the form f(y - hx) where f E A,,. Now let
Et, E, be LCSs and suppose that h is closed. If y E E,, y # 0 then, since (0, y) 4 I,,
there is a functional f E A,, such that f(y) # 0. Conversely, if for each y E E,,
y # 0 there exists a functional f, E A,, such that f,(y) # 0, then r,, = n, Ker f,,
from which the closedness of the graph I, follows immediately. 0
Inverse Homomorphism
Theorem (IHT). Let E,, E, be LCSs with E, barrelled
and E, fully complete. Zf the homomorphism g: E, -+ E, is bijective and continuous
then it is a topological isomorphism, i.e. g-l is continuous.
or
N c E* is w*-dense
5 3. Linear Topology
157
Proof. The space Ker g is closed 53, therefore we can pass to the factor space
G = EJKer g and consider the bijective homomorphism
y: G + E, associated
with g by the relationship g = yj where j: E, --* G is the canonical homomorphism. Full completeness is inherited by factor spaces and y is continuous
(because g is continuous and j is open). According to the IHT, y- is continuous,
i.e. y is open, but then g is also open. 0
As Banach had already shown, all three theorems are true for Frechet spaces
without the assumption of local convexity. Their applications are numerous and
varied (the IHT is used particularly often). The first application which we note
relates to complemented subspaces.
A closed subspace L in an LTS E is said to be complemented if it has a closed
complement s4 M. For examp 1e, a n y closed subspace of finite codimension, in
particular any closed hyperplane, is of this type. In an LCS every tinite-dimensional subspace is complemented. If E is a Hilbert space, any closed subspace
L c E is complemented, the orthogonal complement
Ll = {y: (x, y) = 0 (x E L)}
158
define the same topology. We say further that d, is weaker than d2 (or subordinate
to d,) if the corresponding topologies are in this relationship. If the inequality
d,(x, y) < Cd&, y) holds for some constant C > 0 then d, is weaker than dl.
Theorem. Zf the space E is complete under both metrics d, and d2 and also d,
is weaker than d, then the given metrics are equivalent.
Proof. The identity mapping acts continuously from the d,-topology
d,-topology. According to the IHT we have a homeomorphism.
0
into the
Two norms II.II 1 and II * IIz on a linear space E are said to be equivalent if the
corresponding metrics are equivalent. The relation of subordination is defined
correspondingly and moreover a necessary and sufficient condition for this is that
the inequality IIx II I < C IIx IIz be satisfied for some positive constant C. Thus for
equivalence of the norms it is necessary and sufficient that a llxll z < I/XII 1 <
b llxll z (x E E) where a, b are positive constants. Any two norms on a linitedimensional space are equivalent.
From the preceding theorem we obtain
Corollary. Zf a space E is a Banach space with respect to each of two norms,
one of which is subordinate to the other, then these norms are equivalent.
(x E L).
The space E is said to be super-reflexive if each space which is finitely representable in it is reflexive. Super-reflexivity of E* is equivalent to super-reflexivity of E
(James, 1972). As a corollary to Enflo theorem the complete duality between uniformly convex and uniformly smooth Banach spaces shows that the possibility
0 3. Linear
Topology
159
lIeill
IIXII.
(t E G
The Riesz bases therefore form an equivalence class with respect to the action
of the group of automorphisms of the Hilbert space.
Corollary. Any Riesz basis (ei)$ is unconditional
sense that infi llei 11> 0, sup, (Iei II < CO.
almost normalised
hus
9 is an isomorphism
of the Banach
spaces B, E.
in the
Schauder
160
(18)
Conversely, it follows from (18) that (ei)g is a Schauder basis since if x E Lin(e,}$
there exists a number v = v(x) such that S,x = x for all n > v(x), i.e. (S,,x);
converges to x on a dense subspace and therefore also on the whole of E (because
of (18)).
We now mention two examples which illustrate the effectiveness of the IHT
in concrete analysis. The first of these is due to Banach (1931).
Example 1. Let us consider the system of linear equations
(1%
Any sequence x = (&.)y such that the series (19) converges with sum equal to vi
for each i = 1,2, 3, . . . is called a solution of the system. It turns out that, if the
system (19) has a unique solution for all (rli)~, then the rows of the matrix (4~)
have finite support, i.e. (19) is a Toeplitz system.
For the proof we consider the space R of all x = (&,)y for which all the series
(19) converge and introduce on R a locally convex topology by means of the
seminorms
Pitx)
suP $1
I
(i = 1, 2, 3, . . .).
%ktk
I
(qj)?,
nk
<
ao).
s* cO is the space of sequences converging to zero with the sup norm. We note that co*z I
0 3. Linear Topology
161
The functionals <,(hy) (y E s) are linearly independent since this is true of &(x)
(x E R, k = 1,2,3, . . .). Consequently, the rows of the matrix (/Jr) are linearly
independent. By the theorem of Toeplitz the system of equations
has solution y = (qj)y for any (c&. But then cl, = &(hy) (k = 1,2, 3, . . .) and the
series
kzl f
converge. Because of the arbitrariness
have finite support.
(i = 1,2, 3, . . .)
Example 2. Let (A4.k be a sequence of positive numbers defining a quasianalytic class of functions on [0, 11. Let us consider the Banach space B(A4,)
of infinitely differentiable functions cp(t) (0 < t < 1) for which
1
llcpll = sup max I# (t)l < co.
n M. o<t<l
For each such function the sequence of derivatives at zero satisfies the condition
sup I(P
n M,,
O)1
<*
(W(t)
=n=O
f an;,
.
where the series converges in B(M,)
and so in particular
at the point t = 1.
162
Moreover
Setting a, = &,M,
The last example of the application of the IHT which we present here is
concerned ith the regularisation of Noetherian operators. Let E be any LTS in
which the IHT holds and let T be a Noetherian operator on E whose image Im T
is closed5 (its kernel Ker T is also closed since it is finite-dimensional).
The regularisors for T are defined by the formulae (see Section 1.3) R, = (1 - .4)FT- ,
$ = ? , where A is a projection onto Ker T, P is a projection onto Im T and
T = TIKerA. The projections A, P can be chosen to be continuous and FT- is
continuous in accordance with the IHT. Therefore both regularisors turn out to
be continuous.
A sufficient condition for the preservation of the Noetherian property, normal
resolvability and the value of the index under the change from T to T + V, where
I/ E 2(E), is that the operators 1 + R,V and 1 + VR, be invertible. An effective
formulation is obtained from the following general result: in any Banach algebra
the elements of the form e - x with I(xIJ < 1 are invertible.60 Thus in the Banach
space E it is sufficient to impose on the perturbation V the condition of smallness
of norm (Atkinson, I. Ts. Gokhberg, 1951), namely
IIW-9.
Invertibility
of the operator 1 - U where II UII -C 1 is a simple but very useful
result. In terms of the equation x = Ux + y it implies unique solvability for all
y (this also follows from the contraction mapping principle). Among its many
applications there is, for example, the theorem on stability of buses: if (ei): is a
Schauder basis in a Banach space E and (ci)g is a sequence of coordinate
functionals then each system of vectors (ni); satisfying the condition
X = 2 liei,
UX
igo
rifei
M.A. Rutman,
-
1940). In fact
4)
i=O
163
5 3. Linear Topology
Iliz %(eiI
%)I2
0 i$
Jail2
(0<8<
1).
The homomorphism
g is uniquely defined. It is called the (Frdchet) derivative of
the mapping F at the point x and correspondingly it is denoted by F(x) (further
it can be written in the traditional way: dF(x) = F(x) dx). From standard results
of elementary analysis we note the formula of finite increments in the form
1
Fx, - Fx, =
- x1) dz.
s0
The following are conditions which are sufficient for its validity: 1) [x1, x2] c G;
2) F is differentiable at all points of the interval [xi, x2]; 3) the derivative
F(x, + t(x2 - x1)) is a weakly continuous function of r (0 < z < 1). They are
clearly satisfied in a convex domain G for a differentiable mapping whose derivative is weakly continuous. If moreover q = SUP,,~ IIF(x)I) < co then the mapping satisfies the Lipschitz condition
lb,
- Fx, II G 4 lb2 - XI II
(~1,
x2
E 3.
x,
[@(x,)]- x,
(n=0,1,2,...)
for solving the equation @x = 0. The convergence of this process and of certain
of its modifications was investigated in detail by L.V. Kantorovich (1948). In all
cases, if 1) the root x is simple in the sense that the operator Q(X) has a bounded
164
inverse, 2) the second derivative W(Y) exists and is continuous in the uniform
topology on some ball D(X, 6) = (x: 11x- Xl1 < S}, then for all x0 in a sufficiently
small ball 0, the Newton process will be defined and will converge to the root X
In fact, the formula
Fx = x - [@(x)1- x
defines a differentiable
mapping F: 0, + E, :
F x) dx = [F x)]- F x)
dx)[F x)]-lFx.
x(t)
=XI-J
+sf@(x(4,4
<t<E),
dr(0
0
i.e. the required solution is a fixed point of the mapping defined by the right hand
side on the space of continuous vector-functions on [0, E]. For sufficiently small
Ethis mapping is a uniform contraction on a neighbourhood of the constant x0.
The requirement of continuity of the mapping @ is not satisfied (in the Banach
space setting) if, for example, @ is a differential operator. Therefore the scheme
which has been described is not suitable for partial differential equations. Even
for the investigation of the linear equation 1= Ax where the operator A is unbounded we require special methods which are connected with spectral theory
(see Section 4.5).
$4. Theory
of Operators
165
0 4. Theory of Operators
4.1. Compact Operators. A homomorphism
h: E, + E, of any LTSs is said
to be compact or completely continuous if there exists in E, a neighbourhood of
zero whose image is relatively cpmpact. Any compact homomorphism
is clearly
bounded and so continuous.
The property of compactness is valuable because it is the basis of a satisfactory
generalisation of Fredholm theory. Such a generalisation was effected for normed
spaces by F. Riesz (191Q for LCSs by Leray (1950) and for arbitrary LTSs by
Williamson (1954). The classical Fredholm theory is easily covered by these
generalisations since linear integral operators are compact in the corresponding
function spaces under the usual assumptions (for example, operators on C(S) (S
compact) with continuous kernel or operators on L2(X, p) with Hilbert-Schmidt
kernel are of this type). The compact homomorphisms
form in Homc(E,, E,) a
subspace Ci(E,, E,) which is invariant with respect to the natural action of
!i!(E,) x !i?(E,): h H ahb (a E i?(E,), b E !i?(E,)). Thus (E-(E, E) = CC(E) is a twosided ideal in the algebra L?(E).
If E, is complete c(E,, E,) is closed in the uniform topology. The subspace
B(E,, E,) of continuous, finite rank homomorphisms
and consequently its uniform closure g(E,, E,) are contained in 6(E,, E,). There is considerable interest
in the question of when (I;(E,, E,) and g(E,, E,) coincide, i.e. of the possibility
of uniform approximation
of an arbitrary compact homomorphism
by continuous, finite rank homomorphisms.
In fact Fredholm constructed his theory of
integral equations in this way by approximating their kernels by degenerate ones,
i.e. by kernels of the form
Let E, be a barrelled LCS with a Schauder basis (ei)g and let h be a compact
homomorphism
of any LTS E, into E,. Let us consider the partial sums
S,X= k tiei
(x=2&i).
i=O
if there is
166
Chapter
2. Foundations
and Methods
Proof. Suppose Ker T = 0. Using the compactness of Tit is easy to verify that
the homomorphism
S which is inverse to E 5 Im T is continuous
Let us
suppose that Im T # E and take x 4 T. We consider the n-dimensional subspace
L, = Lin{Tkx};-
(n > 1, Lo = 0). Since AIL, is injective and maps L, n 0 into
a compact set we have that L, n 0 is compact.
Because of the closedness of Im T and the continuity of S we can construct a
neighbourhood
of zero U such that (L, + U n Im T c TU. Now we take a
balanced neighbourhood
of zero u such that u c U n u and choose in
L, n 0 (n 2 1) any point x, & L,-, + U Clearly x, # x, (n # m).
We have x, = TY,-~ + a&, where y,-, E L,-,. Since Ty,-, E L, + t? then
TY,-~ E TU and hence y,-, E U because T is injective. Consequently the set
Under
additional
assumptions
about
E this is guaranteed
by the inverse
homomorphism
theorem.
5 4. Theory of Operators
167
L,)(n > m)
Now let us consider the root subspaces W, = Ker T (n = 1,2,3,. . .). Since
the intersections W, n 0 are compact, if W, # W,-, (n = 2,3, . . .) we can choose
for each n a point x, E (W, A u)\(W,-,
+ U). Then Ax, - Ax, $ U (n > m) in
spite of the relative compactness of {Ax,}. Consequently there exists a number v
such that W, = W,,, = . . * (W, # WV-,). The maximal root subspace W, is tinitedimensional.
The operator T maps the invariant subspace E, = Im T into itself and moreover E,+1 is its image. But Ker T n E, = 0. By the Fredhohn alternative E,+1 = E,
and then E, = E, (n > v), in particular E,, = E,. Therefore for any x E E there
exists y E E such that T = T , i.e. x - T E W,. Thus we have established
Lemma 2. The following
holds:
E = Ker T i Im T
Corollary.
The operator
T is a Fredholm operator.
ind(S + V) = ind S.
Proof. The regularisors R,, R, for S are continuous.
we
have
R,(S+
V)=l+(R,V-A),
Since the parts added to the identity operator are compact we have that R,(S + V)
and (S + V)R, are normally resolvable Fredholm operators, from which all the
assertions follow. 0
168
Chapter
2. Foundations
and Methods
One of the most important areas of application of this theorem is the theory
of singular integral equations 3. The kernel of a singular integral operator is
frequently the sum of a standard singular kernel (for example, the Cauchy kernel
([ - z)-) and a regular4 kernel which defines a compact operator. In this case
the calculation of the index is reduced to that for the standard kernel.
Compactness of integral operators with regular kernels is brought about by
the existence of finite rank approximations.
Moreover the properties of an
operator are better the more rapidly it can be approximated by finite rank
operators. This observation leads us to define the approximation numbers (anumbers) of a compact homomorphism
h: E, + E, of normed spaces as its
deviations
(n = 0, 1,2, . ..)
a,+,@) = g*; Ilh - 911
n
from the subspaces ?j, = &(E,, E,) = {g: g E 7j(E,, E,), rk g < n]. Clearly,
llhll = a,(h) 2 a,(h) 2 . . . and a,(h) + 0 if and only if h E $j(E,, E,). Putting
G,(E,, E,) = &JX,, E,) we define further the class G,(E1, E2) (0 c p < co) by
the requirement
The spaces GJE,, E,) (0 < p < co) are invariant with respect to the action of
f?(E,) x f?(E,); consequently G,(E) = G,(E, E) is a two-sided ideal of the algebra
WI.
For any h E GJE,, E2) the quantity
I(he$
i, j=O
,I,
3 In this context S.G. Mikhlin
in a Hilbert space.
4 Or
eakly singular
(1947) discovered
lclji12< 00.
the stability
ofthe
index under
compact
perturbations
44.
Theory
of
Operators
169
Then it is a Riesz basis and it is called a Bari basis (in honour of N.K. Bari who
established the result in question in 1951). The proof is based on the fact that the
operator defined on E by the equations Aei = e, - ui (i = 0, 1,2, . . .) is a HilbertSchmidt operator and so is compact. If T = 1 - A then Tei = Ui (i = 0, 1,2, . . .).
But Ker T = 0 because of the o-linear independence of the system (Ui)~ . Therefore T is an automorphism
of the Hilbert space E.
Remark. M.G. Krejn (1957) characterised a Bari basis (Ui)gmby means of a
certain collection of intrinsic properties, namely, 1) completeness, 2) o-linear
independence, 3) quadratic closeness of the Gram matrix ((Ui, Uj))Tj=o to the
identity, i.e.
h: E, + E2 is said to be
(21)
where fk E ET, uk E E,, llfkll = 1, )IuJ = 1 and & (~1 < co; thus series (21) converges in the uniform operator topology so that all nuclear homomorphisms
are
compact. The class of nuclear homomorphisms
%(E,, E,) in Homc(E,, E2) is an
L?(E,) x g(E,)-invariant
subspace (in f?(E) it is a two-sided ideal 3(E)).
The nuclear norm in YI(E,, E,) is defined as v(h) = inf ck lcxkl over all representations of the form (21). Clearly llhll < v(h). With respect to the norm just
introduced %(E 1, E2) is a Banach space.
Pietsch (1963) showed that all homomorphisms
of the class 6, are nuclear
and moreover for any p > 0 there exists a nuclear operator on 11 which does not
belong to (Sp. In a Hilbert space the class of nuclear operators coincides with 6,.
If A: E + E is a nuclear operator on a Banach space E we associate with each
of its representations in the form (21) (f. E E*, u, E E) the number
(22)
which we may call the truce5 of the operator A and denote by tr A whenever this
5 More precisely,
not equivalent.
the tensor
truce.
There
are several
other
definitions
of trace which
in general
are
170
(i = 0, 1, 2, . . .).
j=O
+ A)]},
wheres
ln(1 + A)d; zl (- l)n- .
The prior condition v(A) < 1 can be removed by the following means: whatever
the nuclear operator A, the function D(n) = det(1 + LA) of the complex variable
6 Clearly tr A is a linear functional on its domain of definition.
n this situation we can refer to the matrix trace of the operator A.
n general, if cp(l) is an analytic function on the disk (II < r, for any operator A E Q(E) such that
l[All < r we put
m p
0)
d-4 = 1 n!A
*=0
This is one of the principal versions of jimctional caZculus in Q(E).
8 4. Theory
of Operators
171
1 is analytic o\n the disk 111< [v(A)]-. It can be shown that D can be continued
analytically to the whole plane, i.e. it is an entire function. It remains to take as
definition: det(1 + A) = D(1).
Remark.
If
A = k=l2 gk(.bkT
where rk E E, gk E E* and & llgkll llVk[l < Co, then
det(1 + AA) = 1 + F c,A,
n=l
det(gjJujs)):,,,l
(n = 1, 2, 3, . . .) (Koch-Grothendieck
where cn=&,<...<jn
mula - a generalisation of a well-known formula from linear algebra).
for-
(A, B) = tr(AB*),
with respect to which &(E) (with the previous norm) is a Hilbert space. In this
particular case the previous formula follows from the theorem of Riesz.
The concept of a nuclear homomorphism
carries over to the category of LCSs
if we replace the conditions on fk, uk by the following: the family {fk} is equicontinuous, the family {&} is bounded and its closed absolutely convex null U
provides a norm on 0 = Lin U with respect to which 6 is a Banach space. All
nuclear homomorphisms
are compact. The theory of the trace outlined above
extends to nuclear operators on an LCS (G.M. Litvinov, 1979). Starting from the
previous definitions Grothendieck (1955) introduced the important class of nuclear spaces as those9 spaces for which every continuous homomorphism
into a
Banach space is nuclear. The class of nuclear spaces is closed with respect to
completions and passage to subspaces and factor spaces (and certain other
his
is one of several
equivalent
definitions.
172
Chapter
2. Foundations
and Methods
summing homomorphisms
locally
convex
Frkchet
space is separable.
173
u x (VI + VJ - u x Vl - u x v,;
(Ul x u2) x v - u1 x v - u2 x v;
u x uv - a(u x v).
au x u - u(u x v);
For any two seminorms p (I), pt2 on E,, E, respectively we define their tensor
product by
(p
@p
uk)
(w
E E,
C3 E,),
of the form
w=~u,@v,.
k
This is a seminorm
@ pq(u
@ 0) = p (u)p
v).
El,
E,.
Theorem of Grothendieck.
If E,, E, are metrizable
w E E 1 @ E, can be represented in the form
174
w = k=l2 ak(ukC3v,),
where u& E E,, lim&+, u& = 0, u&E E,, lim&,,
z
(
$fk( k
>
from E* @ E to S(E). In the Banach space case the topological tensor product
E* @ E is also a Banach space. Since the inequality v(zw) G IIw 1)(v is the nuclear
norm) holds for all w E E* @ E, then r can be extended by continuity to a
homomorphism
2 E* &I E + %(E) which is clearly surjective. Consequently
%(E) z (E* @ E)/Ker z
This topological
isomorphism
is an isometry.
4.2. The Fixed Point Principle. A famous theorem of Brauer asserts that if X
is a compact convex set in a finite-dimensional
linear space then any continuous mapping F: X + X has a fixed point13. It was extended to inlinitedimensional normed spaces by Schauder (1930) and then to LCSs by A.N.
Tikhonov (1935).
Theorem. Let E be an LCS, X c E a non-empty closed convex set and F: X + X
a continuous mapping such that the set Y = FX is relatively compact. Then there
exists a fixed point: x = Fx.
Proof. If x # Fx for all x E X, there exists a neighbourhood
of zero U such
that x - Fx 4 U (x E X). Let V be an absolutely convex neighbourhood of zero
such that V + V c U. We can construct a finite V-chain (x& + V}; on the set K
The convex hull P c X of the set {xk}T is compact and lies in a subspace of
dimension <n. The mapping FI? is uniformly continuous. We triangulate f
sufficiently finely so that in each open simplex we have Fy - Fy E K Let (yi}T
be the set of all vertices of the triangulation.
We denote by ~yi any one of the
points X~ lor which x& - Fyi E V and extend F affinely in each simplex14. The
mapping F: P + f is continuous. By Brauer theorem it has a fixed point y. But
I3 Since all compact convex sets of dimension p < co are homeomorphic, it sutlicies for the proof to
take X = IY, the closed unit ball in W If x # Fx for all x E P, then each ray (1 - s)Fx + TX
(r 2 0, x E I)p) intersects the sphere 9-i = 8Dp in a uniquely determined point 9x. The mapping
@: Dp + 9-i is a continuous retraction: Cpl,,., = 1. Its existence contradicts homology theory.
i4A mapping $: M + N of convex sets is said to be aJ%re if I&TX + (1 - r)y) = r$x + (1 - r)$y
(0 < T < 1).
$4. Theory
of Operators
175
for some
Zj
> 0
(Cj
Zj
= 1). 0
We
176
Does there exist such a subspace for a pair of commuting compact operators?
Using the fixed point principle, V.I. Lomonosov (1973) obtained the following
general result.
Theorem. Let A be a non-zero compact operator on an infinite-dimensional LCS
E. Then the algebra (A) of all continuous operators which commute
with A is
(topologically) reducible, i.e. there is a non-trivial closed subspace which is invariant
for all TE (A).
independent
significance,
Lemma. Let 52 be an irreducible subalgebra of the algebra of continuous operators on an LCS E which contains a compact operator A # 0. Then there is a compact
operator B E 9l which has a fixed point x # 0.
Proof. For any y E E, y # 0, the subspace L, = {z: z = Ty (T E %)}is different
from the zero subspace, invariant for all S E I and, consequently, dense. We
choose a point x0 such that Ax, # 0 and a convex neighbourhood of zero U such
-.
that the set Q = Ax, + AU 1s compact and does not contain zero (clearly, Q is
convex). Then for each y E Q there is an operator TYE Isuch that T,y - x,, E U.
By continuity Tyv - x0 E U for all v E Q close to y. There exists therefore a
finite collection {K}; c
such that for each point v E Q at least one of the
inclusions TV - x0 E U holds, i.e. at least one of the inequalities q( TV - x,,) < 1
(1 < i < n) holds, where 4 is the gauge of the neighbourhood
U. We put
1(r) = max(1 - z, 0)
4i(v) =
wiV
- x0))
,z 4(4W
The mapping
- x0))
n
@V
2
i=l
4i(V)KV
(0 E
Q)
~ ~i(AX)T,AE~
i=l
and x # 0 since 0 $ x0 + U. 0
I8 For Banach spaces; however the extension to LCSs does not require significant changes.
19(A) is called the centralizer (in general algebraic terminology) or the cmnmutant (in operator theory
terminology) of the operator A. The centralism (commutant) M of any set M c UT(E) is defined as
the intersection of all (A) (A E M). If !!I c f?(E) is a subalgebra then QI n % is the centre of the
algebra 9X. In any case it contains the scalar operators A = 11 (I. E K).
201t is said to be hyperinvariant
for A.
111
178
bility. Invariant means were brought into use by von Neumann (1929) who
established the amenability of abelian groups and gave a simple example of a
non-amenable group: the free group with two generators.
Theorem.
= l,f>
O>
sX
WW
& =
sX
4(x) dp
Z6i.e. a semigroup provided with a HausdorlI topology with respect to which multiplication is
continuous.
* semigroup with an identity is sometimes called a monoid.
281n a topological group not only multiplication but also inversion s I+ s-l is continuous (by
definition).
* or the sake of brevity we sometimes also call them flows.
6 4. Theory
of Operators
179
Bogolyubov-Krylov
Theorem 30. If the phase space X of a dynamical system is
compact and the time semigroup S is left-amenable (in particular, Abelian), there
exists an invariant measure p on X with ,u(X) = 1.
Proof. Let m be a left-invariant mean on S. We take any point x0 E X and
consider the linear functional f (4) = m[q5(T(*)xo)]
on C(X). Since it is nonnegative, by the theorem of Riesz it can be represented in the form jx 4 d,u. This
measure p has the required properties. 0
IM = max IM sup
s IITWII > .
(
A representation of a group G in a Hilbert
operators T(g) (g E G) are unitary.
generalised
form.
180
infs IIW-4
2 ~llxll
then there exists on E an equivalent Hilbert norm with respect to which T is unitary.
Proof. We consider on S the bounded function 4Xx,Y(s)
= (T(s)x, T(s)y) (x, y E E)
and put (x, y) = m[&J.
Then (T(s)x, T(s)y) = (x, y) (s E S) and a2 llx112 <
6, x> G b2 11412(B = sups IIWI).
In many situations this theorem allows us to restrict attention to unitary
representations. The following are classical examples: 1) finite-dimensional
representations of finite groups, whose theory was developed at the end of the 19th
and the beginning of the 20th centuries (Frobenius, Molin, Bumside, Schur); 2)
representations of compact groups in Hilbert space (the theory of unitary representations of compact groups was essentially constructed in the work of Peter
and H. Weyl, 1927).
If T is a finite-dimensional
representation of a semigroup S, the function
x(s) = tr T(s) (s E S) is called its character. This term reflects the fundamental
classical result that, for example, unitary representations with identical characters
are indistinguishable
from the geometric point of view; more precisely, they are
equivalent in the sense defined below.
Let T,, T2 be representations of a topological semigroup S in LTSs E,, E,
respectively. Any continuous homomorphism
V: E, + E, which satisfies the
relation T,(s)V = VT,(s) (s E S) is called an intertwining operator for the pair of
representations T,, T2. For example, V = 0 is such an operator. If among the
intertwining operators there is a topological isomorphism
U (so that T,(s) =
UT,(s)Uwe say that the representations T,, T, are equivalent. If the space is
a Hilbert space and U is an isometry we use the term unitary equivalence.
In a series of situations generalisations of the concept of trace allow us to
consider characters of infinite-dimensional
representations and to apply them
for classification purposes. On the other hand, the characters of one-dimensional
representations are very important and interesting (especially in the Abelian
case). They are the continuous homomorphisms
of a semigroup S into the multiplicative group of the field C and are called (one-dimensional) characters of the
semigroup S. For example, the exponentials are the characters of the additive
semigroup R, (the character which is identically zero is usually excluded from
consideration).
Now let G be a locally compact group and p a right Haar measure. The group
G acts by right shifts on the Hilbert space L2(G, cl). This unitary representation
R of the group G is said to be regular and is the starting point of harmonic
analysis on a group. If G = Iw or G = T (the unit circle) then p is Lebesgue
measure and harmonic analysis on G is Fourier analysis. We now pass to a
31This term is retained for representations by means of shifts in any function spaces on a group
which are invariant with respect to shifts.
$4. Theory
of Operators
181
consideration of the basic aspects of the spectral theory of operators with which
harmonic analysis has a deep connection.
4.4. The Spectrum and Resolvent of a Linear Operator. Many applications
require that spectral theory covers linear operators which are not necessarily
defined on the whole of the underlying LTS E and are not necessarily continuous.
The domain of definition D(A) of an operator A on E can be any subspace of E
and in general it is neither closed nor dense, i.e. A is simply a homomorphism
of D(A) + E. Such a generality has significant implications for the algebra of
linear operators. For example D(A + B) = D(A) n D(B) and it can happen that
D(A + B) = 0 for dense D(A), D(B). We note here that D(U) = B-ID(A), the
B-preimage of the subspace D(A).
In accordance with the usual formal rules, just one change in the domain of
definition changes the operator. If D(A) c D(J) and A aA, = A then A scalled
an extension of the operator A and we may write A c A (in this connection the
operator A is said to be a restriction or part of the operator A .
Let A, B be operators such that BA c AB and B is defined on the whole
space E. Then we say that they commute even if BA # AB, since if BA = AB
and D(AB) = E then D(A) = E, a restriction which is unacceptable for many
applications.
Example. Suppose that D(A) = D(B) = E and A, B commute (AB = BA). If A
is injective, the operator A- (D (A- ) - Im A) commutes with B in the sense
defined above.
The theory of operators is concerned in the first place with their properties
which are invariant with respect to automorhisms of the underlying space. From
this point of view the operator UAW (U is a topological automorphism) is just
as suitable as A; in this connection, just as for representations, we say that they
are equivalent (the term similar operators is also used). In the finite-dimensional
case the problem of classifying linear operators up to similarity reduces to the
theory of elementary divisors. In the infinite-dimensional
case the problem has
on the whole a transcendental nature. However knowledge of certain, if not all,
invariants already allows the possibility of penetrating deeply into the internal
structure of an operator. The spectrum is the most important invariant of this sort.
As the ground field in spectral theory it is expedient both from the internal
point of view and in connection with applications to settle on the field C, which
has a unique combination
of properties: algebraic closure, local compactness,
connectedness.
A complex number I is said to be a regular point of a linear operator A on an
LTS E if A - 11 maps D(A) bijectively onto E and the resoluent R, = R,(A) =
(A - II)- is a continuous operator on E + D(A) (in this connection it is convenient to regard R,(A) as an operator on E + E by using the inclusion of D(A) in
E; in any event Im R,(A) = D(A)).
The set reg A of regular points of an operator A is called the resoluent set and
its complement spec A is the spectrum.
182
The spectrum of an operator contains the set of all eigenvalues and sometimes
reduces to this, for example, if A: E + E is a continuous algebraic32 operator.
Example. Let P be a projection different from 0, 1. Then the set of its eigenvalues is (0, l} and if P is continuous spec P = (0, l} (otherwise, spec P = C).
An eigenvalue 1 is said to be normal if corresponding to it there exists a linitedimensional root subspace with an invariant topological complement L and
A # spec(A IL).
For any linear operator A and all 3, E @we put A1 E A,(A) = Im(A - Al). The
eigenvalues il are characterised by the fact that the homomorphism
(A - 11):
D(A) + A,(A) is not invertible. For all other A the inverse homomorphism
is
defined. As before we can call it the resoloent of the operator A, denote it by
R, = R,(A) and regard it as an operator on E with domain of definition A,(A);
Im R,(A) = D(A).
The eigenvalues of an operator A are contained in the approximate spectrum
spec,A which is defined in the following way: I E spec,A if there is a neighbourhood of zero V such that for any neighbourhood
of zero U a vector x $ V can
be found for which Ax - kx E U.
The set spec,A = spec A\spec,A
is called the residual spectrum. The union
rzg A = reg A u spec,A is characterised by the fact that the resolvent R,: A1 + E
is a continuous operator (but AI # E for I E spec,A in contrast to the situation
for regular points). Points 1 E rzg A (i.e. L $ spec,A) are said to be quasiregular.
If an operator A on a complete LTS E is closed33, the subspace A,(A) is closed
for all 1 E rzg A. In this case the factor space E/A, is called the defect space and
its topological dimension n, is called the defect number of the operator A at the
point A.
The approximate spectrum of an operator A on a normed space E is the set
of those ;I E C for each of which there exists a quasieigensequence of vectors (x.)p :
IlAx - WI 2 c~II4l
Thesetr~gAisthereforeopen(ifI~--l<clwecanputc,=c,-l~--l)and
consequently the approximate spectrum is closed.
Using the technique of gaps, M.G. Krejn, M.A. Krasnosel kij
Mil an (1948) obtained the following general result.
and D.P.
183
5 4. Theory of Operators
= dimt(E/A,)
= n,. 0
2. The resolvent set reg A and the residual spectrum spec,A are open.
In fact, these sets are separated out from rrg A by the respective conditions
n, = 0 and n, # 0.
Corollary
e
ll/(t)dt
s0
is a bounded operator on the whole of L*(O, 1) for ah I E @.
We note further that any non-empty closed set S c @ is the spectrum of some
closed linear operator. For this we consider the Hilbert space L*(S, ,u), where p
is a measure on S which is positive on all open subsets X of S. Let us denote by
A the operator obtained by multiplying by the independent variable [: (A&(c) =
&(C). Its domain of definition consists of those q4E L*(S, p) for which c+(l)
belongs to the same L*. The operator A is closed and if S is compact (and only
in this case) then A is even bounded. It is easy to see that spec A = spec,A = S.
The points with positive measure are the eigenvalues in this example.
For any operator A the resolvent R, is an analytic function of 1 on reg A since
the power series
(R,+)(s) =
2 R;+ L - ,u)
n=O
operator topology for II - ~1 < IIRpll- and its sum
I
34This argument shows once again that reg A is open. Analyticity of the resolvent is the basis for
the application of the methods of complex analysis in spectral theory.
184
Suppose that the operator A is bounded and D(A) = E. The Laurent series
(23)
converges in the domain )A1> l/All under the uniform operator topology and its
sum is R,. Thus spec A lies in the disk 111< 11All. We see that the spectrum of a
bounded operator is compti.,On
the other hand spec A # 0 since otherwise R,
is an entire function which tends to zero as 111+ co and so, by Liouville
theorem, Rt F 0 which is impossible for an inverse operator.
The smallest p = p(A) > 0 for which spec A lies in the disk 111< p is called
the spectral radius of the operator A. Clearly p(A) < /[All but we also have
GelIfand formula3 5
p(A) = lim l IjA = inf ~IIA I.
n+co
n
The existence of the limit and its equality to the inlimum is guaranteed by the
theorem of Fekete that if a sequence (01,) of real numbers is subadditive, i.e.
a ,,, < a, + a,,,, then 5 tends to the infimum of the sequence as n + co. Further,
n
the outer radius of convergence r of the series (23) is equal to the stated limit and
r > p(A) since the domain of convergence of the series is contained in reg A. On
the other hand, r < p(A) since the resolvent is analytic on the domain IA) > p(A).
If p(A) = 0, i.e. spec A = (O}, the operator A is said to be quasinilpotent. In
particular, any bounded nilpotent operator is of this type, however the class of
quasinilpotent operators is more extensive than this.
Example.
is not nilpotent
but it is quasinilpotent.
1
(W)(s) = (n _ l)!
operator
4(t) dt
s0
In fact
; (s - O?(t)
dt
(n = 1,2, 3, . ..).
hence
IlZ = 0l 0; .
Consequently, p(Z) = 0 by Gel and formula.
For any bounded operator A defined on the whole space the circle 111= p(A)
contains points of the approximate spectrum.
Let A be an isometric operator defined on a closed subspace D(A) c E. Its
image Im A is clearly also closed. Since
35Hence it follows that the functional calculus for the operator A operates on the disk )I[ < p(A).
8 4. Theory
of Operators
185
all points L with 111 # 1 are quasiregular. The defect number n, is constant in
the disk IA) c 1 and in its exterior 111> 1. Its value on the disk is equal to
n, = dimt(E/Im
A). We denote its value on the exterior by n,. It follows from
the formula A,(A) = A,-,(A-) (111 # 1) that no0 = dimt(E/D(A)) so that n, =
O-SD(A) = E (clearly n, = OoIm A = E). The pair (n,, n,) is called the deficiency index of the isometric operator. If there exist quasiregular points in the
unit disk, then no0 = n,. In order that an isometric operator map the whole space
onto itself (in a Hilbert space this means that it is unitary) it is necessary and
suflicient that it have deficiency index (0,O).
The approximate spectrum of any isometric operator is unitary, i.e. it lies on
the unit circle. If the operator is defined on the whole space then its spectrum lies
in the disk 111< 1, the circle 111= 1 contains the points of the approximate
spectrum but it can also contain quasiregular points; then they are regular and
n, = 0. If n, # 0, the spectrum is the disk 111 < 1 and the approximate spectrum
is the circle In( = 1. The spectrum of an operator which maps the whole space
isometrically onto itself lies on the unit circle.
To conclude this section let us consider the structure of the spectrum of a
compact operator. This is the next in complexity after the algebraic case. All
linear operators on a finite-dimensional
LT.3 are compact and algebraic; the
spectrum of any of these coincides with its set of eigenvalues.
Theorem. The spectrum of a compact operator A on an infinite-dimensional
LTS E consists of zero and a no more than countable set of eigenvalues diflerent
from zero. The unique limit point of this set, zf it is infinite, is zero.
Proof. Each point 1 E spec A\(O) is an eigenvalue by the Fredholm alternative. The point A = 0 belongs to the spectrum since otherwise the operator A-IA = 1 is compact, which is impossible for an infinite-dimensional
E.
Let V be a balanced neighbourhood of zero whose image is relatively compact.
Let us suppose that there exists a set (A,} r of eigenvalues such that I&l 2 E > 0;
let e, (n = 1,2, . . .) be corresponding eigenvectors. Now consider L, = Lin{e,}l;
clearly dim L, = n. We can choose in L, n 0 a vector x, 4 L,-, + V. Since
Ax, - &x. E L,-, and Ax, E L,-l (n > m) we have Ax, - Ax, 4 EV, which is a
contradiction. 0
Let us enumerate the eigenvalues which are different from zero (if they exist):
. We denote their orders by vl, vl,. . . , and their maximal root subspaces
&,L...
by WI, W,, . . . . They are all finite-dimensional
and have invariant complements Vi = Im(A - A1l)yl, . . . . Moreover spec(Alw,) = {A,} and spec(Aj,J =
spec A\{&}. Hence it follows that on a Banach space E the resolvent R,(A) has
a pole of order v, at the point & and has no other singularities apart from the
points 1, and zero, i.e. it is meromorphic on @\ (0).
186
If a compact operator has no eigenvalues different from zero (i.e. it is quasinilpotent in the Banach situation), it is said to be a volterra operator. The
following example explains the origin of this term.
Example.
integral operator
(0 < s < 1)
R(s, W(t) dt
s0
with continuous kernel R(s, t) is a Volterra operator. In particular the integration
operator I is a Volterra operator.
Let A be a non-negative compact operator on a Banach space E with closed
cone K. Let us suppose that the cone K is total, i.e. Lin
= E. If the spectral
radius p = p(A) is positive, then p is an eigenvalue for A and A* and corresponding to it there are an eigenvector x E K and an eigenfunctional f~ K* with
f(x) > 0 (M.A. Rutman, 1938). The condition p(A) > 0 is satisfied if there exists
a vector u > 0 (u # 0) such that Au 2 cu for some c > 0 (then p(A) > c).
(A&(s) =
4.5. One-Parameter
Semigroups. The representations
T(t) of the additive
semigroup Iw, of non-negative numbers with T(0) = 1 are designated by the term
in the title. The representations of the additive group [ware called one-parameter
groups.
Each bounded linear operator A on a Banach space E generates a oneparameter group
co A?
T(t) = eAt = 1 I
n.
(t E R).
(24)
Moreover
;[T(t)x]
i.e. the vector-function
= AT(t)x = T(t)Ax
(x E E, t E W),
G(t) = Au(t)
(--cO-ct<cO),
u(0) = x
9 4. Theory of Operators
187
bounded, the representation T is continuous not only in the strong topology but
also in the uniform topology:
lim II T(t + z) - T(t)11 = 0
(t E R).
r-10
~ = lim ln
IITWII
-<CC
z-4, t
exists. The Laplace transform
T(n)
=-sm
T(t)e-
dt
is therefore defined on the half-plane II, = {A: re 1 > g}. Since 1)T(t)(I G MeC for
some c > e, M > 0, we have
The values of the operator function ?(A) commute pairwise and satisfy the Hilbert
identity
The subspace L = Im F(A) therefore does not depend on 1. The operator f(A)
maps E bijectively onto L (injectivity follows from the uniqueness theorem for
the Laplace transform, since, if ?(A)x = 0, then f(,u)x = 0 for all p E II,). Let us
put A(1) = 11 + [?$)IThis does not depend on 1 and defines a closed
operator A (D(A) = L) with R,(A) = F(A). The subspace L is dense since
[T(A)]*
= -SW T*(t)e-
dt
Au(t)
(t
> O),
u(0) = x E D(A).
IfpEII,andy=Ax-px,then
x = R,y = -
m (T(s)y)es0
ds,
(25)
188
and hence
m
T(t)x = Consequently
&(t)x]
i.e.
T(t)x
(T(t + s)y)e-
s0
the function
T(t)x
ds = -ee
has a continuous
m (T(s)y)eC
sf
ds.
derivative for
= AT(t)x
2 0 and
(t 2 01,
(25). In particular,
putting
= 0
we obtain Ax = 1 [T(t)x]I1=o.
The Cauchy problem (25) can be considered36 for any operator A. In the
situation described above it does not have solutions other than T(t)x, i.e. the
corresponding homogeneous problem C(t) = Au(t) (t > 0), u(0) = 0 has only the
trivial solution. This is the case (Yu.1. Lyubich, 1960) even under the much weaker
restriction of growth of the resolvent on the ray il > 1, > 0:
-lim lnll&II
--coo
a++m 1
(this cannot be relaxed further). In fact
s
R,u(O) = u(t)edt + e-hR,u(s)
(s > 01,
f0
from which it is obvious that R,u(O) is the local Laplace transform for [-u(t)].
If u(O) = 0 then u(t) = 0 by the uniqueness theorem for the l.L.t.
The subspace I(A) c D(A) of those x for which the Cauchy problem has a
solution (independently of uniqueness37) is called the initial subspace. It can fail
to be dense (and can actually be the zero subspace) even in the case where D(A)
is dense. But if a bound of the form IIRlll = O(lAl
is satisfied in some halfplane II, contained in reg A and D(A) is dense, then I(A) is dense (Yu.1. Lyubich,
1964).
If x E I(A) and u(t) is a corresponding solution, then clearly its values belong
to I(A) for all t (in this sense I(A) is invariant). If the Cauchy problem is
determinate, the evolutionary operator U(t) transforming u(O) into u(t) (t > 0) is
defined on Z(A). Clearly U(0) = 1 and U(t + s) = U(t) U(s). If U(t) is bounded for
each3* t > 0, we have a representation of the semigroup R, in the subspace I(A)
(in general it is not closed). However if sup04tSE II U(t)11 c co for some E > 0,
361t is implicit that the solution is continuously differentiable and its values belong to D(A) for all
t 2 0 (in a relaxed formulation - for t > 0).
37A Cauchy problem for which the solution is unique is said to be determinate (this property does
not depend on x).
3sIn this case the Cauchy problem is said to be pointwise stable on I(A) (simply pointwise stable if
I(A) is dense).
189
then3 the extension by continuity of U(t) onto Z(A) preserves continuity of the
trajectories and we obtain a representation of the semigroup R, in the closed
subspace Z(A). Its infinitesimal operator is the closure of the operator A (S.G.
Krejn - P.E. Sobolevskij, 1958).
The infinitesimal operator of a one-parameter semigroup is also said to be
generating for this semigroup or to be a generator of it since, clearly, the semigroup with a given inhnitesmial
operator is unique. The question arises of
describing the class of generators of all possible one-parameter semigroups.
Miyadera-Feller-Phillips
Theorem. In order that a densely defined linear operutor A on a Bunuch space be the infinitesimal operator of a one-parameter semigroup,
it is necessary and suflicient that reg A contain a half-plane ll, in which the
inequality
M
llR:ll G (re k _ c)
(n=0,1,2,...;
M a positive constant)
is satisfied.
Necessity.
This follows from the bound II T(t)11 < Me and the formula
1 d
,
(-1)
a3
RI! = (n _ I)! dl = (n - l)! 0 T(t)t
s
- dt.
Sufficiency. Under the given conditions the Cauchy problem (25) is determinate and the initial subspace Z(A) is dense. Since
s
u(t)e-
we have by Widder
u(t) = lim
n+m
dt = R,(u(t)e-
s0
inversion formula (0 <
(- LRJ lu(0)
+ I
- u(O)),
t
< s):
kzo ((n
shlik
Ri+ (s)
A=ll/t
Hence it follows that Ilu(t < Me llu(O)ll, i.e. for the evolutionary operator:
II U(t)11 < Meet. Since A is closed under the conditions of the theorem, it is the
infinitesimal operator of the representation U. 0
An immediate consequence of this result (taking account of the connection
between the bounds for the representation and the resolvent) is the
Hille-Yosida Theorem. In order that a densely defined linear operator A on a
Banach space be the infinitesimal operator of a one-parameter semigroup of
contractions, it is necessary and suflicient that the right half-plane II, be contained
in reg A and that the inequality
391n this case the Cauchy problem is said to be stable on I(A) (simply stable if I(A) is dense).
190
be satisfied in II,.
d
ds
which, together with the derivative, belong to Lp. This is a dissipative operator.
The theory of one-parameter groups is reduced to the previous case by considering representations of the semigroups IR, . In particular, the infinitesimal operators of one-parameter groups of isometries can be characterised (I.M. Gel and,
1939) as those operators whose spectrum lies on the imaginary axis and for which
the bound
1
IIWI G Ire
(re 1 # 0)
A regular representation
actually conservative.
Conservative and dissipative operators have many remarkable properties, in
particular the spectral radius of a bounded conservative operator A on a Banach
space E is equal to its norm (V.Eh. Katsnel on, 1970). In fact, the function f(eA )
(x E E,fe E*) is an entire function of exponential type <p(A) and it is bounded
on the real axis. From Bernshtein
inequality we have
IfkfW
f If(eAW
sup
G PCW~I
- Ilxll.
Putting t = 0 we obtain If(Ax)I < p(A)IlfII * (IxIJ and consequently llAl/ < p(A).
Incidentally, Bernshtein inequality itself can be interpreted as the coincidence
of the spectral radius with the norm of the differentiation operator on the space
B, of entire functions of exponential type < 0 which are bounded on the real axis
(it is remarkable that the differentiation operator on B, is bounded).
A general mechanism underlying inequalities of Bernshtein type and results
pertaining to it in spectral theory was brought to light by E.A. Gorin (1980). It
was based on the functional calculus associated with integral representations of
positive functions.
From the point of view of the Cauchy problem there is great interest in the
study of asymptotic (as t + co) properties of one-parameter semigroups. For
example, boundedness (i.e. supt /IT(t)11 < co) is equivalent to Lyapunov stability
4 4. Theory of Operators
191
exists. It is a bounded projection (for semigroups of contractions it is an orthoprojection) on the subspace of lixed points {x: T(t)x = x (t 2 0)} = Ker A (A is
the infinitesimal operator). This theorem owes its origin to the analysis of the
foundations of statistical mechanics, which goes back to Koopman and von
Neumann40.
There is an important class of representations for which the ergodic theorem
is true without restriction on the space - almost-periodic representations.
A representation T is said to be almost-periodic (a.~.) if all of its orbits are
relatively compact41.
Example 1. In the space AP(R) of almost-periodic functions on R the representation of the group Iw by means of shifts (the regular representation) is a.p.
The definition of an a.p.f. carried over without change to functions on the
semiaxis R, . We have correspondingly
a regular a.p. representation of the
semigroup R, in AP(R+).
The general form of an a.p.f. on the semiaxis is 4 = do + h, where boo(t) tends
to zero as t + co, and 4r can be extended to the whole axis as an a.p.f. (Frechet,
1941).
Example 2. Let A be a compact operator on a Banach space E which does
not contain any subspaces isomorphic to the space c of convergent sequences
(5,); (in particular, E may be reflexive). Then, if the representation e (t E R) is
bounded, it is a.p. In fact, for the function u(t) = e% (x E E) the derivative
G(t) = Au(t) is a.p. because of the compactness of A and the function itself is
bounded. Consequently u(t) is a.p. (here we are using a generalisation due to M.I.
Kadets (1969) of the Bohl-Bohr theorem on the integral of an a.p.f.).
By restricting an a.p. representation of the group R to the semigroup R, we
obtain, clearly, an a.p. representation but it is such that all the operators T(t)
(t E W,) are invertible and their inverses are bounded (in this sense it is a group
representation of the semigroup R,). There is a different kind of a.p. representation of the semigroup R, , all the orbits of which tend to zero as t + co. Further,
if the space E can be decomposed into the topological direct sum of the spaces
4oThis author established (1932) the ergodic theorem in Hilbert space. The generalisation given here
is due to Larch. Usually formulations of ergodic theorems are given for semigroups of powers (I )?.
The ontinuous case considered in the text is easily reduced to the iscrete case (such a reduction
occurs quite often). Further, as a rule, we restrict ourselves to only one of these cases.
41 In exactly the same way a.p. is defined for bounded representations of any topological semigroups
in any LTS. The property of boundedness is not required a priori in a barrelled LCS since it follows
from the relative compactness of the orbits by the Banach-Steinhaus theorem.
In general we can define a.p.fs on any topological semigroup; they can be not only scalar-valued
but also vector-valued.
192
Chapter
2. Foundations
and Methods
E,, E, in which a.p. representations T,, T, are given (for example, Tl a group
representation and To having orbits tending to zero) then the natural representation Tl i To defined on E is a.p. We have the boundary spectrum splitting-off
theorem (M.Yu. Lyubich - Yu.1. Lyubich, 1984): let T be an a.p. representation
of the semigroup [w, in a Banach space E; then E can be decomposed into
the topological direct sum of invariant subspaces E,, E, such that Tl = TIE1
is a group representation and the orbits of T, = TIE0 tend to zero. The subspaces E,, E, are called the boundary and interior subspaces respectively and
the projection P associated with them (Im P = E,, Ker P = E,) is called the
boundary projection. It arises in the following way. Corresponding
to an a.p.
representation T there is a Bohr compact set - the strong closure of the family
(semigroup) of operators { T(t)},,o. This is an Abelian semigroup which is compact in the strong topology. In any compact semigroup there is a smallest
two-sided ideal - the Sushkeuich kernel (A.K. Sushkevich, 192843; Numakura,
1952). For an Abelian semigroup (and in certain other cases) the Sushkevich
kernel is a compact group. The boundary projection for an a.p. representation
is the identity of the Sushkevich kernel of the Bohr compact set. Hence it is clear,
incidentally, that [[PII < sup 1)T(t)11 and in particular, if T is a contraction then P
is an orthoprojection.
In the latter case the representation Tl is an isometry.
The theory which has been described carries over to a.p. representations of any
Abelian (and certain non-Abelian) topological semigroups, in particular, to semigroups of powers of an a.p. operator 44. A prototype of the general theory of
a.p. representations, the boundary spectrum splitting-off theorem for bounded
semigroups of operators on a reflexive Banach space, was constructed by Jacobs
(1957), after which De Leeuw and Glicksberg (1961) obtained corresponding
results in the weak topology for any Banach space. In the weak a.p. case the
orbits for T, do not tend to zero in general, even weakly, and they only have zero
as an adherent point. Furthermore To can turn out to be a group representation
(for example, (T(t)f)(s)
= eifsf(s) (s, t 3 0) in L(R+)).
4.6. Conjugation and Closure. Let A be a linear operator on an LTS E and
suppose that its domain of definition D(A) is dense. Then the conjugate operator
A* on E* can be defined by means of the relation (A*f)(x) =f(Ax) for those
f~ E* for which the right hand side of the identity is continuous on D(A) and,
consequently, can be uniquely extended to a continuous linear functional on E.
If E is a Hilbert space the definition of the conjugate operator is usually modified
on the basis of Riesz theorem so that A* acts on the space E itself:
(Ax, Y) = (x9 A*Y)
(x E W),
Y E WA*)).
43The fundamental
work of A.K. Sushkevich
is concerned
with finite semigroups.
441n particular,
any operator
of the form A = V + V, where V is compact,
11Y 11< 1 for some m
and sup,,,
l[A 11< co is of this type (Yosida - Kakutani,
1941).
5 4. Theory
of Operators
193
(~3 Y E WA)),
(27)
then A is said to be a symmetric operator. Selfadjoint operators play a central
role in quantum mechanics, where they are used to give a suitable description of
physical quantities (coordinates, impulses, energy, etc.). The principal credit for
constructing the spectral theory of selfadjoint operators is due to Hilbert and
von Neumann.
We say that an operator A on an LTS E admits closure if the closure of its
graph is the graph of some operator A, the closure of the operator A. Clearly
A c Aand Ais closed. In order that A admit closure, it is necessary and sufficient
that a closed extension exist for it; moreover Ais the least closed extension for A.
A continuous operator always admits closure; this is just its extension by
continuity to D(A). The domain of definition of a closed continuous operator is
therefore closed and if it is also dense then it coincides with the whole space.
Conversely, any continuous operator defined on a closed subspace is closed.
A conjugate operator A* is always closed (even w*-closed). In particular, in a
Hilbert space E any selfadjoint operator is closed and any densely defined
symmetric operator admits closure since A c A*. Moreover it is clear that x
is symmetric. If a symmetric operator A is defined on the whole space E, then
A is a bounded selfadjoint operator (Hellinger-Toeplitz
theorem). In fact, since
A c A* and D(A) = E we have A = A*. But A is then closed and consequently
it is bounded by the closed graph theorem.
Theorem. Zf a densely defined operator A on a Hilbert
then A* is densely defined and A** = A.
194
8 4. Theory
of Operators
195
From what has been said it is also evident that in the case n, = n- the set of
selfadjoint extensions can be parameterised by the isometries of NA onto N;i, in
particular, for differential operators (I.M. Glazman, 1949).
Example 1. Let us consider the operator D = f $ on L2(0, 1) which is defined
Do its restriction under the boundary conditions d(O) = 4(l) = 0. This is a closed
symmetric operator and Dd = D. Each 1 E @ is an eigenvalue for D, the eigen-
- Il)- ,
= 8eMBer.
- %I = e,,
Dev - Iv = tle-8ex,
and hence
V=
e, - 8e-Bex
n-X
.
established by elementary means). From this it is easy to establish the nonemptiness of the spectrum of any selfadjoint operator B. In fact if 0 # spec B then
A = B- is a bounded selfadjoint operator and A # 0. Therefore p(A) = IIAlI > 0.
Now at least one of the numbers &p(A) is in spec A and then its reciprocal must
be in spec B. We note incidentally that p(A) = 0 * A = 0 (i.e. a non-zero self-
196
Chapter
2. Foundations
and Methods
Q ),
Dissipative
by the requirement
(x E D(A));
operators
re(A*y, y) G 0 (y E WI*)).
on a Hilbert
they
otate through
im(Ax, x) > 0). Under this the half-plane re 1 > 0 is quasiregular but we can
have n, > 0. For a dissipative operator in our sense n, = 0. We note in this
connection that symmetric operators with zero defect number and only such
operators are maximal in the sense that they do not admit symmetric extensions
(since their Cayley transforms do not admit isometric extensions).
A symmetric operator A is said to be non-negative (A > 0) if (Ax, x) > 0,
positive (A > 0) if(Ax, x) > 0 (x # 0) and positive definite (A >> 0) if (Ax, x) 2 c llxllz
(c is a positive constant) (in all cases x E D(A)). For example, any orthoprojection
is non-negative, multiplication
by c > 0 is a positive delinite operator. If (e,); is
an orthonormal basis and (A,); is a sequence of positive numbers which tends to
zero, the corresponding diagonal operator
47Dissipativity
is preserved
under
conjugation
space case.
197
Ax = f &,<,e,
n=O
is positive but it is not positive definite.
For a non-negative operator A all points I < 0 are quasiregular and therefore
n - n-. If A* = A > 0, the spectrum of the operator A is non-negative. The
sieirum
of a positive definite selfadjoint operator lies on the corresponding
semiaxis I > c.
In the real Banach space of bounded selfadjoint operators the non-negative
operators form a closed solid cone (1 is an interior point) as a result of which we
can work with inequalities in this space. This approach improves substantially
the norm bounds.
4.8. Spectral Theory of Selfadjoint Operators. The simplest48 spectral theorem
(after the algebraic case) concerns a compact selfadjoint operator A on a Hilbert
space E and asserts that E coincides with the closure of the orthogonal sum of
the eigenspaces E, corresponding to all possible eigenvalues4g (Il,}; (zero is
included if it is an eigenvalue). This theorem is proved just as in the tinitedimensional case. With a view to generalisation it is convenient to express it in
the form of a spectral resolution
Ax = .f Iz,P,x,
n=O
onto E, and
x = 2 P,x.
n=O
198
D(@)={x:xEE,~~
--m IW12 4%
4 < ~0
1
m 44 d&x,
s -00
Y)
(x7 Y E WY)
is satisfied.
It can be shown that in addition
@* =
m 44 6.
s--CO
Consequently it is necessary and suffkient for selfadjointness of the operator @
that the function 4 be real-valued almost everywhere (with respect to each
measure m,). For boundedness of the operator @ it is necessary and sufficient
that the function 4 be bounded in the sense that
ll4l =sy(eywIdl)<
00
(x E ww
The construction which has been described is called the functional calculus for
the given orthogonal resolution of the identity a,. The correspondence 4 H @
is linear and & 42 H Qp,Qz if at least one of the functions 41, 42 is bounded. The
functional calculus is an isometric isomorphism
on the algebra of bounded
functions into the algebra of bounded linear operators.
Example. Let A4 c R be a Bore1 set and let xM be its characteristic function.
Since & = xM and xM = xM, the operator
qf=
is an orthoprojection.
ODx&J
s -m
In particular
c&=
de, =
sM
m del = 1.
s -U3
A=
199
co 1 da,,
s -00
where
(IFAX =
c P&
A<1
(x E E).
* IIdC&
f --oo
holds. Moreover
4(A) zr
m W)
s -a,
d%
In particular
m daL
~
s -mA-P
(P E reg A).
It is easy to see from this that the spectrum of an operator A coincides with the
set of points of increase of the function 52 (IF,. The eigenvalues are the points of
discontinuity of the function (EA.The operator PA = %,+, - GA is an orthoprojection onto the corresponding eigenspace. The dimension of this subspace is called
the multiplicity of the eigenvalue 1. The eigenvalues form the so-called discrete
spectrum spec,A. The continuous spectrum spec,A is defined as the complement
in spec A of the set of eigenvalues of finite multiplicity53. In terms of the resolution
of the identity it can be characterised as the union of the set of non-isolated points
of increase and the set of eigenvalues of infinite multiplicity. This characterisation
combining these two sets is just the result that there exists a normalised quasieigensequence (x,)7 ( llxnll = 1, Ax, - Lx, --f 0) with no convergent subsequence.
The intersection spec,A n spec,A can turn out to be non-empty. However the
whole space E can be decomposed into the orthogonal sum of the closed linear
hull E, of the eigenvectors and the closed invariant subspace E, of vectors with
purely continuous spectrum, i.e. those x E E for which (@,x, x) is continuous. Then
in turn E, = E, @ E,, where the summands are the spaces of singular and absolutely continuous vectors (these concepts are of course defined through the corresponding properties of the function (CE,x, x)). The extraction of the space E,, is
particularly important for the theory of scattering (see the end of this section). If
For rational rp with poles outside of spec A this definition is equivalent to the previous one.
*i.e. with the union of the sets of points of increase of all functions (EAx, x) (x E E).
s3 SpeccA and spec,A are defined for a normal (in particular a unitary) operator in exactly the same
way.
Chapter
200
2. Foundations
and Methods
J. 6,
@,A = A&&w-
s a-o
One of the approaches to the spectral theorem for selfadjoint
through the Cayley transform
V = (A - il)(A
operators is
+ il)-,
a unitary operator for which Im(V - 1) is dense. The point is that the spectral
theorem also holds for unitary operators for which the proof can be effected by
quite simple means.
Theorem. For any unitary operator U there exists a (unique) orthogonal resolution of the identity (3, (0 < a < 274 with
2n
U=
s0
eiO d@01.
Proof. Let us consider the group of powers (U)$, . It is easy to verify that the
Toeplitz matrix with elements (U-5, x) (x E E) is positive. By the Riesz-Herglotz
theorem there exists a non-decreasing function a(~; x) (0 < c1< 24 which is a
solution of the trigonometric moment problem
2n
eina da(a; x)
(n=O, +l, +2 ,... ).
(29
f0
Without loss of generality we can assume that a(a; x) is continuous on the left
on (0,271) and that ~(0; x) = 0. Under these conditions a(a; x) is unique and
(VX,
x)
2n
a(2n; x) =
s0
Consequently a(a; x) < llxll for all a. Performing
polarisation
in (29), we obtain
2%
(U%
Y) =
eina dz(a; x, y)
(n = 0, *l,
+2 ,... ),
s0
(30)
where
z(a; x, y) = t j.
~(a; x + Py)
(x,
E El.
Since there is only one possible function r in (30), it follows that $a; x, y) is a
sesquilinear functional of x, y and hence r(a; x, y) = &x, y), where (5, is a
5 4. Theory
of Operators
201
Zn eia + 1 do =
s oe --l=
co
f --a,
1 d@,,
(im P # 0).
where GA is an orthogonal
representation
of functions of the
m do@; x)
-COTy
m dC&
-mI-q
s
al
s
-Co
I d(E,.
m eiAr de,
s -cc
im(R,x,
x) = (im p)
with
202
In contrast to this the discrete spectrum is not invariant: for any selfadjoint
operator A on a separable Hilbert space there exists a selfadjoint compact
operator B with arbitrarily small norm such that the system of eigenvectors of
the operator A + B is complete (von Neumann, 1935).
Remark. Let A, B be closed operators on a Banach space. The operator B is
said to be A-compact if D(B) 1 D(A) and BIDCA)is compact under the graph norm
of the operator55 A (for this it is necessary and sufficient that the operator BRA(A)
be compact56 for some I which is regular for A). For an A-compact operator B
the set of points of the spectrum of the sum A + B that lie in any connected
component G of the set reg A is at most countable (if it is countable then it
condenses to 8G). All these points are normal eigenvalues.
Let us consider certain classical situations from the point of view of the abstract
spectral theorem.
Example
condi-
tions +( - rt) = d(z) is selfadjoint. Its resolution of the identity is generated by the
Fourier series expansion:
CWM4
= ?A cneins.
The spectrum of the given operator consists of the eigenvalues 1, = in and the
corresponding eigenfunctions are eins (n E Z). This is an example of a purely
discrete spectrum (there is no continuous spectrum).
Example 2. The resolution of the identity for the operator i $ on L2( - co, co)
:m &&@ dp,
s
where 6 is the Fourier-Plancherel
transform. The spectrum of the given selfadjoint operator fills out the whole axis, however there are no eigenvalues in it since
his is llxllA = llxll + llAx]l (x E D(A)). Completeness of the space D(A) under this norm is equivalent to closedngs of the operator A.
56 Hence it is clear that if B is compact then it is A-compact for any A.
5 4. Theory
of Operators
203
the functions eips do not belong to L2. This is an example of a purely continuous
spectrum.
Example 3. Let us consider the Schrodinger operator -y + q(x)y on
L(O, co) with continuous (for simplicity) real potential under the boundary
condition y(0) - hy(0) = 0 (- co < h < co). First of all we define it only on the
functions of finite support that are of class C2 and satisfy the boundary condition;
in this role we denote it by Qo. The operator Q,, is symmetric, densely defined
but not closed. The domain of definition of the adjoint operator Q$ consists of
the functions z E L2 n C1 such that z is absolutely continuous, -z + q(x)z,
z E L2 and z(0) - hz(0) = 0. Moreover Q$z = -z + q(x)z. The domain of delinition of the closure Q0 is extracted from D(Q8) by imposing the additional
condition
lim [y(x)z(x) - y(x)z(x)] = 0
(z E NQ,*))x+cc
This condition can be satisfied automatically (for example, this is the case if the
potential is bounded from below) and then Q. is a selfadjoint operator. In the
contrary case the deficiency index of the operator Q. is (1, 1). This is precisely
the Weyl alternative, since in the oint case the equation Qtz = lz (im L # 0)
has only the trivial solution in L2, while in the isk caseit has a one-dimensional
space of solutions. In any event a selfadjoint extension Q of the operator Q. exists
(it is unique in the first case but not in the second). All these extensions are in
one to one correspondence with the spectral functions, namely, if p(l) is a spectral
function then the resolution of the identity for the corresponding selfadjoint
extension Q is generated by the inversion formula associated with p(l):
P%m)
1. fw~(~~ A dP(PL)
f -0D
of Chapter 1, Section 2.12). From Parseval
identity
i.e. each measure j d((lF,f,f) is absolutely continuous with respect to the measure
Jdp(l), its density being equal to the square of the modulus of the generalised
Fourier transform of the function $ In this sense the resolution of the identity
6, for the Schrodinger operator is absolutely continuous with respect to the
measure jdp(l). A similar result holds for the three-dimensional
Schrbdinger
operator -Au + q(x)u on the whole space (A.Ya. Povzner, 1953) and for more
general operators (Mautner, 1953; Yu.M. Berezanskij, 1956).
Example 4. Let us consider a Jacobian matrix J with positive off-diagonal
elements. Given an orthonormal basis (e$ of a Hilbert space E the matrix J
defines a homomorphism
J^: E + s (s is the space of all complex numerical
sequences). Restricting 3 to vectors of finite support we obtain a densely defined
symmetric operator J,-, on E. It is not closed, therefore we have to pass to its
204
closure IO. The adjoint operator J,* is the restriction of the operator J^to vectors
which satisfy the condition .?u E E. In accordance with the Weyl alternative the
equation J$u = lo either has only the trivial solution for all L (im 1 # 0) and then
the operator .& is selfadjoint (point case) or the space of solutions of this equation
is one-dimensional for all L (im 1 # 0) in which case the deficiency index of the
operator Jo is (1, 1) (disk case). In any event a selfadjoint extension A of the
operator Jo exists (it is unique in the first case but not in the second). All these
extensions are in one to one correspondence with the extremal solutions a(n) of
the power moment problem associated with the given Jacobian matrix. In fact,
if CZi is the resolution of the identity for the operator A, then a(2) = (C?,e,, eo).
Returning to the Schriidinger operator on the semiaxis, we note that progress
in investigating its spectral properties was largely dependent on the discovery
and development5 of the technique of so-called transformation operators (Delsarte,
1938; A.Ya. Povzner, 1948; B.M. Levitan, 1949) which establishes the equivalence
of the operator Q and the elementary operator -y with boundary condition
y(0) = 0. In this connection it is important that the transformation operator has
the form 1 + K, where K is a Volterra integral operator with continuous kernel,
so that the inverse operator has a similar form.
V.A. Marchenko (1950) applied transformation operators to a wide range of
problems on the spectral analysis of the Schriidinger operator. In particular he
obtained in this way the asymptotic expression for the spectral function5*
and established a theorem on the uniqueness of the solution to the inverse problem
of reconstructing the operator from its spectral function. A linear procedure for
solving this problem, which is also based on the use of transformation operators59, was given by I.M. Gel and and B.M. Levitan (1951).
In the physically important inverse problem of scattering theory60 the input
data (to begin with the so-called limiting phases) are linked with the asymptotic
behaviour61 of the solutions of the equation Qy = ly as x + co. An effective
n connection with generalised shift (g.s.). The simplest g.s. (V(t)cp)(s) = (cp(s + t) - cp(s- t))/2
arises from the wave equation &@t2 = a2u/8s2; the general theory of g.s.is in many respects parallel
to the theory of group representations (B.M. Levitan, 1962).
58The right hand side of this relation is the spectral function of the simplest operator, q(x) E 0.
t is a question of the construction of a certan linear integral equation for the kernel K(x, 0, after
which the formulae q(x) = 2 dK(x, x)/dx, h = K(0, 0) are applied. All this goes through under certain
conditions on ~(1) which are close to being necessary. Necessary and sufficient conditions for p(L)
to be the spectral function of a SchrGdinger operator on the semiaxis with continuous potential were
obtained in a different way by M.G. Krejn (1951).
60Recently the inverse problem of scattering theory has been at the centre of attention of many
investigations thanks to its utilisation for integrating the Korteweg-de
Ries equation and other
non-linear equations of mathematical physics.
61 The description of which forms the direct problem. We note that both problems (direct and inverse)
under consideration are concerned with the stationary theory of scattering, which is formulated in
terms associated with the energy spectrum and not with evolution in time.
205
method for its solution (V.A. Marchenko, 1955) therefore required the use of
transformation operators
ttached to infinity:
The existence of such operators for potentials which decrease sulliciently rapidly
as x + cc was established by B.Ya. Levin (1955).
The inverse problem of scattering theory on the whole axis was solved by L.D.
Faddeev (1964), also on the basis of the technique of transformation operators.
Remark. For differential operators of order n > 2 transformation
operators
of the form 1 + K, where K is a Volterra integral operator, do not in general
exist (V.I. Matsaev, 1960), however in the case of analytic coefficients the transformation operator was constructed by L.A. Sakhnovich (1961).
In non-stationary scattering theory the problem is to compare the asymptotic
behaviour as t + + cc of a physical system which is evolving in time t with its
behaviour as t + - co.
Example. The Schriidinger operator Q (just as any other selfadjoint operator)
generates a group of unitary operators eiQt. In quantum mechanics there corresponds to this a non-stationary Schriidinger equation for the wave.function +:
where H is a Hamiltonian
constant.
+ q(x)+
(for example, this represents an a-particle in the field of a massive nucleus62 which
was the situation in the historical experiments of Rutherford). The effect of
scattering is created by the potential q, i.e. it is associated with the deviation of
the actual evolution from the free (unperturbed) evolution which is determined
by the Hamiltonian
which are clearly isometric. The limits here are naturally treated as strong limits
62This is two-particle scattering. N-particle scattering with N > 2 is a much more complicated
process. It was investigated for N = 3 by L.D. Faddeev (1963).
206
and the first question concerns their existence (if not on the whole of L2 then on
subspaces which have to be described). Following that we can define the scattering operatoF3
s = (w+)- C,
Kate-Rosenblum
Theorem. Zf A = A, + B where B is a nuclear operator, the
wave operators W, are defined on the subspace E,,(A,) and map it onto E,,(A).
Thus the scattering operator S is defined and is unitary on E,,(A,).
eiA
E,,(A)*
Therefore the absolutely continuous parts of the operators A,, A are unitarily
equivalent.
In applications the conditions of the Kato-Rosenblum
theorem (1957) are not
usually satisfied, however Kuroda (1960) discovered more effective conditions
for this purpose. They in turn are covered by a theorem of M.S. Birman and M.G.
Krejn (1962), in which the resolvents of the operators A,, A are required to be
nuclear. It is even sufficient that the difference of the resolvents of q5(A,) and &A)
be nuclear, where the function 4 belongs to a certain sufficiently wide class (M.S.
Birman, 1962).
4.9. Spectral Operators.
Functions of a selfadjoint operator are normal operators. However the spectrum of a normal, for example unitary, operator is not in
general real. If
@=
w 44 6,
s --Q,
(31)
63The Heisenberg S-matrix is unitarily equivalent to this operator by means of the Fourier transform
q9H $ (i.e., from the physical point of view, the transition from the coordinate space into the impulse
space).
64The case of commuting A, A,, is not of interest in scattering theory.
8 4. Theory of Operators
201
where el. is an orthogonal resolution of the identity, then spec @ is the closure
of the set of essential values65 of the function 4 (spectral mapping theorem). The
spectral resolution (31) is written more naturally in the form
@=
scPd&I,
(32)
where J d@, is the operator-valued measure in the complex plane C which is the
image of the measure J de, under the mapping 4. The support of the measure in
(32) coincides with the spectrum of the operator @.
The spectral resolution of a normal operator in the form (32) provided the
starting point of the theory of spectral operators on a Banach space which was
established by Dunford and Bade in the 1950s. In this theory it is postulated that
there exists a resolution of the identity G(M) which is a projection-valued
measure on C (E(C) = l), commutes with a given closed densely defined operator A
and is such that spec(Al ,,,,eo,J c M (in this case the operator A is said to be
spectral). The resolution of the identity for any spectral operator is unique and
if A is bounded66 then
A=
where R is a quasinilpotent
sc
Aa
+ R,
sc
Mqdil)
(33)
is also a spectral operator with the same resolution of the identity, spec A = spec A
and moreover the whole spectrum is approximate.
We say that operators of the form (33) are of spectral type. Therefore n is called
the scalar part of the operator A and R is its quusinilpotent part. The representation of a bounded spectral operator as the sum of an operator of scalar type and
a quasinilpotent operator is unique.
The class of spectral operators is rather narrow. For example, the shift operator
on the two-sided lp space (1 < p < co; p # 2) is not spectral (Fixman, 1959). The
Id
i ds
operator - - on Lp( - co, co) or on Lp( - rc, 7~)with periodic boundary conditions
also fails to be spectral if p # 2. We shall describe below a much wider class of
operators (those with so-called separated spectrum).
4.10. Spectral Subspaces. Let A be a closed densely defined operator on a
Banach space E. A closed subspace L c E (L # 0) is said to be spectral for A if
65A value LXis inessential if q(L) # a almost everywhere with respect to each measure m, = j d&x, x).
66For unbounded spectral operators the situation is considerably more complicated than that
described below.
208
p = - &
s t-
R,dl,
where I is a simple closed contour separating Q from its complement Qin spec A.
Moreover 1 - P is the projection onto the spectral subspace corresponding to
the compacP set Q This construction is contained in the functional calculus of
F. Riesz, which we shall describe below under the assumption that A is bounded.
Let us put
67This term requires a more precise definition since A need not be defined everywhere on E.
Inuariance
of a subspace L means that if x E D(A) n L then Ax E L. The restriction Al, has to be
interpreted as the restriction of A to D(A) n L. It is clearly closed along with A, however D(A) n L
may fail to be dense in L when D(A) is dense.
68Thus the spectral subspace L = Im P is completely reducing. If Q consists of a single point n and
L is finite-dimensional, then p is an ordinary eigenvalue and L is the corresponding maximal root
subspace.
6qThe neighbourhood mentioned and the contour r can be disconnected. The integral does not
depend on the choice of contour.
209
For the calculus just described7 Dunford (1943) established the spectral mapping theorem, spec 4(A) = 4 (spec A), and the superposition theorem: ($0 Q)(A) =
$@(A)) under the condition that Q is analytic on spec A and $ is analytic on
hvec 4.
If the function qj is analytic on the disk 1II < r, where r > p(A), then, expanding
it in a Taylor series and using the expansion (23), we obtain from (34)
i.e. we arrive at the previous definition of the function from the operator.
We shall say about an operator A on a Banach space E that its spectrum is
separated if each compact set Q c spec A which is the closure of its set of interior
points relative to spec A is spectral. For example, any spectral operator is of this
type: the spectral subspace L(Q) is the image of the projection
CZQ
=s
@(dl).
IM
G &W,
C))
(A $ Cl,
(35)
In In M,(6) da < co
s0
(36)
is satisfied. Then A is an operator with separated spectrum (Yu.1. Lyubich V.I. Matsaev, 1960). In the proof of this theorem a special functional calculus is
used, with the help of which quasiprojections71 onto the spectral subspaces are
constructed. It also allows us to establish (under conditions (35), (36)) the completeness of the system of spectral subspaces corresponding to any covering of
the spectrum by finite arcs of the curve C.
Example. The Schrodinger operator on L2( - 00, co) with periodic complex
potential has separated spectrum (V.A. Tkachenko, 1964).
For a bounded operator A with real spectrum Levinson condition is equivalent to Ostrowski condition with respect to the weight a(t) = lleiArII (t E R):
his can be developed further in the case of a spectral operator.
These are operators Q, associated with a compact set Q and a neighbourhood U = Q such that
@ILCpi= 1 and cD~,~) = 0 for compact sets K lying outside of U.
certain heoryof duality also holds for spectral subspaces under conditions of type (36) (Bishop,
1959; Yu. I. Lyubich, V.I. Lomonosov, V.I. Matsaev, 1973).
210
(37,)
m qb(t)eeiA dt
(4 E mw)
f -02
independently of the previous approach; this is the operator Fourier transform,
which allows us to obtain functions of compact support of the operator A, in
particular quasiprojections onto spectral subspaces. Moreover L(Q) is constructed
as the intersection of all the subspaces
{x: J(A)x
= x}
(($1, = 1).
The requirement of boundedness of the operator A is not essential, the important thing being rather that it generates a group74 T(t) (t E W). The stated
conditions of growth of the resolvent or exponential cannot be weakened: under
any more rapid growth rate there arise operators for which the spectrum on any
closed invariant subspace coincides with the spectrum on the whole space.
For a bounded operator I/ with spectrum on the unit circle the requirement
cm ln IIv I
n=-m n 2 +l
(37,)
211
(,I,& such that spec(AILk) c Gk (1 < k s n) and c;=i L, = E. The theory of this
class of operators has been developed in depth during the last 25 years in the
works of Foias and his school.
4.11. Eigenvectors of Conservative and Dissipative Operators. Completeness
of the system of eigenvectors (eigenspaces) of a conservative operator A on any
Banach space E is equivalent to the almost-periodicity
of the corresponding
isometric representation
T(t) (t E R).
The necessity of almost-periodicity
is obvious: if (xk}y is a collection of
eigenvectors corresponding to the eigenvalues {i&}y and 11~- En x,II < E, then
T(t)x - k xkeiAkf -c E
(t E R).
k=l
II
/I
Hence it is clear that the orbit of the vector x is relativly compact.
The sufficiency of almost-periodicity
of the representation T for completeness
of the system of eigenvectors of the operator A can be established with the help
of the theory of representations of compact groups.* In particular, the spectral
theorem for compact selfadjoint operators is generalised by this method: if A is
a compact conservative operator on a reflexive Banach space E then its system
of eigenvectors is complete (Yu.1. Lyubich, 1960). The reflexivity condition can
be relaxed: it is sufficient that E should contain no subspace isomorphic to c.
Moreover there is a counterexample in c (Yu.1. Lyubich, 1963).
Example. Let (A,); be a sequence of distinct positive numbers which converges to zero. Let us consider the diagonal operator
4L)
on c. It is compact and conservative:
e
= (%t)
C) = (e n&).
However its eigenvectors S,,,= (6,,,,), (m = 0, 1,2, . . .) only generate the subspace c0 of sequences which converge to zero.
Remark. The ergodic theorem for an almost-periodic representation T(t) (t E R)
follows easily from the completeness of the system of eigenvectors of the generator A: the existence of the limit
1 S
GW.4 dt
5! s s -*
212
Pn is an orthoprojection)
(in the case of a regular representation in AP(R) this is the classical Fourier series
of an a.p.f.) The uniqueness theorem holds: Vk P2x = 0 * x = 0. The proof reduces to the scalar case if we consider the a.p.f. f(T(t)x) (f~ E*) since the Fourier
coefficients of this function are c1 = f(P,x).
Using the FejCr-Bochner method of summation for Fourier series of a.p.fs we
can associate with each subgroup I c T a projection Pr (with bound llP,-ll < M)
on the closure of the sum of the eigenspaces corresponding to eigenvalues 1 E I
(N.K. Nikol kij, 1971). The replacement of subgroups by arbitrary subsets is not
possible (Rudin, 1962).
Thus once again we have evidence of a close connection between harmonic
analysis and the spectral theory of operators. There is however another approach
to the problem of completeness of the system of eigenvectors of a conservative
compact operator, which is based on an analytic technique using a bound for
the resolvent (even weaker than (26)).
We say that an operator A on a Banach space E is quasiconservative if spec A
is purely imaginary and
M
IIJUI G m
5 4. Theory
of Operators
213
the eigenspaces. Hence it follows that any normalised system {e,} of eigenvectors
corresponding to pairwise distinct eigenvalues is distal:
m#n
(38)
normalised eigenvectors
is a Bari basis of
Proof. Without loss of generality we can assume that the sum of the series (38)
is equal to q2 c b. Since im(A(au, + /?u,), au,,, + flu,) 2 0 for all a, B, we have
Ikl~ %)I2 G
(m # n).
Hence by (38)
In addition
- ;11)+
(im 1 > 0)
is a contraction and also Im(V, - 1) = D(A) is dense. Conversely, any contraction V for which Im(V - 1) is dense (since Ker(V - 1) = 0) is the Cayley transform of the dissipative operator iA = (TV - Al)(V - 1)-l. This remark allows
us to transform theorems on dissipative operators into analogous theorems on
contractions and conversely. Thus, for example, we have the analogue of the
Mukminov-G&man
theorem for contractions with condition
214
- IUZ)<co
c (1- ILIZ)(l
11 - n,X z
(39)
m#n
which results from replacing Iz,, 1, in (38) by their imagesso with respect to the
fractional linear transformation I = -i([ + l)(c - 1)-l.
Remark. Condition (39) is only meaningful for eigenvalues which lie in the
disks1 111 < 1. This is not surprising if we take account of the theorem of
Sziikefului-Nagy, Fobs and Lunger on the decomposition of an arbitrary contraction into an orthogonal sum of a unitary operator and a completely non-unitary
(i.e. it has no unitary parts) operator.82
A similar remark can be made with regard to dissipative operators.
Conditions (38) and (39) can be expressed in terms of Blaschke products:
t38) *
jx (s(
0;
A -I, >o;
I I
inf n m
m mnZnI, - I,
(39) * 8 ( f-T,
; / > 0.
ln n
44-Al >0
I n,i, I
inf n
m rn#rl l-
0 4. Theory of Operators
215
4.12. Spectral Sets and Numerical Ranges. The theorem of von Neumann
which has been mentioned is concerned with an aspect of spectral theory which
was discovered by him; this is connected with subtle variants of functional
calculus. We will discuss this topic briefly. Let A be an operator on a Banach
space and S a closed subset of the complex plane C which contains spec A. The
set S is said to be spectral for A if the bound II&t)II
< sup,lql holds for all
rational functions cpwhose poles lie outside of S. Von Neumann
theorem asserts
that the disk 111< 1 is a spectral set for all contractions on a Hilbert space.83
This is false on a Banach space 84, however, as V. Eh. Katsnel on and V.I.
Matsaev (1967) showed, it becomes true for the disk I II < 3 and the radius p = 3
cannot be decreased for the class of all Banach spaces.
For a normal operator on a Hilbert space its own spectrum is a spectral set
(in consequence of the spectral theorem). Conversely, if any function which is
continuous on (spec A) u {co} can be approximated uniformly on this set by
rational functions, then the operator A is normal.
Another useful substitute for the spectrum of an operator A is its so-called
numerical range. In the case of a Hilbert space it is defined as the closure of the
set of values of the quadratic functional (Ax, x) for x E D(A), llxll = 1. According
to a theorem of Hausdorff the numerical range of any operator A is convex.s5 It
contains the approximate spectrum and consequently the convex envelope of
this set, but in general the numerical range is more extensive than the convex
envelope of spec,A. For a normal operator these two sets coincide.
The numerical range of an operator A on a Banach space is defined to be the
closure of the set of values j&4x), where x E D(A), I/XII = 1 and f, is a support
functional at the point x (the mapping x H f, is said to be supporting, as in the
smooth case, but now, in general, it is not uniquely defined). The following
example (Bonsall-Duncan, 1980) shows that the numerical range is not generally
convex: for the space C(S) (S compact) with f, = 6,(,,, where s(x) is any point at
which maxS Ix(s)! is attained, the numerical range of the operator defined by
multiplication
by a function n(s) coincides with the set of values of this function.
The numerical range depends on the choice of support mapping but, if A is a
bounded operator (D(A) = E), all its numerical ranges have the same convex
envelope (Lumer-Phillips,
1961). This set coincides with the intersection of the
half-planes re([e< r(e), where r(0) (0 < 8 < 24 is the minimum of those r
for which IleAc(I G efKl (arg [ = 0), i.e. l&II G l/(lcl - r(0)) (arg 5 = 8, I[1 > z(0)).
The convex envelope of the spectrum is contained in the numerical range.
In all cases the numerical range of an operator A contains its approximate
spectrum.
* onversely, if the unit disk is a spectral set for an operator A (even on a Banach space) then A
must be a contraction. This follows trivially from the definition with q(A) = A.
* f the unit disk is a spectral set for all contractions on some Banach space then the space must be
Hilbert (Foias, 1957).
ssIt is enough to establish this result in 2-dimensional Euclidean space. But in this case the numerical
range is an ellipse (which can degenerate into an interval) or a point (for a scalar operator).
216
O
b <*
for some p > 0.
(41)
Condition (41) is equivalent to the requirement that S belongs to the class GP.
We shall clarify this from a certain more general point of view.
Let T by any compact operator on a Hilbert space E. Then T* T is a selfadjoint
non-negative compact operator. Its positive eigenvalues can be enumerated
according to multiplicity
in decreasing order*: r1 2 rZ > .** . The numbers
s,(T) = 4~1, SZ(T) = ,/rz, . . . are called the singular numbers (s-numbers) of the
operator T (clearly s,(T) = II T II). It turns out that they coincide with the anumbers (D. Eh. Alakhverdiev, 1957). The class GP for a Hilbert space is therefore
determined by the condition
is equivalent to (41).
5 4. Theory of Operators
217
T=cdT)(., GL,
where (u,), (u,) are two orthonormal
systems. The first consists of the eigenvectors of the operator A, = m
while the second is made up of those of the
operator A, = JTT*.
This result
- is closely
- connected
- with polar representations:
T = U,A, = A,&, where U,: Im T* + Im T, U,.: Im T + Im T* are invertible
isometries.
Moving on to a concise account of the proof of Keldysh theorem, we recall
that $, is a two-sided (not closed) ideal in the algebra 5?(E) of all bounded
operators on the Hilbert space** E.
We denote by P the orthoprojection
onto the closed_ linear hull of the system
Z,(A) = Z(A) and put Q = 1 - P. Then the operator A = QAQ = QA turns out
to be a Volterra operator and therefore its Fredholm resoluent F(c) = (1 - CA -
is an entire function of c. Moreover 2 E G,,. We can deduce from this that
F(c) has growth of order no higher than p and of minimal type as l[l+ co:
In 11
F(C)11 = o( Ill . On the other hand, using the given structure of the operator
A, we can show that IIF(LJ/ is bounded on the pair of angles E < (arg (1 < x - E.
Boundedness on the pair of angles larg 51 < E, n - E < larg cl < n follows from
the Phragmen-Lindelof
principle. Finally, F(c) is bounded on the whole plane.
By Liouville
theorem F(c) is constant. Consequently 1 = 0, i.e. QA = 0, which
establishes the required inclusion Im A c Im P = Lin Z,(A).
Corollary. If S, R satisfy
B = (1 + R)S is complete.
the conditions
A which is defined on
*sThe only non-trivial closed two-sided ideal in f?(E) is the ideal of compact operators (theorem of
Calkin).
218
1
k=O
{ajkytk'(a)
let the
bjky'k'(b))
'
(1 <j < 4.
rincipal part
lo=
ff
(
>
define a selfadjoint operator H and let both operators A, H be injective (the latter
does not restrict generality). We put S = HThis is a compact selfadjoint
operator. Its eigenvalues have the form 1= io where o runs through the set 52
of roots of the characteristic determinant
A(o) = det(@j[ei
]);,kzl
of Keldysh
But
is a Hilbert-Schmidt
integral operator and is therefore compact. Consequently
A- = S(l + R), where R is a compact operator. Applying Keldysh theorem we
establish that the system of root functions of the operator ~4 is complete in
L a, b).
5 4. Theory of Operators
219
+ A*),
im A zf &A
- A*).
220
is not only necessary but also sufficient for completeness of the system J&4) (M.S.
Livshits, 1954). In fact, for the invariant subspace E, = Lin C(A) the trace of the
operator re(AIE1) is given by (42). If E, # E then E, E Ef # 0 and then for the
orthoprojection
P on E,, we will have tr(P* re(Al,,) P) = 0; hence re(Al,,) = 0
since re A < 0. Consequently ( l/i)A(eO is selfadjoint (E, is invariant for A*). Since
A IEOis compact it must have an eigenvector - a contradiction.
It is clear from what has just been discussed that the inequality
zt re J., 2 We -4
holds for any compact dissipative operator with nuclear real component. The
system of root vectors of a nuclear dissipative operator is complete (V.B. Lidskij,
1959) since Inn, = tr A and In& = tr A*.
In conclusion we note that completeness of a compact operator is not in
general inherited by its restrictionsg3: any Volterra operator on a Banach space
is the restriction of a complete compact operator acting on some larger Banach
space (N.K. Nikol kij, 1969).
4.14. Triangular Decompositions.
The theorem on the existence of a (nontrivial) closed invariant subspace for any compact operator A on a complex
LCS E (dim E > l), with the help of Zorn lemma, leads to the existence of a
chain of such subspaces which is maximal with respect to inclusion. Any such
chain begins with zero and terminates with the space E. In the finite-dimensional
case it has the form 0 c Lo c L1 c *.* c L, = E, where dim Lk = k (0 < k <
n = dim E). Any system of vectors (~1; chosen so that eL E L,\Lt-I
(1 < k < n)
is a basis in E with respect to which the matrix of the operator is upper triangular.
In the infinite-dimensional
case the chain Z is not in general discrete (for the order
topology) and therefore it cannot generate a basis for a triangular decomposition.
The latter has to be connected directly with the chain Z. For each non-zero
subspace M E Z we consider the closure M- of the union of all those L E Z which
are contained in M but are distinct from it. Then M- E Z since the chain Z is
maximal and clearly M- c M. If M- # M, the factor space M/M- is onedimensional. Those M for which this holds are called points of discontinuity of
the chain Z. In the absence of points of discontinuity the chain Z is said to be
continuous.
If M is a point of discontinuity,
931f an operator is complete along with all of its restrictions, then we say that it allows spectral
synthesis.The problem of spectral synthesis goes back to a work of Wermer (1952) which is concerned
with normal operators. The general situation was studied by A.S. Markus (1970).
8 4. Theory of Operators
221
V,dW
=&s6-V-W)
dt
F(P) dP
sW
222
directed by the partitions PO < PI a* * < P., where all the Pk, Qk belong to the
chain W and Pk-1 < Qk < Pk (1 < k < n). It is clear that not every function F is
integrable in this sense.
If A is a Volterra operator, 2 is a continuous maximal chain of closed invariant
subspaces and W is the corresponding chain of orthoprojections,
then we have
the triangular decomposition (MS. Brodskij, 1961)
P(im A) dP
A = 2i
sW
(the existence of the integral is guaranteed). From this we extract a rich supply
of information on relationships concerning properties of the real and imaginary components of a Volterra operator; for example there is the theorem of
Sakhnooich that, if the imaginary component of a Volterra operator is a HilbertSchmidt operator, then so also is its real component; we also have the more
general theorem of Matsaev: if A is a Volterra operator and im A E E$ then
re A E $, (1 < p < co) and moreover N,(re A) < y,N,(im A), where the constant
yp depends only on p. In the same direction for p = 1 (i.e. for im A nuclear) I. Ts.
Gokhberg and M.G. Krejn (1967) established the deep spectral density theorem
for the operator re A:
lim
n+(R)
1s
-=R a,
N,(dP*im
AodP),
R-W
of
0 4. Theory of Operators
223
e)j,
where the channel vector e and the number j = + 1 are defined by the equation
im A = &(. , e)e. It turned out that for the given class of operators coincidence
of characteristic functions implies unitary equivalence. But for an operator n its
characteristic function can be calculated immediately; in fact with j = 1 (to be
snecific)
(43)
224
Aktk
i .=g+,
&&i&
(Pb)jflk
dt
(k=1,2,3
,... ),
s 0
the
* &--&
l-I-fz=l ;I - &
appears in (43). Developing this approach, MS. Livshits (1954) constructed a
triangular model for a completely non-selfadjoint bounded operator A with
finite-dimensional
imaginary component. In this case the characteristic function
turns out to be a matrix of order r = rk(im A) and the role ofj is taken by a certain
Hermitian matrix .Z such that Jz = 1. The characteristic matrix-function
w,(n)
(im A > 0) is J-expanding, i.e. W(L)JW,(n) > I, and moreover on the real axis
outside of the spectrum it is J-unitary: W(n)JW,(L) = I. The corresponding
analogue of formula (43) (with products of Blaschke type) was obtained by V.P.
Potapov (1950). Passing to the limit as I + cc we can extend the triangular model
further to the case where the imaginary component is nuclear.
It is impossible to get rid of the condition of being completely non-selfadjoint,
since the characteristic function does not change if an arbitrary bounded selfadjoint operator is adjoined orthogonally to A. However this clearly does not
matter from the point of view of classification theory. We can say that for
operators whose imaginary components are nuclear there is a valuable analogue
of Jordan theorem.
Incidentally, although the spectral resolution of a selfadjoint operator
5 4. Theory of Operators
225
component
of an operator is a
A, = ,/(l - AA*).
operator-function
W(A) = (-A
+ 1A,(l - AA*)-lb,)
I Im
is introduced in this context (Yu.L. Shmul an, 1953). The values of this function
are contracting homomorphisms
from I%& into Im A,. It is defined and analytic
on the domain (reg A*)in particular on the disk 111 < 1, where it is usually
considered. If the characteristic functions of completely non-unitary contractions
coincide, then these contractions are unitarily equivalent (A.V. Shtraus, 1960).
Szokefalvi-Nagy and Foias (1963) found a natural (in terms of characteristic
functions) functional model for an arbitrary completely non-unitary contraction
A. Of necessity it is rather complicated, since any contraction is the orthogonal
sum of a completely non-unitary contraction and a unitary operator and any
bounded operator becomes a contraction on normalisation.
The construction of the model is significantly simplified if the sequence of
powers (A ; converges strongly to zero (although even in this case getting
information depends essentially on the measure of non-unitariness, more precisely on the rank of the operator A,). In order to describe the model operator in
this situation we consider an arbitrary Hilbert space G and introduce the Hardy
class H G) of analytic vector-functions
u(A)= jJ t&A
n=O
which satisfy the condition
The shift operator lo3 (7%)(L) = Au(n) acts on it; it is clearly isometric (but not
unitary: Im T = (u: u(O) = O}). The adjoint operator
(T*u)(q
is a contraction
= @) ; No) = 2 u,+lA
n=o
u(A) = 2 U
n=O
>
ro2This construction arose in another way in joint works of Szokefalvi-Nagy and Foias (1962) and
for finite-dimensional A,, A, it is equivalent to the definition of M.S. Livshits (1950). A.V. Shtraus
(1959) extended the concept of the characteristic function to unbounded operators.
lo3 Strictly speaking, the shift is effected on the sequence of Taylor coefficients of the function u(A).
226
for A. The
In fact
11Vxl12 = 2 (AJ ,
n=O
A$
i.e. V maps E isometrically into H(G). In addition it is clear that T*V = VA.
Therefore the subspace L = Im V c H2(G) is invariant for T* and T*I, is
unitarily equivalent to the operator A.
If rk A, = 1, the model operator is the adjoint of a shift on the scalar Hardy
class HZ. It is possible to describe all closed invariant subspaces for this operator
(and hence also those for the initial operator A) (Beurling, 1949). Each such
(non-zero) subspace has the form
8H2 = {cp: cp = e$ ($ E HZ)},
where 0 is a fixed function in H2 whose boundary values on the unit circle have
modulus one almost everywhere (an interior function).
The general approach to models of Szokefalvi-Nagy and Foias is based on
the study of the so-called dilatations 2 of the operator. This term is applied to
any operator A which acts on some Hilbert space E 2 E and which is associated
with A by means of the relations
A=pAI
where P is an orthoprojection
Theorem of Sziikefalvi-Nagy.
(n = 1,2, 3, . . .),
on Ewith Im P = E.
For any contraction A on a Hilbert
space there
where gA = I ,, 6, being an orthogonal resolution of the identity in a larger Hilbext space k and
P an orthoprojection on E onto E.
227
x0 0 Xl 0 x2 CD~~~~Ax~@A,x~~x~~x~~....
= 1 - iJF*R,(A)T
4.16. Indefinite
selfadjoint
lo5 From nternal applications of this theory we note that of necessary and sufficient conditions for
the existence of Riesz and Bari bases composed of root vectors of a contraction on a Hilbert space
(N.K. Nikol kij - B.S. Pavlov, 1970).
ro6Among the channels of communication are entrance and exit channels. Successive unions of
systems are effected by connecting the exits of one system to the entrances of another. This procedure
can be formalised in terms of what has been described.
228
or J-metric,
pseudodistance
P+ + P- = 1;
P+P- = P-P+ = 0.
-x-,
orthogonal
229
and suficient
230
(44
231
5 4. Theory of Operators
(45)
(46)
where C = V-2TV2,
B = V-2RV12. The operators C and B are usually
bounded in applications and after closure they satisfy the conditions formulated
above. Exponential solutions u = uOeQf arise as eigenvectors of the quadratic
pencil of operators S(p) = Cp2 + BP + 1, S(p)uO = 0. If X is an operator root
of equation (44), 1 is any of its eigenvalues (2 # 0) and u0 is a corresponding
eigenvector, then u,, is an eigenvector of the pencil S for the eigenvalue p = 1-l.
A root X gives rise to another root Y = -X* - B, which is also a source of
eigenvalues and eigenvectors of the pencil S. If we now put Z = -X - B, the
pencil factorises:
S(P) = (1 - Pm
- P-w.
(x # 0).
It is easy to see that in this case all the eigenvalues of the pencil are non-negative.
Taking account of multiplicity,
we shall denote by (-0;) those of them which
are generated by the operator X and by (- 0:) those generated by the operator
Y. It turns out that the pencil has no other eigenvaluesl
(and there are no
other roots of equation (44)). Each of the corresponding normalised systems of
eigenvectors {u:}, (0:) is a Riesz basis in H. In this sense the system of all
eigenvectors of the pencil is doubly complete.
The phenomenon of multiple completeness of the system of root vectors of a
polynomial operator pencil had already been discovered by M.V. Keldysh (1951).
Since then the theory of polynomial pencils has developed into an
And even no other points of the spectrum, i.e. the operator S(p) is continuously invertible for p
different from (-CD;), ( - u$ .
232
Chapter
2. Foundations
and Methods
5 4. Theory of Operators
233
Returning to an arbitrary algebra $I, we note that for each closed ideal J c 2I
the factor algebra 2X/Jis a Banach algebra under its usual norm. But all maximal
ideals are closed since the closure of an ideal J # !!I is an ideal .i # $I, as a
consequence of the fact that the set of invertible elements is open. The factor
algebra E = %/M for a maximal ideal M is therefore a Banach field over @. If
E # @ then for any < E E\@ and for all 1 E @ the non-zero elements r - I are
invertible in,E. But then spec 5 = 0. With this we have established the Gel&&Mazur theorem that a complex commutative Banach algebra is spectral.
The characters x of an algebra +Iare continuous because of the closedness of
their kernels, these being maximal ideals. Since x(a) E spec a, we have Ix(a)1 <
p(a) < Ilull. Moreover x(e) = 1 and so llxll = 1.
Now let us consider the set %R= 9R(%)of maximal ideals of the algebra 2I. It
can be regarded as a subset of the unit sphere of the conjugate space ??I*,
in which case it is closed since it is determined by the system of equations
&,a,)
= &,)&)
(al, a, E 2I), x(e) = 1, which are defined by continuous functions of x. Consequently, %I is compact.
Example 1. If ?I = C(S) (S compact) then !Ul@I) = S.
Example 2. If 2I is an algebra with a single topological generator a, i.e.
8 = [a], the uniform closure of the algebra of all polynomials in a, then !IR(%) =
spec a.
We note that the problem of giving an explicit description of the maximal ideal
space of a concrete algebra can be extremely difficult. An example of this is the
famous problem which was solved by Carleson (1962); this is concerned with
the algebra of functions which are analytic and bounded on the unit disk
A = {A: 1 E @, IL1 < l}. The result is that A is dense in 9X.
The canonical functional representation (Gel and transformation)
aH Z
(ii(x) = x(u)) is a morphism118 of 9I + C( I(2I)) which is clearly contracting.
For it to be an isometry (i.e. for the identity p(u) = [lull to hold), it is necessary
and sufficient that the identity llu2 II = IjaIl should hold in %.
The Gel and image of the algebra
Iin C(9R) is a subalgebra (generally not
closed) which separates the points of the space 9.X It can fail to be dense in C(!U2).
In the given context and also for many applications it is important to have an
intrinsic characterisation of dense subalgebras of C(S) (S compact). The effective
solution of this problem includes the Weierstrass approximation
theorem (the
subalgebra of polynomials is dense in C[O, 11). It was obtained by Stone (1937).
Stone-Weierstrass Theorem. Suppose that a subalgebra 9I c C(S) is symmetric
(i.e. along with each function cp it contains its complex conjugate @) and separates
the points of S. Then !JI is dense in C(S).
For the proof (and also independently of this goal) it is useful (following de
Branges, 1959) to deal first of all with the real case.
sIn the category of complex commutative Banach algebras over C (i.e. a continuous algebra
homomorphism) YJI(%)is a contravariant functor into the category of compact spaces.
234
If the algebra
is symmetric,
is said to be symmetric
if its
wJw)).
Remark. The Banach space C,(S) is also a vector lattice with respect to the
usual ordering and moreover the two structures are compatible: if [lx/l < 1 and
IyI < 1x1 (I.1 is the lattice modulus), then llyll < 1. We can say that C,(S) is a
Banach lattice. The following variant of the Stone-Weierstrass theorem holds:
any sublattice of the vector lattice C,(S) which contains 1 and separates the
points of the compact set S is dense in C,(S).
Closed subalgebras of the algebras C(S) (S compact) are called uniform algebras.
An example of such an algebra on the unit circle T is the algebra A(T) offunctions
which are the boundary values of analytic functions on the unit disk A. It is called
the disk algebra and it is maximal among the non-trivial closed subalgebras of
C(T) (Wermer, 1953). The maximal ideal space of the algebra A(T) is naturally
identified with the closed disk A = A u T. In general, if a uniform algebra 2I on
S separates points, then S is homeomorphically
embedded in the maximal ideal
space %Jl(%).
For any algebra
I and any element a E 2I we can construct, just as in the
theory of operators, a functional calculus by putting
5 4. Theory of Operators
235
cpw
=-&srp(l)(a
r
le)- dA,
The point is that, because of the regularity of the algebra, for any finite open
covering of the compact set !JJI we can find a partition of unity as a sum of functions
in a, each of which is concentrated on the corresponding element of the covering.
Example. If on a neighbourhood
of each point a periodic function cp can be
expanded as an absolutely convergent trigonometric
series (depending on the
choice of point), then it has an expansion of this type on the whole period (Wiener,
1933).
An important consequence of the local theorem is a further generalisation of
the Wiener-Levy theorem: suppose that the algebra 9I is regular and that the
function rp on !DI is such that for each point M,, E 2Xthere exists a neighbourhood
U, an element a,, E 2X and a function
. which is analytic on spec a0 and such
that q(M) = cpo(iio(M)); then cp E ai.
In certain cases this theorem allows us to apply multivalent analytic functions lg to elements of the algebra.
Example. If a function
satisfies the conditions
1 - A(1) # 0 (n E R),
transform
A(1)
ind(1 - A(I)) = 0,
1I9 An interesting variation on the topic of multivalent functions of an algebra element is the solution
of algebraic equations with coefficients in the algebra. Non-trivial homomorphisms of the fundamental group nl(%R) into braid groups are global barriers to solvability (E.A. Gorin - V.Ya Lin, 1969).
236
Chapter
2. Foundations
and Methods
then the function ln( 1 - A(1)) is the Fourier transform of some function in L(R).
The role of this result in the theory of equations with a difference kernel is
well-known (see Chapter 1, Section 4.5).
By way of explanation we state that for the Banach algebra L(R) (with identity
adjoined) the Gel and transform and the Fourier transform coincide (1 = 1). The
maximal ideal space of this algebra (or of the Wiener algebra which is isometrically, isomorphic to it) is naturally identified with the one-point compactihcation Iw = IR u (co}.
In any algebra Cuan element a is said to be logarithmic if it can be represented
in the form a = eX, where x E . All such elements are clearly invertible but in
general not all invertible elements are logarithmic.
Example 1. In the algebra C[O, l] all functions that are never zero are logarithmic, i.e. the functions which are invertible in the algebra.
Example 2. In the algebra C(T) the logarithmic
non-vanishing functions cp for which
In the proof of the Arens-Royden theorem (as also in many other questions in
the theory of Banach algebras) we make use of analytic functions of n algebra
elements: a,, . . . , a,. In the corresponding functional calculus12 the role of the
spectrum is played by the combined spectrum spec a of the system a = (ai, . . . , a,),
which is defined as the complement of the combined resolvent set. The latter
consists of regular points (A,, . . . , A,) E C for each of which there exists a system
b = (b,, . . . . b.) of algebra elements such that
= e.
It is easy to see that spec a coincides with the image of the mapping
lzoIt
was developed
in the works
of G.E. Shilov
(1953,
and Calderbn
ii: W + C
(1955).
4 4. Theory
of Operators
231
which is defined by the formula a(M) = (5,(M), . . . , c,(M)). Since this mapping
is continuous, spec a is compact. If it is polynomially
convex, then, by the
well-known Oka-Weil theorem, each function which is analytic on spec a admits
arbitrarily close uniform approximation by polynomials. This allows us to extend
by continuity the intrinsic construction of the element p(a) = rp(a,, a,, . . . , a,)
from the algebra of polynomials in n variables to the algebra of functions which
are analytic on the spectrum. A deeper analysis is required in the absence of
polynomial convexity for the combined spectrum.
Returning to regular algebras, we note their significance for spectral theory.
Example.
In a,
C 7
< *.
n.zn
+1
Then the Banach algebra 1; (with convolution) of scalar sequences that are
summable for the weight a is regular (G.E. Shilov, 1947). Its Gel and image, the
Wiener algebra W, with weight a, consists of those functions on the circle with
Fourier coefficients in 1, The corresponding maximal ideal space is identified
with the unit circle T.
If V is a non-quasianalytic
linear operator on a Banach space (spec V c ),
we can put a, = 11V 11.The functional calculus
q(V)=
II=-al
a,V
q(l) =
-al
a&
111= 1
>
arises on the algebra W,. The spectrum of the operator turns out to be separated
thanks to the fact that, for any compact set Q c T and any neighbourhood
U 3 Q, there exists a function rp E W, such that lo = 1, cp(T\ U) = 0,O < cp < 1
(clearly q is not only non-analytic but it also does not belong to any quasianalytic
class).
We thus have convincing evidence that the methods of the theory of analytic
functions are of paramount importance for the theory of Banach algebras (as
indeed for spectral theory in general). Conversely, the theory of Banach algebras
is a useful instrument for studying various classes of analytic functions, particularly in the setting of approximation
theory. The concept of a boundary of the
algebra
plays an important role here. This term is applied to any subset X of
the maximal ideal space W such that for each function u(M) (a E 2I) its maximum
modulus is attained at some point A4, E X. The trivial boundary is X = W and,
for example, in the case of the algebra C(S) (S compact) it is the only one. The
boundaries for the algebra A(T) are precisely those subsets of the disk A which
contain the circle T. Since in any algebra an extension of a boundary preserves
the boundary property, it is natural to look for a smallest boundary. G.E. Shilov
(1940) proved that any algebra has a smallest closed boundary (subsequently
called the Shilou boundary). Without the requirement of closedness a smallest
boundary may not exist, however such a boundary does exist if the maximal ideal
space is metrizable and in this case it consists of the peak points, i.e. those MO E 9R
(
238
A.A. Gonchar (1963) showed that the following condition is suflicient for a
point s0 of a compact metric space to be a peak point of a uniform algebra
9I c C(S): there exist constants c < 1, N > 1 such that for any neighbourhood
U of s0 there is a function cpE % for which cp(s,) = 1, Iv(s)] < c (s E U) and
llqll < N. It follows from this that the set of peak points of the algebra ?I is of
type G.
Let % be a complex Banach algebra (in general non-commutative).
A mapping
a H a* of the algebra into itself is called an involution if 1) (a + b)* = a* + b*,
2) (au)* = Za*, 3) (ab)* = b*a*, 4) a** = a, 5) Ila*II = Ilull. The element a* is
called the adjoint of a. If a* = a, then a is said to be selfadjoint. If a* commutes
with a, we say that a is normal. In particular, all unitary elements (a*a = aa* = e)
are normal.
Example. The identity e is a selfadjoint unitary element.
If the element a is invertible, then a* is also invertible and (a*)- = (a- *.
Hence it follows that reg a* = i%gZ and spec a* = spec. 1
Any x E rU can be uniquely represented in the form x = a + bi, where a, b are
selfadjoint :a = re x = (x + x*)/2, b = im x = (x - x*)/(2i). The selfadjoint elements of 2I form a closed real subspace re 2I.
For any a E $I the element a*a is selfadjoint. The closure of the non-negative
linear hull of the set {x: x = a*a (a E $I)} is a wedge %Xx,whose elements, as usual,
are said to be non-negative (x > 0). Using $I+ we define in the canonical way the
collection 9Iz*, of non-negative linear functionals. Any f E !!I*, satisfies the inequality If(a)1 < f(e) /la/l and is consequently continuous ( 11f II = f(e)). Moreover, it is
real on re %or equivalently: f (a*) = f(a) (a E 2l). Such functionals are said to be
symmetric. Any functional g E (re $I)* can be uniquely extended to a symmetric
functional g E %I*; also II 811= II 911.
In the category of Banach algebras with involution the continuous homomorphisms h: VIZ,+ a2 which are compatible with the involutions, i.e. (ha)* =
ha* (or, equivalently, h preserves selfadjointness of elements), are called morphisms ((*)-homomorphisms).
A subalgebra of an algebra with involution is said to be symmetric if the
involution does not leave it. The class of symmetric ideals (one- or two-sided) is
defined similarly. Under factorisation by a symmetric two-sided (one-sided) ideal
the involution carries over naturally to the factor algebra (factor space).
Here the bar denotes complex conjugation.
5 4. Theory of Operators
239
is called a C*-algebra if
lb*4 = 11412
for all a E 2I. We mention
involution
(47)
is complex conjugation).
IIA*AII = ,vl
x
(A*&
4 = ,yl
x
space
lbw2 = llAl12.
and we will
240
(cp
*bwd
=scpb+knJMY)
(9
EG)
G
4(x)
=sddx(d
44s)
(x
EG*)
G
(if G is not discrete, the group algebra has another character - the coefftcient of
the identity). The maximal ideal space of the group algebra coincides with
G* if G* is compact (0 G is discrete), while in the general case it is the onepoint compactification of G*. The Gel and representation for the group algebra
L G, ,u) coincides with the Fourier transform, i.e. for G = [wwe have the classical
Fourier transform and for G = T it associates with each function its sequence of
Fourier coefficients. We have correspondingly T* = Z and R* = R. The group
Q4. Theory
of Operators
241
i2
suitably
ormal&l
and correspondingly
also a generalisation
Haar
measure
p* on G* appears
in the inuersion
formula
in the generalised
Plancherel
theorem. In this context it is appropriate
of Bochner
theorem on the representation
of positive functions:
to note
CPM
= x(g)WA
where u is a finite measure (a > 0).
The
original
proof (L.S. Pontryagin,
1934) was based on structure
theory.
The requirement
contained
in it of a countable
base was subsequently
removed (van Kampen,
1935).
242
where for fixed t C(s, r; t) is a Bore1 measure with respect to s and r separately
and for fixed subsets in place of s and r it is a continuous function of t with
compact support; it is required in addition that
s
s
involutory homomorphism
s H s* acts on S, preserving the measure m and
acting on the structure measure so that an involution is defined on the Banach
algebra under consideration by the formula q*(t) = @(t*). From these standpoints, for example, ordinary (group) and generalised (hypergroup) shifts can be
given a unified treatment.
fj 5. Function
Spaces
5.1. Introductory Examples. The development of functional analysis is connected not only with the working out of its bstract intrinsic structures but also
with the expansion of the stock of concrete function spaces. This has both an
applied and a conceptual meaning. The revolutionary introduction into mathematical usage of the language and techniques of generalised functions can serve
as a striking example (S.L. Sobolev, 1936; L. Schwartz, 1950-51). The general
theory of interpolation of linear operators, which came into being right at the
0 5. Function Spaces
243
au& t)
-=
at
a2u(x, t)
ax2
(t
4% 0) = 44x).
is initially set in the class of all sufficiently smooth functions but in this situation
the solution is not unique. In fact in the non-quasianalytic
class C(n
)
(0 < E < 1) on the interval [0, l] there exists a function t+b(t)which together with
all its derivatives vanishes at the end points. Let us put $(t) = 0 for t > 1. Then
cm i) (t)
4% t) = x0 (2n)IxZ
is a solution of the heat equation and U[~=~ = 0. The reason for this phenomenon
lies in the possibility of rapid growth of u(x, t) as [xl+ co. Taklind (1937) showed
that if h(x) is a non-decreasing odd function then the bound
lu(x, t)l < A4exh@)
defines a uniqueness class if and only if
* dx
-=
s 1 h(x)
al.
(48)
Au = 0,
equation
4r = cp
244
Chapter
2. Foundations
D[u] =
sG
and Methods
jVuj2 dx
lbll = JtDbl
with norm
lbll&(G))
WI
= f(O),
and formally
m G(x)f(x)
dx = f(0).
s --a,
By formally integrating by parts we can extract from this equality an heuristic
definition of the derivative 6:
m
sl(x)f(x)
dx = -
m G(x)f(x)
dx = -f(O).
s -00
s -03
Thus we can define a linear functional s f] = -f(O), however C(W) will be the
natural basic space for this (and 6turns out to be continuous). For the simultaneous definition of all derivatives 8()[f] = (- l) ( (O)(n > 1) we have to take
The boundary
sense.
function
cp = I& is defined
almost
everywhere
of u in a certain
integral
8 5. Function Spaces
245
Cm(R) as the basic space. On the other hand, however, the b-function is an
elementary measure, namely the unit measure concentrated at zero.2 If we wish
to regard all measures as generalised functions, then, because a measure can grow
arbitrarily at infinity, we have to take for the basic space C,(R), the space of
continuous functions with compact support. C~(R), the space of infinitely differentiable functions with compact support, is the universal receptacle of all the
formulated requirements. It is clear from what has been said that a generalised
function has to be defined not as an individual object but as an element of some
class.
Let G be an open set in R and E some linear topological space of functions
f: G + @which we will call the basic functions. The continuous linear functionals
on E, i.e. the elements of the conjugate space E* (with its w*-topology), are called
generalised functions over E.
The first and most important example is E = C:(G) = 9(G) (to say that
f: G + c has compact support means that there is a compact set Q/ c G outside of which f(x) = 0; the smallest such compact set, supp f, is the support of
the function f). We introduce the locally convex (non-metrizable) DieudonndSchwartz topology on 9(G) by means of the family of seminorms
ak,+-+k,
p,(f)
k ,,...,
k,=O
suP
xsG
Pkl,...,k
n a.+.
f
. . ax-
for all f E 9(G) with supp f c Q. If sup0 m(Q) = I < cc, we say that I is a
distribution of finite order r. A distribution of order zero is a complex measure.
We can also regard a locally summable function 0 as a distribution by identifying
it with the measure J 8 dx. Such distributions are said to be regular and form a
t is convenient to interpret the &function physically as a unit charge concentrated at zero. Then
6 is a dipole concentrated at zero with unit momentum. However in quantum mechanics the concept
of the S-function is essentially as a functional.
he use of multi-indices in work with functions of n variables is very convenient since it leads
at,+-+k. f
aff
It is also convenient to put Ikl =
to abbreviations of the form: x& = x:1 . . . x2; ax)r =
ax:l... a+'
k, + ... + k,.
246
dense subspace L:,,(G) c g(G)*. In general it is far from being the case that we
can speak of the values of a distribution 1 at points x E G (although the conventional notation A(x) is widely used because of its formal convenience4). If an open
set H c G is given and A[f] = 0 for all f E g(H), we set5 &, = 0. The complement in G of the union of all such H is called the support of the distribution I and
is denoted by supp A. The distributions with compact support form a space which
is the conjugate of Cm(G). The topology on C(G) is defined by the collection of
seminorms appearing in (49), i.e. it is the topology of uniform convergence along
with all derivatives on each compact set. Clearly, Cm(G) is a locally convex
Frechet space.
Remark 1. Without going into the general definitions we note that 5@(G)is the
inductive limit of the family C*(X), where X runs through all closed balls
4For example, in this form it is convenient to make the change of variables x = Fy (F: G + G is a
diffeomorphism):
qx)f(F- )
WY)f(Y) dY =
s0
sG
Consequently, it is necessary to take as definition
dx
det F F-lx)
sThe natural embedding of g(H) into g(G) (extension by zero) is a continuous homomorphism
whose conjugate is the homomorphism which restricts a functional to s(H).
6Continuous operators of multiplication on function spaces by a function (possibly not belonging
to the given space) are called multiplicators.
his circumstance can be regarded as a formal cause of the appearance of divergence in quantum
field theory and the removal of divergence (i.e. normalisation
of the theory) as the regularisation of
multiplication of generahsed functions (N.N. Bogolyubov - O.S. Parasyuk, 1955).
247
0 5. Function Spaces
They are all continuous under the w*-topology, which is taken as the standard
topology on Q(G)*.
As an example to illustrate the flexibility and convenience of the language of
distributions, let us mention the general definition of one of the principal concepts
of the theory of partial differential equations. For any linear differential operator
L with coefficients in P(G) a distribution r,, which depends on the parameter
y E G and satisfies the equations
is called a fundamental solution. For example, the Laplace operator in R has as
a fundamental solution when n > 2 the function9
1
r,(x) = ux9 Y) = -@ _ 2)a,x _ y,n-2v
where O, is the area of the unit sphere of R with the canonical Euclidean norm
IxI=J,$x;.
A fundamental solution r, exists for any operator L with constant coefficients
(Ehrenpreis, 1954; Malgrange, 1955) and clearly it has the formlo T,(x) =
l-(x - y) (x, y E UP). An immediate application of this fundamental solution is
the construction of solutions to the equation L [A] = p by means of the formula
Iz = r * p. Here we are using convolution of distributions. For its introduction it
is necessary to consider first the convolution of the distribution
p with an
arbitrary function f E S+!), i.e. the function
(cl *f)(x) = Am
- .)I-
* 6, is the unit measure concentrated at the point y. In R we can write 6,(x) = 6(x - y). This notation
carries over to any domain G.
9If n = 2 we have l-(x, y) = k lnlxl. If n = 3 we can interpret l-(x, y) physically as the potential of
a point charge.
lo In a more complete formulation of the theorem it is asserted that I is a distribution of finite order.
We note that in theorems of this type not only the formulations are important but also the
constructions which allow us to establish further useful properties of the solutions. The generalised
Fourier transform is usually applied in the construction of fundamental solutions of equations with
constant coeffkients (see below).
ii In the case where L is the Laplace operator this is Poisson equation describing the potential of a
charge with ensity p.
248
then p *f E ~(I?) and, by definition, we can put (I * ,u) [f] = I * (p * f). Thus
the equation L[A] = p is successfully solved (in the class of all distributions) for
any distribution p with compact support. In particular, p can be a summable
function with compact support but the solution 1 nevertheless turns out to be
only a distribution in general (generalised solution). However, if the operator L
is elliptic12 (and only in this case), all generalised solutions will be real analytic
functions whenever the right hand side of the equation is a real analytic function
(I.G. Petrovskij, 1937; Hormander, 1955). An analogous result relating to infinite
differentiability holds for the so-called hypoelliptic operators13 (and only in this
case) (Hormander, 1955). Generalised solutions are entirely natural from the
physical point of view since they satisfy the equation in an integral sense: the
equality L[A] = p means that 1[Laf] = ~[f] for all f E g(G), i.e. in the usual
notation
(L,* = L).
W(Loff)(x)
dx =
PL(Mx) dx
sG
sG
This form of the equation can have even greater physical meaning than the
traditional one.
The onjugate operator principle has already been used repeatedly to transfer
the standard operations of analysis to generalised functions: if a continuous
linear operator A is given on the basic space, then A* is its transfer to the space
of generalised functions. In a more general situation we can be given a continuous
homomorphism
h: E + E, of the basic space into another function space. Since
h * p E E* (,u E E:), it follows that h * p is a generalised function (on E). We can
regard this as the transfer of the homomorphism h to the generalised functions over
the basic space El. The most important example of this type is the generalised
Fourier transform.
f(x)e-i(c*x
s o-4
dx
part
at
is defined by a homogeneous
ax
&,,, a,-
0 5. Function
Spaces
249
exponential type that decrease more rapidly than any power of 1x1.In particular
we can now speak of the Fourier transform of any locally summable function
0(x) without imposing any restrictions as 1x1+ co.
It is clear that the generalised Fourier transform has the usual perational
properties: differentiation Dj passes over to multiplication
by irj, therefore any
linear differential operator with constant coefficients passes over to the operator
of multiplication
by the corresponding polynomial; convolution becomes a product and so on; hence differential, convolution and some other equations are as
usual transformed into algebraic equations. For example, the equation L[n] = p
takes the form P(t)n = ,& and the problem reduces to determining the possibility
of dividing a generalised function by a polynomial (division problem).
The general approach to the Fourier transform was developed gradually
(Bochner, 1932; L. Schwartz, 1951) and took its final form, which has been
described above, in the work of I.M. Gel and and G.E. Shilov (1953), where they
obtained with the help of the Fourier transform uniqueness classes for the
Cauchy problem for systems of differential equations of the form
2 = ktl Ljk(t)CUkl
t1 Gj G m)9
(50)
the Ljk(t) being linear differential operators in R with coefficients which depend
continuously on t (but are constant with respect to x). If p is the least order of
the operators Ljk, then the uniqueness class is defined by the bound
max luj(x, t)I < Mealx14,
j
where15 q = p/(p - 1) and 0, M are positive constants. In this connection the
consideration of generalised functions over suitable spaces of entire functions
turned out to be essential, since it allowed the possibility of establishing uniqueness by using the Phragmen-Lindelof
principle. The application of this sort of
analytic technique requires an intrinsic characterisation of the space16 E* for
those spaces E being used as basic spaces.
We stress the lack of sensitivity of the method described to the type of system
(for parabolic (according to Petrovskij) equations the uniqueness class was
established by O.A. Ladyzhenskaya (1950) and for parabolic systems by SD.
Ehjdel an (1953); in this situation G.H. Zolotarev (1958) obtained a criterion of
Taklind type). The general result is well illustrated by the following example: for
the equation
a~ a%
z=a=
ISIn fact we can take q = ~ PO
PII-1
where
PO c P is the so-called
(51)
formula
for its determination,
from which we can see that p,, is a rational
number
tor c m; in particular,
p,, = p if m = 1 (V.M. Borok, 1957).
leThis is a certain (not very great) cost on account of the restrictions
at infinity.
with
denomina-
250
is defined by the
(52)
For a > 0 this is the heat equation, for a < 0 it represents everseheat flow and
for a = i we have the simplest case of a non-stationary Schriidinger equation.
We note that for a < 0 the Cauchy problem for equation (51) is incorrect in
the sense that we do not have continuous dependence of the solution (in the
uniform metric on an arbitrarily small time interval) on the initial function.
The technique of generalised Fourier transform allows us to identify correctness
classes in the Cauchy problem for the system (50) (G.E. Shilov, 1955). Certainly
they are, in general, no longer the uniqueness classes, although sometimes the
two coincide. For example, the Cauchy problem for the heat equation in the class
(52) is correct.
Now let the basic space E be an extension of the space 9 = g(R
with a
weakened topology and also let 9 be dense in E for the weak topology. Then all
continuous linear functionals on E remain such when restricted to 9 and are
uniquely defined by their restrictions. Thus generalised functions on E can be
regarded in the given case as distributions. The most remarkable example of this
type is the Schwartz space S of functions of class Co3(R which decrease together
with all their derivatives more rapidly than any power 1x1- (v = 0, 1,2, . . .). This
is a locally convex Frechet space defined by the family of seminorms
$5. Function
Spaces
251
(f E 3).
We can say that the functional a is concentrated ml8 Q but this term is also used
in the case where for any neighbourhood
I/ 3 K there exists a constant C, such
that
bCfll G G
Example.
SUP
If(
(f E 4@ ).
(53)
the
condition
lim
Ikl-
;l/lael= 0.
Then
is an analytic functional concentrated at 0 and (54) gives the general form of such
functionals.
Let us denote by ,4*(Q) the set of analytic functionals which are concentrated
on the compact set Q c @ Restricting ourselves now to real compact sets2 (i.e.
such that Q c 08c Q=where the last embedding is the canonical one), let us
consider the linear space
NV
252
this construction is such that hyperfunctions are not generalised functions in the
sense used earlier. However in many respects we can work with them just as with
generalised (and so also as with classical) functions; in particular we can consider
differential equations in hyperfunctions. It is interesting that hyperfunctions were
intended first of all to play the role of limiting values of analytic functions in
situations where these do not exist in any usual sense. Because of this the setting
of hyperfunctions allows us to obtain, for example, adequate generalisations of
Bogolyubou theorem
n the edge of the wedge (Martineau, 1968; Morimoto,
1979; Zharinov, 1980). We note that the original version of this theorem was
obtained by N.N. Bogolyubov (1956) in connection with the construction of
dispersal relations in quantum field theory. The theorem asserts the local analytic
continuability
of one into the other for a pair of functions f+, f- which are
analytic on the domains G + iK (G is a domain and K is a cone in R under the
condition that their limiting values on G coincide. The extended function turns
out to be analytic on some complex neighbourhood V 2 G (we even have V 1 I,
where I is the C-convex hull of the domain G (VS. Vladimirov, 1960)).
5.3. Families of Function Spaces. Although the list of families presented below
contains many of the well-known function spaces it is certainly by no means
complete and illustrates rather the variety of possible constructions. We note
that the nature of a function space is determined to a considerable extent by the
structure which is given on the domain of definition X of those functions from
which the space is composed. For example, if X is a space with a measure, it is
natural to consider measurable functions on it; on a topological space X we
consider continuous functions (if we also look at discontinuous functions, they
should at least be Borel functions); a metric on X allows us to introduce the
modulus of continuity and to define, for example, Lipschitz classes; on a smooth
(analytic) manifold we can consider smooth (analytic) functions and so on.
I. Let M(u) (U > 0) be a convex function such that M(u) > 0 (U # 0) and
lim u- (u)
= 0,
U-+0
lim u- (u)
-+a3
= co.
(u 2 0)
II>0
sX
M*(l44l)
dp < 00.
.e. such that the inverse images of Bore1 sets are of the same type.
23The non-negativity of M*(u) is the classical Young inequality.
5 5. Function Spaces
253
We denote by L;$(X) the set of functions u(x) such that uu E L(X, p) for all
u E J+(X). This linear space has a norm
Ilull =
SUP
M*(u)<1
IS X
uu d,u
I
(it can be shown that the supremum is finite) under which it turns out to be a
Banach space; it is called the Orlicz space with norming function M.
If M(u) = %( p > l), then M*(u) = 4 p + 4 = 1 and the Orlicz
T
l
>
space L;I, coincides with Lp (to within a multiple of the norm).
Now let us assume that M(u) is of slow growth in the sense that
Example.
M&4
fp M(u)
< O
linear functional
on L$ is then
where v E L,,. By the same token L,,. turns out to be topologically isomorphic
to the conjugate space of L$.
II. We associate with each measurable function u(x)(x E X) its indicatrix24
0 2 0).
II4 = oa+h(t))
s
dt < 00
defines the Banach space A,, which is called a Lorentz space. The conjugate of
space M,, which consists of the measurable functions
u(x) (x E X) for which
A, is the Marcinkiewicz
ll4l = sup ~
II/(&a
I4 dp < ~0
sA
(the general form of a continuous linear functional on ,4 ti is standard).
In the space of measurable functions which are described above it is possible
to construct a rich theory of non-linear integral equations, by choosing the space
in accordance with the nature of the non-linearity (power, exponential, etc.). This
24 In the probabilistic situation this is a variant of the distribution function of a random variable. Just
as we speak of identically distributed random variables in probability theory, here we can speak of
equimeasurable
functions (i.e. functions whose indicatrices coincide - this is also an important
concept).
254
programme was realised in the 1950s and 1960s by M.A. Krasnosel kij in
collaboration with several co-authors.
III. Let X be a metric space with metric d and let U(X) (x E X) be a uniformly
continuous function. The quantity
w@) <
+ sup -_
a>0 Q(6)
00
(x0 E X is a fixed point). This family of Banach spaces serves first and foremost
the needs of approximation
theory.
IV. Among the spaces of smooth functions let us mention first of all the Soboleu
spaces which are defined in their general form by the following procedure. Let G
be an open set in R let 1 < p < cc and let m be an integer 2 1. The space WPm(G)
consists of those functions whose generalised partial derivatives up to order m
inclusive are functions of clasP U(G). Correspondingly we introduce the norm
Ilfll, = Ilfllm + J
(Ss
G
>
l/P
Ix
Xyn+rP
dx dx
< 00.
9 5. Function Spaces
255
Finally, we can require that the sequence of norms (e.g. in Lp) of the inverse
Fourier transforms of the terms in (55) decrease at a predetermined rate in
accordance with the growth of the radii r,. An alternative to the last step is to
first calculate a certain norm (taking the radii into account) for the sequence
(qpkf)? and then to require that this function behaves itself sufficiently well at
infinity (e.g. belongs to ~5~).These approaches, which were outlined in the mid1960s (S.M. Nikol kij, P.I. Lizorkin, Peetre), were systematically developed
during the 1970s by Peetre, Triebel and Stijckert and led to the construction of
very extensive families of spaces of smooth (of power order) functions which
include the Nikol kij-Besov
spaces, the Lizorkin spaces associated with derivatives of fractional order and so on. A corresponding theory for the general
modulus of continuity has been constructed during the last ten years (G.A.
Kalyabin, M.G. Gol man).
Function spaces on manifolds are constructed naturally from those on charts
(the same can be said about spaces of differential forms and, in general, about
sections of vector bundles on manifolds). As has already been said above, in the
theory of boundary value problems it is important to know to what space the
boundary values of functions from the given space belong. Moreover it is important to have a collection of embedding theorems for individual spaces into others
and also to know when the embedding mappings are compact. At the present
time extensive information has been accumulated on this account.
5.4. Operators on Function Spaces. We have already touched upon this topic
repeatedly when speaking of integral and differential operators. A formal notation of the type
f(x) = -Jm
66 - YMY) dy
-a,
allows us to regard linear differential operators as integral equations with generalised, highly singular kernels. The kernels of ordinary integral operators are
regular so that these operators (or their powers) are compact on spaces of type*
LJ. An intermediate position between the two cases mentioned is occupied by
singular integral operators whose kernels are functions with discontinuities of
281n particular, kernels with bound K(x, y) = 0(1x - yl-)( m c n) on a bounded domain G c R are
of this type.
256
such a type that, while compactness is lost,2g we still have boundedness on spaces
of type Lp.
Example.
<x<00)
Wd(x)
=& s-1$f$ (-co
(56)
is defined and bounded on Lp( R) (1 < p < co) (theorem of M. Riesz). The integral
(56) is obtained as a principal value. If cpE L(R), it exists for almost all x
(Zygmund, 1959), however the Hilbert operator acts neither on L nor on L: if
(~(5) = 1 (151 < l), rp(O = 0 (ItI > 11, then WV)(~) = ln z
. It is useful for
I
I
the investigation of the operator H to pass to the Fourier transforms 4, eq. Then
it transforms into the operator of multiplication
by $sgn 1. However the Fourier
transform itself acts on Lp only for p = 2, so that for the time being only
boundedness of the Hilbert operator on L2 is clear. From certain considerations of a special character we can obtain in addition its boundedness for
p = 4, 6, . . . , after which the passage to arbitrary p 2 2 is effected with the help
of the following.
M. Riesz-Thorin Interpolation Theorem. Let 1 < pi < ri < co (i = 0, 1) with pi
finite, put pa = [(l - a)p; + ap; - (0 < a G 1) and let r, be defined similarly.
Zf an operator A is defined on the set of piecewise-constant
functions (p(t) of
compact support (--co < 5 < co) and maps them into functions in Lro n L in such
a way that IIAqlI,, < Mi IIrp IIpi (i = 0, l), then
(0 < a < l),
IIA~II,~ G WIMlp,
inf
x=x0+x,
(II%
IlEo
llxl
liEI).
4 5. Function Spaces
251
MMI G ME max(ll&,II,
space fir
a Banach pair
IL%, II),
(57)
constant)3.
Proof. The operator Al, is closed, since it is closed (even bounded) under the
weaker norm induced from E, + E,. It is therefore bounded. Further the space
Z(E,, E,) is a Banach space with respect to both the norm which appears on the
right hand side in (57) and the weaker norm [IA IIEo+s,. These norms are therefore
equivalent. Finally, the restriction homomorphism
of Z(E,, E,) + 2(E) is closed
(which is easily verified with the norm IIAIIE,+E, on Z(E,, E,)). It is therefore
bounded, i.e. (57) holds. 0
Remark. In the more general situation to which, in particular, the RieszThorin theorem relates we have to consider two Banach pairs (E,, E,), (F,, F,),
spaces E, F which are intermediate for these pairs, and an interpolation in the
sense that, for all homomorphisms
A: E, + E, --, F. + F1 such that AE, c F,,
AE, c F1 and AlEo, AlEI are bounded, we have AE c F. There is no problem in
carrying over the interpolation inequality to this situation.
As an example let us consider the Banach pair (L1 (W), Em(~)). The role of the
embedding LTS is taken here by the space @ of measurable functions with
the topology of convergence in measure on each finite interval. The space
Q. = L + L is topologically isomorphic to the space of functions rp E Q, for
which the indicatrix m. f co. A Banach space E which is continuously embedded
in Q. is said to be symmetric if 1) whenever q E E, + E Q. and 11&5)1< Iv(<)1
almost everywhere, it follows that $ E E and jl+llB < llqllE, 2) whenever cpE E,
$ E Q. and cp, II/ are equimeasurable, it follows that I/ E E and ll$llE = IIqIIs.
For example, all Lp (1 < p < co) are of this type, as are Orlicz, Lorentz and
Marcinkiewicz spaces. E.M. Semenov (1964) proved that all symmetric spaces
are intermediate for (L, L) and any interpolation space for a pair of symmetric
spaces is topologically isomorphic to a symmetric space (it is symmetric if the
interpolation
constant is equal to 1). Calderon (1966) obtained a complete
description of interpolation spaces for (L, L). In any event all separable symnder these conditions A is bounded both on E,, + E, and on E, n E 1 for the noms introduced
on these spaces.
31 It can at once be assumed to be the least such constant. If we then have ME = 1, the interpolation
space is said to be exact.
258
<
C(a,
p,
y)
IIxII~~-~)(y-=)IIxII~-=)(y-=)
(x
6 Ey),
where C(a, fi, y) is a positive constant. Such a family is said to be a scale (normal
if IIxIIs, < llxllEb (B > a) and C(a, /I, y) = 1). Many scales (Ea),,4.Gl constructed
under specific conditions turn out to consist of interpolation spaces for the pair
(E,, E,) (Lions, 1958; S.G. Krejn, 1960; S.G. Krejn - Yu.1. Petunin, 1966)33.
Because of this occurrence, scales prove to be a suitable tool for investigating
boundary value problems for elliptic differential operators (Magenes, 1963).
Returning to the consideration of specific operators on function spaces, we
note first of all the multivariable
analogue of the Hilbert operator:
(for simplicity)
and
0(x, 0) d0 = 0.
s lel=l
The integral (58) is obtained as a principal value. The singular integral operator
Hc is defined and bounded on Lp( I?) (1 < p < co) if 52 belongs to Lq with respect
to 13 - + - = 1 and its LP-norm is bounded with respect to x (the case n = 2,
c
:
>
p = 2 is due to S.G. Mikhlin,
1953; the general case is due to Calderon and
Zygmund, 1956).
32 eal and omplex methods differ here. The latter are based on results from complex analysis
such as Hadamard
three-lines theorem. We note that with these methods it is possible to show, for
example, that Nikol kij-Besov spaces are interpolation spaces for a pair of spaces from this same
family or for a pair of Sobolev spaces etc.
33 In particular, the scales which can be constructed by the complex interpolation method.
6 5. Function Spaces
259
eY)
4x,
5)
=ssz(xy
e-i(C.y)
b-4
dy.
IYI
In this case the composition theorem holds: the product of the symbols corresponds to the product of the operators (Calderon-Zygmund,
1952). This result
has a natural connection with the convolution theorem but it is not so obvious
since the integrals concerned are principal values.34 Since the correspondence
between operators and symbols is clearly also linear, we have here a homomorphism of an algebra of operators into an algebra of functions, i.e. a certain
variant of operator calculus. This is an effective route to the solution of multivariate singular integral equations (this was in fact projected in the late 1920s by
Tricomi).
In the general case it is useful to consider along with HB the operator
We naturally expect that the operator Go, whese investigation can turn out to
be simpler than that of Ho, will at the same time imitate HO sufficiently well. The
operator Gc is pseudodifirential.
To explain this term let us consider an operator
of the form
U%)(x) = m
1
s
where p, an arbitrary given function of power growth rate3 with respect to <, is
the symbol of the operator P. If p(x, 5) is a polynomial in r (with coefficients
depending on x), then P is a difirential operator (with the same coefficients). If
we eliminate the intermediate Fourier transform, we can represent the operator
P as a double integral:
u+m)
1
= pq
R Rnp(x, &p(y)ei(e*x-Y)
ss
d( dy.
(59)
34The concept of the symbol was introduced differently at an early stage in the development of the
theory of singular integral equations and correspondingly the composition theorem was proved in
another context (S.G. Mikhlin, 1936; Giraud, 1936; S.G. Mikhlin, 1955). Then the equivalence of the
two definitions of the symbol was established (S.G. Mikhlin, 1956).
Uniformly with respect to x on each compact set. The exponent of the power of growth (not
necessarily integral) is called the order of the operator. Usually p is a function of class C with respect
to all its variables with suitable bounds on the derivatives.
260
Chapter
2. Foundations
and Methods
Bibliography
261
Bibliography*
I. Basic
Texts
Kantorovich, L.V., Akilov, G.P. [1977]: Functional Analysis. Nauka: Moscow. English transl.:
Pergamon Press: Oxford, 1982,Zbl.484.46003
Kolmogorov, A.N., Fomin, S.V. [1972]: Elements of the Theory of Functions and Functional
Analysis, 3rd ed. Nauka: Moscow. French transl.: Mir.: Moscow, 1974, German transl.: VEB
Deutscher Verlag der Wissenschaften: Berlin, 1975, Zbl.57,336; Zbl.90,87; Zbl.167,421; Zbl.299.
46001
Lyustemik, L.A., Sobolev, V.I. [1965]: Elements of Functional Analysis. Nauka: Moscow, Zbl.141,116;
Zbl.44,325. English transl.: J. Wiley & Sons: New York, 1974
Riesz, F., Szokefalvi-Nagy, B. [1972]: Lecons d nalyse Fonctionelle. Akad. Kiado: Budapest,
Zbl.64,354; Zbl.46,331
Rudin, W. [1973]: Functional Analysis. McGraw-Hill: New York, Zbl.253.46001
II. Additional
Akhiezer,
transl.:
Akhiezer,
Transl.
Educational
Literature
*For the convenience of the reader, references to reviews in Zentralblatt fur Mathematik (Zbl.),
compiled using the MATH database, have, as far as possible, been included in this bibliography.
262
Bibliography
Akhiezer, NJ., Glazman, I.M. [1977, 19783: Theory of Linear Operators in Hilbert Space I, II.
Vishcha Shkola: Khar ov. English transl.: Pitman Adv. Publ: Boston, 1981, Zbl.467.47002;
Zbl.143,365
Antonevich, A.B., Radyno, Ya.V. [1984]: Functional Analysis and Integral Equations. Minsk Univ.
Press: Minsk (Russian), Zbl.539.46001
Atiyah, M.F., Macdonald, LG. [1969]: Introduction to Commutative Algebra. Addison-Wesley:
Reading, Mass., Zbl.17536
Berezin, F.A., Shubin, M.A. [1972]: Lectures on Quantum Mechanics. Moscow Univ. Press: Moscow
(Russian)
Birman, M.Sh., Solomyak, M.Z. [1980]: Spectral Theory of Selfadjoint Operators in Hilbert Space.
Leningrad Univ. Press: Leningrad (Russian)
Brudnyj, Yu.A., Gorin, E.A. [1981]: Isometric Representations and Differential Inequalities. Yaroslavl Univ. Press: Yaroslavl (Russian), Zbl.559.34011
Courant, R., Hilbert, D. [1931]: Methoden der Mathematischen Physik, I. Springer-Verlag: Berlin,
Zbl.15
Davis, M. [1977]: Applied Nonstandard Analysis. J. Wiley & Sons: New York, Zbl.359.02060
Edwards, R.E. [ 19651: Functional Analysis. Theory and Applications. Holt: New York, Zbl.182,161
Edwards, R.E. [1979, 1982-j. Fourier Series (a Modern Introduction) I, II. Springer-Verlag: Berlin,
2b1.424.42001,Zb1.599.42001
Frohlicher, A., Bucher, W. [1966]: Calculus in Vector Spaces without Norm. Lect. Notes Math. 30,
Springer-Verlag: Berlin, Zbl.156,383
Gaier, D. [1980]. Vorlesungen iiber Approximationen im Komplexen. Birkhiiuser: Basel, Zbl.442.
30038
Gel and, I.M., Fomin, S.V. [1961]: Calculus of Variations. Phys.-Math. Section of Nauka: Moscow.
English transl.: Prentice Hall: Englewood Cliffs, N.J., 1965 Zbl.127,54
Glazman, I.M., Lyubich, Yu.1. [1969]: Finite-Dimensional
Linear Analysis. Nauka: Moscow,
Zbl.188,74. English transl.: MIT Press: London, 1974
Halmos, P.R. [1950]: Measure Theory. Van Nostrand: New York, Zbl.40,168
Halmos, P.R. [1956]: Lectures on Ergodic Theory. Math. Sot. Japan: Tokyo, Zbl.73,93
Halmos, P.R. [1974]: A Hilbert Space Problem Book. Springer-Verlag: New York, Zbl.144,387
Helson, H. [1964]: Lectures on Invariant Subspaces. Academic Press: New York, Zbl.119,113
Kadets, V.M., Kadets, M.I. [ 19883: Rearrangements of Series in Banach Spaces. Tartu Univ. Press:
Tartu (Russian), Zbl.685.46005. English transl.: Transl. Math. Monogr. 86: Providence, 1991
Kirillov, A.A. [1978]: Elements of the Theory of Representations. (2nd ed.) Nauka: Moscow. English
transl.: Springer-Verlag: Berlin, 1976,Zbl.264.22001
Kirillov, A.A., Gvishiani, A.D. [1979]: Theorems and Problems from Functional Analysis. Nauka:
Moscow. English transl.: Springer-Verlag Berlin, 1982,Zbl.481.46002
Koosis, P. [1980]: Introduction to HP Spaces. Cambridge Univ. Press: Cambridge, Zbl.435.30001
Krejn, S.G. [1971]: Linear Equations in Banach Space. Nauka: Moscow. English transl.: Birkhluser:
Base], 1982,Zbl.233.47001
Kutateladze, S.S. [1983]: Foundations of Functional Analysis. Nauka: Novosibirsk (Russian),
Zbl.581.46001
Loomis, L.H. [1953]: An Introduction to Abstract Harmonic Analysis. Van Nostrand: New York,
Zbl.52117
Markus, A.S., Sementsul, A.A. [1985]: Series in Banach Spaces. Kishinev Univ. Press: Kishinev
(Russian)
Matsaev, V.I., Palant, A.A. [1975]: Spectral Theory of Operator Bundles. Donetsk Univ. Press:
Donetsk (Russian)
Maurin, K. [1959]: Methods of Hilbert Spaces. Warszawa. English transl.: Polish Scientific Publishers: Warsaw [1967], Zbl.86,309,Zbl.166,102
Natanson, I.P. [1949]: Constructive Theory of Functions. State Science Press: Moscow-Leningrad
(Russian), Zbl.41,186
Nikol kij, N.K. [1980]: Lectures on the Shift Operator. Nauka: Moscow, Zbl.508.47001. English
transl.: Springer-Verlag: Berlin, 1986,Zbl.587.47036
Bibliography
263
Phelps, R.R. [1966]: Lectures on Choquet Theorem. Van Nostrand: Princeton, Zbl.l35,362
Plesner, A.I. [1965]: Spectral Theory of Linear Operators. Nauka: Moscow. English transl.: F. Ungar
Publ.: New York, 1969, Zbl.147,343
Reed, M., Simon, B. [1972-19791: Methods of Modem Mathematical Physics I-IV. Academic Press:
New York, Zbl.242.46001; 2b1.308.47002,2b1.405.47007,2b1.401.47001
Rickart, C. [1960]: General Theory of Banach Algebras. Van Nostrand: Princeton, Zbl.95.97.
Robertson, A.P., Robertson, W.J. [1964]: Topological Vector Spaces. Cambridge Univ. Press:
Cambridge, Zbl.123,302, Zbl.251.46002
Rockafellar, R.T. [1970]: Convex Analysis. Princeton Univ. Press: Princeton, Zbl.l93,184
Sadovnichij, V.A. [1986]: Theory of Linear Operators. Moscow Univ. Press: Moscow (Russian)
Saks, S. [ 19301: Theory of the Integral. Monografie Matematyczne: Warsaw-Lwbw. English transl.:
Stechert & Co: New York, 1937,Zbl.17,300; Zbl.7,105
Schaefer, H.H. [1966]: Topological Vector Spaces. Macmillan: New York, Zbl.l41,305
Shilov, G.E. [1965]: Mathematical Analysis. Second Special Course. Nauka: Moscow. English transl.:
Generalized Functions and Partial Differential Equations, Gordon and Breach: New York, 1968,
Zbl.l77,363; Zbl.591.46030
Shilov, G.E., Fan, D.T. [1967]: Integral, Measure and Derivative in Linear Spaces. Phys.-Math.
Section of Nauka: Moscow (Russian)
Shilov, G.E., Gurevich, B.L. [1967]: Integral, Measure and Derivative. General Theory. Nauka:
Moscow. English transl.: Prentice Hall: Englewood Cliffs, N.J., 1966, Zbl.l38,275; Zbl.l31,52
Smimov, V.I. [1959]: Course of Higher Mathematics, V. Phys.-Math. Section of Nauka: Moscow,
Zbl.l14,30. English transl.: Pergamon Press: Oxford, 1964
Trenogin, V.A. [1980]: Functional Analysis. Nauka: Moscow (Russian), Zbl.571.46001
Vladimirov, V.S. [1979]: Generalised Functions in Mathematical Physics. Nauka: Moscow. English
transl.: MIR: Moscow, Zbl.515.46033,Zbl.403.46036
Vulikh, B.Z. [1961]: Introduction to the Theory of Partially Ordered Spaces. Phys.-Math. Section
of Nauka: Moscow, Zbl.l01,85. English transl.: Wolters-Noordhoff Sci. Publ.: Groningen, 1967
Wermer, Y. [1976]: Banach Algebras and Several Complex Variables. Springer-Verlag: New York,
Zbl.281.46052
Yosida, K. [1965]: Functional Analysis. Springer-Verlag: Berlin, Zbl.126,115,Zbl.286.46002
III. Monographs
Akhiezer, N.I. [1961]: The Classical Moment Problem. Phys.-Math. Section of Nauka: Moscow,
Zbl.l24,62. English transl.:Edinburgh, London, 1965
Akhiezer, N.I., Krejn, M.G. [1938]: On Certain Questions in the Theory of Moments. Ukrainian
State Science Press: Khar ov (Russian)
Azizov, T.Ya., Iokhvidov, I.S. [1986]: Basic Theory of Linear Operators on Spaces with Indefinite
Metric. Nauka: Moscow (Russian), Zbl.607.47031
Banach, S. [1932]: Theorie des Operations Lineaires. Polish Scientific Publishers: Warsaw, Zbl.5,209
Banach, S. [1987]: Theory of Linear Operations (with a survey by Pelczyhski, A. and Bessaga, C.,
Some Aspects of the Present Theory of Banach Spaces).North-Holland: Amsterdam, Zb1.613.46001;
Zbl.5,209
Berezanskij, Yu.M. [1965]: Eigenfunction Expansion of Selfadjoint Operators. Naukova Dumka:
Kiev, Zbl.142,372. English transl.: Providence, 1968
Berezanskij, Yu.M. [1978]: Selfadjoint Operators on Spaces of Functions with an Infinite Number
of Variables. Naukova Dumka: Kiev, Zbl.449.47001. English transl.: Transl. Math. Monogr. 63:
Providence, 1986
Besov, O.V., Bin, V.P., Nikol kij, S.M. [1975]: Integral Representations of Functions and Embedding Theorems. Nauka: Moscow, Zbl.352.46023. English transl.: J. Wiley: New York, Vol I,
1978, Vol II, 1979
Bessaga, S., Pelczynski, A. [1975]. Selected Topics in Infinite-Dimensional Topology. Polish Scientific Publishers: Warsaw, Zbl.304.57001
264
Bibliography
Bonsall, F.F., Duncan, J. [1971]: Numerical Ranges of Operators on Normed Spaces and of Elements
of Normed Algebras. London Math. Sot. Lecture Note Series, No. 2,Zbl.207.448
Bourbaki, N. [1952, 1957, 19593: Integration. Hermann: Paris. Zbl.49,317,Zbl.80,268,Zbl.l15,49
Bourbaki, N. [1953,1955]: Espaces Vectoriels Topologiques. Hermann: Paris, Zbl.50,107,Zbl.67,83
Bourbaki, N. [1967]: Theories Spectrales. Hermann: Paris, Zbl.152,326
Branges, L.De [1968]: Hilbert Spaces of Entire Functions. Prentice Hall Inc.: Englewood Cliffs, New
Jersey, Zbl.157,433
Brodskij, M.S. [1969]: Triangular and Jordan Representations of Linear Operators. Nauka: Moscow,
Zbl.182,458. English transl.: Providence, 1971
Butzer, P.L., Berens, H. [1967]: Semigroups of Operators and Approximation. Springer-Verlag
Berlin, Zbl.164,437
Collatz, L. [1964]: Funktionalanalysis und Numerische Mathematik. Grundlehren Math. Wiss. 120,
Springer-Verlag: Berlin, Zbl.139,98
Colojoara, I., Foias, C. [1968]: The Theory of Generalised Spectral Operators. Gordon and Breach:
New York, Zbl.189442
Cooke, R.G. [1950]: Infinite Matrices and Sequence Spaces. Macmillan: London, Zbl.40,25
Cornfel , I.P., Sinai, Ya.G., Fomin, S.V. [1980]: Ergodic Theory. Nauka: Moscow, Zbl.508.28008.
English transl.: Springer-Verlag: Berlin, 1982, Zbl.493.28007
Daletskij, Yu.L., Fomin, S.V. [1983]: Measures and Differential Equations in Infinite Dimensional
Spaces. Nauka: Moscow (Russian), Zbl.536.46031
Daletskij, Yu.L., Krejn, M.G. [1970]: Stability of Solutions of Differential Equations in a Banach
Space. Nauka: Moscow. English transl.: Transl. Math. Monogr. Vol. 43, Providence, 1974,Zbl.233.
34001
Diestel, J. [1975]: Geometry of Banach Spaces. Springer-Verlag: Berlin, Zbl.307.46009
Diestel, J. [1984]: Sequences and Series in a Banach Space. Springer-Verlag: Berlin, Zbl.542.46007
Diestel, J., Uhl, J.J. [1977]: Vector Measures. Math. Surv. 15 Providence, RI., Am. Math. Sot.,
Zbl.369.46039
Dieudonne, J. [1981]: History of Functional Analysis. North-Holland: Amsterdam, Zbl.478.46001
Dixmier, J. [1969]: Les C*-algebres et leurs Representations. Gauthier-Villars: Paris, Zbl.174,186
Dragilev, M.M. [1983]. Bases in Kijthe Spaces. Rostov Univ. Press: Rostov (Russian), 2b1.556.46005
Dulst, D. van. [1978]: Reflexive and Superreflexive Banach Spaces. Math. Centre Tracts. Mathematisch Centrum: Amsterdam, Zbl.412.46006
Dunford, N., Schwartz, J.T. [1958]: Linear Operators I. General Theory. Intersci. Publ.: New York,
Zbl.84,104
Dunford, N., Schwartz, J.T. [1963]: Linear Operators II. Spectral Theory. Self Adjoint Operators in
Hilbert Space. Intersci. Pub].: New York, Zbl.128,348
Dunford, N., Schwartz, J.T. [1971]: Linear Operators III. Spectral Operators. I. Wiley and Sons: New
York, Zbl.243.47001
Dzhrbashyan, M.M. [1966]: Integral Transforms and Representation of Functions on a Complex
Domain. Phys.-Math. Section of Nauka: Moscow (Russian), Zbl.154,377
Ehfros, A.M., Danilevskij, A.M. [1937]: Operational Calculus and Contour Integrals. Ukrainian
State Science Press: Khar ov (Russian)
Faddeev, L.D. [1963]: Mathematical Questions of the Quantum Theory of Scattering for Systems
of Three Particles. Tr. Mat. Inst. Steklova 69, 1-122. English transl.: Works Steklov Inst. Math.,
Vol. 69, 1965,Zbl.125,215
Floret, K. [1980]: Weakly Compact Sets. Springer-Verlag: Berlin, Zbl.437.46006
Friedrichs, K.O. [1965]: Perturbation of Spectra in Hilbert Space. Am. Math. Sot.: Providence,
RI, Zbl.142,110
Gakhov, F.D. [1958]: Boundary Value Problems. Phys.-Math. Section of Nauka: Moscow. English
transl.: Pergamon Press: Oxford, 1966,Zbl.141,80
Gamelin, T.W. [1969]: Uniform Algebras. Prentice Hall: New York, Zbl.213404
1 Gantmakher, F.P., Krejn, M.G. [1950]: Oscillation Matrices and Kernels and Small Oscillations of
Mechanical Systems. State Science Press: Moscow-Leningrad (Russian), Zbl.41,355
Bibliography
265
Girding, L. [1954]: Applications of the Theory of Direct Integrals of Hilbert Spaces to some Integral
and Differential Operators. Lect. Ser. Inst. Fluid Dynam. and Appl. Math., No. 11. Univ. Maryland, Zbl.55,342
Garnett, J.B. [1981]: Bounded Analytic Functions. Academic Press: New York, Zbl.469.30024
Ciel and, I.M., Graev, M.I., Pyatetski-Shapiro, 1.1.[1966]: General&d Functions Vol. 6. Theory of
Representations and Automorphic Functions. Nauka: Moscow; Zbl.138,72. English transl.: Philadelphia, 1969
Gel and, I.M., Graev, M.I., Vilenkin, N.Ya. [1962]: Generalised Functions, Vol. 5. Integral Geometry
and Questions Connected with it in the Theory of Representations. Phys.-Math. Section of Nauka:
Moscow, Zbl.l15,167. English transl.: Academic Press: New York-London, 1966
Gel and, I.M., Rajkov, D.A., Shilov, G.E. [1960]: Commutative Normed Rings. Phys.-Math. Section
of Nauka: Moscow, Zbi.134,321. French transl.: Gauthier: Paris, 1964
Gel and, I.M., Shilov, G.E. [1958a]: General&d Functions, Vol. 2. Spaces of ordinary and generalised functions. Phys.-Math. Section of Nauka: Moscow, Zb1.91,111, English transl.: Academic
Press: New York-London, 1968
Gel and, I.M., Shilov, G.E. [1958b]: General&d Functions, Vol. 3. Some Questions in the Theory
of Differential Equations. Phys.-Math. Section of Nauka: Moscow, Zb1.91,111. English transl.:
Academic Press: New York-London, 1967
Gel and, I.M., Shilov, G.E. [1959]: Generalised Functions, Vol. 1. Generalised Functions and
Operations on them. Phys.-Math. Section of Nauka: Moscow, Zb1.91,111. English transl.: Academic Press: New York-London, 1964.
Gel and, I.M., Vilenkin, N.Ya. [1961]: Generalised Functions, Vol. 4. Some Applications of Harmonic Analysis. Rigged Hilbert Spaces. Phys.-Math. Section of Nauka: Moscow, Zb1.103,92.
English transl.: Academic Press: New York-London, 1964
Glazman, I.M. [1963]: Direct Methods of Qualitative Spectral Analysis of Singular Differential
Operators. Phys.-Math. Section of Nauka: Moscow, Zbl.143,365. English transl.: Oldboume Press:
London, 1965
Gokhberg, I.Ts., Fel man, LA. [1971]. Convolution Equations and Projection Methods for their
Solution. Nauka: Moscow (Russian), Zbl.214.385
Gokhberg, LTs., Krejn, M.G. [1965]: Introduction to the Theory of Non-selfadjoint Linear Operators. Nauka: Moscow, Zbl.138,78. English transl.: Providence, 1969
Gokhberg, LTs., Krejn, M.G. [1967]: Theory of Volterra Operators on Hilbert Space and their
Applications. Nauka: Moscow, Zb1.168,120. English transl.: Providence, 1970
Gokhberg, I.Ts., Krupnik, N.Ya. [1973]: Introduction to the Theory of One-dimensional Singular
Integral Operators. Shtiintsa: Kishinev, Zbl.271.47017. German transl.: Birkhauser: Base& 1979
Gould, S.H. [1966]: Variational Methods for Eigenvalue Problems. Univ. Press: Toronto, Zbl.156,124
Greenleaf, F.P. [1969]: Invariant Means on Topological Groups and their Applications. Van
Nostrand: New York, Zb1.174,190
Gross, L. [1963]: Harmonic Analysis on Hilbert Space. Mem. Am. Math. Sot. 46: New York,
Zbl.l18,322
Grothendieck, A. [1955]: Produits Tensoriels Topologiques et Espaces Nucleaires. Mem. Am. Math.
Sot. 16,Zb1.55,97
Halmos, P.R., Sunder, V.S. [1978]: Bounded Integral Operators on L*-spaces. Springer-Verlag:
Berlin, Zb1.389.47001
Hardy, G.H. [1949]: Divergent Series. Oxford, Clarendon Press: Oxford, Zbl.32,58
Hewitt, E., Ross, K.A. [1963, 19703: Abstract Harmonic Analysis I, II. Springer-Verlag: Berlin,
Zb1.115,106,2b1.213,401
Hille, E., Phillips, R.S. [1957]: Functional Analysis and Semi-groups. Am. Math. Sot.: Providence,
R.I,Zb1.78,100
Hirschmann, I.I., Widder, D.V. [1955]: The Convolution Transform. Princeton Univ. Press: Princeton,
Zb1.65,93
Hiirmander, L. [1983, 19851: The Analysis of Linear Partial Differential Operators, I-IV. SpringerVerlag: Berlin, Zb1.521.35001,2b1.521.35002,Zb1.601.35001,Zb1.612.35001
266
Bibliography
Kuo, H-H. [1975]: Gaussian Measures in Banach Spaces. Lect. Notes Math. 463, Springer-Verlag:
Berlin, Zbl.306.28010
Kutateladze, S.S.,Rubinov, A.M. [1976]: Minkowski Duality and its Applications. Nauka: Novosibirsk
(Russian)
Lacey, H.E. [1974]: The Isometric Theory of Classical Banach Spaces. Springer-Verlag: Berlin,
Zbl.285.46024
Ladyzhenskaya, O.A. [1970]: Mathematical Questions of the Dynamics of Viscous Fluid. Nauka:
Moscow (Russian)
Landkof, N.S. [1966]: Foundation of Modem Potential Theory. Nauka: Moscow, Zb1.148,103.
English transl.:Springer-Verlag: Berlin, 1972
Bibliography
261
Levin, B.Ya. [1956]: The Distribution of the Roots of Entire Functions. State Science Press: Moscow.
German transl.: Berlin, 1962., Zbl.l11,73
Levitan, B.M. [1950]: Factorisation of Second Order Differential Equations by means of Eigenfunctions. State Science Press: Moscow-Leningrad (Russian)
Levitan, B.M. [1953]: Almost-periodic Functions. State Science Press: Moscow (Russian)
Levitan, B.M. [1962]: Generalised Shift Operators and Some of their Applications. Phys.-Math.
Section of Nauka: Moscow. English transl.: Jerusalem, 1964,Zb1.192,489
Levitan, B.M,, Zhikov, V.V. [1978]: Almost-periodic Functions and Differential Equations. Moscow
Univ. Press: Moscow. English transl.: Cambridge Univ. Press., 1982,Zb1.414.43008
Levy, P. [1951]: Les Problemes Concrets d nalyse Fonctionelle. Gauthier-Villars: Paris, Zbl.43,323
Light, W.A., Cheney, E.W. [1985]: Approximation Theory of Tensor Product Spaces. SpringerVerlag: Berlin, Zb1.575.41001
Lindenstrauss, J. [1969]: Geometric Theory of Banach Spaces. State Univ. Press: Florida
Lindenstrauss, J., Tzafriri, L. [1977]: Classical Banach Spaces I. Sequence Spaces. Springer-Verlag:
Berlin, Zbl.362.46013
Lindenstrauss, J., Tzafriri, L. [1979]: Classical Banach Spaces II. Function Spaces. Springer-Verlag:
Berlin, Zbl.403.46022
Livshits, M.S. [1966]: Operators, Oscillations, Waves. Open Systems. Nauka: Moscow (Russian),
Zb1.143,367
Luxemburg, W.A.J., Zaanen, A.C. [1971]: Riesz Spaces. North-Holland: Amsterdam, Zb1.231.46014
Lyubich, Yu.1. [1985]: Introduction to the Theory of Banach Representations of Groups. Vishcha
Shkola: Khar ov, English transl.: Birkhluser: Basel, 1988,Zbl.623.22001
Maak, W. [1950]: Fastperiodische Funktionen. Springer-Verlag: Berlin, Zbl.37,57
Mackey, G.W. [1963]: The Mathematical Foundations of Quantum Mechanics. Benjamin: New
York, Zbl. 114,440
Mackey, G.W. [1965]: Group Representations and Commutative Harmonic Analysis with Applications to Analysis, Number Theory and Physics. Univ. Press: Berkeley
Mandelbrojt, S. [1935]: Series de Fourier et Classes Quasi-analytiques des Fonctions. GauthiersVillars: Paris, Zb1.13,llO
Marchenko, V.A. [1977]: Sturm-Liouville Operators and their Applications. Naukova Dumka: Kiev,
Zbl.399.34022. English transl.: Birkhiiuser: Base], 1986,Zbl.592.34011
Markus, AS. [1986]: Introduction to the Spectral Theory of Polynomial Operator Pencils. Shtiintsa:
Kishinev. English transl.: Transl. Math. Monogr. 71: Providence, 1988,Zbl.678.47005
Maslov, V.P. [1965]: The Theory of Perturbations and Asymptotic Methods. Moscow Univ. Press:
Moscow. French transl.: Gauthiers-Villars: Paris, 1972, Zbl.247,47010
Maslov, V.P. [1973]: Operator Methods. Nauka: Moscow, Zbl.288.47042. English transl.: MIR:
Moscow, 1982
Massera, J.L., Schaffer, J.J. [1966]: Linear Differential Equations and Function Spaces. Academic
Press: New York, Zb1.243.34107
Michael, E. [1952]: Locally Multiplicatively-convex Topological Algebras. Am. Math. Sot.: Providence, R.I,Zbl.47,355
Mikhlin, S.G. [1949]: Integral Equations and their Applications. State Science Press: Moscow,
Zb1.39,113. English transl. Pergamon Press: London, 1957
Mikhlin, S.R. [1952]: The Minimum Problem for a Quadratic Functional. State Science Press:
Moscow-Leningrad, Zb1.49,205. English transl.: Holden-Day: San Francisco, 1965
Mikhlin, S.R. [1962]: Multi-dimensional Singular Integrals and Integral Equations. Phys.-Math.
Section of Nauka: Moscow, Zb1.105,303. English transl.: Pergamon Press: Oxford, 1965
Mikusinski, J. [1953]: Rachunek Operatorow (The Calculus of Operators). Polskie Towarzystow
Matematyczne: Warsaw, Zb1.53,85. English transl. Pergamon Press: London, 1959,Zbl.88,330
Milman, V.D., Schechtman, G. [1986]: Asymptotic Theory of Finite Dimensional Normed Spaces.
Springer-Verlag: Berlin, Zbl.606.46013
Morris, S.A. [1977]: Pontryagin Duality and the Structure of Locally Compact Abelian Groups.
London Math. Sot. Lect. Notes Series. 29,Zb1.446.22006
268
Bibliography
Nagy, K.L. [1966]: State Vector Spaces with Indefinite Metric in Quantum Field Theory. Akad.
Kiado: Budapest, Zb1.151441
Najmark, M.A. [1968]: Normed Rings. 2nd. ed. Nauka: Moscow. English transl.: Noordhoff:
Groningen, 1964,Zb1.89,101
Najmark, M.A. [1969]: Linear Differential Operators. 2nd ed. Nauka: Moscow, Zb1.193,41. English
transl.: F. Ungar Publ.: New York, Vol. I, 1967,Zbl.219,34001, Vol. II, 1968,Zb1.227.34020
Nakano, H. [1950]: Modulared Semi-ordered Linear Spaces. Mazuren: Tokyo
Neumann, J. von [1932]: Mathematische Grundlagen der Quantemnechanik. Springer-Verlag: Berlin,
Zb1.5,91
Nikol kij, N.K. [1974]: Selected Problems of Weighted Approximation and Spectral Analysis. Tr.
Mat. Inst. Steklova 120. English transl.: Proc. Steklov Inst. Math. 120, 1976,Zb1.342.41027
Pekzynski, A. [1968]: Linear Extensions, Linear Averagings and their Applications to Linear
Topological Classification of Spaces of Continuous Functions. Diss. Math.: Warzawa, Zb1.165146
Pelczynski, A. [1977]: Banach Spaces of Analytic Functions and Absolutely Summing Operators.
Am. Math. Sot.: Providence, R.I., Zbl.384.46015
Petunin, Yu.I., Plichko, A.N. [1980]: The Theory of Characteristic Subspaces and its Applications.
Vishcha Shkola: Kiev (Russian), Zb1.467.46016
Pietsch, A. [1965]: Nukleare Lokalkonvexe Rlume. Schriftenr. Inst. Math. Dtsch. Akad. Wiss., A,
Berlin, Zbl.152,323
Pietsch, A. [1978]: Operator Ideals. VEB Deutscher Verlag der Wiss.: Berlin, Zb1.399.47039
Przeworska-Rolewicz, D, Rolewicz, S. [1968]: Equations in Linear Spaces. Polish Scientific Publishers: Warsaw, Zb1.181,405
Radjavi, H., and Rosenthal, P. [1973]: Invariant Subspaces. Springer-Verlag: Berlin, Zbl.269.47003
Rieffel, M.A. [1967]: Dentable Subsets in Banach Spaces with Application to a Radon-Nikodym
Theorem in Functional Analysis. Thompson: Washington; appeared in Proc. Conf., Univ. California, Irvine 1966,71-77, 1967,Zb1.213,137
Rolewicz, S. [1972]: Metric Linear Spaces. Polish Scientific Publishers: Warsaw, Zb1.226.46001
Rudin, W. [1962]: Fourier Analysis on Groups. Interscience: New York, Zb1.107,96
Rudin, W. [1969]: Function Theory in Polydiscs. Benjamin: New York, Zbl.177,341
Rudin, W. [ 19801: Function Theory in the Unit Ball of C Springer-Verlag: Berlin, Zb1.495.32001
Samojlenko, Yu.S. [1984]: Spectral Theory of Collections of Selfadjoint Operators. Naukova Dun&a:
Kiev (Russian), Zb1.543.47040
Schaefer, H.H. [1974]: Banach Lattices and Positive Operators. Springer-Verlag: Berlin, Zb1.296.
47023
Schapira, P. [1970]: Thtorie des Hyperfonctions. Lect. Notes Math. 226. Springer-Verlag: Berlin.
Zb1.192,473
Schatten, R. [1950]: A Theory of Cross-spaces. Princeton Univ. Press: Princeton, Zbl.41,435
Schatten, R. [1960]: Norm Ideals of Completely Continuous Operators. Springer-Verlag: Berlin,
Zb1.9494
Schmets, J. [1983]: Spaces of Vector-valued Continuous Functions. Lect. Notes Math. 1003. SpringerVerlag: Berlin, Zb1.511.46033
Schwartz, L. [1950, 19511: Theorie des Distributions I, II. Hermann: Paris, Zb1.37,73,Zb1.42,114
Schwartz, L. [1961]: Application of Distributions to the Study of Elementary Particles in Relativistic
Quantum Mechanics. Univ. Press: California, Zb1.124,230
Segal, I.E. [1963]: Mathematical Problems of Relatavistic Physics. Am. Math. Sot.: Providence, R.I.,
Zb1.112,453
Semadeni, Z. [1982]: Schauder Bases in Banach Spaces of Continuous Functions. Lect. Notes Math.
918. Springer-Verlag: Berlin, Zb1.478.46014
Shohat, J.A., Tamarkin, J.D. [1943]: The Problem of Moments. Am. Math. Sot.: New York,
Zb1.41,433
Shubin, M.A. [1978]: Pseudodifferential Operators and Spectral Theory. Nauka: Moscow, Zbl.451.
47064. English transl.: Springer-Verlag: Berlin, 1987,Zb1.616.47040
Sigal, J.M. [1983]: Scattering Theory for Many-bodies Quantum Mechanical Systems. SpringerVerlag: Berlin, Zb1.522.47006
Bibliography
269
270
Bibliography
Kadets, M.I., Mityagin, B.S. [1973]: Complemented subspaces in Banach spaces. 28, No. 6,77-94.
English transl.: Russ. Math. Surv. 28, No. 6,77-95 (1973X Zbl.287.46029
Kantorovich, L.V. [1948]: Functional analysis and applied mathematics. 3, No. 6,89-185 (Russian),
Zbl.34,212
Keldysh, M.V. [1971]: On completeness of the eigenfunctions of certain classes of non-selfadjoint
linear operators. 26, No. 4, 15-41. English transl.: Russ. Math. Surv. 26, No. 4, 15-44 (1972),
Zbl.225,47008
Khaplanov, M.G. [1956]: Infinite matrices in an analytic space. 11, No. 5,37-44,Zbl.73,334. English
transl.: Am. Math. Sot., Transl., II. Ser. 13, 177-183 (1960)
Krejn, M.G., Krasnosel kij, M.A. [1947]: Fundamental theorems on the expansion of linear operators and some of their applications to the theory of orthogonal polynomials and the moment
problem. 2, No. 3,60-W (Russian)
Krejn, M.G., Rutman, M.A. [1948]: Linear operators leaving invariant a cone in a Banach space. 3,
No. 1,3-95 (Russian), Zbl.34129
Kuptsov, N.P. [1968]: Direct and converse theorems of the theory of approximations and semigroups
of operators. 23, No. 4, 117-178. English transl.: Russ. Math. Surv. 23, No. 4, 115-177 (1968),
Zbl. 159,434
Kutateladxe, S.S. [1975]: Choquet boundaries in K-spaces. 30, No. 4,107-146. English transl.: Russ.
Math. Surv. 30, No. 4, 115-155 (1975), Zbl.331.46011
Lyubich, Yu.1. [1966]: Classical and local Laplace transforms in an abstract Cauchy problem. 21,
No. 3, 3-51. English transl.: Russ. Math. Surv. 21, No. 3, l-52 (1966). Zbl.173,120
Madxhenes, Eh. (=Magenes, E.) [1966]. Interpolation spaces and partial differential equations. 21,
No. 2, 169-218 (Russian), Zbl.173,158
Markus, AS. [1964]: Eigenvalues and singular values of the sum and composition of linear operators.
19, No. 4,93-123. English transl.: Russ. Math. Surv. 19, No. 4,91-120 (1964), Zbl.133,72
Mautner, F.I. [1955]: On eigenfunction expansions. 10, No. 4, 127-132. English transl.: Proc. Natl.
Acad. Sci. USA 39,49-53 (1953), Zbl.50,119
Mil an, V.D. [1970]: Geometric theory of Banach spaces I. Theory of bases and minimal systems.
25, No. 3, 113-174. English transl.: Russ. Math. Surv. 25, No. 3, 111-170 (1970), Zbl.198,165
Mil an, V.D. [1971]: Geometric theory of Banach spaces II. Geometry of the unit sphere. 26,
No. 6,73-149. English transl.: Russ. Math. Surv. 26, No. 6,79-163 (1972), Zbl.229.46017
Minlos, R.A. [1968]: Lectures on statistical physics. 23, No. 1,133-190. English transl.: Russ. Math.
Surv. 23, No. 1, 137-196 (1968), Zbl.165,585
Mishchenko, A.S. [1976]: Hermitian K-theory. Theory of characteristic classes, methods of functional analysis. 31, No. 2, 69-134. English transl.: Russ. Math. Surv. 31, No. 2, 71-138 (1976),
Zbl.427.55001
Mishchenko, A.S. [1979]: Banach algebras, pseudodifferential operators and their applications to
K-theory. 34, No. 6,67-79. English transl.: Russ. Math. Surv. 34, No. 6,77-91(1979),Zbl.432.46050
Mityagin, B.S. [1961]: Approximate dimension and bases in nuclear spaces. 16, No. 4, 63-132.
English transl.: Russ. Math. Surv. 16, No. 4, 59-128 (1961), Zbl.104.86
Mityagin, B.S. [1970]: Homotopy structure of the linear group of a Banach space. 25, No. 5,63-106.
English transl.: Russ. Math. Surv. 25, No. 5, 59-103 (1970), Zbl.232.47046
Mityagin, B.S., Shvarts, A.S. [1964]: Functors on categories of Banach spaces. 19, No. 2, 65-130.
English transl.: Russ. Math. Surv. 19, No. 2,65-127 (1964), Zbl.129.82
Nikishin, E.M. [1970]: Resonance theorems and superlinear operators. 25, No. 6,129-191. English
transl.: Russ. Math. Surv. 25, No. 6, 125-187 (1970), Zbl.222.47024
Opojtsev, V.I. [1976]: A converse of the principle of contraction mappings. 31, No. 4, 169-198.
English transl.: Russ. Math. Surv. 31, No. 4, 175-204 (1976), Zbl.335.54025
Pelcxynski, A. [1973]: On certain problems of Banach. 28, No. 6,67-75. English transl.: Russ. Math.
Surv. 28, No. 6,67-76 (1973), Zbl.281.46014
Pelcxynski, A., Figel T. [1973]: On Entlo method of constructing Banach spaces without the
approximation property. 28, No. 6,95-108. English transl.: Russ. Math. Surv. 28, No. 6,95-108
(1973), Zbl.305.46027
Peller, V.V., and Khrushchev, S.V. [1982]: Hankel operators, best approximations and stationary
Bibliography
271
Gaussian processes. 37, No. 1,53-124. English transl.: Russ. Math. Surv. 37, No. 1,61-144 (1982),
Zbl.497.60033
Radzievskij, G.V. [1982]: The problem of completeness of root vectors in the spectral theory of
operator-functions. 37, No. 2,81-145. English transl.: Russ. Math. Surv. 37, No. 2,91-164 (1982),
Zbl.524.47009
Rubinov, A.M. [1977]: Sublinear operators and their applications. 32, No. 4,1133177. English transl.:
Russ. Math. Surv. 32, No. 4, 115-175 (1977), Zbl.355.47002
Sadovskij, B.N. [1972]: Limit-compact and condensing operators. 27, No. 1,81-146. English transl.:
Russ. Math. Surv. 27, No. 1,85-155 (1972), Zbl.232.47067
Sakhnovich, L.A. [1980]: Equations with a difference. kernel on a finite interval. 35, No. 4, 69-l 12.
English transl.: Russ. Math. Surv. 35, No. 4,81-152 (1980), Zbl.444.45008
Tsekanovskij, Eh.R., Shmul an, Yu.L. [1977]: Theory of biextensions of operators on rigged Hilbert
spaces. Unbounded operator colligations and characteristic functions. 32, No. 5,69-124. English
transl.: Russ. Math. Surv. 32, No. 5,73-132 (1977), Zbl.447.47011
Ul anov, P.L. [1961]: Divergent Fourier series. 16, No. 3,61-142. English transl.: Russ. Math. Surv.
26, No. 3, l-75 (1961), Zbl.103,290
Ul anov, P.L. [1964]: Solved and unsolved problems in the theory of trigonometric and orthogonal
series. 19, No. 1, 3-52. English transl.: Russ. Math. Surv. 19, No. 1, l-62 (1964), Zb1.125,312
Vershik, A.M., Gel and, I.M., Graev, MI. [1975]: Representations of the group of diffeomorphisms.
30, No. 6, 3-50. English transl.: Russ. Math. Surv. 30, No. 6, l-50 (1975), Zbl.317.58009
Vishik, MI. [1971]: Parametrix of elliptical operators with infinitely many independent variables.
26, No. 2, 155-174. English transl.: Russ. Math. Surv. 26, No. 2,91-112 (1972), Zbl.221.35070
Vishik, MI., Ehskin, G.I. [1967]: Elliptical convolution equations on bounded domains and their
applications. 22, No. 1, 15-76. English transl.: Russ. Math. Surv. 22, No. 1, 13-75 (1967), Zbl.167,
448
Vishik, M.I., Ladyzhenskaya, O.A. [1956]: Boundary value problems for partial differential equations
and certain classesof operator equations. If, No. 6,41-97. English transl.: Am. Math. Sot., Transl.,
II. Ser. 10,223-281 (1958), Zbl.73,99
Vlasov, L.P. [1973]: Approximative properties of sets in normed linear spaces. 28, No. 6, 3-66.
English transl.: Russ. Math. Surv. 28, No. 6, l-66 (1973), Zbl.291.41028
Volevich, L.P., Gindikin, S.G. [1972]: The Cauchy problem and problems associated with it for
convolution equations. 27, No. 4, 65-143. English transl.: Russ, Math. Surv. 27, No. 4, 71-160
(1973), Zbl.244.35015
Volevich, L.R., Paneyakh, B.P. [1965]: Certain spaces of general&d functions and embedding
theorems. 20, No. 1, 3-74. English transl.: Russ. Math. Surv. 20, No. 1, l-73 (1965), Zbl.135,165
Zajdenberg, M.G., Krejn, S.G., Kuchment, P.A., Pankov, A.A. [1975]: Banach bundles and linear
operators. 30, No. 5, 101-157. English transl.: Russ. Math. Surv. 30, No. 5, 115-175 (1975),
Zbl.317.47018
Zharinov, V.V. [1979]. Compact families of locally convex topological vector spaces, FrechetSchwartz and dual Frechet-Schwartz spaces.34, No. 4,97-131. English transl.: Russ. Math. Surv.
34, No. 4,97-131 (1979), Zbl.412.46003
V. Articles from Previous Volumes of the Encyclopaedia of Math. Sciences
Aleksandrov, A.B. [1985]: Theory of functions on a ball. In: Khenkin, Vitushkin (eds.): Several
Complex Variables II. Vol. 8, 115-19O,Zbl.614.32005. English transl.: Enc. Math. Sc. 8, SpringerVerlag: Berlin (to appear)
Arkhangel kij, A.V., Fedorchuk, V.V. [1988]: Basic concepts and constructions of general topology.
In: Arkhangel kij. Pontryagin (eds.): General Topology I. Vol. 17. English transl.: Enc. Math. SC.
17, Springer-Verlag: Berlin, 1990, l-90
Cornfel , I.P., Sinai, Ya.G., Vershik, A.M. [1985]: General ergodic theory of groups of measure
preserving transformations. In: Arnol (ed.): Dynamical Systems II. Vol. 2,5-l 12. English transl.:
Enc. Math. SC.2, Springer-Verlag: Berlin, 1989, l-98
212
Bibliography
Dubrovin, B.A., Krichever, I.M., Novikov, S.P. [1985]: Integrable systems. In: Amol , Novikov
(eds.): Dynamical Systems IV. Vol. 4, 179-248. English transl.: Enc. Math. SC.4 Springer-Verlag:
Berlin, 1989, 173-28O,Zb1.591.58013
Dyn in, E.M. [1987]: Methods of the theory of singular integrals (Hilbert transform and CalderonZygmund theory). In: Khavin, Nikol kij (eds.): Commutative Harmonic Analysis I. Vol. 25,
197-202. English transl.: Enc. Math. SC.15, Springer-Verlag: Berlin, 199l,Zbl.661.42009
Evgrafov, M.A. [1986]: Series and integral representations. In: Gamkrelidze (ed.): Analysis I. Vol.
13,5-92. English transl.: Enc. Math. SC.13 Springer-Verlag Berlin, 1989, 1-82,2b1.655.40001
Fedoryuk, M.V. [1986]: Integral transforms. In: Gamkrelidze (ed.): Analysis I. Vol. 13, 211-253.
English transl.: Springer-Verlag, 1989, 193-232,Zbl.658.44001
Khavin, V.P. [1987J: Methods and structure of commutative harmonic analysis. In: Khavin, Nikol kij
(eds.): Commutative Harmonic Analysis I. Vol. 15, 6-133. English transl.: Enc. Math. SC. IS,
Springer-Verlag: Berlin, 199l,Zb1.656.43001
Kirillov, A.A. [1985]: Geometric quantization. In: Amol , Novikov (eds.) Dynamical Systems IV.
Vol. 4, 141-178. English transl.: Enc. Math. SC.4, Springer-Verlag: Berlin, 1989,137-172,Zb1.591.
58014
Kislyakov, S.V. [1987]: Classical problems of Fourier analysis. In: Khavin, Nikol kij (eds.): Commutative Harmonic Analysis I. Vol. 15, 135-195. English transl.: Enc. Math. SC.15, Springer-Verlag:
Berlin, 199l,Zb1.655.42006
Sergeev, A.T., Vladimirov, V.S. [1985]: The future cone of complex analysis. In: Khenkin, Vitushkin
(eds.): Several Complex Variables I. Vol. 8, 191-266. English transl.: Enc. Math. SC.8, SpringerVerlag: Heidelberg (to appear); Zb1.614.32001
Shafarevich, I.R. [1986]: Fundamental concepts of algebra. In: Kostrikin, Shafarevich (eds.): Algebra
I. Vol. 11, 290 pp. English transl.: Enc. Math. SC. 12, Springer-Verlag: Heidelberg, 1989, l-258,
Zb1.655.00002,2b1.674.12001
Tikhomirov, V.M. [1987]: Convex analysis. In: Gamkrelidze (ed.): Analysis II. Vol. 14,5-101. English
transl.: Enc. Math. SC.14, Springer-Verlag: Heidelberg, 1990, l-92
Tikhomirov, V.M. [1987]: Approximation theory. In: Gamkrelidze (ed.): Analysis II. Vol. 14, 103260. English transl.: Enc. Math. SC.14, Springer-Verlag: Heidelberg, 1990,93-244,Zb1.655.41002
Vitushkin, A.G. [1985]: Remarkable facts of complex analysis. In: Vitushkin (ed.): Several Complex
Variables I. Vol. 7,5-22. English transl.: Enc. Math. SC.7, Springer-Verlag: Heidelberg, 1989,1-18
VI. Reference
Rooks
from
Matematichesknya
Bibiioteka
Babich, V.M. et al. [1964]: Linear Equations of Mathematical Physics. (Mikhlin, S.G. (ed.)) Nauka:
Moscow, Zbl.122,330
Birman, MSh. et al. [1972]: Functional Analysis. (2nd ed.) (Krejn, S.G. (ed.)) Nauka: Moscow,
Zb1.261.47001
Brychkov, Yu.A., Prudnikov, A.P. [1977]: Integral Transforms of General&d Functions. Nauka:
Moscow, Zbl.464.46039
Ditkin, V.A., Prudnikov, A.P. [1971]: Integral Transforms and Operational Calculus. Nauka: Moscow,
Zb1.99.89. English transl.: Pergamon Press: Oxford, 1965
Guter, R.S.,Kudryavtsev, L.D., Levitan, B.M. [1963]: Elements of the Theory of Functions. (Ul anov,
P.L. (ed.)) Phys.-Math. Section of Nauka: Moscow, Zb1.115,47. English transl.: Pergamon Press:
Oxford, 1966
Zabrejko, P.P. et al. [1968]: Integral Equations. Nauka: Moscow
Author Index
Adamyan,
V.M.
227
Akhiezer,
NJ.
29,61
Alakhverdiev,
D.Eh.
216
Alaoglu, L. 145
Aleksandrov,
P.S. 175
Amir, D.
143
Appel, P.E.
65
Arens, R.F.
148,236
Aronszajn,
N.
175,258
Arov, D.Z.
227
Arzela, C. 154
Ascoli, G.
113, 154
Atiyah, M.F.
95,260
Atkinson,
F.V.
162, 167
Babenko,
K.I.
135
Babich, V.M.
244
Bade, W.G.
207
Baire, R. 122
Banach, S. 14, 112,122, 123, 145,146, 151153, 157-160,261
Bari, N.K.
159,169
Berezanskij,
Yu.M.
203,242,260
Bernoulli,
D.
19
Bernoulli,
Johann 10
Bemshtein,
S.N.
26, 29, 30,43, 55, 57, 63
Besov, O.V.
243
Bessel, F.W.
24,28, 52
Beurling,
A. 226
Birkhoff,
G.A.
121
Birman, MS.
206
Bishop, E. 130,209,238
Blaschke, W. 68,224
Bochkarev,
S.V. 134
Bochner, S. 35,52,53,
141,241,249
Bogolyubov,
N.N.
179,246,252
Bohl, P. 191
Bohr, H. 51,52,191
Bonsall, F.F.
215
Borel, E. 53,69, 151
Borok, V.M.
5,249
Bourbaki,
N. 261
Brauer, A. 174
Brodskij,
MS.
221-223
Brudnyj,
Yu.A.
5,258
Bunyakovskij,
V.Ya.
13
Burnside, W.
177,180
Calderon,
A.P.
236,243,257-259
Calkin, J.V.
217
Caratheodory,
K.
108
Carleman,
T. 30, 70
Carleson,
L. 23,214,233
Cauchy, A. 12, 13, 144
Chebyshev,
P.L.
53,54,58,59
Choquet,
G.
143
Clairaut,
AC.
20
Clarkson,
J.A.
130, 158
Cotes, R. 64
Courant,
R. 47
D lembert,
J. 20
Daletskij,
Yu.L.
260
Day, M.M.
131, 158
De Leeuw, K. 192
Delsarte, J. 44, 204
Denjoy, A. 26
Dieudonne,
J. 152,261
Dirac, P.A.M.
11
Dirichlet,
P.G.L.
17,22,42
Dolzhenko,
E.P.
66
Du Bois-Reymond,
P. 23
Duncan, J. 215
Dunford,
N.
143,207,209
Dvoretzki,
A. 136, 172
Dynin, A.S.
172
Eberlein, W.F.
151
Etimov, N.V.
147
Ehjdel an,
SD.
249
Ehrenpreis,
L. 247
Eidelheit,
M.
113
Enflo, P. 134, 158, 166
Esclangon,
E. 93
Euclid 4
Euler, L. 10, 19,20
Faber, G. 63,66
Faddeev,
L.D.
205,227
Fedoryuk,
M.V.
260
Fejer, L. 23,53, 117
Fekete, M.
184
Feller, W. 189
Fermat, P. 10
Fischer, E. 14,47
Fixman, V. 207
Foias, C. 211,214,215,225,226
Fomin, S.V. 260
Fourier,
J. 19,20, 50
Frechet, M.
191
274
Author Index
228,230
Jackson, D. 55,57
Jacobs, K. 192
James, R.C. 130,147,158
Jordan, G. 22,97,98,224
Jordan, P. 130
Kadets, M.I. 5, 122, 135, 158, 191
Kadets, V.M. 5, 136
Kakutaoi, S. 121, 177, 192
Kalisch, G.K. 223
Author Index
Lumer, G. 215
Luzin, N.N. 23,26
Lyubarskij, Yu.1. 5
Lyubich, M.Yu. 5, 192
Lyubich, Yu.1. 43,44,92, 188, 192,209,211
Mackey, G.W. 148,149
Magenes, E. 258
Mairhuber, J.C. 56
Malgrange, B. 247
Mal sev, AI. 97
Marchenko, V.A. 204,205
Marcinkiewicz, J. 256
Markov, A.A. 57-60
Markov, A.A. (junior)
177
Markus, A.S. 5, 138, 170,212,219,220
Markushevich, A.I. 135
Martineau, A. 251,252
Maslov, V.P. 48,260
Matsaev, V.I. 170,205,209,210,215,217,
219,222
Mautner, F.I. 203
Mazur, S. 107,113,122,131,146,233
Mellin, H. 39
Mercer, T. 78
Mergelyan, S.N. 29,65,66
Mikhlin, S.G. 168,258,259
Mil an, D.P. 130,137-139,162, 182
Milyutin, A.A. 122
Minkowski, H. 14, 138
Mityagin, B.S. 172
Miyadera, I. 189
Molchanov, A.M. 51
Molin, F.Eh. 180
Montel, P. 151
Morimoto, M. 252
Morris, P. 143
Mukminov, B.R. 213
Miintz, C.H. 133
Najmark, M.A. 226,239,241
Nevanlinna, R. 31
Newton, I. 6
Nikodjrm, 0. 142
Nikol kaya, L.N. 210
Nikol kij, N.K. 5,210,212,220,221,227
Nikol kij, S.M. 243,255
Noether, E. 89
Noether, M. 82
Numakura, K. 192
Ostrowski, A.M.
70,210
Paley, R.E.A.C.
38, 163
215
276
Sieklucki, K. 56
Singer, I.M. 95,260
Slobodetskij, L.N. 244
Smith, K.T. 175
Sobolev, S.L. 242-244
Sobolevskij, P.E. 189
Sokhotskij, Yu.V. 83
Stechkin, SB. 147
Steinhaus, H. 123
Steinitx, E. 136
Steklov, V.A. 47
Stieltjes, K. 12, 28
Stiickert, B. 255
Stone, M.H. 104,201,233,234,261
sturm, c. 46,47
Sudakov, V.N. 260
Sushkevich, A.K. 192
Sylvester, J.J. 99
Szokefalvi-Nagy, B. 214,225,226
Taklind, S. 243
Tannaka, T. 241
Thorin, G.O. 256
Tikhonov, A.N. 145,174,243
Titchmarsh, E.C. 45,48,223
Tkachenko, V.A. 5,44,209
Toeplitz, 0. 96, 123, 159, 193
Tonelli, L. 56
Tricomi, F.G. 259
Triebel, H. 255
Tzafriri, L. 157
Author Index
Van Kampen, E.R. 241
Vershik, A.M. 5,260
Vitushkin, A.G. 65
Vladimirov, V.S. 252
Volterra, V. 10
VonNeumann, J. 124,130,178,179,191,
193,194,202,214,215,261
Walsh, J.L. 65
Weierstrass, K. 24,28,65,68, 151,233,
234
Weil, A. 241
Wermer, J. 220,234
Weyl, H. 31,48,50, 51, 180,202
Widder, D.V. 40,189
Wiener, N. 27, 38, 39, 79, 80, 163, 235,
236
Williamson, J.H. 165
Wintner, A. 51
Wold, H. 214
Yang, C.T. 57
Yentsch, R. 175
Yosida, K. 189, 192
Young, L.C. 252
Zharinov, V.V. 252
Zobin, N.M. 172
Zolotarev, E.I. 58
Zolotarev, G.N. 249
Zygmund, A. 256,259
Subject Index
Action 10, 178
Aftine manifold 88
Algebra, Banach 27,232
- - commutative 27
- Boolean
04
- function 101
- group 240
- local 101
- of sets 12
- operator 239
- - reducible 176
- semisimple 102
- spectral 102
- symmetric 234
- uniform 234
- universally enveloping 242
- Wiener 27
- - with weight 237
- with involution
238
Algebraic interior 108
Almost-period
52
Annihilator of a subspace 89
Ball, q-unit 111
Banach limit 178
Barrel 122
Bases, quadratically close 169
Basis, Bari 169
- Hamel 85
- Riesz 159
- Schauder 134
- unconditional
135
C*-algebra 239
Capacity, analytic 65
- - continuous 65
Central symmetry 108
Centralizer of a subalgebra 176
Centre of an algebra 176
-0fmass
143
Character of a representation
180
- of a semigroup 180
- of an algebra 104
Charge 131
Chebyshev alternant 53
Class, Hardy 38,225
- Holder 26
- Lipschitz 26
- Nevanlinna 32
- quasianalytic 69
Levinson
85
209
- Lipschitz 9, 163
- of regularity of a functional 16
Cone 109
- minihedral
118
-total
186
Constant, interpolation
257
- Lebesgue 55
Convergence, absolute 101,135
- mean square 13
- unconditional
135
Convolution
22,240
- of sequences 27
Coordinates of a vector 86, 134
Corank of a homomorphism
90
Cross-property of a seminorm 173
Decomposition, triangular 220,222
- Wold 214
Defect of a homomorphism
90
Deficiency index of a symmetric operator
-of an isometric operator 185
Definiteness, procedure for 230
S-function 11
Derivative, Frechet 163
- generalised 244
- of fractional order 42
- Radon-Nikodym
142
Determinant, Fredholm 77
- of a linear operator 170
Deviation 53, 147
194
278
Differential of a function 6
- of a mapping 163
Dilatation
226
Dimension, algebraic 86
- topological
136
Direct sum of linear operators
Disk algebra 234
Dissipation 18
Distribution
245
- of compact support 246
- of finite order 245
- probability
12
- regular 245
- tempered 250
Duality of linear spaces 147
Subject Index
92
Subject Index
Gap between subspaces 137
Generalised sum of a series 100
Group, amenable 177- 178
-dual
240
- one-parameter
186
- unimodular
179
Half-open interval 108
Half-space 113
Hamiltonian
46,205
Harmonic oscillator 17
Homomorphism, absolutely summit xg 172
- bounded 153
-closed
155
- compact 165
- conjugate
-algebraic 90
- - topological
155
- finite rank 93
- Hilbert-Schmidt
168
- Noetherian 93
- nuclear 169
- *- 238
Hull, convex 108
- - absolutely 108, 138
- - closed 138
- linear 88
Hyperfunction
251
Hyperplane 88
- of support 115
- separating 113
-almost 114
- - strictly 128
Ideal, maximal 101
- - standard 106
- simple 102
- symmetric 238
Index of a homomorphism
93
- of an indefinite metric 228
- of nilpotence 97
Inequality, interpolation
257
- triangle 13, 109
Intimum 118
Integral, Bochner 142
- Brodskij 221
- Dirichlet 243
- Fourier 36
- Gel and-Pettis 140
- Lebesgue 12
- Lebesgue-Stieltjes 15
- of Cauchy type 82
- of fractional order 42
Invariant mean 177
279
93
280
Subject Index
- Laplace 18
- linear 11,91
- metric 228
- non-quasianalytic
209
- normal 193
- normally resolvable 162, 166
- pseudoditferential
259
- quasiconservative 212
- quasidissipative 212
- quasinilpotent
184
- scattering 206
- selfadjoint 73,193
- shift 99, 225
- - weighted 221
- symmetric 46,193
- - maximal 196
- - non-negative 196
- unicellular 92,221
- Volterra 186
-wave
205
Order of a pseudodifferential operator
- of an eigenvalue 97
215
202
259
Subject Index
- of uniform boundedness 123,154
- variational 10
Problem, Cauchy 164,186
- - pointwise stable 188
- - stable 189
- Dirichlet 243
- of scattering theory
-direct 204
- - inverse 204
- Riemann-Hilbert
82
- - additive 82
- Sturm-Liouville
45
Process, interpolation
- - of Lagrange type 63
- - of Newton type 63
- Newton
8,163
- quadrature
-Cotes 64
- - Gaussian 64
Product, Blaschke 68
- of measures 15
- scalar 13
- tensor 173
Projection 91
- boundary 192
-root
98
Property, approximation
165
- of reciprocity of polars 148
- Radon-Nikodym
142
Pseudodistance 228
Pseudonorm
120
Quasimodel of a linear operator
Quasinilpotent
103
Quasinorm
121
Quasipolynomial
40
Quasiprojection
209
Radical of an algebra 102
Radius, spectral 184
- Weyl 31,SO
Random events 12
- variables 12
Rank of a homomorphism
90
- of a set of vectors 88
Ray 109
Regularisor 94
Representation 179
- almost-periodic
191
- bounded 179
- canonical functional 103
- contracting 179
- isometric 179,233
- regular 180
222
282
- general&d
248
Space, Banach 14
- barrelled 122
- configuration
16
- conjugate
-algebraic 88
- - topological
129
- coordinate
134
- countably Hilbert 173
- defect 182
-factor
88
- Frechet 122
- fully complete 152
- function 11
- Hilbert 14
- intermediate 256
- interpolation
257
- linear topological
119
- - complete 122
- locally bounded 126
- - convex 124
- Lorentz 253
- Mackey 149
- Marcinkiewicz
253
- maximal ideal 103
- Monte1 150
- Nikol kij-Besov 254
- normed 14
-nuclear
171
- of Lipschitz type 254
- ordered 117
- Orlicz 253
- phase 178
- pre-Hilbert
14
- probability
12
- Ptak 152
- reflexive algebraically 90
- - topologically
130, 150
- Riesz 118
- Schwartz 250
- second conjugate 150
- Slobodetskij 254
- smooth 130
- Sobolev 254
- strictly convex 131
- super-reflexive 158
- symmetric 257
- uniformly convex 130
- - smooth 130
Spectrum 51,97, 103,181
- absolutely continuous 200
- approximate
182
-Bohr
52
- combined 236
Subject Index
- continuous 51, 199
- discrete 51, 199
- residual 182
- simple 224
- singular 199-200
Subordination of metrics 158
- of norms 158
Sum of spaces, external direct 89
- - - orthogonal
87
Sum of subspaces, direct 87
---topologicaf
157
- - - orthogonal
157
Support of a distribution
246
- of a function 86
- of a measure 143
- of a vector 85
- of an analytic functional 251
Supremum 118
Symbol of a differential operator 248,259
- of a pseudodifferential operator 259
System, biorthogonal
61
- Chebyshev 56
- complete 27, 132-133
- distal 136
i dynamical 178
- Markov 57
- - normalised 59
- minimal 135
- w-linearly independent
134
- orthogonal
13
- orthonormal
13,134
Theorem, boundary spectrum splitting-off
192
- closed graph 155
- convolution 22
- ergodic 191
- expansion 40
- implicit function 9
- inverse homomorphism
156
- local 235
- open mapping 156
- separating hyperplane 113,114,128
- spectral 98,197,200,230
- spectral mapping 99,104,207,209,235
- stability of the index 95
- superposition 209
Theory, Fredholm 76
- Hilbert-Schmidt
77
- Noether 84
- scattering, non-stationary
205
- - stationary 204
Topological isomorphism 121
Topology, consistent with duality 148
Subject Index
- DieudonnC-Schwartz
245
- locally convex 124
- Mackey 149
-norm
150
- of convergence in measure 120
- of uniform convergence 125
- - on bounded sets 150
- - on compact sets 125
- strong 145
- tensor product 173
- uniform 150, 154
-weak
145
- weakened 145
Total family of linear functionals 124
- - of projections 92
- - of seminorms 124
Trace 170
-matrix
170
- spectral 219
- tensor 169
Transform, Cayley 194
- Fourier 36
- - generalised 49, 248
- - operator 210
- - -Plancherel 37
- - -Stieltjes 37
- Gel and 233
- Laplace 39
-finite 44
-local 43
283
- - -Stieltjes 42
- Legendre 252
Travelling wave 18
Ultradistribution
250
Ultrapower
106
Unit sphere 127
Variation of a functional 16
Vector, channel 223
- cyclic 224
-root
97
Vector-function, absolutely integrable
- analytic 144
- Bochner integrable 142
-entire
144
- Gel and-Pettis integrable 140
-simple
141-142
w*-topology
144
Wedge 109
- generating 109
Weight 28
Weights of a quadrature formula
Weyl disk 32, 50
- point 32, 50
- radius 31,50
Young
relation
132
59
141