Mathematical Notes, Voi. 64, No.

3, 1998

T h e B e s t A p p r o x i m a t i o n to a Class o f Functions o f S e v e r a l V a r i a b l e s
by A n o t h e r Class and R e l a t e d E x t r e m u m P r o b l e m s
V. V. Arestov UDC 517.518, 517.983

ABSTRACT. We study the relationship between several extremum problems for unbounded linear operators of
convolution type in the spaces L-y = L-r(R'n), m > 1, 1 < 3' _<oo. For the problem of calculating the modulus
of continuity of the convolution operator A on the function class Q defined by a similar operator and for the
Stechkln problem on the best approximation of the operator A on the class Q by bounded linear operators,
we construct dual problems in dual spaces, which are the problems on, respectively, the best and the worst
approximation to a class of functions by another class.

K E Y WORDS: function space, convolution operator, Stechkin problem, linear operator, function class, best
approximation, modulus of continuity.

In this paper we study the relation between several extremum problems for u n b o u n d e d linear operators
of convolution type in the spaces L~ = Lz(Rm), m _> 1, 1 _< 7 -< co. For the problem of calculating the
modulus of continuity of the convolution operator A on the function class Q defined by a similar operator
and for the Stechkin problem [1] on the best approximation of the operator A on t h e class Q by bounded
linear operators, we construct dual problems in dual spaces, which are the problems on, respectively, the
best and the worst approximation to a class of functions by another class. Similar results for differentiation
operators on classes of functions of one variable in the spaces L ~ ( I ) on the axis I -----( - c o , co) and on the
half-axis I = [0, co) were already obtained by the author in [2] and for periodic functions by Klots [3]. At
present, there are many papers dealing with the problem studied here; a survey of these problems is given
in [4, 51.
w Introduction. Statement of the problems
For 1 _< 3' < co, let L.r = L ~ ( R " ) , m >_ 1, be the space of real measurable functions x on R '~ with
finite norm

For 7 - co (unless otherwise specified), Lz is the space Co -- C0(R 'n) of continuous functions on R 'n
that tend to zero at infinity; this space is equipped with the uniform norm

Ilxll** = Ilxllc0 = max{lxCt)l : t e Rm}.

Next, let 8 = S ( R '~) be the space of rapidly decreasing infinitely differentiable functions on R '~ (the
space of test functions), and let 8 ' -- 8 ' ( R m) be the corresponding space of distributions (e.g., see [6-8]).
By (0, ~) we denote the value of a functional (distribution) 0 e 8 ' on an element x E 8 . The function
y(~/) -- (0, a~z), where the operation a~, 1/e R " , is defined by the formula (~r~x)(t) = x(T/- t), is called
the convolution 0 * x of elements 0 E S' and z E 8 . For each 0 E 8 ' , by # we denote the functional
acting by the formula

thus the operation 0 ~ 0 just changes the sign of the argument of a distribution.

Translated from Malematicheskie Zametki, Vol. 64, No. 3, pp. 323-340, September, 1998.
Original article submitted September 1, 1997.
0001-,13,16/98/6,13,1-0279520.00 C)1999 Kluwer Academic/Plenum Publishers 279
 In what follows, we discuss several extremum problems for linear unbounded o p e r a t o r s of the convolution
type (with some distribution); these operators are constructed by the following unified scheme. For an
element 0 E S ~, we define an operator e on S and the formally adjoint operator O* b y the formulas

Oz = O . z , O*z = 0 . = , zES.

We always assume that the condition 8*x E S holds for any function = E S , i.e., 0 2 C S (or, which is the
same, O*S C S); for instance, this condition is satisfied if 8 is compactly s u p p o r t e d . Let us extend the
operator 0 to a wider set by using the Sobolev scheme. By E = E(R m) we d e n o t e the set of measurable
locally integrable functions on R m for each of which there exists an exponent # = #(x) < 0 such that

/ I=(t)l(1 + IriS)" dt < oo;

here and in the following we do not specify the integration domain in integrals o v e r R m . One can readily
see that ~ C S ' . For a pair of functions x, y E ~ , we say that x belongs to t h e domain D ( O ) of the
operator 0 and y = Ox if the relation

f xO* f dt = f yCdt (i)

holds for any ~b E ,~. In a similar way, we can extend the operator 0 " . One c a n readily see that a
function x belongs to the domain D(O*) of the operator O* if and only if ~ E D ( O ) , and moreover,
(O*x)(~) = ( O ~ ) ( - ~ ) , where ~ E R " . The operators 0 and O* axe translation invariant on D ( O ) and
on D(O*), respectively, i.e., r,~O = Or~ and r~O* = O*r~ for any 1/E R m, w h e r e r, I is the translation
operator defined on E by the relation (v,x)(t) = x(t - rl).
Suppose that a pair of (formally adjoint) operators A, A* with domains D ( A ) and D(A*) is con-
structed by the above scheme from an element a E S ' . We assume that one more p a i r of operators B , B*
with domains D(B) and D(B*) is determined in the same way for another e l e m e n t b E S ' .
Suppose that p, q, r, and s are parameters satisfying the conditions 1 < p, q, r , s < oo. We introduce
the spaces

Wp(B) = {x E D ( B ) : Bz 6 L , } , W r , , ( B ) = Lr N Wp(B) = {x 6 Lr f'l D ( B ) : Bx 6 L,}.

On the set W~,p(B) we define the class

Q = Q,.,,,(B) = e W,,,(B): IIB II, <- t}.

By fir" we denote the set of bounded linear operators from Lr into L , . Let fl~.(N) be the set of operators
T E B~ whose norm is bounded by a number N >__0, that is, IITll = IITllL,--L. < I v .
One of the problems discussed in this paper is the Stechkin problem on the b e s t approximation to an
operator A on the class Q in the metric of the space Lq by the set of bounded linear operators T E fl~(N),
i.e., the problem of studying the variable

g(N) = E(A, B ; r , s ; p , q ; N) = i n f { U ( T ) : T E B~(N)}, (2)

where
U(T) = sup{l[Ax - T = [ I , : x e Qr,p(B)}

and if Wr,p(B) ~ D(A), then we assume by convention that U(T) = do for a n y operator T E B~(N).
In the classical version of this problem [1], the parameters s and q coincide: s = q. Then we denote
the variable (2) by E ( N ) ; thus we have E(N) = $(A, B ; r , q;p, q; N). P r o b l e m (2) was studied in
considerable detail; in particular, the relationship between this problem and o t h e r extremum problems
was explored and the solution of this problem was found for several special o p e r a t o r s in classical function

280
spaces (see the references in [5]). In what follows, we use the results of the author [9-13] concerning this
problem. The function
w(,~) = suP{llAzll, : z e Q,. p ( B ) , Ilzllr ___6} (3)
of a variable 6 E [0, oo) is called the" modulus of continuity of an operator A on the class Q = Q , . r ( B ) .
We set

R(N) = sup{w(6) - N 6 : 6 > 0} = sup{llAxli, - N[[xllr : x e Q r , , ( B ) } ,

A(6) = i n f { E ( N ) + g 6 : N E 0}.

Stechkin [1] noticed that the following inequalities are satisfied:

E(N) > R(N), N > O, (4)

,,(6) < ~(6), 6 > 0;

moreover, if there exists an element x 6 Q and a bounded linear operator T 6 B~ such that

IIA~ll = U(T) + IITII. II~ll,

then the relations
w([[x[[) = [[Az[[, E([[T[[) = U ( T ) = [[Ax[[ -[[T[[- [[z[[
hold and hence the extremum is attained at the operator T in the problem E ( N ) for N = [[T[[ =
[[T[[L,-.L, and at the element x in problem (3) for 6 = [[x[[. This statement yields a simple but effective
and frequently used lower bound for the best approximation to an operator via its modulus of continuity.
So far problems (2) and (3) have been studied most completely for the kth-order differentiation operators
D k x = x (k) on the class

of n times differentiable functions (0 _< k < n) in the spaces L-~ = L T ( I ) , 1 _< 7 -< oo, and of functions
of one ~ - ~ a b l e on the real = i s I = ( - c o , r or on the h ~ - ~ s I = [0, oo) (see the references in [4, 51).
In this case we have

E ( N ) = E.,~(N) = inf{U(T): IITIE,-.L, _< N}, (5)

U(T) = sup{[lx(k) - Txl[, : z e Q~',p}.

It is well known that if the condition k + 1/r - 1/q > 0 is satisfied, then the relation

n - k + 1/q - l / p
E,,~(g) = g-'rE,,k(1), 7 = k + 1/r - 1/q

holds. In this case the function (3) has the form

~ ( 6 ) = sup{ll~(k)ll, : II~ll,- ~ 6, II~(")llp <__ 1};

if n - 1/p + 1/r > 0, then this function satisfies the relation

n - k1/p + 1/q
-
w(6) = K 6 ~, K =w(1), a =
,~ - 1 / v + 1/,-

The variable K = w(1) is the least (the best) constant in the following well-known inequality relating the
norms of derivatives of functions (the Kolmogorov inequality):

n - k - 1 / . + 1/q
= ~ = 1 - ~. (6)
n-1/p + l/r '

281
If k + 1 / r - 1/q > 0, then the Stechkin inequality (4) turns into the following relation between the variable
En,k(N) and the best constant K in inequality (6):

E,~,k(N) > ~a~/~K1/#N -~1~, N > O.

Problems (3) and (2) are related to the problem on the approximation of a f u n c t i o n class by another
function class and to the corresponding problem of the linear approximation of a class by another class.
Suppose that pt, q~, and r ~ are parameters satisfying the conditions 1 < p*, q~, r * < c~ and so far in no
way related to the parameters p, q, and r. By using the above scheme, we define one more operator H
for an element h E S t 9 In the spaces

We(B) = {V C D ( B ) : Bqo C Le}, We(H) = ( r C D ( H ) : Hg, C Lq,}

we introduce the classes

Qe(B) = c We(B): IlB lle _< 1},

Q2 = NQr,(B) = {q0 C We(B): IIB lle <__ N},
Qa = Qr = {r 6 We(H): IIHr162 < 1}.

We set

F(N) -- F(Q1, Q2)v' = F(Qr NQe(B)) v, = sup{F(0, NQr,(B))v, : ~b E Q u (7)

where
F(r NQ,~(B))p, = inf{llr - qoll~,' : qo e NQe(B)}.

The variable (7) is called the variable of the best approximation (or simply the bes~ approximation) of the
class Q1 = Qu (H) by the class Q2 = g Q e (B) in the space Lv,.
The variable

G(N) = G(Q1, Q2)v' = G(Q,,(H), NQe(B)) v, = inf{a(Qr P)v' : P C 7~(N)}, (8)

where
G(Q,,(H); P)v' = sup{lie - Pr : r e Qr
and the infimum is taken over the set 79(N) of all linear mappings P of the class Q1 into the class Q2,
is called the best linear approximation of the class QI by the class Q2 in Lv,.
The fact that problems (7) and (8) are related to problems (3) and (2) was first p o i n t e d out in 1965-1967
by participants of the S. B. Stechkin seminar at the Sverdlovsk Division of the V. & . Steldov Mathematics
Institute (presently, the Institute for Mathematics and Mechanics of the Ural Division of Russian Academy
of Sciences) in connection with the Stechkin problem (5) for the differentiation o p e r a t o r s and with the
Kolmogorov inequalities (6) for classes of functions of one variable on the real a x i s and on the half-axis.
The central idea was that the upper bound for the variable G(N) of the best l i n e a r approximation of a
class by another class, and hence for the variable F(N) of the best approximation of a class by another
class can be expressed in terms of the variable E(N). The results obtained in solving these problems are
contained in [14-18]. Let us outline the essence of this theory as applied to p r o b l e m s studied here. We
assume that the following assumptions are satisfied.
1) The parameters of the problem satisfy the conditions p' = q, q' = p, and r' ~---q.
2) The operator B is the composition of the operator A and the operator H , i.e., B = H o A; more
precisely, we have the relations

D(B) C D(A), AD(B) C D(H), HAx = Bz, x e D(I~).

282
In the following this assumption is called the factorization condition (F).
3) The set A W v ( B ) is a sufficiently rich subset of the space Wp(H) ; for simplicity, we restrict ourselves
to the case
AWr(B ) = Wv(H ). (9)
4) In problem (2) we can restrict ourselves to the set B' of operators T E Bp,q that are defined on the
set Wv(B), are linear there, and commute with the operator B (for instance, in problem (2) there exists
an extremum operator T with such properties). Moreover, for any such operator T E B t , the variable
U(T) = sup{HAx - TxHq : x G Qv(B)} coincides with

U(T) = sup{llAz - Tzllq : z e Q = Qp,v(B)}.

Under these assumptions, we have the inequality

G(N) = G(Qv(H), NQq(B))q < E ( N ) = F..(A, B; p, q; p, q; N), (lO)

and hence the inequality F ( N ) < E ( N ) is satisfied. Indeed, let us assign an operator P E T~(N),
N = IITIIL,--L,, to each operator T E B' according to the following rule. In view of property (9), for each
function C q Wv(H) there exists a function z E Wv(B) such that Az = r We introduce the operator P
by setting P C = T x , C E Wv(H ) . We have

BPC = B T z = T B x = T H e .

Therefore, IIBPr _< IITIIL,-L, IIHOlIp and hence P G V ( N ) , N = IITllL,--L,. Next, we see that

I1r - POll, = IIAx - Txllq < Y(T)IIBxlIp = (T)IIHr

and hence inequality (10) holds by virtue of the fourth assumption.
In fact, there exists another natural relationship between the problems introduced above. Namely
(under some specific assumptions on the operators and for appropriate relations between the parameters),
the problem of approximating a class by another class is dual to problem (3), and the linear problem
of approximating a class by another class is dual to problem (2). This fact was first pointed out by the
author for the differentiation operators D k on the axis and on the half-axis [2]; a similar result for periodic
functions was obtained by Klots [3].
Problems of approximating a class by another class are considered in dual spaces with respect to
problems (3) and (2). From now on, we assume that the problem parameters satisfy the HSlder conjugation
condition (H); more precisely, we assume that

1 1 1 1 1 1

Now it is necessary to define some function classes more precisely (that is, slightly change their defini-
tion). In the case 7 = co, together with the space L o o = Co, we shall consider the following two classes:
the space Lor = L o o ( R ' ) of essentially bounded measurable functions on R " and the space C = C ( R m)
of bounded continuous functions equipped with the uniform norm. Sometimes, for convenience, we shall
also denote the spaces L. t = L.t(Rm), 1 < 7 < co, by L-t = L't(Rm) 9 Let M = M ( R m) be the space of
signed a-additive finite Lebesgue measures on R " ; we use the standard notation V E for the variation of
a measure Z E M . For 1 < p < co the space Lpat of bounded linear functionals on Lp coincides with Lp,,
1/p + 1/p' = 1, if 1 < p < co and with Loo if p = 1 (i.e., with the space Lp, for all p, 1 < p < co); more
precisely, the functional T E L; has the form

Tz = / z(t)~(t) dr, (11)

283
where ~ e L a, and IITIIL; = I1~11,,- If p = co, then the space L~o = C~ is t h e space of measures
M = M ( R m) and each functional T 6 C~ can be written in the form

T27 f =(t)~_(t), (12)

where -- e M ; in this case V - = IITIIc; 9 In the following we shall denote the integrals on the right-hand
side of (11) and (12) by (x, ~) = (x, ~)a and (x, dE) = ix, dE)co, respectively, so t h a t

(=, ~) = (=, ~). = =(t)~(t) at, 1 <_ p < co,

9 f
(=, dE) = (=, dE)co = f =(t) dE(t), a = co.

Let s 1 < 7 < _ c o , b e t h e s p a c e L,oo for 7 = c o , t h e s p a c e L.y for 1 < 7 < co (hence the space L.r
for 1 < 7 -< co), and finally, the space M for 7 = 1. For the norm of elements z 6 s we shall use the
notation I1=11~,, as welt as I1=11,; in particular, we write t1~111 = V ~ , -= e M = C x . In this notation we
have L~ =/~v', 1 <_ p _< co.
Let us extend the domain of the operators in question. Let ~ = r be t h e space of signed a-
additive a-finite Lebesgue measures .~ on R " (with alternating signs) for each o f which there exists an
exponent /z =/~(-~.) _< 0 with properties

f(1 + It12)" dl=_l(t) < co ;

here I=1 is the modulus of the measure E, i.e., the (nonnegative) measure defined by the relation I=1 =
E + - E - , where E + and E - are the negative and positive components of the m e a s u r e E. In what follows,
sometimes it is convenient to denote a measure E 6 :~4 also by ~ = dE. Obviously, M C .M C S ' . For
a locally absolutely continuous measure E 6 ~ 4 , its derivative ~ (with respect to t h e classical Lebesgne
measure dr) belongs to the space ~2. Conversely, each function ~ 6 E generates a. measure E 6 ~ in a
natural way. In this sense, we have ~ C jb4. Suppose that (9 is any of the o p e r a t o r s A, A*, B , B*, H ,
and H*. If a function a: 6 ~ and a measure E 6 A4 are such that the relation

f.(o(o',)(O f,(,)dE(O=

holds for any function ~b 6 S , then we assume that x is contained in the domain Z ) ( O ) of the operator O
and the measure E, or, which is the same, ~ = dE is the value Oz of the operator O , on the function x.
For a pair of spaces 2d and 3} each of which is either one of the function s p a c e s ~ , Lp, Lp, and C
introduced above or one of the measure spaces fie/ and M , by ~D(X, y ; O) we d e n o t e the set of elements
(i.e., functions or measures) x 6 X such that the value Oz of the operator O on the e l e m e n t x, determined
by the Sobolev scheme, belongs to 3}. Thus the set D(O) just introduced is t h e set ~D(~, A4 ; 0 ) .
Moreover, for example, in this notation we have

D ( e ) = D(2, 52,; e ) , wp(e) = v ( ~ , Lv; e), w~,p(e) = Z~(L~,/~v ; e).

Let us introduce the spaces

W,.,(B') = {~o 6 D ( B ' ) : B*~ e s = D(Z, s ; B*),

Wq,(H*) = {r e D ( H * ) : H * r c s = v ( ~ , Lq, ; H*)
and define the following classes in these spaces:

Qr,(B*) = {~ e Wr,(B*): IIB*~II~ - 1},

Q2 = NQr,(B*) = {~ e W~,(n*): IIB*~llr' ~ N},
Q, = Qq,(H*) = {r 6 Wq,(n*): IIH*r162 ___1}.

284
For an element r E "~r we set

.T'(r NQ,-,(B*))p, = inf{llr - ~llp' : ~ e NQ~,(B*)}, r 6 ]d;q,(H*) ; (13)

in the space Lp, this is the best approximation of the function r E 14;r by the class Q2 = NQ~,.,(B*).
Then in Lp, the variable

~'(N) = 9v(Qr NQ,.,(B*))p, = sup{~(r NQ.,-,(B*))p,: r E Qr (14)

is the best approximation of the class Q~ = Qr by the class Q2 = NQ,.,(B*). Next, by

"P(N) = P( Q1, Q2) = :P(Qr NQ.,.,(B*))

we denote the set of linear mappings of the class Q1 into the class Q2 or, more precisely, the set of linear
mappings P : ],Yq,(H*) -.--*],V,.,(B*) with the property

HB*Pr < NHH*OHr r E Wq,(H*),

or, which is the same, the property PQr C NQ,.,(B*). For an operator P 6 ~~ we set

3"(p) = ~(Q,,(H*); P),,, = sup{lie - POllp, : r e Qr

Then the variable

G(N) = Q(Q~, Q2)p, = Q(Qr NQ,~(B*))p, = inf{G(Qr P)p, : P 6 P ( N ) } (15)

is the best linear approximation in the space Lp, of the class Q~ = Q r by the class Q2 = NQ.,.,(B*).
Probably, it is more natural to study the best approximation *~'(N) and the best linear approximation
~ ( N ) of the class *QI = Qr by the class *Q2 = NQ,.,(B) in the space Lp,. However, one can
readily see that
*.T(N) = ~'(N), ~ ( N ) = G(N),
and moreover, just the problems .T'(N) and g(N) arise in our further considerations in a natural way.
In what follows, in studying our problems (in the case p = oo) we need to impose one more condition
on the operator B . The domain of the operator B can be extended to the set :D(.A4, .A4 ; B ) . Namely,
for two measures X 6 9,4 and -- 6 A4 we say that X is contained in the domain T)(.A4, .A4 ; B) of the
operator B and the measure E is the value BX of the operator B the function X , that is, E = BX, if
for any function ~b 6 , 9 we have the relation

f (B'r =f
By (M) we denote the condition saying that any measure S E T)(fl4, .h4; B) is locally absolutely con-
tinuous and, consequently, the derivative ~ of each such measure with respect to the classical Lebesgue
measure dt is contained in ~ , i.e.,

v ( M , M ; B) = z)(Z, ~ ; U).

For instance, a fairly wide class of differentiation operators with constant coefficients possesses this prop-
erty. However, in the following we shall need a somewhat weaker (compared with (M)) condition on the
operator B and the parameter p, 1 <_ p <_ c~. This condition, denoted by (M, p), states that any measure
E E T~(.hd, Ep ; B) is locally absolutely continuous, i.e.,

z)(M, s B) = z)(~, L~; B).

285
 w A u x i l i a r y statements
The space
Wr,,, = W~,p(B) = {x E L r n ~ ( B ) : Bx E Lp}
is a Banach space with respect to the norm

II llw,., = 11 11 + IIB IIp.

This fact readily follows from the relation

f zB*r f CBxdt, ceS,

that determines the element B x and from the assumption that B S C S . By I ~ ~,p(B) we denote the
closure of the set S in W~,,(B). Obviously, ff'~,p(B) C Wr,p(B); this embedding can be strict (see [10]
for an example). In what follows, it is important to study the case where the set S is dense in W~,p(B),
i.e.,
= (16)

So far there are many results concerning similar problems (see the monographs [19-21] and the survey [5]).
The following assertion is contained in the author's papers [12] and [10]. To m a k e this statement and
further references to it more convenient, we impose the following condition on t h e parameter p and the
operator B ; this condition will be denoted by (B, p).
Condition (B, p) states that in the case 1 < p < co the operator B has the p r o p e r t y
{gEC0, Bg=0} :. g = 0 ,
and in the case p = 1, the property
{gE(7, Bg=0} ~ g=0. (17)
L e m m a 1. /.f condition (B, p) is satisaed, then the set S is dense in W r , p ( B ) , i.e., property (16)
holds.
It is a consequence of the foUowing very simple statement, whose proof can b e found in [12], that
condition (B, p) also implies that the set B S isdense in the space Lp.
L e m m a 2. The set B S is dense in Lp ff and only ff the operator B possesses the property
{.qeLf riCo, Bg=0} ===>g=0
in the case p > 1 and property (17) in the case p = 10
It is natural to study problems (3) and (2) for the case in which the following c o n d i t i o n is satisfied:
(E) the embedding
W,.,,(B) C Wr,q(A) (18)
takes place, and moreover, there exists a (finite) constant K such that

IIAxliq <_ K(lizll~ + IIBzllp), 9 E Wr,,(B).

In what follows, we always assume that this condition is satisfied. There are m a n y papers dealing
with the study of this property (see the monographs [19-21], the papers [22, 23], and the references
therein). As is known [22], for m = 1 (i.e., on the axis, as well as on the half-axis [0, co)) and for
the differentiation operators of order k and n , 0 <_ k < n, except for some d e g e n e r a t e values of the
parameters, inequality (18) or, which is the same, inequality (6) with a finite c o n s t a n t holds if a n d only if
the condition
-n--qk- - > - k n
r p q
is satisfied.

286
 L e m m a 3. Suppose that the factorization condition (F), condition (B, p), and condition (E) are
satisfied. If for a pair of functions d2 E W e ( H * ) and ~o E VI)r,(B*) the difference X = r - ~ belongs to
the space Lp, , then the representation

(Az, H'r - (x, B*~),. = f dt (19)

holds for any function x E W , , p ( B ) .

P r o o f . First we assume that x E S and hence Ax E $ . In view of assumption (F), it follows from the
definition of the function H * r that

(Ax, H * r = = fw=)(0O(*)d*.
Furthermore, we have
f
dr.

These relations yield representation (19) for functions x E $ .

Now let x E W r , p ( B ) . It follows from the assumption (B, p) that there exists a sequence {xn},~~1761 C $
such that []x - xn[[w.., --~ 0 as n ---* co, i.e.,

I[x -- xn[[r --* 0 and [[Bx - Bx,,[[p ---* 0 as n .--* co.

In view of condition (E), the property that flAx - Ax,,[[, -* 0 as n -* co also holds. By writing (19) for
the functions x,, and passing to the limit as n --* co in the relation obtained, we arrive at (19) for the
function x. The lemma is thereby proved. []

For a function r E YVq,(H*) and a bounded linear functional T E L$, let us consider the variable

v ( r T) = sup{(Ax, H*r - T x : x E Q,,p(B)}. (20)

L e m m a 4. Suppose that the factorization condition (F), condition ( B , p), and condition (E) are
satisfled, and moreover, condition (M,r') is satis~ed in the case p = co. I r a function r E 14~,,(H*) and
a functional T E L r represented in the form (11) for 1 <_ r < co and in the form (12) for r -- co satisfy
v(r T) < co, then there exists a function

~o E W r , ( B * ) (21)

with the properties

B*~o -- ( ~ 1 <_ r < co,

forr--co,f~ (22)

I1r - llp, -- v), (23)

and the representation

(24)
is valid for any function x E B~,p(B).

287
 P r o o f . By S v ( B ) we denote the set of functions x E ,5 such that IJBxllp _< 1. T h e n we set

/z(r T) = sup{<Ax, H ' r - Tx: x E Sv(B)}.

Since S p ( B ) C Q ~ , v ( B ) , we obtain
~(r T) < ~(r T ) < ~ . (2~)
Let us show that there exists a function ~0 E IV,, (B*) such that property (22) holds, t h e representation (24)
holds for functions z E S , mad the relation (similar to (23))

I1r - ~o11r = ~ ( r T)

is satisfied. In this proof, the above statement will be referred to as the s t a t e m e r t t of the lemma on the
function set S .
By using well-known arguments due to Sard (e.g., see Lemma 2 and its application in [24]), we conclude
that the representation
( A x , H*r - Xx = H B x , x E S, (26)
where II is a linear functional on the linear space Y = B S , holds on the set S .
By assumption, Y = B S C S and thus Y C L v 9 We endow the space Y with t h e n o r m of the space L v .
We have [[IIl[y. = / ~ ( r T ) . Let us linearly continue II from Y to the entire space L v so that the norm
remains preserved. For the functional II we have the following representation on L v :

IIy = (!1, X)v, 11 E L v, (27)

where, according to the value of p,

x ~ Lp,, ~(r T) = Ilxllv for X < v < oo, (28)

x=dX, XeM, ~(r for v=o~. (29)
Since r E W e ( H * ) and z E 8 , we have

eA~,H'~), = f C~Ax)(O~Ct)at = fw~)(0~Ct)dt. (30)

Our further considerations depend on the values of the parameters r mad p. F i r s t , let us consider the
case 1 < r, p < oo. The relation

f f

follows from (26), (30), (27), (28), and (11). This relation means that the difference ~ = r is contained
in the domain Z~(B*) of the operator B* and that ~ = B*~0. IIence we have ~0 E ld]r,(B*). In this case,
in view of (28), we obtain
I1r - ~11r = Ilxlb, = ~ ( r T).
Relation (26) coincides with (24) for x E S . Thus we have verified that, for 1 _< r , p < oo, the statement
of the lemma holds on the function class S .
Now let r = oo and 1 < p < oo. In this case the relation

Tx = f x(t)~(t) = f (~(~)- x(O)B~(Odt, 9 ~ s,

follows from (26), (30), (27), (28), and (12). This relation means that the difference ~ = r - X is contained
in the domain ~D(B*) of the operator B* and that ~ = B*~v. Hence, just as a b o v e , we conclude that
for r = cx~ and 1 < p < oo the statement of the lemma also holds for the f u n c t i o n s x E S .

288
 It remains to consider the case in which p = co and 1 < r < co. For 1 <__r < co, by =. we denote
the locally absolutely continuous measure whose derivative is the function ~ in representation (11) of the
functional T . Then (here p = co and 1 < r < co) relations (26), (30), (27), (29), (11), and (12) imply
the representation # m

T= = J =J x(t)), 9 s, (31)
where ~ is a locally absolutely continuous measure whose derivative is the function ~b. Relation (31)
means that the measure q~ - ~ - X is contained in the set D(.A4, L~. ; B*) and that -~.= B*~. In view of
the assumption (M,r'), the measure @ is locally absolutely continuous. W e denote the derivative of this
measure by ~. The measure X = ~ - ~, being the difference of locally absolutely continuous measures,
is also locally absolutely continuous. It follows from (29) that the derivative X of the measure X is
integrable and
~,(r T) = V x = Ilxll, = I1~ - ~o11,.

Thus the statement of the lemma on the set S is valid for all values of the parameters. Under the
assumptions of Lemma 4, the assumptions of Lernma 3 concerning the parameters and operators are
satisfied. Moreover, the functions ~b, 7~, and X = ~b - ~o also satisfy the assumptions of Lemma 3.
According to Lemma 3, the representation (24) holds on the set Wr,p(B). This representation implies
that
v(@, T ) = s u p { ( A z , H * r - rx: z E Q,-,p(B)} _< Ilxll,, = I1r - ~11,, = ~,(,~, T ) .
In view of (25), this proves that property (23) is satisfied.
Let us show that, under the assumptions of this lemma, for any function ~b E PPr there exists
a unique function ~o satisfying properties (21)-(23). Assume that for some function ~b there exist two
functions ~01, ~p2 satisfying the properties (21)-(23). Then their difference ~o = ~o~ - ~oz is contained in
the space D(LW, s ; B*) and satisfied the homogeneous equation B*~o = 0. In view of Definition (1),
this implies that
f (B~)(t)~p(t) dt = O, ( E S. (32)

According to Lemma 2, it follows from condition (B, p) that the set B S is dense in the space L~.
Therefore, it follows from (32) that ~o = 0. The proof of Lemma 4 is complete. []
For a function ~b E ~4~r we set

~r(~b, N) = inf{v(~b, T ) : T E L;, IITII < N}, (33)

where the variable v(r T) is defined by formula (20). Now we shall show that cr(4;, N ) coincides with

~-(~b, N) = ~'(r NQ~,(B*))r = inf{ll@ - ~)11,' : ~ E NQ~,(B')}.

C o r o l l a r y 1. Suppose that the assumptions of Lemma 4 are satis/~ed. Then the relation

~(r N) = 3:(r NQ,~(B*)),, (34)

holds for any function ~b E PVq,(B*).

Indeed, the inequality cr(~b, N) < ~'(~b, N) follows from Lemma 3. The opposite inequality ~-(~b, N) _<
cr(~b, N) follows from Lemma 4.
R e m a r k 1. We can see from the proofs of Lemmas 3 and 4 and from Corollary 1 that, aside from
relation (34), we have the following statement: to each extremum functional T of problem (33) there
corresponds a function ~o E NQr,(B*) for which II~-~llp' = J:(~, N), and conversely, each such function
generates an extremum functional.

289
 w D u a l i t y b e t w e e n t h e p r o b l e m o f calculating the m o d u l u s o f c o n t i n u i t y o f a n
operator and t h e p r o b l e m on t h e best a p p r o x i m a t i o n o f a class by a n o t h e r class
In this section we use the following well-known result due to Gabushin [25] on t h e best approximation
of unbounded linear functionals by bounded linear functionals.
L e m m a 5. Suppose that X is a Banach space, A is a linear (not necessarily bounded) functional
on X , and IC is a centrally symmetric convex set contained in the domain of the functional A. Then for
any N > 0 the variable
e(N) = inf sup(Ax - Tx)
IITIIx- <N ~ K
satisfies the relation
~(N) = sup(A~. - Nil~lix).
z6K:

In the statements treated below, we impose the following additional restriction ( H , p) on the opera-
tor H: the set HWp(H) is dense in the space Lp.
Obviously, condition (H, p) is satisfied if the set H S is dense in Lp; therefore, L e m m a 3 implies
sufficient conditions for the property (H, p) to be satisfied.
T h e o r e m 1. Suppose that the factorization condition (F), condition (B, p), condition (E), and con-
dition ( H, q') are satisfied, and moreover, condition (M,r') is satisfied in the case 1o = oo. Then for any
N > 0 problems (3) and (14) are related by R ( N ) = ~ ( N ) .
P r o o f . We set
a ( g ) = sup{a(r g ) : ~b 9 Qr
where the variable tr(r N ) is defined by formulas (33) and (20). By (34) we have

~(N) = sup{~(r N ) : r 9 Q,,(H')}

= sup{~'(r NQ,~(B*))f : r 9 Qu = ~-(N).

On the other hand, according to Lemma 5, we have the relation

~(r N) = sup{(A~, H ' r - NII~II, : 9 e Q,.,CB)}.

Hence we obtain
a(N) = sup sup {(Az, H*r - NII~II,-}.
z~Q,,p (B) ,PC~,, (H')
Next, it follows from condition (H, q') that

sup{(Az, H*r : r 6 Qu = IIAxll,,

and finally, we arrive at the relation

a(N) = sup{llA~llq - NIl~llr : z e Qr,p(B)} = R(N).

Theorem 1 is thereby proved. []

w D u a l i t y b e t w e e n t h e S t e c h k i n p r o b l e m o f a p p r o x i m a t i n g a n u n b o u r t d e d o p e r a t o r by
b o u n d e d o p e r a t o r s and t h e p r o b l e m o n t h e b e s t a p p r o x i m a t i o n o f a c l a s s b y a n o t h e r class
L e m m a 6. For any values of the parameters p, q, and r (1 < p , q , r < c o ) if the factorization
condition (F), condition ( B , p ) , and condition (E) are satisfied, and moreover, i f condition (M,r') is
satisfied in the case p = oo, then problems (3) and (14) are related by the inequality

~(N) < E(N). (35)

290
 P r o o f . It suffices to study the case E(N) < co. Assume that U(T) is finite for a bounded linear
operator T 9 Bqr(N). To each function r 9 Wq,(H*) we assign a functional T = T o by setting

Tx = T 0 x = (Tx, H*~b)q.
One can readily see that T o is a bounded linear functional on Lr, i.e., T o 9 L~, and moreover,

[[T0[ILr ~ IIH*r 'N,

sup{(Az, H*r - T 0 z : x 6 Qr,p} = sup{(Ax - Tx, H * r : z 6 Qr,p}
< U(T)IIH*r162
It follows from Lemma 4 that there exists a function ~0 6 W e ( B * ) satisfying properties ( l l ) , (12), (22),
and (23). Hence this function has the properties

lie - ~ll.' < U ( T ) I I H * r 1 6 5 IIB*~II. < N I I H * r '- (36)

The function ~o is uniquely determined by the function r and thus we have %o = P C , where P is a
(single-valued) mapping of the space ]4;r into the space W e ( B * ) . It follows from (36) that the
variable
J ( P ) = g ( Q , , ( H * ) ; P ) , , = sup{lit - Pr : r 9 Q,,(H*)}
satisfies the estimate
J ( P ) <_U(T). (37)
Let us verify that P is a linear mapping of the space 14~q,(H*) into the space W e ( B * ) . Assume that tx
and 'kz are two functions from W u and cx and c2 are two real constants. We introduce the function
r = cl,h + c2,k2 ; this function also belongs to the space W , , ( / / * ) . We set

~0, = P , / , , , ~oz = P r ~ = PC, ~ = ~o - (c,~o, + cz~o2) = P C - ( C l P t x + c z e c h ) .

The function ~ belongs to the space ~,p, and satisfies the equation B * ~ = 0. Hence, just as in the proof
of Lemma 4, we see that ~ = 0, which means that the operator P is linear and hence P 9 ~ ( N ) . Now
inequality (35) follows from inequality (37). The proof of the lemma is complete. []
In the following statement, under restrictions that are stronger than those in Lemma 6, we prove the
inequality that is opposite to (35).
Lernma 7. Suppose :hat 1 < q < co, 1 < p, r < co, the factorization condition (F), condition ( B , p),
condition (E), and condition (tt, q') are satisfied, and moreover, condition ( M , r ' ) is sat/st~ed it~p = co.
Then for any N > 0 the inequality
E(N) _<g(N) (38)
holds.
P r o o f . Assume that ~ ( N ) < oo and an operator P 6 P ( N ) is such that J ( P ) < oo. The fact
that P 6 J~ means that P is a linear mapping of the space M;q,(H*) into the space M e ( B * ) . This
mapping has the property
IIB*Pr _< NIIH*r , r 9 Wq,(H*).
Hence the function ~o = B ' P C is uniquely determined by the function H * r r 9 Wq,(H*), i.e., ~0 =
B ' P C is the value of a (single-valued) mapping 7" on the function H * r One can readily see that 7" is
a bounded linear operator from the space

= HWr = {/t*r : r 9 We(H*)},

equipped with the norm of the space s162= Lq,, into the space s and moreover, [[7"[[~--z., _< N . It
follows from assumption ( H , q') that the set "H is dense in the space L r Therefore, we can continue the

291
operator T from 7"/ to Lr so that it remains continuous (and hence the norm is preserved). We shall
denote this continuation by the same symbol T .
Let T = T* be the adjoint of the operator T . Then T is a bounded linear o p e r a t o r from the space s
into the space L;, = Lq (recall the assumption that 1 < q < co). We consider t h e operator T only on
the subspace Lr C s The norm of the operator T satisfies the estimate
IITII = IITIIL,-.L, --< IITII ',--L, = <--- N.
Next, in view of the definition of the operator T , for any functions r E I~Yr a n d z 6 Lr we have
(x, B*Pr = (z, T H * r = (Tx, H * r
and hence if r e W e ( H * ) and x e Wr,p, then
(Az - T x , H * r = (Az, H*r - (Tz, H* ~b)q ---- (Ax, H*r (x, B ' P C ) , - . - -

The pair of functions r and ~ = B ' P C satisfies the assumptions of Lemma 3. By this lemma, the
functions z E Wr,p satisfy the relation
(Az - T z , H*r = (Bx, r - qa)p.
Hence we have
I(Az - Tz, H*C)d < IIBxllpllr - 11r
By using assumption ( H , q'), we see that the estimate U(T) < J ( P ) holds, which implies inequality (38).
The proof of the lemma is complete. [:3
Combining the last two lemmas, we obtain the following statement.
T h e o r e m 2. Suppose that 1 < q < c0, 1 < p, r < co, the factorization condition (F), condi-
tion ( B , p ) , condition (E), and condition ( H , q') are sat/sfied, and moreover, condition ( M , r ' ) is satisfied
if p = co. Then problems (3) and (14) are related by E ( N ) = ~ ( N ) .
R e m a r k 2. In the proof of the last theorem, more exactly, in the proofs o f the last two lemmas
(at least in the case 1 < q < co), we have established a one-to-one correspondence B * ( P r = T H * r
between the sets of operators P E :P(N) and T E Brq(N) for which the corresponding deviations if(P)
and U(T) are finite, where T is the adjoint of the operator T; in this case, U(T) ~- i f ( P ) . Hence if one
of the operators P or T is an extremum operator, then the other is also an extrexnum operator (in the
corresponding problem).

w T h e r e l a t i o n s h i p b e t w e e n t h e p r o b l e m s in t h e c a s e q ---- co
If an element 0 E S' is such that 6 9 r E S or, which is the same, 0 9 ~b E S f o r any function ~ E S ,
then one can define the convolution ~ * 0 of the elements ~ E ~qt and 0 by the r e l a t i o n
(r 0, r = ( r ~), eES.
Recall that, by assumption, the elements a, b, h E S ~, which determine the o p e r a t o r s A, B , and H and
their adjoints A*, B*, and H*, possess the same property.
Assume that there exists an eigenfunction t9 of the operator H*, i.e., there exists an element 0 E S t
such that the convolution t9 9 h is the g-function. In this case for any function ~b E S we have
O*b.~b=tg.h.a.qb=tg*h,A*r = A*r = a * r
and hence the representation a = t9 9 b holds. For instance, if there exists an eigenfunction ~ of the
operator B*, then the function t9 = a * )r = A*x has the properties listed above.
In addition, assume that the function t9 is a classical function; more exactly, assume that t9 E ~ .
Obviously, then we have t9 E l/Yx(H*), and H't9 is an atomic measure c o n c e n t r a t e d at the point t = 0;
the variation of this measure is equal to 1 ; thus we obtain ~ E Q~(H*). Let us c o n s i d e r the variable (see
notation (13))
~ ( g ) = e(tg; r, p; N) = inf{llO - :~ e = ~(~, NQ~,(B*))f. (39)
In I,p, this is the best approximation of the function t9 by the set N Q , . , ( B * ) . In connection with
problem (2), a variable more general than (39) was studied by the author in [12]. I n particular, Theorem 2
in [12] contains the following statcment.

292
 L e m m a 8. If q = c o , 1 <_ p , r <_ co, the factorization condition (F) and condition ( B , p) are satisfied,
then
E ( N ) = e(N). (40)

The next statement supplements Theorems 1 and 2 in the case q = co. The last relation in (41) below
means that, in this situation, the function 0 E Qu is the maximum function in problem (14), i.e.,
this function provides the upper bound in (14). The first relation in (41) is a result due to Gabushin [25]
and follows from Lemma 5 considered above.

Theorem 3. If for q = co and 1 _< p, r < co the factorization condition (F), condition (B, p),
and condition (E) are satisfied and moreover, condition (M,r') is satisfied in the case p = co, then the
equalities
E ( N ) = R ( N ) = ~ ( N ) = ~ ( N ) = ~(N) (41)

hold.

P r o o f . Since 0 E Qq,(H*), by using definitions (14) and (15), inequalities (35), and relation (40), we
obtain the chain of inequalities

~(N) < .F(N) < ~ ( N ) <_ E ( N ) = e(N),

which implies all statements of Theorem 3. []

w C o n c l u d i n g remarks
The set of unbounded linear operators A and B of convolution type studied in the present paper con-
tains the important class of differential operators with constant coei~icients. So far problems (2), (3), (13),
and (15) have been studied most completely for the differentiation operators Ax = x (k) and Bx -- x ('0 ,
0 _< k < n, in the spaces of functions of one variable on the real axis (and on the half-axis); for functions of
several variables these problems are significantly less studied (see the references in [4, 5]). Undoubtedly, in
spaces of functions of several variables, all four problems are of great interest in the case where A and B
are powers (iterations),
A = A k, B = A n, O <_ k < n,

of the Laplace operator

m
A z = ~_, 02x
i=1 o ~ "

There are some cases in which the problems studied in this paper have exact solutions for some classes
of functions of several variables. For example, a multidimensional generalization can readily be obtained
for the one-dimensional results due to Subbotin and Taikov [16] in the case p = q = r = 2 and for the
one-dimensional results due to Taikov [26] in the case q = co and p = r = 2. In these papers the authors
studied problems (5), inequality (6) (in [26]), and the corresponding problems of approximating a class
of differentiable functions by another class of smoother functions. Some exact multidimensional results
related to problem (2) and obtained with the help of the invariance concept can be found in [9, 10, 12].
Recently there appeared a number of interesting results concerning the exact solution of problems close
to (2), (3) in classes of functions of several variables, where the multidimensional n a t u r e of these problems
manifests itself completely (see [5, 27, 28] and the bibliography therein). The results obtained in the
present paper permit one to obtain the solution of problems (13) and (15) in a number of new cases and,
as a consequence of the corresponding results, of problems (2), (3).

This research was supported by INTAS under grant No. 94-4070.

293
 References
1. So B. Stechkin, "The best approximation of linear operators," Mat. Zametki [Math. Notes], 1, No. 2, 137-148 (1967).
2. V. V. Arestov, "On some extremum problems for differentiable functions of one variable," Tl-udy Mat. Inst. Steldov
[Proc. Steklov Inst. Math.], 138, 3-28 (1975).
3. B. E. Klots, "Approximation of differentiable functions by more smooth functions," Mat. Z a m e t k l [Math. Notes], 21,
No. I, 21-32 (1977).
4. V. V. Arestov and V. N. Gabushin, "The best approximation of unbounded operators by bounded operators," Izv.
Vyssh. Uchebn. Zaved. Mat. [Russ/an Math. (Iz. VUZ)], No. 11, 42--68 (1995).
5. V.V. Arestov, "Approximation of unbounded operators by bounded operators and related e x t r e m u m problems," Uspekhl
Mat. Nauk [Russian Math. Surveys], 51, No. 6, 89-124 (1996).
6. G. E. Shilov, Mathematical Analysis. The Second Special Course [in Russian], Nauka, Moscow (1965).
7. E. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton Univ. Press, Princeton (1971).
8. R. Larsen, An Introduction to the Theory of Multlpliers, Springer, Berlin (1971).
9. V. V. Arestov, "Approximation of translation invariant operators,~ Trudy Mat. Inst. Steklov [Proc. Steklov Inst. Math.],
138, 43-70 (1975).
10. V. V. Arestov, "Approximation of convolution type operators by bounded linear operators, ~ Trudy Mat. Inst. Steklov
[Proc. Steklov Inst. Math.], 145, 3-19 (1980).
11. V. V. Arestov, "Approximation of invariant operators," Mat. Zametki [Math. Notes], 34, No. 1, 9-29 (1983).
12. V~ V. Arestov, "The best approximation of unbounded translation invariant operators by b o u n d e d linear operators,"
Trudy Mat. Inst. Steklov [Proc. Steklov Inst. Math.], 198, 3-20 (1992).
13. V. V. Arestov, "The best reconstruction of operators and related problems," Trudy Mat. Inst. Steklov [Proc. Steklov
Inst. Math.], 189, 3-20 (1989).
14. L. V. Talkov, "T)n the best approximation in the mean for some classes of analytic functions," Mat. Zametki [Math.
Notes], 1, No. 2, 155-162 (1967).
15. Yu. N. Subbotin, "The best approximation of a function class by another function class," MaP. Zametkl [Math. Notes],
2, No. 5, 495-504 (1967).
16. Yu. N. Subbotin and L. V. Talkov, "The best approximation of the differentiation operator in the space La ," Mat.
Zametki [Math. Notes], 3, No. 2, 257-264 (1968).
17. V. V. Arestov and V. N. Gabushin, "On the approximation of classes of differentiable functions, ~ Mat. Zametkl [Math.
Notes], 9, No. 2, 105-112 (1971).
18. Yu. N. Subbotin, "Relation between the spllne-fitting and the problem of approximation of o n e class by another class,"
Mat. Zametki [Math. Notes], 9, No. 5, 501-510 (1971).
19. S. M. Nikol%kii, Approximation of Functions of Several Variables and Embedding T h e o r e m s [in Russian], Nauka,
Moscow (1977).
20. O. V. Besov, V. P. IVin~ and S. M. NikoVskii, Integral Representations of Functions and Embedding Theorems [in
Russian], Nauka, Moscow (1975).
21. E. Stein, Singular Integrals and Differentlability Properties of Functions, Princeton Univ. Press, Princeton (1970).
22. V. N. Gabushin, "Inequalities for the norms of a function and of its derivatives in the Lp-metrlc, ~ Mat. Zametki [Math.
Notes], 1, No. 3, 291-298 (1967).
23. O. A. Timoshin, "On the best approximation of partial differential operators," Mat. Zametkl [Math. Notes], 46, No. 1,
78-87 (1989).
24. V. V. Arestov, "Approximation of linear operators and related extremum problems," Ttudy ~dat. Inst. Steklov [Proc.
Steklov Inst. Math.], 138, 29-42 (1975).
25. V. N. Gabushin, "The best approximation of functionals on some sets," Mat. Zametkl [Mat/l. Notes], 8, No. 5, 551-
562 (1970).
26. L. V. Taikov, "Kolmogorov-type inequalities and the best formulas of numerical differentiation," Mat. Zametkl [Math.
Notes], 4, No. 2, 233-238 (1968).
27. G. G. Magaril-IVyaev and V. M. Tikhomirov, ~On the inequalities of Kolmogorov type for derivatives," Mat. Sb. [Russian
Acad. Sci. Sb. Math.], 188, No. 12, 73-106 (1997).
28. V. F. Babenko, V. A. Kofanov, and S. A. Pichugov, "Exact inequalities of Kolmogorov type f o r multivariable functions
and their applications," East J. Approx., 3, No. 2, 155-186 (1997).

A. M. GOPJKII UP~ALSTATE UNIVEP~SITY,EKATERINEURG

E-mail address: [email protected]

Translated by M. A. Shishkova

294