Systems of random variables equivalent in distribution to the Rademacher system and -

closed representability of Banach couples

This article has been downloaded from IOPscience. Please scroll down to see the full text article.
2000 Sb. Math. 191 779
(http://iopscience.iop.org/1064-5616/191/6/A01)
Download details:
IP Address: 150.254.77.50
The article was downloaded on 19/06/2009 at 13:18
Please note that terms and conditions apply.
The Table of Contents and more related content is available
HOME | SEARCH | PACS & MSC | JOURNALS | ABOUT | CONTACT US
Sbornik: Mathematics 191:6 779807 c 2000 RAS(DoM) and LMS
Matematicheski Sbornik 191:6 330 DOI: 10.1070/SM2000v191n06ABEH000481
Systems of random variables equivalent
in distribution to the Rademacher system
and KKK-closed representability of Banach couples
S. V. Astashkin
Abstract. Necessary and sucient conditions ensuring that one can select from a
system {fn}

n=1
of random variables on a probability space (, , P) a subsystem
{
i
}

i=1
equivalent in distribution to the Rademacher system on [0, 1] are found. In
particular, this is always possible if {fn}

n=1
is a uniformly bounded orthonormal
sequence. The main role in the proof is played by the connection (discovered in
this paper) between the equivalence in distribution of random variables and the
behaviour of the Lp-norms of the corresponding polynomials.
An application of the results obtained to the study of the K-closed representabil-
ity of Banach couples is presented.
Bibliography: 26 titles.
Introduction
We say that systems {f
n
}

n=1
and {g
n
}

n=1
of random variables dened on prob-
ability spaces (, , P) and (

, P

), respectively, are equivalent in distribution

(we write {f
n
}
P
{g
n
}) if there exists C > 0 such that for arbitrary m N, a
n
R
(n = 1, 2, . . . , m), and z > 0 we have
C
1
P

n=1
a
n
f
n
()

> Cz

n=1
a
n
g
n
(

> z

C P

n=1
a
n
f
n
()

> C
1
z

.
If only the left-hand side of this relation holds, then we say that the system
{f
n
}

n=1
is majorized in distribution by the system {g
n
}

n=1
. In a similar way one
can introduce these concepts in the case of nite collections {f
n
}
N
n=1
and {g
n
}
N
n=1
.
As shown in [1], from a uniformly bounded and orthonormal (in L
2
) sequence of
random variables {f
n
}

n=1
one can always select a subsequence {f
nk
}

k=1
majorized
in distribution by the Rademacher system {r
n
}

n=1
, where
r
n
(x) = signsin(2
n1
x) (x [0, 1]).
AMS 1991 Mathematics Subject Classication. Primary 28A20, 60E99.
780 S. V. Astashkin
The main aim of the present paper is to prove that, under the same assumptions,
it is possible to select a subsequence that is not merely majorized, but also equiv-
alent to the system r
n

n=1
. Moreover, we nd necessary and sucient conditions
ensuring that a system of random variables contains a subsystem equivalent in dis-
tribution to the Rademacher system. In the case of nite systems we are interested
in the density of their subsystems with a similar property.
Problems related to the selection of almost independent subsystems (that is,
subsystems with a certain property characteristic for systems of independent func-
tions) are discussed in detail in Gaposhkins well-known paper Lacunary series
and independent functions [2], which we shall repeatedly refer to in what follows.
We recall the most important results relating to the integration of the sum of a
series in a xed system of functions and to its absolute convergence.
In view of a classical inequality established by Khintchine [3], the sum of a
series

n=1
a
n
r
n
(x) (x [0, 1]) belongs to all L
p
-spaces (p > 2), provided that
the sequence of coecients a = (a
n
)

n=1
belongs to l
2
. For lacunary trigonometric
series a similar result was obtained by Zygmund [4]. In this connection, but slightly
later, Banach and Sidon introduced the concept of a lacunary system of order p,
or briey, S
p
-system. A sequence of random variables f
n

n=1
, f
n
L
p
(p > 2), is
called an S
p
-system if
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_
p
K
p
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_
2
, (0.1)
where the constant K
p
> 0 is independent of m N and a
n
R (n = 1, 2, . . . , m).
If f
n
is an S
p
-system for each p > 2, then it is called an S

-system.
Banach proved that one can select an S
p
-subsystem from each orthonormal
system f
n
of measurable functions on [0, 1] such that limsup
n
|f
n
|
p
<
([5] or [6], 7.2). This, in particular, means that one can select an S

-subsystem
from each uniformly bounded system of functions. Here is a more precise result
due to Stechkin [2], Theorem 1.3.1.
Theorem A. Let f
n

n=1
be a system of measurable functions on [0, 1] and assume
that p > 2. Then one can select from f
n

n=1
an S
p
-subsystem if and only if there
exists a subsequence f
nk
such that
(1) |f
nk
|
p
D (k = 1, 2, . . . );
(2) f
nk
0 weakly in L
2
.
We now proceed to a brief discussion of another property of lacunary series, their
absolute convergence. Well known for trigonometric series is Sidons theorem on
the absolute convergence of a lacunary (in the sense of Hadamard) Fourier series of
a bounded function. Zygmund established a local version of this result [7]. Similar
results were later obtained for the Rademacher system and certain other special
lacunary systems of functions [8]. The following concepts were introduced as a
natural generalization of these results. A sequence of random variables f
n

n=1
is
called a Sidon system if
m

n=1
[a
n
[ C
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_

, (0.2)
where the constant C > 0 is independent of m N and a
n
R (n = 1, 2, . . . , m).
Systems of random variables 781
A system f
n

n=1
of measurable functions on [0, 1] is called a SidonZygmund
system if there exists a subset E of [0, 1] of positive Lebesgue measure [E[ such that
for each interval I, [I E[ > 0, the condition

n=1
a
n
f
n
(x) L

(I) implies that

(a
n
)

n=1
l
1
. If one can set E = [0, 1] in this denition, then we say that f
n

n=1
is a SidonZygmund system in the narrow sense.
Broad sucient conditions ensuring the existence of Sidon and SidonZygmund
subsystems have been obtained by Gaposhkin [2], Theorems 1.4.1 and 1.4.2.
Theorem B. Let f
n

n=1
be a system of measurable functions on [0, 1] such that
(1) |f
n
|
2
= 1 (n = 1, 2, . . . );
(2) [f
n
(x)[ D (n = 1, 2, . . .; x [0, 1]);
(3) there exists a subsequence f
nk
f
n
such that f
nk
0 weakly in L
2
.
Then one can select from f
n

n=1
a Sidon subsequence.
Theorem B

. (a) Under the assumptions of Theorem B one can select from

f
n

n=1
a SidonZygmund subsystem.
(b) if the subsequence in assumption (3) of Theorem B satises additionally the
condition
liminf
k
_
F
f
2
nk
(x) dx > 0 for each F [0, 1], [F[ > 0, (0.3)
then one can select from f
n

n=1
a SidonZygmund subsystem in the narrow sense.
It follows from the above denitions and the properties of the Rademacher system
that if f
n

P
r
n
, then
(1) f
n
is an S

-system and the constant K

p
in inequality (0.1) has the same
growth order in p, as p , as in the case of the Rademacher system, that
is, K
p

p;
(2) f
n
is a Sidon system.
Thus, the results established below in 3 ensure nice properties of the selected
subsystem in both respects, improving and complementing the above-stated results.
Alongside the problems of the selection of innite subsequences with proper-
ties characteristic for systems of independent functions, problems of another kind,
treating nite collections of random variables, are also of interest. Let f
n

N
n=1
be
a collection of random variables. We are looking for a possibly large s = s(N) such
that there exists a collection f
ni

s
i=1
f
n

N
n=1
satisfying a certain condition of
lacunarity with constant independent of N. We state here just one result on the
density of nite Sidon subsystems, which was proved by Kashin (see [9] or [10], 8.4,
Theorem 9).
Theorem C. Let f
n

N
n=1
be an orthonormal collection of functions on [0, 1]
such that [f
n
(x)[ D (n = 1, . . . , N; x [0, 1]). Then there exists a collection
f
ni

s
i=1
f
n

N
n=1
(1 n
1
< < n
s
N) such that s max[
1
6
log
2
N]; 1 and
for all a
i
R (i = 1, . . . , s) the following inequality holds:
D
1
_
_
_
_
s

i=1
a
i
f
ni
_
_
_
_

i=1
[a
i
[ 4D
_
_
_
_
s

i=1
a
i
f
ni
_
_
_
_

.
782 S. V. Astashkin
In the present paper, under the same assumptions, we prove a result on the selec-
tion of a subsystem f
ni

s
i=1
of the same density that satises a stronger condition:
it is equivalent to the system r
i

s
i=1
with constant dependent only on D.
This paper consists of ve sections. In 1 we collect concepts of operator inter-
polation theory used in what follows (see [11] for greater detail). For completeness
we also present there the proof of a formula from [12] for the K-functional cor-
responding to the pair (l
1
, l
2
). In 2 we prove a criterion for the equivalence in
distribution of an arbitrary sequence of random variables and the Rademacher sys-
tem and derive consequences for some particular cases. The central results of this
paper, on the selection of subsystems equivalent in distribution to the Rademacher
system, are the subject of 3. The next section is devoted to similar problems for
nite collections of functions. Finally, in 5 we apply these results to the study of
the K-closed representability of Banach couples.
In conclusion let us introduce our notation. An expression of the form F
1
F
2
will mean that there exists C > 0 such that C
1
F
1
F
2
CF
1
; moreover, the
constant C will, as a rule, be independent of all or some of the variables of the
functions F
1
and F
2
. If 1 p and f = f() is a random variable on a
probability space (, , P), then
|f|
p
= (E[f[
p
)
1/p
=
__

[f()[
p
dP()
_
1/p
;
while if a = (a
n
)

n=1
is a number sequence, then
|a|
p
=
_

n=1
[a
n
[
p
_
1/p
(the formulae must be modied in the natural way for p = ). As usual, the
space L
p
consists of all random variables f such that |f|
p
< , and l
p
consists of
all sequences a = (a
n
)

n=1
such that |a|
p
< .
We denote by [E[ the Lebesgue measure of the subset E of [0, 1].
1. Petre KKK-functional and real interpolation method
Let (X
0
, X
1
) be a Banach couple, let x X
0
+X
1
, and assume that t > 0. We
consider the Petre K-functional
K(t, x; X
0
, X
1
) = inf
_
|x
0
|
X0
+ t|x
1
|
X1
: x = x
0
+ x
1
, x
0
X
0
, x
1
X
1
_
,
which plays an important role in operator interpolation theory.
It is easy to show that for xed x X
0
+ X
1
the K-functional is an increasing
continuous concave function of t (see [11], 3.1).
In what follows we require only the functional K
1,2
(t, a) = K(t, a; l
1
, l
2
) con-
structed for the Banach couple (l
1
, l
2
), two approximations of which are discussed
in the lemmas below. The rst lemma, due to Montgomery-Smith (see [12]), will
be essential in the proof of Theorem 4. For this reason we present its proof here
for completeness.
Systems of random variables 783
For arbitrary a = (a
n
)

n=1
and t N we dene the norm
|a|
Q(t)
= sup
_
t

j=1
_

nAj
a
2
n
_
1/2
_
, (1.1)
where the supremum is taken over all partitionings A
j

t
j=1
of the set of positive
integers N.
Lemma 1. If a = (a
n
)

n=1
l
2
and t
2
N, then
|a|
Q(t
2
)
K
1,2
(t, a)

2|a|
Q(t
2
)
. (1.2)
Proof. It follows from the denition of Q(t) that
|a|
Q(t
2
)
|a|
1
and |a|
Q(t
2
)
t|a|
2
.
Hence
K
1,2
(t, a) = inf|b|
1
+ t|c|
2
: b + c = a, b l
1
, c l
2

inf|b|
Q(t
2
)
+|c|
Q(t
2
)
: b +c = a, b l
1
, c l
2
|a|
Q(t
2
)
which proves the left-hand side of (1.2).
To prove the right-hand side of (1.2) we note rst of all that
K
1,2
(t, a) = sup
_

n=1
a
n
b
n
: b = (b
n
)

n=1
l
2
, J
,2
(t
1
, b) 1
_
,
where J
,2
(s, b) = max|b|

; s|b|
2
(see [11], 2.7).
Hence, for each > 0 there exists a sequence b l
2
such that
(1 )K
1,2
(t, a)

n=1
a
n
b
n
and J
,2
(t
1
, b) = 1.
We select n
0
, n
1
, n
2
, . . . , n
t
2 0, 1, . . ., by induction as follows: if we have
already selected n
0
, n
1
,..., n
m
, where 0 = n
0
< n
1
< < n
m
, then
n
m+1
= 1 + sup
_
k :
k

n=nm+1
b
2
n
1
_
.
Since |b|

1, it follows that

nm+1
n=nm+1
b
2
n
2. In view of the inequality |b|
2
t,
we have n
t
2 = .
Hence
(1 )K
1,2
(t, a)

n=1
a
n
b
n

t
2

m=1
_
nm

n=nm1+1
b
2
n
_
1/2
_
nm

n=nm1+1
a
2
n
_
1/2

2|a|
Q(t
2
)
.
The last relation holds for all > 0, which proves the lemma.
Another relation, Holmstedts formula is well known (see [13] or [11], 5.7).
784 S. V. Astashkin
Lemma 2. There exists a constant >0 independent of a =(a
n
)

n=1
l
2
and t > 0
such that

1
_[t
2
]

i=1
a

i
+t
_

i=[t
2
]+1
(a

i
)
2
_
1/2
_
K
1,2
(t, a)
_[t
2
]

i=1
a

i
+ t
_

i=[t
2
]+1
(a

i
)
2
_
1/2
_
,
(1.3)
where (a

i
)

i=1
is a decreasing rearrangement of the sequence ([a
n
[)

n=1
and [z] is the
integer part of the quantity z.
Corollary 1. The function (t, a) = K
1,2
(

t, a) is continuous and increases for

t 0. For each sequence a = (a
n
)

n=1
l
2
,
(1) lim
t0+
(t, a) = 0;
(2)
1
|a|
1
lim
t+
(t, a) |a|
1
.
To conclude the section we recall the denition of the real interpolation method.
For an arbitrary Banach couple (X
0
, X
1
), 0 < < 1, 1 q , we consider the
norm
|x|
,q
=
_

n=
_
2
n
K(2
n
, x; X
0
, X
1
)

q
_
1/q
, x X
0
+ X
1
(and its natural modication for q = ). The spaces
(X
0
, X
1
)
,q
= x X
0
+ X
1
: |x|
,q
<
are interpolation spaces for the couple (X
0
, X
1
) (that is, each linear operator
bounded in X
0
and X
1
is bounded also in (X
0
, X
1
)
,q
); they are called the spaces
of the real interpolation method.
2. Systems of random variables equivalent
by distribution to the Rademacher system
Theorem 1. Let f
n

n=1
be a sequence of random variables on a probability space
(, , P). Then the following conditions are equivalent:
(1) f
n

P
r
n
(r
n

n=1
is the Rademacher system on [0, 1]);
(2) the equivalence
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_
t

_
_
_
_
m

n=1
a
n
r
n
_
_
_
_
t
holds with constant independent of t [1, ), m N, and a
n
R (n =
1, . . . , m);
(3) the equivalence
_
_
_
_

n=1
a
n
f
n
_
_
_
_
t
K
1,2
(

t, a) (2.1)
holds with constant independent of t [1, ) and a = (a
n
)

n=1
l
2
.
Systems of random variables 785
Proof. As shown in [14], with constant independent of t [1, ) and a =(a
n
)

n=1
l
2
we have
_
_
_
_

n=1
a
n
r
n
_
_
_
_
t
K
1,2
(

t, a) (2.2)
(this relation is already implicit in [12]). Hence the equivalence (2) (3) is obvious.
The implication (1) (2) is a consequence of the denition of systems equivalent
in distribution, therefore it remains to establish the implication (3) (1).
We x a = (a
n
)
m
n=1
, m N (where not all the a
n
are equal to zero) and show
that the distribution of the absolute value of the random variable
f() =
m

n=1
a
n
f
n
() ( )
is determined by the behaviour of its moments |f|
t
and therefore, by assumption,
by the function (t, a) = K
1,2
(

t, a) (t 1).
Let 1 be the constant involved in the equivalence (2.1). In view of the Paley
Zygmund inequality [15], 1.6 and the concavity of the K-functional for t 1, we
obtain
P[f()[ (2)
1
(t, a) P[f()[
t
2
t
|f|
t
t

(1 2
t
)
2
|f|
2t
t
|f|
2t
2t

4t
(1 2
t
)
2
_
(t, a)
(2t, a)
_
2t
(2)
4t
.
The function (t, a) increases in t, therefore
P[f()[ (2)
1
(t, a) (2)
4t4
(2.3)
also for all t > 0.
Conversely, by Chebyshevs inequality and assumption (2.1) we obtain
P[f()[ (t, a)
t
_
|f|
t
(t, a)
_
t

_
t
for arbitrary t 1 and > 0. In particular, if = e, then
P[f()[ e(t, a) e
t
(t 1)
or
P[f()[ e(t, a) e
1t
(t > 0). (2.4)
Following [14] we now dene the functionals
F(s) = supt > 0 : (t, a) s, G(s) = inft > 0 : (t, a) s.
786 S. V. Astashkin
If 0 < z < (4)
1
|a|
1
, then, by Corollary 1, there exists t > 0 such that
(t, a) 4z. It then follows from (2.3) that
P[f()[ > z P[f()[ (2)
1
(t, a) (2)
4t4
.
Hence, by the denition of the functional G,
P[f()[ > z (2)
4G(4z)4
,
so that
P[f()[ > z C
1
1
e
C2G(4z)
(z < (4)
1
|a|
1
), (2.5)
where C
1
= (2)
4
and C
2
= 4 ln(2).
Setting C
3
= 4

2C
2
we claim that
C
2
G(4z) F(C
3
z) (z > 0).
For, in view of the concavity of the K-functional and the denition of G,
(C
2
G(4z), a)

2C
2
(2
1
G(4z), a) 4

2 C
2
z = C
3
z,
and it remains to use the denition of the functional F. Thus, (2.5) can be written
as follows:
P[f()[ > z C
1
1
e
F(C3z)
(z < (4)
1
|a|
1
). (2.6)
Note that by (2.1) we obtain
|f
n
|

= lim
t+
|f
n
|
t
,
so that |f|

|a|
1
. Hence
P[f()[ > z = 0 (z |a|
1
). (2.7)
Further, by Corollary 1 we can nd t

> 0 such that 0 < (t

, a) (2e)
1
z. It
now follows from (2.4) that
P[f()[ > z P[f()[ e(t

, a) e
1t

,
so that, by the denition of F,
P[f()[ > z e
1F(C
1
4
z)
(z > 0), (2.8)
where C
4
= 2e.
In view of (2.2), relations similar to (2.6)(2.8) hold also for the Rademacher
system. Namely, if

1 is the constant involved in the equivalence (2.2), then

the function r(x) =

m
n=1
a
n
r
n
(x) satises the relations
[[r(x)[ > z[ (C

1
)
1
e
F(C

3
z)
(z < (4

)
1
|a|
1
), (2.6

)
[[r(x)[ > z[ = 0 (z |a|
1
), (2.7

)
[[r(x)[ > z[ e
1F((C

4
)
1
z)
(z > 0). (2.8

)
Systems of random variables 787
Let
A = max(C
1
e; C
3
C

4
; C

1
e; C

3
C
4
; 4

). (2.9)
Note that, since the constants and

are universal, the dependence of A on the

system f
n

n=1
reduces to its dependence on , the constant involved in (2.1).
If z < (4)
1
|a|
1
, then relations (2.8

) and (2.6) and our choice of A show

that
[[r(x)[ > Az[ e
1F((C

4
)
1
Az)
e
1F(C3z)
C
1
e P[f()[ > z AP[f()[ > z. (2.10)
On the other hand if z (4)
1
|a|
1
, then it follows from (2.9) that Az |a|
1
,
so that [[r(x)[ > Az[ = 0 in view of (2.7

). Hence inequality (2.10) also holds in

that case.
Conversely, if z/A < (4

)
1
|a|
1
, then it follows from (2.8), (2.6

), and our
choice of A that
P[f()[ > z e
1F(C
1
4
z)
e
1F(C

3
(z/A))
C

1
e

_
[r(x)[ >
z
A
_

_
[r(x)[ >
z
A
_

. (2.11)
If z/A (4

)
1
|a|
1
, then by (2.9) we obtain the inequality z |a|
1
. Hence
P[f()[ > z = 0 in view of (2.7), and (2.11) holds again.
Inequalities (2.10) and (2.11) mean that f
n

P
r
n
, which completes the proof
of Theorem 1.
Remark 1. Of course, a similar result holds also for nite collections of random vari-
ables. Moreover, it is clear from the proof that the constants involved in the equiva-
lences in conditions (1)(3) of Theorem 1 depend only on one another. For instance,
if (2.1) holds for collections f
n

N
n=1
with constant independent of N = 1, 2, . . . ,
then f
n

N
n=1
P
r
n

N
n=1
also with constant independent of N = 1, 2, . . . .
We now apply the above theorem to two more special cases.
Recall that a system of random variables f
n

n=1
is said to be multiplicative if
for arbitrary, pairwise distinct n
1
, n
2
, . . . , n
k
(k N) we have
E(f
n1
f
n2
f
nk
) = 0.
If, in addition,
E(f
n1
f
n2
f
nk
f
2
n
) = 0
for arbitrary, pairwise distinct n
1
, n
2
, . . . , n
k
and n ,= n
s
(s = 1, 2, . . ., k), then we
say that the system f
n
is strongly multiplicative. In the case of nite collections
the denitions are completely similar.
We now discuss the most important examples of such systems [16].
Example 1. Let f
n
(x) = sin(2k
n
x) (x [0, 1]). If k
n+1
/k
n
2, then this is
a multiplicative system and if k
n+1
/k
n
3, then it is a strongly multiplicative
system.
788 S. V. Astashkin
Example 2. A sequence f
n

n=1
of independent random variables such that
f
n
L
2
and E(f
n
) = 0 (n = 1, 2, . . . ) is a strongly multiplicative system.
Kwapien and Jakubowski have established the following result [16].
Theorem D. Let
n

N
n=1
be a multiplicative and
n

N
n=1
a strongly multiplica-
tive system of random variables. If for each n = 1, . . . , N,
|
n
|

|
n
|

E(
2
n
), (2.12)
then there exist a probability space (

, P

), a -subalgebra
0
of the -algebra

,
and a random vector (

1
,

2
, . . . ,

N
) on (

, P

) such that this vector itself is

equidistributed with (
1
,
2
, . . . ,
N
), while (
1
,
2
, . . . ,
N
) is equidistributed with
the conditional expectations E
_
(

1
,

2
, . . . ,

N
) [
0
_
.
In particular, for an arbitrary convex function H: R
n
R,
E
_
H(
1
, . . . ,
N
)
_
E
_
H(
1
, . . . ,
N
)
_
.
From Theorems D and 1 we obtain the following result.
Theorem 2. Let f
n

n=1
be a strongly multiplicative system of random variables
such that [f
n
()[ D (n = 1, 2, . . . ; ) and d = inf
n=1,2,...
E(f
2
n
) > 0. Then
f
n

P
r
n
, and the constant involved in the equivalence depends only on D and d.
Proof. Assumption (2.12) of Theorem D holds for the systems
n
= f
n
/D and

n
= r
n
. Hence for all t 1, m N, and a
n
R (n = 1, 2, . . . , m) we have
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_
t
D
_
_
_
_
m

n=1
a
n
r
n
_
_
_
_
t
.
Conversely, if
n
= (d/D)r
n
,
n
= f
n
, then (2.12) holds again, so that
_
_
_
_
m

n=1
a
n
r
n
_
_
_
_
t

D
d
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_
t
.
To complete the proof it remains to use Theorem 1.
Corollary 2. Each sequence of independent random variables f
n

n=1
such that
f
n
L
2
, E(f
n
) =0, [f
n
()[ D (n=1, 2, . . . ; ), and d = inf
n=1,2,...
E(f
2
n
) >0
is equivalent in distribution to the Rademacher system.
We consider now another situation. Let G be a compact Abelian group, the
group of its characters, and the Haar measure on G. In accordance with the
above denition a subset F of is called a Sidon set if

f()[ C |f|

for some C = C(F) and each f C(G) such that

f() =
_
G
f d = 0 ( / F).
Pisier [17] has proved the following result (we present its formulation only in the
scalar case).
Systems of random variables 789
Theorem E. If F =
n
is a Sidon set, then the equivalence
_
_
_
_
m

n=1
a
n

n
_
_
_
_
t

_
_
_
_
m

n=1
a
n
r
n
_
_
_
_
t
(t 1)
holds with constant depending only on the Sidon constant C(F).
Theorems E and 1 immediately give us the following result, which was also
proved (by another method, but also on the basis of Theorem E) in [18].
Theorem 3. Each innite Sidon system F =
n

n=1
of characters of a com-
pact Abelian group is equivalent in distribution to the Rademacher system. The
corresponding equivalence constant depends only on the Sidon constant C(F).
Corollary 3. Sequences
f
n
(x) = sin(2k
n
x) and g
n
(x) = cos(2k
n
x) (x [0, 1])
in which k
n+1
/k
n
> 1 are equivalent in distribution to the Rademacher system.
Remark 2. We point out the fact that it was the similarity between the behaviour
of Rademacher series and lacunary trigonometric series (see the introduction) that
was the starting point of the study of deeper connections between these series
and led eventually to the investigation of systems equivalent in distribution to the
model Rademacher system. Among papers considering these subjects, besides
[16][18], we also mention a note [19] by Rodin and Belov, where the authors used
Theorem D to carry over the results of [20] on the behaviour of Rademacher series
in symmetric function spaces to lacunary trigonometric series. As mentioned in that
note, in 1988 one of its authors reported at the North-Caucasian workshop that in
many symmetric spaces the norms of polynomials in the Rademacher system and
of lacunary trigonometric polynomials are the same up to equivalence. Corollary 3
shows that this actually holds in all symmetric spaces, not just in many, because
the norm of a function in such a space is dened in terms of the distribution of its
absolute value (see [21], 2.4). This allows us, in particular, to improve some results
of [19]. For instance, in Theorem 2 from that paper, which establishes necessary
and sucient conditions for the equivalence of the norms of lacunary trigonometric
polynomials in a symmetric space to the norms of the sequences of their coecients
in the spaces l
1
and l
2
, the assumption that the symmetric space is an interpolation
space for the couple (L
1
, L

) is superuous.
3. Selection of subsystems equivalent in
distribution to the Rademacher system
Theorem 4. Let f
n

n=1
be a system of random variables on a probability space
(, , P) containing a subsequence f
nk

k=1
such that
(1) [f
nk
()[ D (k = 1, 2, . . . ; );
(2) f
nk
0 weakly in L
2
;
(3) d = inf
k=1,2,...
|f
nk
|
2
> 0.
790 S. V. Astashkin
Then there exists a subsystem
i

i=1
f
n

n=1
such that
_
_
_
_

i=1
a
i

i
_
_
_
_
t
K
1,2
(

t, a) (a = (a
i
)

i=1
l
2
, t 1) (3.1)
with constant dependent only on D and d.
Proof. As shown by Gaposhkin (see [2], Theorem 1.3.2, a full proof can be found
in [1]), if conditions (1) and (2) are satised, then one can select a subsequence
g
i
f
nk
such that for some C
1
= C
1
(D) and all a = (a
i
)

i=1
l
2
and t 1 we
have the inequality
_
_
_
_

i=1
a
i
g
i
_
_
_
_
t
C
1

t |a|
2
. (3.2)
Hence, for each expansion (a
i
) = (b
i
) + (c
i
) with (b
i
) l
1
and (c
i
) l
2
we have
_
_
_
_

i=1
a
i
g
i
_
_
_
_
t

_
_
_
_

i=1
b
i
g
i
_
_
_
_

+
_
_
_
_

i=1
c
i
g
i
_
_
_
_
t
D

i=1
[b
i
[ + C
1

t
_

i=1
c
2
i
_
1/2
max(C
1
; D)
_
|b|
1
+

t |c|
2
_
.
Consequently, by the denition of the K-functional we obtain
_
_
_
_

i=1
a
i
g
i
_
_
_
_
t
max(C
1
; D) K
1,2
(

t, a) (a = (a
i
)

i=1
l
2
, t 1). (3.3)
In the proof of the reverse inequality we use the equivalence of the K-functional
K
1,2
(t, a) and the norm Q(t
2
) (see (1.1) and (1.2)), and also the upper estimate
of the L
q
-norms (q > 1) of modied Riesz products (for applications of classical
products in similar cases, see [2], 1.4 or [10], 8.4).
First, assumptions (1) and (3) of the theorem show that 0 < d |g
i
|
2
D
(i = 1, 2, . . . ), therefore, in view of the positive homogeneity of both sides of (3.1)
we can assume that |g
i
|
2
= 1 (i = 1, 2, . . . ).
Let (
i
)

i=1
be a number sequence such that

i
0, 0 <
i
<
1
16
min(1; D),

k=i+1

k
<
i
(i = 1, 2, . . . ). (3.4)
By the assumptions of the theorem the sequence g
2
i
is weakly compact in L
2
.
Hence there exists h
k
g
i
such that h
2
k
h weakly in this space, where
0 h() D
2
and E(h) = 1. Together with g
i
, the sequence of h
k
also converges
to 0 weakly in L
2
. Hence we can nd an index k
1
such that the function
1
= h
k1
satises the relation
[E(
1
)[ +[E(h
1
)[ +[E(
2
1
h)[

1
2D
.
Systems of random variables 791
Assume that we have already selected the indices k
1
< k
2
< < k
i1
and the
functions
1
= h
k1
,
2
= h
k2
, . . . ,
i1
= h
ki1
(i 2). We now nd k
i
> k
i1
such that for
i
= h
ki
we have the relation

_
[E(
j1

js

i
)[ +[E(h
j1

js

i
)[ +[E[
j1

js
(
2
i
h)][
+
s

l=1
[E(
j1

jl1

2
jl

jl+1

js

i
)[
_
2
i
D
1

i
, (3.5)
where
j0
=
js+1
= 1 and the summation proceeds over all collections of indices
such that 1 j
1
< j
2
< < j
l
< < j
s
i 1 (s = 1, 2, . . . , i 1).
We claim that the sequence
i

i=1
satises (3.1) for arbitrary t 1.
Let t N rst, and let A
j

t
j=1
be an arbitrary partitioning of the positive
integers into t sets. For N N and j = 1, 2, . . . , t let A
N
j
= i = 1, . . . , N : i A
j

(some of these sets may be empty). We consider now the Riesz products formed
by blocks corresponding to the sets A
N
j
:
R
N
() =
N

i=1
(1 +b
i

i
()) =
t

j=1

iA
N
j
(1 + b
i

i
()),
where the b
i
are real coecients such that

iAj
b
2
i
D
2
. (3.6)
Reasoning along the lines of Sidons classical method, for an arbitrary function
() =

i=1
a
i

i
(), a = (a
i
)

i=1
l
2
,
we nd a lower estimate of the integral
I
N
=
_

R
N
()() dP() =

i=1
a
i
E(R
N

i
) =

i=1
a
i

i,N
. (3.7)
Here

i,N
= E(R
N

i
) =
_

_
1 +
N

k=1
b
k

k
() +

1k1<k2N
b
k1
b
k2

k1
()
k2
()
+ +b
1
b
2
b
N

1
()
N
()
_

i
() dP() = b
i

i,N
+
i,N
, (3.8)
where
i,N
= 1 (i N),
i,N
= 0 (i > N), and

i,N
= E(
i
) +
N

k=1,k=i
b
k
E(
k

i
) +

1k1<k2N
b
k1
b
k2
E(
k1

k2

i
)
+ +

1k1<<ksN
b
k1
b
ks
E(
k1

ks

i
) + + b
1
b
N
E(
1

N

i
).
792 S. V. Astashkin
Then

i,N
= S
i
1
+S
i
2
+S
i
3
, (3.9)
where the sums S
i
k
(k = 1, 2, 3) are dened as follows. Setting
E(
k1

ks

2
i
) = E(
k1

ks
h) +E[
k1

ks
(
2
i
h)],
we collect in S
i
1
all terms containing integrals of the form E(
k1

ks

i
) or
E[
k1

ks
(
2
i
h)], where k
1
< k
2
< < k
s
< i. The terms containing inte-
grals E(
k1

kl1

kl

ks
), where k
1
< k
2
< < k
l1
i < k
l
< < k
s
,
1 l s, we collect in S
i
2
, and in S
i
3
we include the terms containing integrals
E(
k1

ks
h).
In view of (3.5), (3.4), and (3.6) we obtain the estimates
[S
i
1
[

i
D
, [S
i
2
[
1
D

k=i+1

k
<

i
D
. (3.10)
Since each term in S
i
3
contains a factor b
i
, it follows by (3.4) that
[S
i
3
[
[b
i
[
D

i=1

i

[b
i
[
8
(3.11)
(we have not used the coecient 2
i
in (3.5) so far). Moreover, for i > N there are
no sums S
i
2
and S
i
3
in (3.9), therefore

i=N+1
[
i,N
[
1
D

i=N+1

i

1
16D
. (3.12)
Assume now that 1 i N. For each j = 1, 2, . . . , t, in view of (3.9)(3.11),
(3.4), (3.6), we have
_

iA
N
j

2
i,N
_
1/2

k=1
_

iA
N
j
(S
i
k
)
2
_
1/2

2
D

iA
N
j

i
+
1
8
_

iA
N
j
b
2
i
_
1/2

3
8D
. (3.13)
Relations (3.7) and (3.8) show that
I
N
=
N

i=1
a
i
b
i
+
N

i=1
a
i

i,N
+

i=N+1
a
i

i,N
=
t

j=1
_

iA
N
j
a
i
b
i
_
+
t

j=1
_

iA
N
j
a
i

i,N
_
+

i=N+1
a
i

i,N
.
Systems of random variables 793
For each j = 1, 2, . . ., t we now select b
i
(i A
N
j
) such that (3.6) holds and

iA
N
j
a
i
b
i
=
1
D
_

iA
N
j
a
2
i
_
1/2
;
we also select N N such that
|a|
2
2
_
N

i=1
a
2
i
_
1/2
2
t

j=1
_

iA
N
j
a
2
i
_
1/2
.
Then it follows by the above equality and (3.12), (3.13), and (3.4) that
I
N

1
D
t

j=1
_

iA
N
j
a
2
i
_
1/2

j=1
_

iA
N
j
a
2
i
_
1/2
_

iA
N
j

2
i,N
_
1/2

i=N+1
[a
i
[ [
i,N
[

5
8D
t

j=1
_

iA
N
j
a
2
i
_
1/2

i=N+1
[a
i
[
2
_
1/2
_

i=N+1

2
i,N
_
1/2

1
2D
t

j=1
_

iA
N
j
a
2
i
_
1/2
.
The last inequality shows that for each t N and an arbitrary partitioning A
j

t
j=1
of positive integers there exist suciently large N and coecients b
i
(i =1, 2, . . . , N)
satisfying (3.6) such that
I
N
=
_

R
N
()() dP()
1
3D
t

j=1
_

iAj
a
2
i
_
1/2
. (3.14)
The next part of the proof is the estimate of the integral I
N
by an expression of
the form C||
t
, with constant C > 0. By Holders inequality
[I
N
[ |R
N
|
t
||
t
, t

=
t
t 1
, (3.15)
and the non-negativity of Riesz products it is sucient for this to nd an estimate
of the quantity
L
N
= |R
N
|
t

t
=
_

[R
N
()]
t

dP() (3.16)
(for the proof that
i

i=1
is a Sidon system we require only the estimate of the
L
1
-norms of Riesz products (see [2], Theorem 1.4.1)).
We consider t N, t 3. As before, let A
j

t
j=1
be a partitioning of N, let
A
N
j
= i = 1, . . . , N : i A
j
, and let b
i
be coecients satisfying (3.6).
794 S. V. Astashkin
By the elementary inequality (1 + x)
y
1 + yx (x 1, 0 < y 1),
[R
N
()]
t

i=1
(1 +b
i

i
())[1 + (t 1)
1
b
i

i
()]

i=1
_
1 +
D
2
t 1
b
2
i
+
t
t 1
b
i

i
()
_
=
t

j=1

iA
N
j
_
1 +
D
2
t 1
b
2
i
+
t
t 1
b
i

i
()
_
. (3.17)
Let m(B) be the cardinality of a subset B of N. Then the inner product in the last
expression is equal to the sum
_
1 +
D
2
t 1

iA
N
j
b
2
i
+
_
D
2
t 1
_
2

i1<i2
i1,i2A
N
j
b
2
i1
b
2
i2
+ +
_
D
2
t 1
_
m(A
N
j
)

iA
N
j
b
2
i
_
+

CjA
N
j
_
1 +
_
D
2
t 1
_
m(A
N
j
)m(Cj)

iA
N
j
\Cj
b
2
i
__
t
t 1
_
m(Cj )

iCj
b
i

i
(),
where the last sum is taken over all non-empty subsets C
j
of the set A
N
j
. In view
of (3.6), for t 3 the expression in the curly brackets has the estimate (t1)/(t2).
Hence, by (3.16), (3.17), and, again, by (3.6) we obtain
L
N

_
t 1
t 2
_
t
+
_
t 1
t 2
_
t1 t

j=1

CjA
N
j
_
t
t 1
_
m(Cj )+1

E
_

iCj

i
()
_

+
_
t 1
t 2
_
t2

1j1<j2t

Cj
1
A
N
j
1

Cj
2
A
N
j
2
_
t
t 1
_
m(Cj
1
)+m(Cj
2
)+2

E
_
2

k=1

iCj
k

i
()
_

+ +
_
t 1
t 2
_
ts

1j1<<jst

Cj
1
A
N
j
1

Cjs
A
N
js
_
t
t 1
_

s
k=1
m(Cj
k
)+s

E
_
s

k=1

iCj
k

i
()
_

+ +

C1A
N
1

CtA
N
t
_
t
t 1
_

t
j=1
m(Cj)+t

E
_
t

j=1

iCj

i
()
_

. (3.18)
If t 3, then
t
t 1
2 and
_
t 1
t 2
_
ts
_
t
t 1
_
s
4e (s = 0, 1, . . . , t). (3.19)
Systems of random variables 795
Moreover, since the sets A
N
j
(j = 1, 2, . . . , t) are pairwise disjoint, the inte-
grands in the last expression are distinct products of pairwise distinct functions
i
(i = 1, 2, . . . , N). The number of these functions is m(C
j
) (j = 1, 2, . . . , t) for terms
in the rst sum, m(C
j1
)+m(C
j2
) (1 j
1
< j
2
t) for terms in the second, . . . , and

t
j=1
m(C
j
) for terms in the last sum. Hence the largest index of functions
i
in
each integrand is at least m(C
j
), m(C
j1
) +m(C
j2
), . . . ,

t
j=1
m(C
j
), respectively.
As a result, from (3.18) and (3.19) we deduce
L
N
4e
_
N

i=1
2
i

1j1<<jsi1
[E(
j1

js

i
)[
_
, (3.20)
where the inner sum is taken over all collections of indices 1 j
1
< < j
s
i 1.
Hence it follows by (3.4) and (3.5) that
L
N

4e
D

i=1

i
< 2,
and therefore, by (3.15) and (3.16) for all N N we obtain
[I
N
[ 2||
t
.
In view of (3.14), the last inequality shows that
t

j=1
_

iAj
a
2
i
_
1/2
6D||
t
.
Hence, by Lemma 1 for positive integers t 3 and arbitrary a = (a
i
)

i=1
l
2
we
obtain the inequality
K
1,2
(

t, a) 6

2 D||
t
. (3.21)
We shall also verify this inequality for t = 1, 2.
We claim that the above-constructed system
i

i=1
is a Riesz basis sequence,
that is,
_
_
_
_

i=1
a
i

i
_
_
_
_
2
|(a
i
)|
2
. (3.22)
In fact, we assume that |
i
|
2
= 1, therefore for each m N it follows by (3.5) that
_
_
_
_
m

i=1
a
i

i
_
_
_
_
2
2
=
m

i=1
a
2
i
+ 2

1i<jm
a
i
a
j
E(
i

j
)

i=1
a
2
i
_
1 2
_

1i<jm
(E(
i

j
))
2
_
1/2
_

1
2
|(a
i
)|
2
2
. (3.23)
In a similar way,
_
_
_
_
m

i=1
a
i

i
_
_
_
_
2
2

3
2
|(a
i
)|
2
2
and the proof of (3.22) is complete.
796 S. V. Astashkin
Relation (3.2) shows that g
i

i=1
, and therefore also
i

i=1
, is an S
p
-system
for each p < . Hence it is a Banach system ([2], Corollary 1.3.1), that is, for some
M = M(D) we have
|a|
2
M
_
_
_
_
m

i=1
a
i

i
_
_
_
_
1
(a = (a
i
)

i=1
l
2
). (3.24)
Since |a|
2
|a|
1
, it follows that K(1, a; l
1
, l
2
) = |a|
2
. By the concavity of the
K-functional we can now deduce from (3.23) and (3.24) that
K
1,2
(

2, a)

2 K
1,2
(1, a) 2||
2
(3.25)
and
K
1,2
(1, a) M||
1
. (3.26)
For arbitrary t 1 we can nd a positive integer t
0
such that t
0
t < t
0
+ 1.
Now, in view of the concavity of the K-functional again and also by (3.21), (3.25),
and (3.26) we obtain
K
1,2
(

t, a)
_
t/t
0
K
1,2
(

t
0
, a)

2 K
1,2
(

t
0
, a) C||
t
,
where C =

2 max(2; M; 6

2 D) depends only on D.
Thus, relation (3.3) holds also for the system
i

i=1
, which completes the proof
of the theorem.
The next two results are immediate consequences of Theorems 1 and 4.
Theorem 5. From a system f
n

n=1
of random variables on a probability space
(, , P) one can select a subsystem
i

i=1
equivalent in distribution to the
Rademacher system on [0, 1] if and only if there exists a subsequence f
nk
of f
n

such that
(1) [f
nk
()[ D (k = 1, 2, . . . ; );
(2) f
nk
0 weakly in L
2
;
(3) d = inf
k=1,2,...
|f
nk
|
2
> 0.
Moreover, the constant involved in the equivalence will depend only on D and d.
Proof. By Theorem 4 we can select a subsequence
i

i=1
f
n

n=1
such that
the equivalence (3.1) holds. Then, however,
i

P
r
n
by Theorem 1 and the
proof of the suciency is complete.
The necessity of conditions (1) and (3) is a consequence of the denition of equiv-
alence in distribution and the fact that the Rademacher system has these properties.
As pointed out in the introduction, the Rademacher system is an S
p
-system for each
p < . Hence a system equivalent to it in distribution must also have this property.
Thus, the necessity of condition (2) is a consequence of Theorem A stated in the
introduction.
Remark 3. In connection with Theorem 5 we recall a problem posed by Alexits
[22], 3.2. Is it always possible to select an innite multiplicative or even strictly
multiplicative subsystem from an arbitrary complete orthogonal system of func-
tions? The result of Theorem 5 answers (in the armative) a closely related ques-
tion: from an arbitrary system of random variables with properties (1)(3) one can
select a subsystem that is equivalent in distribution to the strictly multiplicative
(more than that: consisting of independent functions) Rademacher system.
Systems of random variables 797
Theorem 6. If f
n

n=1
is an orthonormal system of random variables on a prob-
ability space (, , P), [f
n
()[ D (n = 1, 2, . . .; ), then one can select from
this system a subsequence
i

i=1
equivalent in distribution to the Rademacher
system. The constant involved in the equivalence depends only on D.
It is clear that a sequence
i

i=1
satisfying (3.1) is a Sidon system (and also
an S
p
-system for all p < , with constant K
p

p ). It turns out that its Sidon

property holds in a certain strong sense.
Corollary 4. Let f
n

n=1
be a system of random variables that has a subsequence
f
nk
satisfying conditions (1)(3) from Theorem 5. Then f
n

n=1
contains a
subsystem
i

i=1
such that for some constants
1
,
2
,
3
> 0 dependent only on D
and d and each polynomial
T() =
m

i=1
a
i

i
()
there exists a set E = E(T) , P(E) >
1
2
m
, such that
[T()[
2
|T|

3
m

i=1
[a
i
[ ( E).
Proof. By Theorem 5 there exists a subsequence
i
f
n
such that for some
C > 0 and all z > 0 we have the inequality
C
1

T(x)[ > Cz

P[T()[ > z C

T(x)[ > C
1
z

,
where

T(x) =

m
i=1
a
i
r
i
(x) (x [0, 1]). In particular, this shows that
C
1
|T|

T|

C|T|

.
Hence it follows by the denition of the Rademacher system that
P[T()[ > (2C
2
)
1
|T|

P[T()[ > (2C)

1
|

T|

C
1

T(x)[ > 2
1
|

T|

C
1
2
m
.
We set
1
= C
1
,
2
= (2C
2
)
1
, and E = E(T) = : [T()[ >
2
|T|

.
Then for E we have
[T()[ >
2
|T|

2
C
1
|

T|

=
3
m

i=1
[a
i
[,
where
3
=
2
C
1
.
As already pointed out in the introduction (see Theorem B

), under the same

assumptions about a sequence of measurable functions on [0, 1] as in Theorem B
one can select a subsystem of this sequence that has the SidonZygmund property
and, under the additional assumption (0.3), the SidonZygmund property in the
narrow sense. The following two theorems show that even somewhat stronger results
actually hold.
798 S. V. Astashkin
Theorem 7. Let f
n

n=1
be a sequence of measurable functions on [0, 1] that
contains a subsequence f
nk
with properties (1)(3) in Theorem 5. Then there
exists a subsystem
i
f
n
with the following property.
For some subset E of [0, 1], [E[ > 0, and each interval I [0, 1], [I E[ > 0,
there exists an index k
0
= k
0
(I) such that the sequence
i

i=k0
(
I
(x) = 1,
x I, and
I
(x) = 0, x / I) is equivalent in distribution to the Rademacher
system on [0, 1]. The constant involved in this equivalence depends only on D, d,
and the interval I.
Proof. By Theorem 1 it is sucient to nd a subsystem
i
of f
n
and a subset E
of [0, 1], [E[ > 0, such that for each interval I [0, 1], [I E[ > 0, there exists an
index k
0
= k
0
(I) such that
_
_
_
_

i=1
a
i

k0+i1

I
_
_
_
_
t
K
1,2
(

t, a) (a = (a
i
)

i=1
, t 1)
(the constant involved in this equivalence depends on D, d, and I).
The arguments bringing us to this relation are mostly similar to the ones used
in the proof of Theorem 4. We merely make several observations.
First of all, inequality (3.2) remains valid, of course, if we replace the functions
g
i
by g
i

I
(here I is an arbitrary subinterval of [0, 1]).
Next, as in the proof of Theorem 1.4.2 in [2], one can select a subsequence

i
g
i
such that relations of type (3.5) hold for integrals over each interval
I [0, 1].
Finally, the well-known Marcinkiewiczs lemma ([2], Lemma 1.2.5) allows us to
assume that there exists a subset E of [0, 1], [E[ > 0, such that for each F E,
[F[ > 0, the inequality
liminf
i
_
F

2
i
(x) dx > 0 (3.27)
holds. After that one virtually repeats the proof of Theorem 4. The only dierence
is that one must replace integrals over [0, 1] by integrals over an interval I such that
[I E[ > 0.
Theorem 8. Let f
n

n=1
be a sequence of measurable functions on [0, 1] that
contains a subsequence f
nk
satisfying conditions (1) and (2) of Theorem 5 and
also condition (0.3).
Then there exists a subsystem
i
of f
n
such that for each interval I [0, 1],
[I[ > 0, there exists an index k
0
= k
0
(I) such that the sequence
i

i=k0
is
equivalent in distribution to the Rademacher system on [0, 1].
The proof of Theorem 8 is perfectly similar to the proof of the preceding theo-
rem because, in view of (0.3), we can assume that the selected subsystem
i

i=1
satises (3.27) for each set F [0, 1] of positive measure.
4. Density of subsystems equivalent in
distribution to the Rademacher system
We consider now the problem of the selection of nite subsystems of uniformly
bounded orthonormal systems of functions f
n

N
n=1
on [0, 1]. We claim that such a
Systems of random variables 799
system always contains a subsystem f
ni

s
i=1
of logarithmic density (s C log
2
N)
that is equivalent in distribution to the system of the rst s Rademacher functions
with constant independent of s. This is an improvement on Theorem C in the
introduction established by Kashin [9].
Theorem 9. Let f
n

N
n=1
be an orthonormal system of functions on [0, 1] such that
[f
n
(x)[ D (n = 1, 2, . . ., N). Then it contains a collection of functions f
ni

s
i=1
,
where s max[
1
6
log
2
N]; 1 and 1 n
1
< n
2
< < n
s
N, such that for some
C > 0 dependent only on D and all a
i
R (i = 1, 2, . . . , s) and z > 0 the following
inequality holds:
C
1

i=1
a
i
r
i
(x)

>Cz
_

i=1
a
i
f
ni
(x)

>z
_

i=1
a
i
r
i
(x)

>C
1
z
_

.
We require two lemmas for the proof. The rst is well known ([10], Lemma 8.4.1).
Lemma 3. Each system of functions f
n

N
n=1
(log
2
N 6) satisfying the assump-
tions of Theorem 9 contains functions f
ni

s
i=1
(1 n
1
< n
2
< < n
s
N),
where s [
1
6
log
2
N], such that

As
_
E
_
s

i=1
_
f
ni
(x)
D
_
i
__
2
10
s
. (4.1)
Here A
s
is the set of collections = (
i
)
s
i=1
such that
(a)
i
= 0, 1 or 2; (b)

i: i=1
1 1; (c)

i: i=2
1 1.
The second lemma enables one to extend under certain conditions a uniformly
bounded system of functions on [0, 1] to a larger set so that it becomes a multiplica-
tive system remaining at the same time uniformly bounded. For similar results on
the extension to an orthonormal collection, see [10], 7.1.
Lemma 4. Let g
i

s
i=1
be a collection of functions on [0, 1] such that [g
i
(x)[ D
(i = 1, 2, . . . , s; x [0, 1]) and assume that
max
A

E
_
s

i=1
_
g
i
(x)
D
_
i
_

< 2
s
, (4.2)
where A

s
is the set of all collections = (
i
)
s
i=1
A
s
such that
i
= 0 or 1.
Then there exists a multiplicative system h
i

s
i=1
of functions on [0, 2] such that
h
i
(x)
[0,1]
(x) = g
i
(x) and |h
i
|
L[0,2]
D (i = 1, 2, . . ., s).
Proof. We introduce the partitioning
[1, 2) =
2
s
_
k=1

k
,
k
= [a
k1
, a
k
), a
k
= 1 +k2
s
(k = 1, 2, . . . , 2
s
),
800 S. V. Astashkin
and consider some (arbitrary) one-to-one correspondence between the set A

s
and
the family of intervals
k

2
s
1
k=1
. We shall dene the functions h
i
on each interval

k
in this family.
Let = (
i
)
s
i=1
be an element of A

s
corresponding to
k
. We consider the set
i
j

m
j=1
= i = 1, 2, . . ., s :
i
= 1, i
1
< i
2
< < i
m
(m = 1, 2, . . . , s).
For
k
let
u

(x) =
[ak1,)
(x)
[,ak]
(x) (x
k
).
Then the quantity
v() =
_
k
u

(x) dx = 2 a
k1
a
k
(a
k1
a
k
)
ranges over the entire interval [2
s
, 2
s
]. By (4.2) the functions g
i
(x) = g
i
(x)/D
satisfy the inequality [E( g
i1
g
i2
g
im
)[ < 2
s
, therefore there exists
k
(a
k1
, a
k
)
such that
v(
k
) = E( g
i1
g
i2
g
im
). (4.3)
Assume rst that m 2 and consider a family of piecewise constant functions
d
j
(x) (x [a
k1
,
k
) and j = 1, 2, . . . , m1) such that
(1) [d
j
(x)[ 1;
(2)
_
k
ak1
d
j1
d
j2
d
jl
dx = 0 (1 j
1
< j
2
< < j
l
m1)
(4.4)
(one can construct such a family, for instance, by carrying over the Rademacher
functions to [a
k1
,
k
) from [0, 1]). Let c
j
(x) (x [
k
, a
k
] and j = 1, 2, . . . , m1)
be a similar system on the interval [
k
, a
k
].
If x
k
, then
h
i
(x) 0 (i ,= i
j
, j = 1, 2, . . . , m),
h
ij
(x) = D[d
j
(x)
[ak1,k)
(x) +c
j
(x)
[k,ak]
(x)] (j = 1, 2, . . . , m1)
and
h
im
(x) = D
_
m1

j=1
d
j
(x)
[ak1,k)
(x)
m1

j=1
c
j
(x)
[k,ak]
(x)
_
.
By (4.4) and (4.3),
_
k
h
i1
h
i2
h
im
dx = D
m
_
k
u
k
(x) dx
= D
m
v(
k
) = E(g
i1
g
i2
g
im
). (4.5)
At the same time, for any other collection of indices 1 i

1
< < i

k
s it follows
from the denition of the functions h
i
(i = 1, 2, . . . , s) and from (4.4) that
_
k
h
i

1
h
i

2
h
i

l
dx = 0. (4.6)
Systems of random variables 801
On the other hand, if m = 1, then we set
h
i
(x) 0 (i ,= i
1
), h
i1
(x) = Du
k
(x),
and relations (4.5) and (4.6) again hold.
We now dene the functions h
i
(x) on the entire interval [0, 2] by setting h
i
(x) =
g
i
(x) for x [0, 1] and h
i
(x) = 0 for x
2
s. In view of (4.5) and (4.6), h
i

s
i=1
is
a multiplicative system on [0, 2] and [h
i
(x)[ D.
Proof of Theorem 9. In view of Theorem 1 and Remark 1 after it, it suces to
show that there exists a collection f
ni

s
i=1
(s max[
1
6
log
2
N]; 1) such that for
some constant dependent only on D,
_
_
_
_
s

i=1
a
i
f
ni
_
_
_
_
t
K
1,2
(

t, a) (t 1, a = (a
i
)
s
i=1
), (4.7)
where, as before, K
1,2
(

t, a) = K(

t, a; l
1
, l
2
).
Using Lemma 3 we select a subsystem f
ni

s
i=1
(s max[
1
6
log
2
N]; 1) satis-
fying (4.1). Since A

s
A
s
, the functions g
i
= f
ni
satisfy at the same time (4.2).
Hence, by Lemma 4, we can extend the functions f
ni

s
i=1
to a multiplicative sys-
tem h
i

s
i=1
on [0, 2], [h
i
(x)[ D. Then, however, the functions h

i
(x) = h
i
(2x)
make up a multiplicative system on [0, 1], [h

i
(x)[ D, and therefore by Corollary 3
in [16], for arbitrary t 1 and a = (a
i
)
s
i=1
we obtain
_
_
_
_
s

i=1
a
i
h

i
_
_
_
_
Lt[0,1]
C
1

t |a|
2
,
where C
1
> 0 depends only on D. Hence
_
_
_
_
s

i=1
a
i
f
ni
_
_
_
_
t
2C
1

t |a|
2
, (4.8)
which in a similar way to the proof of Theorem 4 (see inequality (3.20)) shows that
_
_
_
_
s

i=1
a
i
f
ni
_
_
_
_
t
CK
1,2
(

t, a) (t 1), (4.9)
where C = max(D; 2C
1
) depends only on D.
For the proof of the reverse inequality we set
a = (a
i
)
s
i=1
and P(x) =
s

i=1
a
i
f
ni
(x).
Let t N rst, t 3. If s 16D
2
, then by the denition of the K-functional we
have
|P|
t
|P|
2
= |a|
2

s
s

i=1
[a
i
[
1
4D
K
1,2
(

t, a). (4.10)
802 S. V. Astashkin
Hence it suces to consider the case when
s > 16D
2
. (4.11)
We consider now the Riesz products
R
s
(x) =
s

i=1
_
1 +
b
i
D
f
ni
(x)
_
.
For an arbitrary partitioning A
j

t
j=1
of the set 1, 2, . . . , s into t disjoint subsets
(some of which may be empty) we assume that

iAj
b
2
i
1 (j = 1, 2, . . . , t). (4.12)
We now nd a lower estimate for the integral
I
s
=
_
1
0
P(x)R
s
(x) dx
=
s

i=1
a
i
_
1
0
f
ni
(x) dx +
s

i=1
a
i
_
1
0
f
ni
(x)

s
s

k=1
_
b
k
D
f
nk
(x)
_
k
dx.
Rearranging the terms we obtain
I
s
= D
1
s

i=1
a
i
b
i
+
s

i=1
a
i

i
, (4.13)
where

i
=
_
1
0
f
ni
(x) dx +
_
1
0
f
ni
(x)

s
(i)
s

k=1
_
b
k
D
f
nk
(x)
_
k
dx,
A

s
(i) = A

s

i
,
i
= (
i
j
),
i
i
= 1,
i
j
= 0 (j ,= i).
In view of (4.1) and the condition [b
i
[ 1 (i = 1, 2, . . . , s), after applying the
CauchySchwarzBunyakovski inequality we obtain
s

i=1
[
i
[ D

As

_
1
0
s

k=1
_
f
nk
(x)
D
_
k
dx

D(s2
s
)
1/2
10
s/2
= D(s5
s
)
1/2
.
In view of (4.11), (s5
s
)
1/2
8/s < 2
1
D
2
, therefore
s

i=1
[
i
[
1
2D
. (4.14)
Systems of random variables 803
For each j = 1, 2, . . ., t we now nd b
i
(i A
j
) such that (4.12) holds and

iAj
a
i
b
i
=
_

iAj
a
2
i
_
1/2
.
Then by (4.13) and (4.14) we obtain
I
s

1
D
t

j=1
_

iAj
a
2
i
_
1/2

j=1

iAj
a
i

1
D
t

j=1
_

iAj
a
2
i
_
1/2

j=1
_

iAj
a
2
i
_
1/2
_

iAj
[
i
[
_

1
2D
t

j=1
_

iAj
a
2
i
_
1/2
.
The partitioning A
j

t
j=1
is arbitrary, therefore by Lemma 1 we nally obtain
I
s

1
2

2D
K
1,2
(

t, a) (t N, t 3). (4.15)
Next, we nd an upper estimate for I
s
. By Holders inequality,
[I
s
[ |R
s
|
t
|P|
t
, t

=
t
t 1
.
As in the proof of Theorem 4, we claim that for t 3 we have
|R
s
|
t
4e2
s

E
_
s

i=1
_
f
ni
(x)
D
_
i
_

,
therefore by (4.1), the inclusion A

s
A
s
, and the CauchySchwarzBunyakovski
inequality we obtain the relation |R
s
|
t
4e. Hence [I
s
[ 4e|P|
t
and it follows
from (4.15) that
K
1,2
(

t, a) 8

2 eD|P|
t
(t N, t 3). (4.16)
Similar relations hold also for t = 2 and t = 1. In fact, f
n

N
n=1
is an orthonormal
system and K
1,2
(1, a) = |a|
2
, so that
K
1,2
(

2, a)

2K
1,2
(1, a) =

2 |P|
2
. (4.17)
On the other hand, if t = 1, then, in view of (4.8) and as in the proof of Theorem 4,
we can use [2], Corollary 1.3.1, and therefore
K
1,2
(1, a) = |a|
2
M|P|
1
(4.18)
with M > 0 dependent only on D.
804 S. V. Astashkin
Finally, passing from positive integers (see (4.16)(4.18)) to arbitrary t 1 we
obtain in the standard way
K
1,2
(

t, a) C|P|
t
,
where the constant C =

2 max(M;

2; 8

2 eD) depends only on D. Hence, in

view of (4.9), the equivalence (4.7) holds with constant dependent only on D, which
completes the proof.
Remark 4. The example of a trigonometric system shows that the result of Theo-
rem 9 (similarly to Theorem C in the introduction) is best possible in order. For
Stechkin [23] has shown that if a sequence

2 cos 2n
k
x

k=1
(x [0, 1]) is a Sidon
system, then

k: nk<N
1 C lnN (N = 2, 3, . . . ).
Remark 5. The results on the selection of both innite and nite subsystems equiv-
alent in distribution to the Rademacher system proved in this section are best pos-
sible in the following sense: no further thinning can improve the distributions of
the corresponding polynomials. It is easy to see that for each subsequence r
ni

i=1
of the Rademacher system and arbitrary m N, a
i
R (i = 1, 2, . . ., m), and z > 0
we have

i=1
a
i
r
ni
(x)

> z
_

i=1
a
i
r
i
(x)

> z
_

.
5. K-closed representability of Banach couples
We say that a Banach couple (X
0
, X
1
) is K-closed representable in a Banach
couple (Y
0
, Y
1
) if there exists a linear operator T : X
0
+X
1
Y
0
+Y
1
such that
(a) T is a bounded injective operator from X
0
into Y
0
and from X
1
into Y
1
;
(b) for some constant C > 0 independent of x X
0
+X
1
and t > 0,
K(t, Tx; Y
0
, Y
1
) K(t, x; X
0
, X
1
).
As shown in [24] (see also [25]), a couple (l
1
, l
2
) is K-closed representable in the
couple (L

, G) of spaces of measurable functions on [0, 1]. Here we denote by G

the closure of the space L

in the Orlicz space L

N
with N(u) = e
u
2
1 and we
can dene the operator T as follows:
Ta(x) =

n=1
a
n
r
n
(x) (a = (a
n
)

n=1
l
2
).
As before, r
n

n=1
is the Rademacher system on [0, 1].
Assume that a sequence f
n

n=1
of measurable functions on [0, 1] has properties
(1)(3) in Theorem 5. Then we can select a subsystem
i
of it that is equivalent
in distribution to the Rademacher system. All the proofs in [24] use only estimates
Systems of random variables 805
of the distributions of the absolute values of polynomials in the Rademacher system,
therefore the corresponding results hold also for
i
. In particular, the operator
T

a(x) =

i=1
a
i

i
(x) (a = (a
i
)

i=1
l
2
)
also brings about a K-closed representability of the couple (l
1
, l
2
) in (L

, G).
We now prove a negative result on the K-closed representability of couples, which
is related to the above example.
Theorem 10. The couple (l
1
, l
2
) is not K-closed representable in the couple of
spaces (L

, L
2
) of random variables on a probability space (, , P).
Proof. Assume that, on the contrary, there exists a linear operator T : l
2
L
2
such
that
|Ta|

|a|
1
(a l
1
), |Ta|
2
|a|
2
(a l
2
) (5.1)
and for some C
1
> 0,
C
1
1
K
1,2
(t, a) K(t, Ta; L

, L
2
) C
1
K
1,2
(t, a) (a l
2
, t > 0), (5.2)
where, as before, K
1,2
(t, a) = K(t, a; l
1
, l
2
).
Let f
n
= Te
n
, where e
n
= (
j
n
),
n
n
= 1,
j
n
= 0 (j ,= n). Then it follows
from (5.1) that there exist D > 0 and d > 0 such that
D
1
|a|
1

_
_
_
_

n=1
a
n
f
n
_
_
_
_

D|a|
1
(a l
1
)
and
d|a|
2

_
_
_
_

n=1
a
n
f
n
_
_
_
_
2
d
1
|a|
2
(a l
2
). (5.3)
Hence, in particular,
[f
n
()[ D (n = 1, 2, . . .; ) and |f
n
|
2
d. (5.4)
Moreover, it follows from (5.3) that f
n
,= 0 (n = 1, 2, . . .) and for all k, m N,
k m, we have
_
_
_
_
k

n=1
a
n
f
n
_
_
_
_
2
d
2
_
_
_
_
m

n=1
a
n
f
n
_
_
_
_
2
.
Hence f
n

n=1
is a basis in the closed (in the L
2
-norm) linear hull Z of this system
([26], part I, 1). In other words, f
n

n=1
is a Riesz basis sequence, so that for
g Z we see that E(gf
n
) 0 as n [2], 1.1. If g L
2
is arbitrary, then
g = g
1
+ g
2
, where g
1
Z and g
2
Z

(Z

is the orthogonal complement of Z)

and E(gf
n
) = E(g
1
f
n
) 0 as n . Thus,
f
n
0 weakly in L
2
(as n ). (5.5)
806 S. V. Astashkin
Relations (5.4) and (5.5) show that the system f
n

n=1
has properties (1)(3) in
Theorem 5. Hence there exists a subsequence
i
f
n
equivalent in distribution
to the Rademacher system on [0, 1]. Hence, by Khintchines inequality (see [3] or
[6], 4.5) we obtain
C
1
2
|a|
2

_
_
_
_

i=1
a
i

i
_
_
_
_
p
C
2
|a|
2
(a = (a
i
)

i=1
l
2
), (5.6)
where the constant C
2
> 0 depends only on D, d, and p [1, ).
At the same time, it follows from (5.2) that
C
1
1
K
1,2
(t, a) K
_
t,

i=1
a
i

i
; L

, L
2
_
C
1
K
1,2
(t, a) (a l
2
, t > 0).
Hence, applying to the couples (L

, L
2
) and (l
1
, l
2
) the interpolation functor (,)
,p
with parameters 0 < < 1, p = 2/ (see 1), in view of [11], Theorem 5.2.1 we
obtain
_
_
_
_

i=1
a
i

i
_
_
_
_
p
|a|
r,p
=
_

i=1
(a

i
)
p
i
p/r1
_
1/p
,
where r = 2/(2 ) < 2 (the constant involved in the last equivalence depends
only on C
1
and ). Since |a|
r,p
, |a|
2
, the last relation is in contradiction with
inequalities (5.6) for p = 2/.
The author is grateful to B. S. Kashin, who posed the problem whose solution is
the subject of 4.
Bibliography
[1] S. V. Astashkin, Selection of subsystems majorized by the Rademacher system, Mat.
Zametki 65:4 (1999), 483495; English transl. in Math. Notes 65 (1999).
[2] V. F. Gaposhkin, Lacunary series and independent functions, Uspekhi Mat. Nauk 21:6
(1966), 382; English transl. in Russian Math. Surveys 21 (1966).
[3] A. Khintchine,

Uber dyadische Br uche, Math. Z. 18 (1923), 109116.

[4] A. Zygmund, Sur les series trigonometriques lacunaires, J. London Math. Soc. (2) 5:2
(1930), 138145.
[5] S. Banach, Sur les series lacunaires, Bull. Int. Acad. Polon. Sci. A. 1933:4/8, 149154.
[6] S. Kaczmarz and G. Steinhaus, Theorie der Orthogonalreihen, Subwencji Funduszu Kultury
Narodowej, Warsaw 1935.
[7] A. Zygmund, On lacunary trigonometric series, Trans. Amer. Math. Soc. 34 (1932),
435446.
[8] S. Kaczmarz and G. Steinhaus, Le syst`eme orthogonal de M. Rademacher, Studia Math.
2 (1930), 231247.
[9] B. S. Kashin, On certain properties of the space of trigonometric polynomials with uniform
norm, Trudy Mat. Inst. Steklov. 145 (1980), 111116; English transl. in Proc. Steklov Inst.
Math. 145 (1981).
[10] B. S. Kashin and A. A. Saakyan, Orthogonal series, Nauka, Moscow 1984; English transl.,
Amer. Math. Soc., Providence, RI 1989.
[11] J. Bergh and J. Lofstr om, Interpolation spaces. An introduction, Springer-Verlag, Berlin
1976.
[12] S. J. Montgomery-Smith, The distribution of Rademacher sums, Proc. Amer. Math. Soc.
109 (1990), 517522.
Systems of random variables 807
[13] T. Holmstedt, Interpolation of quasi-normed spaces, Math. Scand. 26 (1970), 177199.
[14] E. D. Gluskin and S. Kwapien, Tail and moment estimates for sums of independent random
variables with logarithmically concave tails, Studia Math. 114 (1995), 303309.
[15] J.-P. Kahane, Some random series of functions, Heath, Lexington, MA 1968.
[16] J. Jakubowski and S. Kwapien, On multiplicative systems of functions, Bull. Acad. Pol.
Sci. Ser. Sci. Math. 27 (1979), 689694.
[17] G. Pisier, Les inegalites de KhintchinKahane dapr`es C. Borell, Semin. Geom. des
Espaces de Banach, vol. 7, 1978, pp. 114.
[18] N. H. Asmar and S. Montgomery-Smith, On the distribution of Sidon series, Ark. Mat.
31:1 (1993), 1326.
[19] A. S. Belov and V. A. Rodin, Norms of lacunary polynomials in function spaces, Mat.
Zametki 51 (1992), 483495; English transl. in Math. Notes 51 (1992).
[20] V. A. Rodin and E. M. Semyonov [Semenov], Rademacher series in symmetric spaces,
Anal. Math. 1:3 (1975), 207222.
[21] S. G. Kren, Yu. I. Petunin, and E. M. Semenov, Interpolation of linear operators, Nauka,
Moscow 1978; English transl., Amer. Math. Soc., Providence, RI 1982.
[22] G. Alexits, Convergence problems of orthogonal series, Pergamon Press, New York 1961.
[23] S. B. Stechkin, Absolute convergence of Fourier series, Izv. Akad. Nauk SSSR Ser. Mat.
20 (1956), 385412. (Russian)
[24] S. V. Astashkin, Interpolation of subspaces of symmetric spaces spanned by the Rademacher
system, Izv. Ross. Akad. Estestv. Nauk Ser. Mat. Mat. Model. Inform. Upr. 1:1 (1997),
1835.
[25] S. V. Astashkin, On series with respect to the Rademacher system in rearrangement
invariant spaces close to L, Funktsional. Anal. i Prilozhen. 32:3, 6265; English transl.
in Functional Anal. Appl. 32 (1998).
[26] J. Lindenstrauss and L. Tzafriri, Classical Banach spaces, Springer-Verlag, Berlin 1977.
Samara State University
E-mail address: astashkn@ssu.samara.ru
Received 12/AUG/99
Translated by IPS(DoM)
Typeset by A
M
S-T
E
X