STUDIA MATHEMATICA 205 (1) (2011)

Rademacher functions in BMO

by
Sergey V. Astashkin (Samara), Mikhail Leibov (New York)
and Lech Maligranda (Lule)
Abstract. The Rademacher sums are investigated in the BMO space on [0, 1]. They
span an uncomplemented subspace, in contrast to the dyadic BMO
d
space on [0, 1], where
they span a complemented subspace isomorphic to l
2
. Moreover, structural properties of
innite-dimensional closed subspaces of the span of the Rademacher functions in BMO
are studied and an analog of the KadecPeczyski type alternative with l
2
and c
0
spaces
is proved.
1. Introduction. In 1961, when studying some problems concerning
partial dierential equations, F. John and L. Nirenberg introduced the space
BMO of functions of bounded mean oscillation. In 1971 Feerman [6] an-
nounced that the dual to the real Hardy space H
1
on R
n
is BMO. Next
year, the proof was published by Feerman and Stein in their paper [7, The-
orem 2] (see also Garnett [8, Theorem 4.4], Grafakos [12, Theorem 7.2.2],
Kashin and Saakyan [15, Theorem 5.5], and Stein [30, pp. 142144]). This
duality result of Feerman called considerable attention to the BMO space
and after 1971 many results were proved about this space (see e.g. Garnett
[8, Chapter VI], Grafakos [12, Chapter 7] and Stein [30, Chapter IV]).
There is also a larger dyadic counterpart BMO
d
of the space of functions
of bounded mean oscillation, BMO
d
BMO. This dyadic space related to
BMO was studied already by Garnett and Jones [9]. Since BMO is trans-
lation invariant and BMO
d
is not, BMO is more important in analysis. On
the other hand, it is much easier to work with BMO
d
because of the fact
that dyadic cubes are nested (if two open dyadic cubes intersect then one of
them is contained in the other).
Consider the Rademacher functions on [0, 1] dened by
r
k
(t) = sign[sin(2
k
t)], k N, t [0, 1],
2010 Mathematics Subject Classication: 46E30, 46B20, 46B42.
Key words and phrases: Rademacher functions, BMO space, dyadic BMO space, sub-
spaces, complemented subspaces.
DOI: 10.4064/sm205-1-6 [83] c Instytut Matematyczny PAN, 2011
84 S. V. Astashkin et al.
and the set of Rademacher sums
R
n
(t) =
n

k=1
a
k
r
k
(t), a
k
R, for k = 1, . . . , n and n N.
The behaviour of Rademacher sums in the spaces L
p
= L
p
[0, 1] is well known
and it is described by the classical Khintchine inequalities: there exist con-
stants A
p
, B
p
> 0 such that for every sequence a
k

n
k=1
of real numbers and
any n N we have
(1) A
p
_
n

k=1
[a
k
[
2
_
1/2
|R
n
|
L
p
[0,1]
B
p
_
n

k=1
[a
k
[
2
_
1/2
, 0 < p < .
Hence, the Rademacher functions r
n
span an isomorphic copy of l
2
in
L
p
for every 0 < p < . Moreover, the subspace [r
n
]

n=1
is complemented
in L
p
for 1 < p < and it is not complemented in L
1
since no com-
plemented innite-dimensional subspace of L
1
can be reexive. In L

we
have |R
n
|
L

[0,1]
=

n
k=1
[a
k
[ and so the Rademacher functions span an
isometric copy of l
1
, which is known to be uncomplemented in L

. Inves-
tigations of Rademacher sums in general symmetric (rearrangement invari-
ant) spaces are well presented in the books by LindenstraussTzafriri [21],
KrenPetuninSemenov [17] and Astashkin [2], where also the denition
and several properties of symmetric spaces can be found.
The purpose of this paper is to investigate sums of Rademacher functions
in the BMO space on [0, 1]. Some results are contained in Leibovs disserta-
tion [18] (Proposition 2 with some estimates of type (7), which we correct in
this paper, and partly Theorem 5) but, in fact, they are not known to a wide
audience since they were not published in any journal and the dissertation
is not available.
The paper is organized as follows: after the introduction in Section 1,
we collect some necessary denitions, notation, and auxiliary results in Sec-
tion 2. The main result in Section 3 is Theorem 2 describing the behaviour of
Rademacher sums in BMO[0, 1]. In Section 4 we discuss the complementabil-
ity of Rademacher subspaces
d
and in BMO
d
and in BMO, respectively.
Namely, it is well-known that
d
is complemented in BMO
d
(Theorem 3). At
the same time we prove that is not complemented in BMO (Theorem 4).
Finally, in Section 5, we investigate the structure of innite-dimensional
subspaces of . In particular, in Theorem 5 we state the following analogue
of the KadecPeczyski type alternative for : every innite-dimensional
closed subspace X is either isomorphic to l
2
and complemented in
BMO, or contains a subspace Y isomorphic to c
0
and complemented in .
Then, in Examples 1 and 2, we construct block bases of the Rademacher
system whose span is l
2
and c
0
, respectively.
Rademacher functions in BMO 85
2. Preliminaries and auxiliary results. For any function f L
1
[0, 1]
and arbitrary interval I [0, 1] we denote
f
I
=
1
[I[

I
f(s) ds,
where [I[ is the Lebesgue measure of I. Then, as usual, the space BMO =
BMO[0, 1] consists of all f L
1
[0, 1] such that
(2) |f|
BMO
:= sup
I
1
[I[

I
[f(s) f
I
[ ds < .
The quantity |f|
BMO
is only a seminorm, since |f|
BMO
= 0 if f equals a
constant a.e. To turn BMO into a Banach space we can either restrict (2)
to the linear space
_
f L
1
[0, 1] :
1

0
f(x) dx = 0 with |f|
BMO
<
_
(with identication of functions equal a.e.) or consider in BMO one of the
norms |f|

BMO
= |f|
BMO
+[

1
0
f(x) dx[ or |f|

BMO
= |f|
BMO
+|f|
L
1
[0,1]
.
We also introduce a dyadic version of BMO. If I
k
n
= (k/2
n
, (k + 1)/2
n
],
k = 0, 1, . . . , 2
n
1, n = 0, 1, 2, . . . , are dyadic intervals in [0, 1], then the
space BMO
d
= BMO
d
[0, 1] consists of all f L
1
[0, 1] such that
|f|
d
= |f|
BMO
d
:= sup
k,n
1
[I
k
n
[

I
k
n
[f(s) f
I
k
n
[ ds < .
It is clear that BMO BMO
d
and |f|
d
|f|
BMO
for all f BMO.
Moreover, L

BMO and, for f L

[0, 1], we have

|f|
BMO
sup
I
_
1
[I[

I
(f(s) f
I
)
2
ds
_
1/2
sup
I
_
1
[I[

I
[f(s)[
2
ds
_
1/2
= |f|
L

.
At the same time, BMO ,= L

and BMO
d
,= BMO. For example, we have
ln [s 1/2[
[0,1]
(s) BMO L

and ln [s 1/2[
[1/2,1]
(s) BMO
d
BMO.
To nd a connection between the BMO- and BMO
d
-norms, we introduce
the functional
A(f) := sup
I
1
,I
2
[f
I
1
f
I
2
[, f L
1
[0, 1],
where the supremum is taken over all adjacent dyadic intervals I
1
, I
2
of the
same length.
86 S. V. Astashkin et al.
The following assertion is an exercise from Garnetts book (cf. [8, Problem
12(b), p. 266]). We present the proof with concrete constants for the sake of
completeness.
Proposition 1. For any f L
1
[0, 1] we have
(3)
1
3
[|f|
d
+ A(f)] |f|
BMO
32[|f|
d
+ A(f)].
Proof. It is clear that the left-hand inequality of (3) is an immediate
consequence of the estimate
(4) A(f) 2|f|
BMO
.
To prove (4), take two adjacent dyadic intervals I
1
and I
2
of the same length.
Then
(5) f
I
1
I
2
=
1
2
_
1
[I
1
[

I
1
f ds +
1
[I
2
[

I
2
f ds
_
=
1
2
(f
I
1
+ f
I
2
).
Therefore, for I := I
1
I
2
we have
1
[I
1
[

I
1
[f f
I
[ ds =
1
2[I
1
[

I
1
[(f f
I
1
) + (f f
I
2
)[ ds

1
2

1
[I
1
[

I
1
(f f
I
1
) ds +
1
[I
1
[

I
1
(f f
I
2
) ds

=
1
2
[f
I
1
f
I
2
[,
and similarly
1
[I
2
[

I
2
[f f
I
[ ds
1
2
[f
I
1
f
I
2
[.
Thus,
1
[I[

I
[f f
I
[ ds =
1
2
_
1
[I
1
[

I
1
[f f
I
[ ds +
1
[I
2
[

I
2
[f f
I
)[ ds
_

1
2
[f
I
1
f
I
2
[,
which implies (4).
Let us prove the right-hand inequality of (3). For any I [0, 1] we can nd
adjacent dyadic intervals I
1
and I
2
of the same length such that I I
1
I
2
and
1
2
[I
1
[ [I[ 2[I
1
[.
Rademacher functions in BMO 87
Then
1
[I[

I
[f(s) f
I
[ ds =
1
[I[

f(s)
1
[I[

I
f(t) dt

ds

1
[I[
2

I
[f(s) f(t)[ dt ds

16
[I
1
I
2
[
2

I
1
I
2

I
1
I
2
[f(s) f(t)[ dt ds

16
[I
1
I
2
[
2

I
1
I
2

I
1
I
2
[f(s) f
I
1
I
2
[ ds dt
+
16
[I
1
I
2
[
2

I
1
I
2

I
1
I
2
[f(t) f
I
1
I
2
[ ds dt
=
32
[I
1
I
2
[

I
1
I
2
[f(s) f
I
1
I
2
[ ds.
The above estimate and equality (5) imply that
1
[I[

I
[f(s) f
I
[ ds
16
[I
1
[

I
1
[f(s)
1
2
(f
I
1
+ f
I
2
)[ ds
+
16
[I
2
[

I
2
[f(s)
1
2
(f
I
1
+ f
I
2
)[ ds

16
[I
1
[

I
1
[f(s) f
I
1
[ ds +
16
[I
2
[

I
2
[f(s) f
I
2
[ ds
+ 16[f
I
1
f
I
2
[ 32|f|
d
+ 16A(f).
Hence,
|f|
BMO
32|f|
d
+ 16A(f) 32[|f|
d
+ A(f)].
3. Rademacher sums in BMO spaces. The main purpose of this
paper is to investigate the behaviour of Rademacher sums in the BMO and
BMO
d
spaces.
Proposition 2. For any a
k
R, k = 1, . . . , n, we have
(6)
1

2
|(a
k
)
n
k=1
|
l
2

_
_
_
n

k=1
a
k
r
k
_
_
_
d
|(a
k
)
n
k=1
|
l
2
and
(7)
2
3
max
0j<mn

k=j+1
a
k

A
_
n

k=1
a
k
r
k
_
4 max
0j<mn

k=j+1
a
k

.
88 S. V. Astashkin et al.
Proof. Set f =

n
k=1
a
k
r
k
and let I be a dyadic interval of length 2
m
,
that is, I = I
i
m
= (i/2
m
, (i + 1)/2
m
]. Then
(r
k
)
I
=
_
sgn(r
k
[
I
) if k m,
0 if k > m,
and so
f
I
=
n

k=1
a
k
(r
k
)
I
=
min(m,n)

k=1
a
k
sgn(r
k
[
I
).
Thus, if m n, then f[
I
is a constant and therefore the oscillation of f on
I vanishes, i.e., O
I
(f) := [I[
1

I
[f(x) f
I
[ dx = 0. Otherwise, if m < n, we
have
O
I
(f) =
1
[I[

I
[f(x) f
I
[ dx =
1
[I[

k=1
a
k
r
k
(x)
m

k=1
a
k
sgn(r
k
[
I
)

dx
=
1
[I[

k=m+1
a
k
r
k
(x)

dx =
1

k=m+1
a
k
r
km
(x)

dx
=
_
_
_
n

k=m+1
a
k
r
km
_
_
_
L
1
.
Using Khintchines inequality (1) for the space L
1
[0, 1] with the sharp con-
stant A
1
= 1/

2 (cf. [31]), we obtain

1

2
_
n

k=m+1
a
2
k
_
1/2

_
_
_
n

k=m+1
a
k
r
km
_
_
_
L
1

_
n

k=m+1
a
2
k
_
1/2
.
Thus,
1

2
_
n

k=1
a
2
k
_
1/2

_
_
_
n

k=1
a
k
r
k
_
_
_
d

_
n

k=1
a
2
k
_
1/2
.
Let now I
1
and I
2
be adjacent dyadic intervals of length 2
m
each. Then by
the above observation,
f
I
1
f
I
2
=
min(m,n)

k=1
a
k
sgn(r
k
[
I
1
)
min(m,n)

k=1
a
k
sgn(r
k
[
I
2
).
Let I be the smallest dyadic interval containing I
1
I
2
; let I have length 2
j
.
Of course, j < m and r
k
[
I
1
= r
k
[
I
2
if k j. Then for j > n we have f
I
1
= f
I
2
,
and for j n,
f
I
1
f
I
2
=
min(m,n)

k=j+1
a
k
[sgn(r
k
[
I
1
) sgn(r
k
[
I
2
)].
Rademacher functions in BMO 89
From the denition of I it follows that I
1
I
2
is in the middle of I. Suppose
that I
1
lies to the left of I
2
. Then it is easy to see that r
j+1
[
I
1
= 1, r
j+1
[
I
2
=
1 and r
k
[
I
1
= 1, r
k
[
I
2
= 1 if j + 2 k m. Thus,
[f
I
1
f
I
2
[ = 2

min(m,n)

k=j+2
a
k
a
j+1

and
A(f) = 2 max
0j<mn

k=j+2
a
k
a
j+1

.
It is not hard to check that
1
3
max
0j<mn

k=j+1
a
k

max
0j<mn

k=j+2
a
k
a
j+1

2 max
0j<mn

k=j+1
a
k

.
Combining this inequality with the previous equality, we obtain (7).
The following well-known assertion is an immediate consequence of in-
equalities (6) from Proposition 2. It was already proved by Garsia [10], [11]
and even for martingale BMO spaces. It was also obtained by Mller and
Schechtman [25, Theorem 1] and Kochne, Sagher and Zhou [16, Theorem 1],
who also gave an example showing that the similar result for BMO is not
true.
Theorem 1. The sequence r
k

k=1
of Rademacher functions is equiva-
lent in BMO
d
to the standard unit basis in l
2
.
From Propositions 1 and 2 and the elementary observation that
max
1mn

k=1
a
k

max
0j<mn

k=j+1
a
k

2 max
1mn

k=1
a
k

,
we obtain the following assertion:
Theorem 2. For any a
k
R, k = 1, . . . , n, and n N we have
2
9
__
n

k=1
a
2
k
_
1/2
+ max
1mn

k=1
a
k

_

_
_
_
n

k=1
a
k
r
k
_
_
_
BMO
256
__
n

k=1
a
2
k
_
1/2
+ max
1mn

k=1
a
k

_
.
In particular, the following equivalence holds:
(8)
_
_
_

k=1
a
k
r
k
_
_
_
BMO
|(a
k
)

k=1
|
l
2
+ sup
nN

k=1
a
k

.
90 S. V. Astashkin et al.
Corollary 1. The sequence r
k

k=1
of Rademacher functions is equiv-
alent in BMO to each of its subsequences.
Corollary 2. The system r
k

k=1
does not contain an unconditional
basic subsequence in BMO.
For example,

k=1
(1)
k
r
k
/k and

k=1
r
k
/k / .
Corollary 3. L

[0, 1] is a unique (up to equivalence of norms) sym-

metric space on [0, 1] which is embedded into BMO
d
.
Proof. For arbitrary (0, 1) let us introduce the -translation of the
dyadic BMO, that is, the space BMO
d
() with the norm |f|
d,
:= |f

|
d
,
where
(9) f

(s) := f(s )
[,1]
(s) + f(s + 1)
[0,]
(s), s [0, 1].
Mei proved in [23] that there exists a
0
(0, 1) such that
BMO = BMO
d
BMO
d
(
0
).
If X is a symmetric space on [0, 1] such that X BMO
d
, then X
BMO
d
(
0
) as well. Thus, X BMO. Next, since r
k

k=1
is an uncondi-
tional basic sequence with constant 1 in an arbitrary symmetric space (see,
for example, [2, Corollary 1.7]), by Theorem 2 we obtain
_
_
_

k=1
a
k
r
k
_
_
_
X
=
_
_
_

k=1
[a
k
[r
k
_
_
_
X
c
1
_
_
_

k=1
[a
k
[r
k
_
_
_
BMO
c
2
|(a
k
)|
l
1
.
On the other hand, we have
_
_
_

k=1
a
k
r
k
_
_
_
X
C
_
_
_

k=1
a
k
r
k
_
_
_
L

= C|(a
k
)|
l
1
.
Therefore, r
k

k=1
is equivalent in X to the standard unit basis in l
1
and,
hence, by the RodinSemenov theorem [27, Theorem 7], we conclude that
X = L

with equivalent norms.

Corollary 4. The sequence r
k

k=1
of Rademacher functions is not
weakly convergent to zero in BMO.
Proof. Dene a linear functional on by

0
_
n

k=1
a
k
r
k
_
=
n

k=1
a
k
, a
k
R, k = 1, . . . , n, n N.
Then, by the HahnBanach theorem, it can be extended to a functional

, because in view of Theorem 2 we have

0
_

k=1
a
k
r
k
_

k=1
a
k

C
_
_
_

k=1
a
k
r
k
_
_
_
BMO
.
Since
0
(r
n
) = 1 0, we see that r
n
0 weakly in BMO.
Rademacher functions in BMO 91
Remark 1. Astashkin and Maligranda [3] recently proved an equiva-
lence completely similar to (8) for Cesro function spaces K
p
on [0, 1] dened
by the norms |f|
K
p
= sup
0<x1
(x
1

x
0
[f(t)[
p
dt)
1/p
(1 p < ):
_
_
_
n

k=1
a
k
r
k
_
_
_
K
p
|a
k

n
k=1
|
l
2
+ max
1mn

k=1
a
k

.
It is worth noting that the spaces K
p
and BMO are not comparable, that is,
neither is embedded in the other one.
4. On complementability of Rademacher subspaces in BMO and
BMO
d
. We investigate the geometrical properties of the subspaces
d
:=
[r
k
]
BMO
d
and := [r
k
]
BMO
generated by the Rademacher system in BMO
d
and BMO, respectively. In fact, complementability of
d
in BMO
d
is well-
known (see, for example, Garsia [10], [11], Mller and Schechtman [25]).
However, we present a simple proof.
Theorem 3. The subspace
d
is complemented in BMO
d
.
Proof. In view of a dyadic version of the JohnNirenberg theorem, which
can be proved in the same way as the classical JohnNirenberg theorem (see,
for example, [12, pp. 124127]), for an arbitrary f BMO
d
, any dyadic
interval I
k
n
and > 0 we have
mx I
k
n
: [f(x) f
I
k
n
[ > e[I
k
n
[ exp
_

2e|f|
d
_
.
This inequality implies that, for any 1 p < ,
|f|
d
|f|
d,p
:= sup
k,n
_
1
[I
k
n
[

I
k
n
[f(x) f
I
k
n
[
p
dx
_
1/p
,
with a constant depending on p (see [12, p. 128] and [26, p. 525]). More
precisely,
|f|
d
|f|
d,p
2e[p(p)e]
1/p
|f|
d
.
Therefore, for every 1 p < , we have BMO
d
L
p
[0, 1]. It is well-known
that the orthogonal projector P generated by the Rademacher system is
bounded from L
p
[0, 1] onto [r
k
]

k=1
if 1 < p < . Then, by (6) and the
Khintchine inequality (1), for all p (1, ) and f BMO
d
,
|Pf|
d
|Pf|
L
p
|P| |f|
L
p
C
p
|P| |f|
d
.
Thus, P is bounded in BMO
d
and the proof is complete.
In contrast with Theorem 3 and with the remarkable theorem of Maurey
[22] (cf. also [24, pp. 229242]) which states that BMO[0, 1] is isomorphic
to dyadic BMO
d
[0, 1], the subspace is an uncomplemented subspace of
BMO. To prove this, we will need an auxiliary assertion.
92 S. V. Astashkin et al.
Denote by | = u
n

n=1
a block basis of the Rademacher system, that is,
u
n
=
m
n+1

k=m
n
+1
a
k
r
k
(n = 1, 2, . . .),
where 1 m
1
< m
2
< and a
k
R. Moreover, let

n
(|) =
m
n+1

k=m
n
+1
a
k
, n = 1, 2, . . . .
Proposition 3. The subspace contains a complemented subspace E
isomorphic to c
0
.
Proof. Take a block basis | = u
n

n=1
satisfying the following assump-
tions:
(a) |u
n
|
BMO
= 1, n = 1, 2, . . . .
(b) |u
n
|
d
(

m
n+1
k=m
n
+1
a
2
k
)
1/2
2
n
, n = 1, 2, . . . .
(c) [
n
(|)[ 2
n
, n = 1, 2, . . . .
It is clear that such a block basis exists. Let us prove that [u
n
]
BMO
is
isomorphic to c
0
.
If f =

n=1

n
u
n
with
n
R, then
f =

n=1
_
m
n+1

k=m
n
+1

n
a
k
r
k
_
=

k=1

k
r
k
,
where
k
=
n
a
k
if k = m
n
+ 1, . . . , m
n+1
. Assuming that p, q N satisfy
m
n1
p < m
n
< m
n+l
< q m
n+l+1
with some positive integers n and l, we will estimate the sum

q
k=p

k
. Using
(c), (a), Proposition 2 and inequality (4), we have

k=p

m
n

k=p

k
+
m
n+l

k=m
n
+1

k
+
q

k=m
n+l
+1

m
n

k=p

n1
a
k
+
n+l1

i=n
m
i+1

k=m
i
+1

i
a
k
+
q

k=m
n+l
+1

n+l
a
k

[
n1
[

m
n

k=p
a
k

+
n+l1

i=n
[
i
[

m
i+1

k=m
i
+1
a
k

+[
n+l
[

k=m
n+l
+1
a
k

Rademacher functions in BMO 93

sup
n
[
n
[
_

m
n

k=p
a
k

+
n+l1

i=n

m
i+1

k=m
i
+1
a
k

k=m
n+l
+1
a
k

_
sup
n
[
n
[
_
3
2
A(u
n1
) +
3
2
A(u
n+l
) +
n+l1

i=n
2
i
_
(3|u
n1
|
BMO
+ 3|u
n+l
|
BMO
+ 1)|
n
|
c
0
= 7|
n
|
c
0
.
Then from Theorem 2 and (b) it follows that |f|
BMO
C|
n
|
c
0
.
On the other hand, by (a), (b) and Theorem 2, there is a > 0 such that
A(u
n
) for all n N. Therefore, by (4) and Proposition 2, we obtain
|f|
BMO

1
2
A(f)
1
3
sup
nN
sup
m
n
+1p<qm
n+1

k=p

n
a
k

1
12
sup
nN
[
n
[A(u
n
)

12
|
n
|
c
0
.
Thus, we have proved that E := [u
n
]
BMO
c
0
. Since is separable, the
Sobczyk theorem (see, for example, [1, Corollary 2.5.9]) implies that E is a
complemented subspace in .
Theorem 4. The subspace is not complemented in BMO.
Proof. Assume on the contrary that is complemented in BMO and let
P
1
: BMO be a bounded linear projection whose range is . By Propo-
sition 3, there is a subspace E complemented in and such that E c
0
. Let
P
2
: E be a bounded linear projection. Then P := P
2
P
1
is a linear
projection bounded in BMO with image E. Thus, BMO contains a comple-
mented subspace E c
0
. Since BMO is a conjugate space (more precisely,
BMO = (Re H
1
)

see, for example, [15, p. 195]), this contradicts the well

known result due to BessagaPeczyski that a conjugate space cannot con-
tain a complemented subspace isomorphic to c
0
(see [5, Corollary 4], which
follows from Theorem 4 and its proof in [4]). This contradiction proves the
theorem.
5. Structure of subspaces of . Sarason [29] introduced the VMO =
VMO[0, 1] space (space of vanishing mean oscillation in [0, 1]) consisting of
all f BMO[0, 1] for which lim
|I|0
[I[
1

I
[f(x) f
I
[ dx = 0. This is a
closed subspace of BMO containing the space C[0, 1] of continuous func-
tions and is equal to the BMO-closure of C[0, 1]. The space VMO was in-
vestigated by several authors. For example, it is known that VMO is not
complemented in BMO (see, e.g., [13]). Structural properties of closed sub-
spaces of VMO were considered by Leibov [18], [19] who proved an analog
of the KadecPeczyski theorem for VMO (KadecPeczyski type alter-
94 S. V. Astashkin et al.
native [14]): an innite-dimensional closed subspace of VMO is either com-
plemented in BMO and isomorphic to l
2
, or, for every > 0, it contains a
subspace which is complemented in VMO and (1 + )-isometric to c
0
(cf.
[18, Theorem 3.4] and [19]). A similar dichotomy for the dyadic VMO space
was proved by Mller and Schechtman [25].
Our purpose here is to prove the KadecPeczyski type alternative for
the Rademacher subspace of BMO.
Theorem 5. Every innite-dimensional closed subspace X is ei-
ther isomorphic to l
2
and complemented in BMO, or contains a subspace Y
isomorphic to c
0
and complemented in .
In the proof of this theorem we will need some auxiliary result. Since
r
n
0 weakly in BMO (cf. Corollary 4), it follows that the corresponding
system of functionals biorthogonal to r
k

k=1
is not a basis of the space

.
Nevertheless, the following assertion holds.
Proposition 4. The space

has a basis.
Proof. Consider the sequence
s
n
= r
n
r
n1
, n = 1, 2, . . . , with r
0
= 0.
If f =

n=1

n
s
n
, then
f =

n=1

n
(r
n
r
n1
) =

n=0
(
n

n+1
)r
n
, where
0
= 0.
Therefore, by Theorem 2,
(10) |f|
BMO

_

n=0
(
n

n+1
)
2
_
1/2
+ sup
0m<n<
[
m

n
[
with a constant c > 0. This implies, in particular, that
(11) [
n
[ [
n

1
[ +[
1
[ 2c|f|
BMO
, n = 1, 2, . . . .
Let us prove that s
n

n=1
is a shrinking basis in , that is, for any

,
(12) |[
[s
n
]

n=m
| 0 as m .
Assume that (12) does not hold. Then there exist (0, 1), a functional

with || = 1, and a sequence of functions f

n
=

k=m
n
a
m
n
k
s
k
, where
m
1
< m
2
< , such that |f
n
|
BMO
= 1 and (f
n
) > 0 (n = 1, 2, . . .).
We construct two sequences of positive integers, q
i

i=1
and p
i

i=1
,
1 < q
1
< p
1
< q
2
< p
2
< , in the following way: q
1
= m
1
and p
1
is chosen
so that |

k=p
1
+1
a
q
1
k
s
k
|
BMO
< /2; q
2
is the least m
n
> p
1
and p
2
is such
that |

k=p
2
+1
a
q
2
k
s
k
|
BMO
< /2; q
3
is the least m
n
> p
2
and p
3
is such
that |

k=p
3
+1
a
q
3
k
s
k
|
BMO
< /2; and so on.
Rademacher functions in BMO 95
Set a
q
i
k
= 0 if p
i
< k < q
i+1
, i = 1, 2, . . . . Then | = u
i
, where
u
i
=

q
i+1
1
k=q
i
a
q
i
k
s
k
, is a block basis of s
n

n=1
. Moreover, by the denition
of u
i
,
(13) sup
iN
|u
i
|
BMO
2
and
(14) (u
i
) = (f
i
)
_

k=p
i
+1
a
q
i
k
s
k
_
(f
i
)
_
_
_

k=p
i
+1
a
q
i
k
s
k
_
_
_
BMO

2
.
Let us show that for every nonnegative sequence
n

n=1
such that

n=1

2
n
< and

n=1

n
=
the series

n=1

n
u
n
converges in BMO.
Let b
k
= a
q
i
k

i
for q
i
k < q
i+1
, i = 1, 2, . . . . If q
i
k < q
i+1
, then, in
view of (11) and (13), we obtain [b
k
[ 4c
i
, and hence
(15) lim
k
b
k
= 0.
Moreover,

k=1
(b
k
b
k+1
)
2
=

i=1
q
i+1
2

k=q
i
(a
q
i
k

i
a
q
i
k+1

i
)
2
+

i=1
(a
q
i
q
i+1
1

i
a
q
i+1
q
i+1

i+1
)
2
= A
1
+ A
2
.
We will estimate A
1
and A
2
separately. By (10),
A
1
=

i=1

2
i
q
i+1
2

k=q
i
(a
q
i
k
a
q
i
k+1
)
2
c
2

i=1

2
i
,
and, by (11),
A
2
2

i=1
_
(a
q
i
q
i+1
1
)
2

2
i
+ (a
q
i+1
q
i+1
)
2

2
i+1
_
4 sup
iN
max
q
i
k<q
i+1
[a
q
i
k
[
2

i=1

2
i
16c
2

i=1

2
i
.
Therefore, according to (10) and (15), the series

n=1

n
u
n
=

k=1
b
k
s
k
96 S. V. Astashkin et al.
converges in BMO. On the other hand, taking into account (14), we have

n=1

n
u
n
_
=

n=1

n
(u
n
)

2

n=1

n
= ,
which contradicts

. Thus, (12) is proved.

Finally, by Proposition 1.b.1 in [20], the biorthogonal system s

n=1
of
functionals is a basis in the space

and the proof is complete.

Proof of Theorem 5. Assume that there is an > 0 such that |f|
d

|f|
BMO
for every f X. Then, according to (6) and (1), for every 1 p
< and all f =

k=1
a
k
r
k
X,
(16) |f|
BMO

1
|f|
d

1
_

k=1
a
2
k
_
1/2

1
A
1
p
|f|
L
p
[0,1]
.
On the other hand, BMO L
p
for every 1 p < (see the proof of
Theorem 3). Therefore, the BMO-norm on X is equivalent to the L
p
-norm for
every 1 p < . In particular, this implies that X is isomorphic to l
2
. Since
the subspace
p
generated by the Rademacher system is complemented in L
p
if 1 < p < , and
p
is isomorphic to l
2
, it follows that X is complemented
in L
p
as well. Denote by P a linear projection bounded from L
p
onto X.
Then, by (16),
|Pf|
BMO

1
A
1
p
|Pf|
L
p
A
1
p

1
|P|
L
p
L
p
|f|
L
p
C
p
()|P|
L
p
L
p
|f|
BMO
.
Therefore, X is complemented in BMO.
Now, assume that there is a sequence f
n

n=1
X with |f
n
|
BMO
= 1
(n = 1, 2, . . .) such that |f
n
|
d
0 as n . Then f
n

n=1
does not
contain any subsequence converging in BMO. In fact, if lim
k
f
n
k
= f
BMO for some f
n
k
f
n
, then we have both |f|
BMO
= 1 and |f|
d
= 0
(and therefore f = 0), which is impossible.
Hence, passing to a subsequence if necessary, we may assume that for
some > 0,
(17) |f
m
f
n
|
BMO
> 0 for all m, n = 1, 2, . . . , m ,= n.
Let s

k=1
be the basis of

constructed in the proof of Proposition 4.

Using the diagonal process, it is not hard to choose a subsequence f
n
i

i=1

f
n
such that for any k N we have
s

k
(f
n
i+1
f
n
i
) 0 as i .
Since s

k=1
is a basis of

and |f
n
|
BMO
= 1 (n = 1, 2, . . . ), it follows
that
f
n
i+1
f
n
i
0 weakly in BMO.
Rademacher functions in BMO 97
Then, by (17) and by Proposition 1.a.12 in [20], there is a block basis | =
u
k

k=1
of the Rademacher system, which is equivalent to some subsequence
of f
n
i+1
f
n
i

i=1
(denoted f
n
i+1
f
n
i
as well). In particular,
|u
k
(f
n
k+1
f
n
k
)|
BMO
2
k
, k = 1, 2, . . . .
This implies that both u
k
0 weakly in BMO and |u
k
|
d
0 as k .
Let
0

be as in the proof of Corollary 4. If u

k
=

m
k+1
i=m
k
+1
a
i
r
i
,
1 m
1
< m
2
< , then

0
(u
k
) =
m
k+1

i=m
k
+1
a
i
0 as k .
Therefore, using the same arguments as in the proof of Proposition 3, we
may choose a subsequence u
k
i

i=1
such that [u
k
i
]

i=1
c
0
and [u
k
i
]

i=1
is
complemented in .
Using Theorem 5, we are able to describe the structure of subspaces of
in the following way.
Corollary 5. Let X be an innite-dimensional closed subspace
of BMO. The following conditions are equivalent:
(1) X does not contain a subspace isomorphic to c
0
.
(2) X is isomorphic to a dual space.
(3) X is reexive.
(4) X is isomorphic to l
2
.
(5) X is complemented in BMO.
(6) The BMO-norm on X is equivalent to the L
1
-norm.
Proof. By Theorem 5, condition (1) implies either of conditions (2)(6).
Conversely, it is obvious that each of the conditions (3), (4) and (6) im-
plies (1). The implications (2)(1) and (5)(1) are consequences of the
above mentioned results of BessagaPeczyski.
Recall that the function f

was dened in (9).

Corollary 6. There is a (0, 1) such that no bounded linear pro-
jection P in BMO
d
with range
d
commutes with the -shift on K

:=
BMO
d
BMO
d
(). This means that for every such projection there is a
function f K

such that P(f

) ,= (Pf)

.
Proof. Suppose, on the contrary, that for any (0, 1) there exists a
bounded linear projection P : BMO
d

d
such that P(f

) = (Pf)

for
every f K

. By Meis theorem (cf. [23]) there is

0
(0, 1) such that
K

0
= BMO and
|f|
BMO
max(|f|
d
, |f|
d,
0
).
98 S. V. Astashkin et al.
Then for any f BMO, by assumption,
|Pf|
BMO
max(|Pf|
d
, |Pf|
d,
0
) max(C|f|
d
, |(Pf)

0
|
d
)
= max(C|f|
d
, |P(f

0
)|
d
) C max(|f|
d
, |f

0
|
d
)
= C max(|f|
d
, |f|
d,
0
) |f|
BMO
,
which implies that P is bounded in BMO, contrary to Theorem 4.
To end the paper, we consider some examples of block bases of the
Rademacher system whose span is l
2
and c
0
, respectively.
Example 1. A block basis of the Rademacher system which spans l
2
in
BMO. Let u
k
:= r
2k+1
r
2k
and f =

n
k=1
a
k
u
k
, n = 1, 2, . . . . Then
f =
n

k=1
a
k
(r
2k+1
r
2k
) =
n

k=1
a
k
r
2k+1

n

k=1
a
k
r
2k
and, by (7),
A(f) max
1kn
[a
k
[ = |a
k

n
k=1
|
c
0
.
On the other hand, according to (6),
1

2
|a
k

n
k=1
|
l
2
|f|
d
2|a
k

n
k=1
|
l
2
.
Combining these relations with Theorem 2, we obtain
_
_
_
n

k=1
a
k
u
k
_
_
_
BMO
|a
k

n
k=1
|
l
2
, n = 1, 2, . . . .
Example 2. A block basis of the Rademacher system which spans c
0
in
BMO. Take a block basis | = u
n

n=1
, where
u
n
=
m
n+1

k=m
n
+1
a
k
r
k
with m
n+1
m
n
= 2
2n
, n = 1, 2, . . . ,
and
a
k
=
_
2
2n
if m
n
+ 1 k (m
n
+ m
n+1
)/2,
2
2n
if (m
n
+ m
n+1
)/2 + 1 k m
n+1
.
Then, by (6) and (8),
|u
n
|
d

_
m
n+1

k=m
n
+1
2
4n
_
1/2
= 2
n
, n = 1, 2, . . . ,
and
|u
n
|
BMO
2
2n
m
n+1
m
n
2
+ 2
n

1
2
, n = 1, 2, . . . .
Rademacher functions in BMO 99
Moreover,

n
(|) =
(m
n
+m
n+1
)/2

k=m
n
+1
2
2n

m
n+1

k=(m
n
+m
n+1
)/2+1
2
2n
= 0.
Then (see the proof of Proposition 3) [u
n
]
BMO
c
0
.
Acknowledgements. Research of S. V. Astashkin was partially sup-
ported by RFBR grant no. 10-01-00077.
Research of L. Maligranda was partially supported by the Swedish Re-
search Council (VR) grant 621-2008-5058.
References
[1] F. Albiac and N. J. Kalton, Topics in Banach Space Theory, Grad. Texts in Math.
233, Springer, New York, 2006.
[2] S. V. Astashkin, Rademacher functions in symmetric spaces, Sovrem. Mat. Fundam.
Napravl. 32 (2009), 3161 (in Russian); English transl.: J. Math. Sci. (New York)
169 (2010), 725886.
[3] S. V. Astashkin and L. Maligranda, Rademacher functions in Cesro type spaces,
Studia Math. 198 (2010), 235247.
[4] C. Bessaga and A. Peczyski, On bases and unconditional convergence of series in
Banach spaces, ibid. 17 (1958), 151164.
[5] , , Some remarks on conjugate spaces containing subspaces isomorphic to the
space c
0
, Bull. Acad. Polon. Sci. Sr. Sci. Math. Astronom. Phys. 6 (1958), 249250.
[6] C. Feerman, Characterizations of bounded mean oscillation, Bull. Amer. Math.
Soc. 77 (1971), 587588.
[7] C. Feerman and E. M. Stein, H
p
spaces of several variables, Acta Math. 129 (1972),
137193.
[8] J. B. Garnett, Bounded Analytic Functions, Springer, New York, 2007.
[9] J. B. Garnett and P. W. Jones, BMO from dyadic BMO, Pacic J. Math. 99 (1982),
351371.
[10] A. M. Garsia, The Burgess Davis inequalities via Feermans inequality, Ark. Mat.
11 (1973), 229237.
[11] , Martingale Inequalities: Seminar Notes on Recent Progress, Benjamin, Reading,
MA, 1973.
[12] L. Grafakos, Modern Fourier Analysis, 2nd ed., Springer, New York, 2009.
[13] D. Hadwin and H. Youse, A general view of BMO and VMO, in: Contemp. Math.
454, Amer. Math. Soc., Providence, 2008, 7591.
[14] M. I. Kadec and A. Peczyski, Bases, lacunary sequences and complemented sub-
spaces in the spaces L
p
, Studia Math. 21 (1962), 161176.
[15] B. S. Kashin and A. A. Saakyan, Orthogonal Series, Nauka, Moscow, 1984 (in Rus-
sian); English transl.: Amer. Math. Soc., Providence, 1989.
[16] E. Kochne, Y. Sagher and K. C. Zhou, BMO estimates for lacunary series, Ark.
Mat. 28 (1990), 301310.
100 S. V. Astashkin et al.
[17] S. G. Kren, Yu. I. Petunin, and E. M. Semenov, Interpolation of Linear Operators,
Nauka, Moscow, 1978 (in Russian); English transl.: Amer. Math. Soc., Providence,
1982.
[18] M. V. Leibov, The Geometry of the Function Space BMO, Candidates Dissertation,
Rostov-na-Donu, 1985, 133 pp. (in Russian).
[19] , Subspaces of the space VMO, Teor. Funktsi Funktsional. Anal. i Prilozhen. 46
(1986), 5154 (in Russian); English transl.: J. Soviet Math. 48 (1990), 536538.
[20] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces, I. Sequence Spaces, Sprin-
ger, Berlin, 1977.
[21] , , Classical Banach Spaces, II. Function Spaces, Springer, Berlin, 1979.
[22] B. Maurey, Isomorphismes entre espaces H
1
, Acta Math. 145 (1980), 79120.
[23] T. Mei, BMO is the intersection of two translates of dyadic BMO, C. R. Math.
Acad. Sci. Paris 336 (2003), 10031006.
[24] P. F. Mller, Isomorphisms between H
1
Spaces, IMPAN Monogr. Mat. 66, Birk-
huser, Basel, 2005.
[25] P. F. Mller and G. Schechtman, On complemented subspaces of H
1
and VMO, in:
Lecture Notes in Math. 1376, Springer, Berlin, 1989, 113125.
[26] J. Pipher and L. A. Ward, BMO from dyadic BMO on the bidisc, J. London Math.
Soc. (2) 77 (2008), 524544.
[27] V. A. Rodin and E. M. Semyonov, Rademacher series in symmetric spaces, Anal.
Math. 1 (1975), 207222.
[28] , , The complementability of a subspace that is generated by the Rademacher
system in a symmetric space, Funktsional. Anal. i Prilozhen. 13 (1979), no. 2, 9192
(in Russian); English transl.: Funct. Anal. Appl. 13 (1979), no. 2, 150151.
[29] D. Sarason, Functions of vanishing mean oscillation, Trans. Amer. Math. Soc. 207
(1975), 391405.
[30] E. M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscil-
latory Integrals, Princeton Univ. Press, Princeton, 1993.
[31] S. J. Szarek, On the best constants in the Khinchin inequality, Studia Math. 58
(1976), 197208.
Sergey V. Astashkin
Department of Mathematics and Mechanics
Samara State University
Akad. Pavlova 1
443011 Samara, Russia
E-mail: astashkn@ssu.samara.ru
Lech Maligranda
Department of Engineering Sciences and Mathematics
Lule University of Technology
SE-971 87 Lule, Sweden
E-mail: lech.maligranda@ltu.se
Mikhail Leibov
Horton Point LLC
1180 Avenue of the Americas
New York, NY 10036, U.S.A.
E-mail: mleybov@gmail.com
Received April 3, 2011
Revised version June 9, 2011 (7156)