THE DISTRIBUTION OF

VECTOR-VALUED RADEMACHER SERIES

S.J. Dilworth and S.J. Montgomery-Smith
Abstract. Let X =

nxn be a Rademacher series with vector-valued coecients.

We obtain an approximate formula for the distribution of the random variable ||X||
in terms of its mean and a certain quantity derived from the K-functional of inter-
polation theory. Several applications of the formula are given.
1. Results
In [6] the second-named author calculated the distribution of a scalar Rademacher
series

n
a
n
. The principal result of the present paper extends the results of [6]
to the case of a Rademacher series

n
x
n
with coecients (x
n
) belonging to an
arbitrary Banach space E. Its proof relies on a deviation inequality for Rademacher
series obtained by Talagrand [9]. A somewhat curious feature of the proof is that
it appears to exploit in a non-trivial way (see Lemma 2) the platitude that every
separable Banach space is isometric to a closed subspace of

. The principal re-

sult is applied to yield a precise form of the Kahane-Khintchine inequalities and to
compute certain Orlicz norms for Rademacher series.
First we recall some notation and terminology from interpolation theory (see e.g.
[1]). Let (E
1
, ||.||
1
) and (E
2
, ||.||
2
) be two Banach spaces which are continuously
embedded into some larger topological vector space. For t > 0, the K-functional
K(x, t; E
1
, E
2
) is the norm on E
1
+E
2
dened by
K(x, t; E
1
, E
2
) = inf{||x
1
||
1
+t||x
2
||
2
: x = x
1
+x
2
, x
i
E
i
}.
For a sequence (a
n
)
2
, we shall denote the K-functional K((a
n
), t;
1
,
2
) by
K
1,2
((a
n
), t) for short. For 1 p < , a sequence (x
n
) in a Banach space (E, ||.||) is
said to be weakly-
p
if the scalar sequence (x

(x
n
)) belongs to
p
for every x

.
The collection of all weakly-
p
sequences is a Banach space, denoted
w
p
(E), with the
norm given by l
w
p
((x
n
)) = sup
||x

||1
||(x

(x
n
))||
p
(where ||(a
n
)||
p
= (

|a
n
|
p
)
1/p
).
If (x
n
)
w
2
(E), we make the following denition:
K
w
1,2
((x
n
), t) = sup
||x

||1
K
1,2
((x

(x
n
)), t).
Observe that K
w
1,2
((x
n
), t) is a continuous increasing function of t. In fact, it is a
Lipschitz function with Lipschitz constant at most
w
2
((x
n
)).
1991 Mathematics Subject Classication. Primary 46B20; Secondary 60B11, 60G50.
Key words and phrases. Rademacher series, K-functional, Banach space.
The second author was supported in part by NSF DMS-9001796
Typeset by A
M
S-T
E
X
1
2 S.J. DILWORTH AND S.J. MONTGOMERY-SMITH
Next we set up some function space notation. Let (, , P) be a probability
space. A Rademacher (or Bernoulli) sequence (
n
) is a sequence of independent
identically distributed random variables such that P(
n
= 1) = P(
n
= 1) =
1
2
.
For a random variable Y dened on , its decreasing rearrangement, Y

, is the
function on [0, 1] dened by Y

(t) = inf{s > 0 : P(|Y | > s) t}. For 0 < p < ,

the weak-L
p
norm of Y , denoted ||Y ||
p,
, is given by ||Y ||
p,
= sup
0<t<1
t
1
p
Y

(t).
As usual, ||Y ||
p
denotes (E|Y |
p
)
1/p
. Let be an Orlicz function on [0, ). The
Orlicz norm, ||Y ||

, is given by ||Y ||

= inf{c > 0 : E(|Y |/c) 1}. We shall

be particularly interested in the Orlicz functions
q
(t) = e
t
q
1 for 2 < q < .
The weak-
p
norm of the scalar sequence (a
n
) is dened by ||(a
n
)||
p,
= supn
1
p
a

n
,
where (a

n
) is the decreasing rearrangement of (|a
n
|).
Finally, we shall write A B to mean that there is a constant C > 0 such that
1
C
A B CA. We shall try to indicate in each case whether the implied constant
is absolute or whether it depends on some parameter, typically p [1, ), entering
into the expressions for A and B.
Now we can state the principal result of the paper.
Main Theorem. Let X =

n
x
n
be an almost surely convergent Rademacher
series in a Banach space E. Then, for t > 0, we have
(1) P(||X|| > 2E||X|| + 6K
w
1,2
((x
n
), t)) 4e
t
2
/8
,
and, for some absolute constant c, we have
(2) P
_
||X|| >
1
2
E||X|| +cK
w
1,2
((x
n
), t)
_
ce
t
2
/c
.
The proof of the Main Theorem will be deferred until the end of the paper in
order to proceed at once with the applications.
Corollary 1. Let X =

n
x
n
be an almost surely convergent Rademacher series
in a Banach space. Then, for 0 < t
1
10
, we have
(3) S

(t) E||X|| +K
w
1,2
((x
n
),
_
log(1/t)),
where S denotes the real random variable ||X||. The implied constant is absolute.
Proof. (1) and (2) give rise to the inequalities S

(4e
t
2
/8
) 2E||X||+6K
w
1,2
((x
n
), t)
and S

(ce
t
2
/c
)
1
2
E||X|| + cK
w
1,2
((x
n
), t), respectively, whence (3) follows for all
suciently small t by an appropriate change of variable. To see that the lower
estimate implicit in (3) is valid in the whole range 0 < t <
1
10
, we recall from [2]
that E||X||
2
9E
2
||X||. Hence, by the Paley-Zygmund inequality (see e.g. [4,p.8]),
for 0 < < 1, we have
P(||X|| > E||X||) (1 )
2
E
2
X
EX
2

1
9
(1 )
2
,
whence P(||X|| > (1
3

10
)E||X||)
1
10
, which easily implies (3).
In [4] Kahane proved that if P(||X|| > t) = , where X is a Rademacher
series in a Banach space, then P(||X|| > 2t) 4
2
. By iteration this implies
P(||X|| > st)
1
4
(4)
s
for s = 2
n
. According to our next corollary the exponent s
in the latter result may be improved to be a certain multiple of s
2
.
THE DISTRIBUTION OF VECTOR-VALUED RADEMACHER SERIES 3
Corollary 2. Let X =

n
x
n
be an almost surely convergent Rademacher series
in a Banach space. Then, for t > 0 and s 1, we have
P(||X|| > st)
_
1
c
1
P(||X|| > t)
_
c1s
2
for some absolute constant c
1
.
Proof. By choosing c
1
< c, where c is the constant which appears in (2), the
result becomes trivial whenever P(||X|| > t) c. Hence we may assume that
P(||X|| > t) < c. Choose > 0 such that P(||X|| > t) = ce

2
/c
. Then (2) gives
t
1
2
E||X|| +cK
w
1,2
((x
n
), ). Thus,
st
s
2
E||X|| +scK
w
1,2
((x
n
), )
2E||X|| +K
w
1,2
((x
n
), cs)
provided s max(4, 1/c). Now (1) gives
P(||X|| > st) 4e
(cs)
2
/8
= 4
_
1
c
(ce

2
/c
)
_
c
3
s
2
/8
= 4
_
1
c
(P(||X|| > t)
_
c
3
s
2
/8
,
which gives the result.
Our next corollary, which is the vector-valued version of a recent result of
Hitczenko [3], is a rather precise form of the Kahane-Khintchine inequalities.
Corollary 3. Let X =

n
x
n
be a Rademacher series in a Banach space. Then,
for 1 p < , we have
(E||X||
p
)
1/p
E||X|| +K
w
1,2
((x
n
),

p).
The implied constant is absolute.
Proof. We may assume that p 2. It follows from a result of Borell [2] that
(E||X||
2p
)
1/2p

3(E||X||
p
)
1/p
. Since
1
2
||Y ||
p
||Y ||
2p,
||Y ||
2p
for every ran-
dom variable Y (as is easily veried), it follows (letting S denote the random
variable ||X||) that
1
2
||S||
p
||S||
2p,

3||S||
p
. So it suces to prove that
||S||
p,
ES + K
w
1,2
((x
n
),

p) to obtain the desired conclusion. By Corollary 1,

we have
||S||
p,
ES + sup
0<t<1
t
1/p
K
w
1,2
((x
n
),
_
log(1/t))
= ES + sup
0<t<1
_
t
1/p
sup
||x

||1
K
1,2
((x

(x
n
)),
_
log(1/t))
_
= ES + sup
||x

||1
_
sup
0<t<1
t
1/p
K
1,2
((x

(x
n
)),
_
log(1/t))
_
.
4 S.J. DILWORTH AND S.J. MONTGOMERY-SMITH
To evaluate the expression in brackets we shall make use once more (see Corollary
2) of the elementary inequality K
1,2
((a
n
), s) max(1, s/t)K
1,2
((a
n
), t). Thus,
sup
0<te
p
t
1/p
K
1,2
((x

(x
n
)),
_
log(1/t))
_
sup
0<te
p
t
1/p

log(1/t)
p
_
K
1,2
((x

(x
n
)),

p)
= e
1
K
1,2
((x

(x
n
)),

p).
Moreover,
sup
e
p
<t<1
t
1/p
K
1,2
((x

(x
n
)),
_
log(1/t)) K
1,2
((x

(x
n
)),

p).
Finally, we obtain
1
e
K
w
1,2
((x
n
),

p) sup
||x

||1
_
sup
0<t<1
K
1,2
((x

(x
n
)),
_
log(1/t))
_
K
w
1,2
((x
n
),

p),
which gives the desired result.
Our nal application is to the calculation of the Orlicz norms ||S||
q
for 2 <
q < . The proof will use the scalar version of the result, which was obtained by
Rodin and Semyonov [8] (see also [7]). (Recall that by a result of Kwapien, [5],
||S||
q
E||X|| in the range 0 < q 2.)
Corollary 4. Let X =

n
x
n
be an almost surely convergent Rademacher series
in a Banach space. Then, for 2 < q < , we have
||S||
q
E||X|| + sup
||x

||1
||(x

(x
n
))||
p,
,
where
1
p
+
1
q
= 1 and S denotes ||X||. The implied constant depends only on q.
Proof. It is easily veried that ||f||
q
sup
0<t<1
(log(1/t))
1/q
f

(t). Hence, by
Corollary 1, we have
||S||
q
E||X|| + sup
0<t<1
(log(1/t))
1/q
K
w
1,2
((x
n
), t)
E||X|| + sup
0<t<1
_
(log(1/t))
1/q
sup
||x

||1
K
1,2
((x

(x
n
)), t)
_
E||X|| + sup
||x

||1
_
sup
0<t<1
(log(1/t))
1/q
K
1,2
((x

(x
n
)), t)
_
E||X|| + sup
||x

||1
||

n
x

(x
n
)||
q
E||X|| + sup
||x

||1
||(x

(x
n
))||
p,
,
where the last line follows from the result of Rodin and Semyonov.
THE DISTRIBUTION OF VECTOR-VALUED RADEMACHER SERIES 5
2. Proof of main result
The principal ingredient in the proof of the Main Theorem is the following de-
viation inequality of Talagrand [9].
Theorem A. Let X =

N
n=1

n
x
n
be a nite Rademacher series in a Banach
space and let M be a median of ||X||. Then, for t > 0, we have
P
_
|||
N

n=1

n
x
n
|| M| > t
_
4e
t
2
/8
2
,
where =
w
2
((x
n
)
N
n=1
).
Lemma 1. Let X =

N
n=1

n
x
n
be a nite Rademacher series in a Banach space
E. Then, for t > 0, we have
P(||X|| > 2E||X|| + 3K((x
n
)
N
n=1
, t;
w
1
(E),
w
2
(E))) 4e
t
2
/8
.
Proof. It follows from Theorem A that for all y
1
, . . . , y
N
in E, we have
(4) P(||

n
y
n
|| > 2E||

n
y
n
|| +t
w
2
((y
n
))) 4e
t
2
/8
.
On the other hand, since max ||

n
y
n
|| =
w
1
((y
n
)), we have the trivial estimate
(5) P(||

n
y
n
|| >
w
1
((y
n
))) = 0.
Let x
n
= x
(1)
n
+x
(2)
n
for 1 n N, let X
(1)
=

n
x
(1)
n
, and let X
(2)
=

n
x
(2)
n
.
Then

w
1
((x
(1)
n
)) +t
w
2
((x
(2)
n
)) + 2E||X
(2)
||
w
1
((x
(1)
n
)) +t
w
2
((x
(2)
n
)) + 2E||X
(1)
|| + 2E||X||
3
w
1
((x
(1)
n
)) +t
w
2
((x
(2)
n
)) + 2E||X||
2E||X|| + 3(
w
1
((x
(1)
n
)) +t
w
2
((x
(2)
n
))).
Let Q denote 2E||X|| + 3(
w
1
((x
(1)
n
)) + t
w
2
((x
(2)
n
))). Then, by (4) and (5) and by
the above inequality, we have
P(||X|| > Q) P(||X
(1)
|| +||X
(2)
|| >
w
1
((x
(1)
n
)) +t
w
2
((x
(2)
n
)) + 2E||X
(2)
||)
P(||X
(1)
|| >
w
1
((x
(1)
n
))) +P(||X
(2)
|| > 2E||X
(2)
|| +t
w
2
((x
(2)
n
)))
< 0 + 4e
t
2
/8
The desired conclusion now follows from the denition of the K-functional.
Lemma 2. Let x
1
, . . . , x
N
be elements of the Banach space

. Then
K((x
n
)
N
n=1
, t;
w
1
(

),
w
2
(

)) 2K
w
1,2
((x
n
)
N
n=1
, t).
6 S.J. DILWORTH AND S.J. MONTGOMERY-SMITH
Proof. For 1 n N, let x
n
= (x
n,j
)

j=1

. A simple convexity argument gives

||(x
n
)||

w
p
()
= sup
1j
_
N

n=1
|x
n,j
|
p
_
(1/p)
.
It follows that the mapping which associates an element (y
n
)

n=1

w
p
(

) with
the element in

(
p
) whose jth coordinate equals (y
n,j
)

n=1
is an isometry. Hence
K((x
n
), t;
w
1
,
w
2
) = K(((x
n
)), t;

(
1
),

(
2
)). Let (y
n
)

n=1

(
2
) and let
> 0. For each n there exists a splitting y
n
= z
(1)
n
+z
(2)
n
such that
||(z
(1)
n,j
)

j=1
||
1
+t||(z
(2)
n,j
)

j=1
||
2
K
1,2
((y
n,j
)

j=1
, t) +.
It follows that
||(z
(1)
n
)||
(1)
+t||(z
(2)
n
)||
(2)
= sup
1n<
||(z
(1)
n,j
)

j=1
||
1
+t sup
1n<
||(z
(2)
n,j
)

j=1
||
2
2 sup
1n<
K
1,2
((y
n,j
)

j=1
, t) + 2
2K
w
1,2
((y
n
), t) + 2.
Since is arbitrary, the result now follows from the denition of the K-functional.
Proof of Main Theorem. First we prove (1) for a nite Rademacher series

N
n=1

n
x
n
.
Since every separable Banach space embeds isometrically into

, we may as-
sume that E is a closed subspace of

. Recall that K
w
1,2
((x
n
), t) was dened
as sup
||x

||1
K
1,2
((x

(x
n
)), t). By the Hahn-Banach Theorem, the supremum is
the same whether it is taken over elements of E

or over elements of

. Hence
(1) follows by combining Lemmas 1 and 2. The result for an innite series fol-
lows from the result for

N
n=1

n
x
n
by taking the limit as N . To prove
(2), we use the result from [6] that there exists an absolute constant d such that
P(

n
a
n
> dK
1,2
((a
n
), t)) de
t
2
/d
for every sequence (a
n
)
2
. Hence
P
_
||

n
x
n
|| >
d
2
K
w
1,2
((x
n
), t)
_
inf
||x

||1
P(||

n
x
n
|| > dK
1,2
((x

(x
n
)), t))
inf
||x

||1
P(

n
x

(x
n
) > dK
1,2
((x

(x
n
)), t))
de
t
2
/d
.
The Paley-Zygmund inequality now gives
P
_
||X|| >
1
2
E||X|| +
d
6
K
w
1,2
((x
n
), t)
_
min
_
P
_
||X|| >
3
4
E||X||
_
, P
_
||X|| >
d
2
K
w
1,2
((x
n
), t)
__
min
_
1
144
, de
t
2
/d
.
_
.
THE DISTRIBUTION OF VECTOR-VALUED RADEMACHER SERIES 7
References
1. J. Bergh and J. Lofstrom, Interpolation Spaces. An Introduction, Grundlehren der mathema-
tischen Wissenschaften, Vol. 223, Springer-Verlag, Berlin- Heidelberg-New York, 1976.
2. C. Borell, On the integrability of Banach valued Walsh polynomials, Seminaire de Probabilites
XIII Lecture Notes in Mathematics, Vol. 721, Springer-Verlag, Berlin-Heidelberg- New York,
1979, pp. 13.
3. P. Hitczenko, Domination inequality for martingale transforms of a Rademacher sequence,
preprint (1992).
4. J.P. Kahane, Some Random Series of Functions, Cambridge Studies in Advanced Mathemat-
ics Vol. 5, Cambridge University Press, Cambridge, 1985.
5. S. Kwapie n, A theorem on the Rademacher series with vector valued coecients, Probability
in Banach spaces, Oberwolfach 1975, Lecture Notes in Mathematics, Vol. 526, Springer-Verlag,
Berlin-Heidelberg-New York, 1976, pp. 157158.
6. S.J. Montgomery-Smith, The distribution of Rademacher sums, Proc. Amer. Math. Soc. 109
(1990), 517522.
7. G. Pisier, De nouvelles caracterisations des ensembles de Sidon, Mathematical Analysis and
Applications, Adv. Math. Supp. Stud. 7B (1981), 686725.
8. V.A. Rodin and E.M. Semyonov, Rademacher series in symmetric spaces, Analyse Math. 1
(1975), 207222.
9. M. Talagrand, An isoperimetric theorem on the cube and the Khintchine-Kahane inequalities,
Proc. Amer. Math. Soc. 104 (1988), 905909.
Department of Mathematics, University of South Carolina, Columbia, South Car-
olina 29208, U.S.A.
Department of Mathematics, University of Missouri, Columbia, Missouri 65211,
U.S.A.