3

Sums of independent random variables

This lecture collects a number of estimates for sums of independent random
variables with values in a Banach space E. We concentrate on sums of the
form

N
n=1

n
x
n
, where the
n
are real-valued Gaussian variables and the
x
n
are vectors in E. As we shall see later on such sums are the building
blocks of general E-valued Gaussian random variables and, perhaps more
importantly, stochastic integrals of E-valued step functions are of this form.
Furthermore, they are used in the denition of various geometric properties
of Banach spaces, such as type and cotype.
The highlights of this lecture are the Kahane contraction principle (The-
orem 3.1), a covariance domination principle (Theorem 3.9) and the Kahane-
Khintchine inequalities (Theorems 3.11 and 3.12).
3.1 Gaussian sums
We begin with an important inequality for sums of independent symmetric
random variables, due to Kahane.
Theorem 3.1 (Kahane contraction principle). Let (X
n
)

n=1
be a se-
quence of independent symmetric E-valued random variables. Then for all
a
1
, . . . , a
N
R and 1 p < ,
E
_
_
_
N

n=1
a
n
X
n
_
_
_
p

_
max
1nN
[a
n
[
_
p
E
_
_
_
N

n=1
X
n
_
_
_
p
.
Proof. For all (
1
, . . . ,
N
) 1, +1
N
the E
N
-valued random variables
(X
1
, . . . , X
N
) and (
1
X
1
, . . . ,
N
X
N
) are identically distributed and therefore
E
_
_
_
N

n=1

n
X
n
_
_
_
p
= E
_
_
_
N

n=1
X
n
_
_
_
p
.
30 3 Sums of independent random variables
For the general case we may assume that [a
n
[ 1 for all n = 1, . . . , N. Then
a = (a
1
, . . . , a
N
) is a convex combination of the 2
N
elements of 1, +1
N
,
say a =

2
N
j=1

(j)

(j)
. Hence,
E
_
_
_
N

n=1
a
n
X
n
_
_
_
p
= E
_
_
_
2
N

j=1

(j)
N

n=1

(j)
n
X
n
_
_
_
p
E
_
2
N

j=1

(j)
_
_
_
N

n=1

(j)
n
X
n
_
_
_
_
p
E
2
N

j=1

(j)
_
_
_
N

n=1

(j)
n
X
n
_
_
_
p
=
2
N

j=1

(j)
E
_
_
_
N

n=1
X
n
_
_
_
p
= E
_
_
_
N

n=1
X
n
_
_
_
p
,
where the third step follows from the convexity of the function t t
p
(or an
application of Jensens inequality).
As an application of the Kahane contraction principle we shall prove an
inequality which shows that Rademacher sums have the smallest L
p
-norms
among all random sums. Rademacher sums are easier to handle than the
Gaussian sums in which we are ultimately interested, and, as we shall see, there
are various techniques to pass on results for Rademacher sums to Gaussian
sums.
Let us begin with a denition. An 1, +1-valued random variable r is
called a Rademacher variable if
Pr = 1 = Pr = +1 =
1
2
.
Throughout these lectures, the notation (r
n
)

n=1
will be used for a Rademacher
sequence, that is, a sequence of independent Rademacher variables.
Theorem 3.2 (Comparison). Let (
n
)

n=1
be a sequence of independent
symmetric integrable real-valued random variables satisfying E[
n
[ 1 for all
n 1. Then for all x
1
, . . . , x
N
E and 1 p < we have
E
_
_
_
N

n=1
r
n
x
n
_
_
_
p
E
_
_
_
N

n=1

n
x
n
_
_
_
p
.
The proof of this theorem relies on an auxiliary lemma, for which we need
two denitions based on the following easy observation: if X
1
, . . . , X
N
are
random variables with values in E
1
, . . . , E
N
, then (X
1
, . . . , X
N
) is a random
variable with values in E
1
E
N
.
3.1 Gaussian sums 31
Denition 3.3. Two families of random variables (X
i
)
iI
and (Y
i
)
iI
, where
I is some index set and X
i
and Y
i
take values in a Banach space E
i
, are
identically distributed if for all choices of i
1
, . . . , i
N
I the random variables
(X
i1
, . . . , X
iN
) and (Y
i1
, . . . , Y
iN
) are identically distributed.
Note that by Proposition 2.13, if (X
i
)
iI
and (Y
i
)
iI
are families of inde-
pendent random variables such that X
i
and Y
i
are identically distributed for
all i I, then (X
i
)
iI
and (Y
i
)
iI
are identically distributed.
Denition 3.4. Two families of random variables (X
i
)
iI
and (Y
j
)
jJ
, where
I and J are index sets, X
i
takes values in E
i
for all i I and Y
j
takes values in
F
j
for all j J, are independent of each other if for all choices i
1
, . . . , i
M
I
and j
1
, . . . , j
N
I the random variables (X
i1
, . . . , X
iM
) and (Y
j1
, . . . , Y
iN
) are
independent.
Lemma 3.5. Let (
n
)

n=1
be a sequence of independent symmetric real-valued
random variables and let (r
n
)

n=1
be a Rademacher sequence independent of
(
n
)

n=1
. The sequences (
n
)

n=1
and (r
n
[
n
[)

n=1
are identically distributed.
Proof. By independence and symmetry we have
Pr
n
[
n
[ B
= Pr
n
= 1,
n
0,
n
B +Pr
n
= 1,
n
< 0,
n
B
+Pr
n
= 1,
n
0,
n
B +Pr
n
= 1,
n
< 0,
n
B
=
1
2
P
n
0,
n
B +
1
2
P
n
< 0,
n
B
+
1
2
P
n
0,
n
B +
1
2
P
n
< 0,
n
B
=
1
2
P
n
0,
n
B +
1
2
P
n
> 0,
n
B
+
1
2
P
n
0,
n
B +
1
2
P
n
< 0,
n
B
= P
n
B.
Since (
n
)

n=1
and (r
n
[
n
[)

n=1
are sequences of independent random vari-
ables, the lemma now follows from the observation preceding Denition 3.4.

Proof (Proof of Theorem 3.2). We may assume that the sequences (

n
)

n=1
and (r
n
)

n=1
are dened on distinct probability spaces

and
r
. By consid-
ering the
n
and r
n
as random variables on the probability space

r
,
we may assume that (
n
)

n=1
and (r
n
)

n=1
are independent of each other.
Since E

[
n
[ 1, with the Kahane contraction principle and Jensens
inequality we obtain
E
r
_
_
_
N

n=1
r
n
x
n
_
_
_
p
E
r
_
_
_E

n=1
r
n
[
n
[x
n
_
_
_
p
E
r
E

_
_
_
N

n=1
r
n
[
n
[x
n
_
_
_
p
= E

_
_
_
N

n=1

n
x
n
_
_
_
p
,
32 3 Sums of independent random variables
where the last identity follows from Lemma 3.5.
A real-valued random variable is called standard Gaussian if its distri-
bution has density
f

(t) =
1

2
exp(
1
2
t
2
)
with respect to the Lebesgue measure on R. For later reference we note that
is standard Gaussian if and only if its Fourier transform is given by
Eexp(i) = exp(
1
2

2
), R. (3.1)
The only if statement follows from the identity
1

2
_

exp(it
1
2
t
2
) dt = exp(
1
2

2
)
which can be proved by completing the squares in the exponential and then
shifting the path of integration fromi+R to R by using Cauchys formula; the
if part then follows from the injectivity of the Fourier transform (Theorem
2.8).
For a standard Gaussian random variable we have
E[[ =
1

2
_

[t[ exp(
1
2
t
2
) dt =
2

2
_

0
t exp(
1
2
t
2
) dt =
_
2/. (3.2)
From this point on, (
n
)

n=1
will always denote a Gaussian sequence, that is,
a sequence of independent standard Gaussian variables.
From (3.2) and Theorem 3.2 we obtain the following comparison result.
Corollary 3.6. For all x
1
, . . . , x
N
E and 1 p < ,
E
_
_
_
N

n=1
r
n
x
n
_
_
_
p
(/2)
p
2
E
_
_
_
N

n=1

n
x
n
_
_
_
p
. (3.3)
The geometric notions of type and cotype will be introduced in the exer-
cises. Without proof we state the following important converse to Corollary
3.6 for Banach spaces with nite cotype. Examples of spaces with nite cotype
are Hilbert spaces, L
p
-spaces for 1 p < , and the UMD spaces which will
be introduced in later lectures.
Theorem 3.7. If E has nite cotype, there exists a constant C 0 such that
for all x
1
, . . . , x
N
E,
E
_
_
_
N

n=1

n
x
n
_
_
_
2
C
2
E
_
_
_
N

n=1
r
n
x
n
_
_
_
2
.
3.1 Gaussian sums 33
The Kahane-Khintichine inequalities (Theorems 3.11 and 3.12 below) can
be used to extend this inequality to arbitrary exponents 1 p < .
The proof of Theorem 3.7 is beyond the scope of these lectures; we refer to
the Notes at the end of the lecture for references to the literature. When taken
together, Corollary 3.6 and Theorem 3.7 show that in spaces with nite cotype,
Gaussian sequences and Rademacher sums can be used interchangeably.
Without any assumptions on E, Theorem 3.7 fails. This is shown by the
next example.
Example 3.8. Let E = c
0
and let (u
n
)

n=1
be the standard unit basis of c
0
.
Then
E
_
_
_
N

n=1
r
n
u
n
_
_
_
c0
= E
_
max
1nN
[r
n
[
_
= 1.
Next we estimate E
_
_

N
n=1

n
u
n
_
_
c0
from below. First, if is standard Gaus-
sian, the inequality 1 x e
x
implies
P
_
max
1nN
[
n
[ r
_
=
_
1 P[[ > r

N
exp(NP[[ > r).
For r =
1
2

log N we estimate
P[[ >
1
2
_
log N
2

2
_

log N
1
2

log N
e

1
2
x
2
dx

1
2
_
log N e

1
2
log N
=
_
log N
2N
.
Hence, using the integration by parts formula of Exercise 2.1,
E
_
_
_
N

n=1

n
u
n
_
_
_
c0
= E
_
max
1nN
[
n
[
_
=
_

0
P
_
max
1nN
[
n
[ > r
_
dr

_ 1
2

log N
0
_
1 exp(NP[[ > r)

dr

1
2
_
log N
_
1 exp
_

_
N log N
2
__

1
2
_
log N as N .
Similar estimates show that the bound O(

log N) for N is of the

correct order.
We conclude this section with an important comparison result for Gaussian
sums.
34 3 Sums of independent random variables
Theorem 3.9 (Covariance domination). Let (
m
)

m=1
and (

n
)

n=1
be
Gaussian sequences on probability spaces and

, respectively, and let

x
1
, . . . , x
M
and y
1
, . . . , y
N
be elements of E satisfying
M

m=1
x
m
, x

)
2

n=1
y
n
, x

)
2
x

.
Then, for all 1 p < ,
E
_
_
_
M

m=1

m
x
m
_
_
_
p
E

_
_
_
N

n=1

n
y
n
_
_
_
p
.
Proof. Denote by F the linear span of x
1
, . . . , x
M
, y
1
, . . . , y
N
in E. Dene
Q L(F

, F) by
Qz

:=
N

n=1
y
n
, z

)y
n

m=1
x
m
, z

)x
m
, z

.
The assumption of the theorem implies that Qz

, z

) 0 for all z

,
and it is clear that Qz

1
, z

2
) = Qz

2
, z

1
) for all z

1
, z

2
F

. Since F is nite-
dimensional, by linear algebra we can nd a sequence (x
j
)
M+k
j=M+1
in F such
that Q is represented as
Qz

=
M+k

j=M+1
x
j
, z

)x
j
, z

.
We leave the verication of this statement as an exercise for the moment and
shall return to this issue from a more general point of view in the next lecture.
Now,
M+k

m=1
x
m
, z

)
2
=
N

n=1
y
n
, z

)
2
, z

. (3.4)
It follows from (3.1) that the random variables X :=

M+k
m=1

m
x
m
and Y :=

N
n=1

n
y
n
have Fourier transforms
Eexp(iX, x

)) =
M+k

m=1
Eexp(i
m
x
m
, x

))
=
M+k

m=1
exp
_

1
2
x
m
, x

)
2
_
= exp
_

1
2
M+k

m=1
x
m
, x

)
2
_
and similarly E

exp(iY, x

)) = exp(
1
2

N
n=1
y
n
, x

)
2
). Hence by (3.4) and
Theorem 2.8, X and Y are identically distributed. Thus, for all 1 p < ,
3.2 The Kahane-Khintchine inequality 35
E
_
_
_
M+k

m=1

m
x
m
_
_
_
p
= E

_
_
_
N

n=1

n
y
n
_
_
_
p
.
By Proposition 2.16,
E
_
_
_
M

m=1

m
x
m
_
_
_
p
E
_
_
_
M+k

m=1

m
x
m
_
_
_
p
,
and the proof is complete.
3.2 The Kahane-Khintchine inequality
The main result of this section states that all L
p
-norms of an E-valued Gaus-
sian sum are comparable, with universal constants depending only on p. First
we prove the analogous result for Rademacher sums; then we use the central
limit theorem to pass it on to Gaussian sums.
The starting point is the following inequality, which is a consequence of
Levys inequality.
Lemma 3.10. For all x
1
, . . . , x
N
E and r > 0 we have
P
__
_
_
N

n=1
r
n
x
n
_
_
_ > 2r
_
4
_
P
__
_
_
N

n=1
r
n
x
n
_
_
_ > r
__
2
.
Proof. Let us write S
n
:=

n
j=1
r
j
x
j
. As in the proof of Lemma 2.18 we put
A
n
:= |S
1
| r, . . . , |S
n1
| r, |S
n
| > r.
If for an A
n
we have |S
N
()| > 2r, then |S
N
()S
n1
()| > r. Now the
crucial observation is that (r
1
, . . . , r
N
) and (r
1
, . . . , r
n
, r
n
r
n+1
, . . . , r
n
r
N
)
are identically distributed; we leave the easy proof as an exercise. From this
and the fact that [r
n
[ = 1 almost surely we obtain
P(A
n
|S
N
S
n1
| > r) = P
_
A
n

__
_
_
N

j=n
r
j
x
j
_
_
_ > r
__
= P
_
A
n

__
_
_r
n
N

j=n
r
j
x
j
_
_
_ > r
__
= P
_
A
n

__
_
_x
n
+
N

j=n+1
r
n
r
j
x
j
_
_
_ > r
__
= P
_
A
n

__
_
_x
n
+
N

j=n+1
r
j
x
j
_
_
_ > r
__
= P(A
n
|x
n
+ (S
N
S
n
)| > r),
36 3 Sums of independent random variables
and similarly P|S
N
S
n1
| > r = P|x
n
+ (S
N
S
n
)| > r. Hence, by
the independence of A
n
and S
N
S
n
,
P(A
n
|S
N
| > 2r) P(A
n
|S
N
S
n1
| > r)
= P(A
n
)P|x
n
+ (S
N
S
n
)| > r
= P(A
n
)P|S
N
S
n1
| > r 2P(A
n
)P|S
N
| > r,
where the last step follows from Levys inequality after changing the order
of summation. Summing over n = 1, . . . , N and using Levys inequality once
more we obtain
P|S
N
| > 2r =
N

n=1
P(A
n
|S
N
| > 2r) 2
N

n=1
P(A
n
)P|S
N
| > r
= 2P
_
max
1nN
|S
n
| > r
_
P|S
N
| > r 4[P|S
N
| > r]
2
.

We are now ready to prove the following result, which is the Banach space
generalisation due to Kahane of a classical result for scalar random variables
of Khintchine.
Theorem 3.11 (Kahane-Khintchine inequality - Rademacher sums).
For all 1 p, q < there exists a constant K
p,q
, depending only on p and
q, such that for all nite sequences x
1
, . . . , x
N
E we have
_
E
_
_
_
N

n=1
r
n
x
n
_
_
_
p
_1
p
K
p,q
_
E
_
_
_
N

n=1
r
n
x
n
_
_
_
q
_1
q
.
Proof. By H olders inequality it suces to consider the case p > 1 and q = 1.
Fix vectors x
1
, . . . , x
N
E. Writing X
n
= r
n
x
n
and S
N
=

N
n=1
X
n
, we
may assume that E|S
N
| = 1.
Let j 1 be the unique integer such that 2
j1
< p 2
j
. By successive
applications of Lemma 3.10 for r > 0 we have
P|S
N
| > 2
j
r 4
2
j
1
(P|S
N
| > r)
2
j
.
Chebyshevs inequality gives rP|S
N
| > r E|S
N
| = 1. Hence,
E|S
N
|
p
=
_

0
pt
p1
P|S
N
| > t dt
= 2
jp
_

0
pr
p1
P|S
N
| > 2
j
r dr
2
jp
4
2
j
1
_

0
pr
p1
(P|S
N
| > r)
2
j
dr
(2p)
p
4
2p1
_

0
pr
p1
(P|S
N
| > r)
p
dr
(2p)
p
4
2p1
_

0
pP|S
N
| > r dr
(2p)
p
4
2p1
p.
3.2 The Kahane-Khintchine inequality 37

The best possible constants K

p,q
in this inequality are called the Kahane-
Khintchine constants. Note that K
p,q
= 1 if p q by H olders inequality.
The bound on K
p,1
produced in the above proof is not the best possible: for
instance it is known that K
p,1
= 2
1
1
p
; see the Notes at the end of the lecture.
By an application of the central limit theorem, the Kahane-Khintchine
inequality extends to Gaussian sums:
Theorem 3.12 (Kahane-Khintchine inequality - Gaussian sums). For
all 1 p, q < and all nite sequences x
1
, . . . , x
N
E we have
_
E
_
_
_
N

n=1

n
x
n
_
_
_
p
_1
p
K
p,q
_
E
_
_
_
N

n=1

n
x
n
_
_
_
q
_1
q
,
where K
p,q
is the Kahane-Khintchine constant.
Proof. Fix k = 1, 2, . . . and dene
(k)
n
:=
1

k
j=1
r
nk+j
. For each k we have
_
E
_
_
_
N

n=1

(k)
n
x
n
_
_
_
p
_1
p
=
_
E
_
_
_
N

n=1
k

j=1
r
nk+j
x
n

k
_
_
_
p
_1
p
K
p,q
_
E
_
_
_
N

n=1
k

j=1
r
nk+j
x
n

k
_
_
_
q
_1
q
= K
p,q
_
E
_
_
_
N

n=1

(k)
n
x
n
_
_
_
q
_1
q
.
The proof is completed by passing to the limit k and using the central
limit theorem.
The attentive reader has noticed that we are cheating a bit in the above
proof, as the usual formulation of the central limit theorem only asserts that
lim
k
(
(k)
1
, . . . ,
(k)
N
) = (
1
, . . . ,
N
) in distribution, that is,
lim
k
Ef(
(k)
1
, . . . ,
(k)
N
) = Ef(
1
, . . . ,
N
)
for all bounded continuous functions f : R
N
R. We will show next how,
in the present situation, the convergence of the L
r
-norms (with r = p, q) of
the sums can be deduced from this. The main idea is contained in the next
lemma.
Lemma 3.13. Suppose
0
,
1
, . . . and are R
N
-valued random variables
such that for all bounded continuous functions f : R
N
R we have
lim
k
Ef(
k
) = Ef().
38 3 Sums of independent random variables
Let : R
N
R be a Borel function such that sup
k1
E[(
k
)[ < and
E[()[ < . If g : R
N
R is a continuous function satisfying
[g(t)[ [c(t)[[(t)[, t R
N
,
where c : R
N
R is a bounded function satisfying lim
|t|
[c(t)[ = 0, then
lim
k
Eg(
k
) = Eg().
Proof. Let g
R
:= g 1
{|g|<R}
+R 1
{gR}
R 1
{gR}
denote the truncation
of g at the levels R. By assumption we have
lim
k
Eg
R
(
k
) = Eg
R
(). (3.5)
Furthermore, by dominated convergence,
lim
R
Eg
R
() = Eg(). (3.6)
Fix > 0 and choose R
0
> 0 so large that sup
|t|>R0
[c(t)[ < . Choose
R
1
> 0 so large that [g(t)[ > R
1
implies [t[ > R
0
. Then, for all R R
1
,
sup
k0
E[g(
k
) g
R
(
k
)[ sup
k0
E(1
{|g|>R}
(
k
)[g(
k
)[)
sup
k0
E(1
{|g|>R}
(
k
)[c(t)[[(
k
)[)
sup
k0
E[(
k
)[,
(3.7)
Combined with (3.6) and (3.5), this gives the desired result.
Now we can nish the proof of Theorem 3.12:
Lemma 3.14. With the notations of Theorem 3.12, for all 1 r < and
x
1
, . . . , x
N
E we have
lim
k
E
_
_
_
N

n=1

(k)
n
x
n
_
_
_
r
= E
_
_
_
N

n=1

n
x
n
_
_
_
r
,
where
1
, . . . ,
N
are independent standard Gaussian variables.
Proof. Without loss of generality we may assume that max
1nN
|x
n
| 1.
We x 1 r < and check the condition of Lemma 3.13 for the functions
: R
N
R and g : R
N
R dened by
(t) := exp
_
N

n=1
[t
n
[|x
n
|
_
, g(t) :=
_
_
_
N

n=1
t
n
x
n
_
_
_
r
,
where
k
:= (
(k)
1
, . . . ,
(k)
N
) and := (
1
, . . . ,
N
).
3.3 Exercises 39
If is a symmetric real-valued random variable, then
Eexp([[) = Eexp(1
{<0}
) +Eexp(1
{0}
)
= Eexp(1
{<0}
) +Eexp(1
{0}
) 2Eexp().
Hence, since max
1nN
|x
n
| 1,
E(
k
)
N

n=1
Eexp([
(k)
n
[) 2
N
N

n=1
Eexp(
(k)
n
)
= 2
N
N

n=1
k

j=1
Eexp
_
r
nk+j

k
_
= 2
N
_
1
2
exp
_
1

k
_
+
1
2
exp
_
1

k
_
_
kN
= 2
N
O
_
1 +
1
2k
_
kN
= 2
N
exp(N/2) O(1) as k .
3.3 Exercises
1. Let (X
n
)
N
n=1
be a sequence of independent symmetric E-valued random
variables, and let (r
n
)
N
n=1
be a Rademacher sequence which is indepen-
dent of (X
n
)
N
n=1
. Prove that the sequences (X
n
)
N
n=1
and (r
n
X
n
)
N
n=1
are
identically distributed.
Hint: As in the proof of Theorem 3.2 it may be assumed that (X
n
)
N
n=1
and
(r
n
)
N
n=1
are dened on distinct probability spaces. Use Fubinis theorem
together with the result of Exercise 2.2.
Remark: This technique for introducing Rademacher variables is known
as randomisation. It enables one to apply inequalities for Rademacher
sums in E to sums of independent symmetric random variables in E.
2. (!) Let (r

n
)

n=1
and (r

n
)

n=1
be independent Rademacher sequences on
probability spaces (

, F

, P

) and (

, F

, P

). Prove that on the prod-

uct (, F, P) = (

, F

, P

), the sequence (r

m
r

n
)

m,n=1
consists of Rademacher variables, but as a (doubly indexed) sequence it
fails to be a Rademacher sequence (that is, the random variables r

m
r

n
fail to be independent).
3. (!) We continue with the notations of the previous exercise. Prove that
for 1 p < the following version of the contraction principle holds
for double Rademacher sums in the spaces L
p
(A), where (A, A, ) is a
-nite measure space: there exists a constant C
p
0 such that for all
nite sequences (f
mn
)
N
m,n=1
in L
p
(A) and all scalars (a
mn
)
N
m,n=1
we have
E
_
_
_
N

m,n=1
a
mn
r

m
r

n
f
mn
_
_
_
p
C
p
p
_
max
1m,nN
[a
mn
[
p
_
E
_
_
_
N

m,n=1
r

m
r

n
f
mn
_
_
_
p
.
40 3 Sums of independent random variables
Hint: Proceed in three steps: (i) the result holds for E = R with exponent
2; (ii) the result holds for E = R with exponent p; (iii) the result holds
for E = L
p
(A) with exponent p.
4. Let 1 p 2. A Banach space E is said to have type p if there exists a
constant C
p
0 such that for all nite sequences x
1
, . . . , x
N
in E we have
_
E
_
_
_
N

n=1
r
n
x
n
_
_
_
2
_1
2
C
p
_
N

n=1
|x
n
|
p
_1
p
.
Let 2 q . The space E is said to have cotype q if there exists a
constant C
q
0 such that for all nite sequences x
1
, . . . , x
N
in E we have
_
N

n=1
|x
n
|
q
_1
q
C
q
_
E
_
_
_
N

n=1
r
n
x
n
_
_
_
2
_1
2
.
For q = we make the obvious adjustment in the second denition.
Prove the following assertions:
a) Every Banach space has type 1 and cotype (accordingly, a Banach
space is said to have non-trivial type if it has type p (1, 2] and nite
cotype if it has cotype q [2, )).
b) Every Hilbert space has type 2 and cotype 2.
c) If a Banach space has type p for some p [1, 2], then it has type p

for all p

[1, p]; if a Banach space has cotype q for some q [2, ],

then it has cotype q

for all q

[q, ].
d) Let p [1, 2]. Prove that if E has type p, then the dual space E

has
cotype p

,
1
p
+
1
p

= 1.
Hint: For each x

n
E

choose x
n
E of norm one such that |x

n
|
1
2
[x
n
, x

n
)[. Then use H olders inequality to the eect that for all scalar
sequences (b
n
)
N
n=1
one has
_
N

n=1
[b
n
[
p

_ 1
p

= sup
_
N

n=1
a
n
b
n
:
_
N

n=1
[a
n
[
p
_1
p
1
_
.
Remark: The analogous result for spaces with cotype fails. Indeed, the
reader is invited to check that l
1
has cotype 2 while its dual l

fails to
have non-trivial type.
5. Let p [1, 2]. Prove that a Banach space E has type p if and only if it
has Gaussian type p, that is, if and only if there exists a constant C 0
such that for all nite sequences x
1
, . . . , x
N
in E we have
_
E
_
_
_
N

n=1

n
x
n
_
_
_
2
_1
2
C
_
N

n=1
|x
n
|
p
_1
p
.
3.3 Exercises 41
Hint: One direction follows from Corollary 3.6. For the other direction
use a randomisation argument.
Remark: The corresponding assertion for cotype is also true but much
harder to prove; see the Notes.
Notes. The results of this lecture are classical and can be found in many
textbooks. Our presentation borrows from Albiac and Kalton [1] and Di-
estel, Jarchow, Tonge [35]. Both are excellent starting points for further
reading.
The Kahane contraction principle is due to Kahane [54], who also ex-
tended the classical scalar Khintchine inequality to arbitrary Banach spaces.
It is an open problem to determine the best constants K
p,q
in the Kahane-
Khintchine inequality; a recent result of Latala and Oleszkiewicz [67] as-
serts that the constant K
p,1
= 2
1
1
p
is optimal for 1 p 2.
For a proof of Theorem 3.7 see, e.g., [35]. The proofs of Theorems 3.9 and
3.11 are taken from Albiac and Kalton [1]. The central limit argument in
Lemma 3.14 is adapted from Tomczak-Jaegermann [102].
The contraction principle for double Rademacher sums of Exercise 3 has
been introduced by Pisier [92]. This property, nowadays known under the
rather unsuggestive name property () plays an important role in many
advanced results in Banach space-valued harmonic analysis. It can be shown
that the Rademachers can be replaced by Gaussians without changing the
class of spaces under consideration. Not every Banach space has property ();
a counterexample is the space c
0
.
The notions of type and cotype were developed in the 1970s by Maurey
and Pisier. As we have seen in Exercise 4, Hilbert spaces have type 2 and
cotype 2. A celebrated theorem of Kwapie n [64] asserts that Hilbert spaces
are the only spaces with this property: a Banach space E is isomorphic to a
Hilbert space if and only if E has type 2 and cotype 2. Another class of spaces
of which the type and cotype can be computed are the L
p
-spaces. For the
interested reader we include a proof that the spaces L
p
(A), with 1 p <
and (A, A, ) -nite, have type minp, 2. A similar argument can be used
to prove that they have cotype maxp, 2.
Let f
1
, . . . , f
N
L
p
(A) and put r := minp, 2. Using the Fubini theo-
rem, the scalar Kahane-Khintchine inequality, the type p inequality, H olders
inequality, and the triangle inequality in L
p
r
(A), we obtain
42 3 Sums of independent random variables
_
E
_
_
_
N

n=1
r
n
f
n
_
_
_
p
L
p
(A)
_1
p
=
_
_
A
E

n=1
r
n
f
n
()

p
d()
_1
p
K
p,2
_
_
A
_
E

n=1
r
n
f
n
()

2
_
p
2
d()
_1
p
= K
p,2
_
_
A
_
N

n=1
[f
n
()[
2
_
p
2
d()
_1
p
K
p,2
_
_
A
_
N

n=1
[f
n
()[
r
_
p
r
d()
_1
p
= K
p,2
_
_
_
N

n=1
[f
n
[
r
_
_
_
1
r
L
p
r (A)
K
p,2
_
N

n=1
_
_
_[f
n
[
r
_
_
_
L
p
r (A)
_1
r
= K
p,2
_
N

n=1
|f
n
|
r
L
p
(A)
_1
r
.
An application of the Kahane-Khintchine inequality for L
p
(A) to replace the
L
p
-moment in the left hand side by the L
2
-moment nishes the proof.
It was noted in Exercise 4 that if E has type p, then E

has cotype p

(where
1
p
+
1
p

= 1) and that the analogous duality result for cotype fails. It

is a deep result of Pisier [93] that if E has cotype q [2, ) and non-trivial
type, then E

has type q

,
1
q
+
1
q

= 1.
The fact that a Banach space has cotype q if and only if it has Gaussian
cotype q can be deduced from a deep result of Maurey and Pisier (see
[1, Chapter 11]) which gives a purely geometric characterisation of type and
cotype. For the details we refer to [35].