VOL. 71 1996 NO. 1
AN EXTENSION OF AN INEQUALITY DUE TO STEIN AND LEPINGLE
BY
FERENC W E I S Z (BUDAPEST)
Hardy spaces consisting of adapted function sequences and generated by the q-variation and by the conditional q-variation are considered. Their dual spaces are characterized and an inequality due to Stein and Lepingle is extended.
1. Introduction. It is known that the dual of the martingale Hardy space H
1S2generated by the quadratic variation is BMO
2−and that of the Hardy space H
1s2generated by the conditional quadratic variation is BMO
2(see Garsia [4], Herz [5]). The first result is extended by Dellacherie and Meyer [3] to the space h
S12containing adapted function sequences. The inequality
(1)
X
∞n=0
|E
Fn−1X
n|
q1/qp
≤ C
pX
∞n=0
|X
n|
q1/qp
(1 < p < ∞) was proved by Stein [9] for q = 2 and by Asmar and Montgomery-Smith [1] for 1 ≤ q ≤ ∞, where X
n(n ∈ N) are arbitrary measurable functions.
Using the latter duality result Lepingle [8] verified (1) for p = 1, q = 2 and for adapted functions. The two-parameter analogue of Lepingle’s result can be found in Weisz [10].
Lepingle [7] proved that the dual of the martingale Hardy space H
1sqis BMO
q0and more recently the author [12] verified that the dual of H
1Sqis BMO
q−0, where 1 ≤ q < ∞, 1/q + 1/q
0= 1 and s
q(resp. S
q) denotes the conditional q-variation (resp. the q-variation).
In this paper the Hardy spaces of adapted function sequences are em- bedded isometrically in martingale Hardy spaces and so the dual of h
Spqgenerated by the q-variation and, moreover, the dual of h
spqgenerated by the conditional q-variation are characterized (1 ≤ p, q < ∞). Applying the
1991 Mathematics Subject Classification: Primary 60G42; Secondary 42B30.
This research was supported by the Hungarian Scientific Research Funds (OTKA) No F4189.
[55]
duality result with respect to h
S1qwe extend inequality (1) to 1 = p ≤ q < ∞.
Moreover, if (F
n) is regular then (1) holds also for 0 < p < 1 ≤ q < ∞.
2. Preliminaries and notations. Let (Ω, A, P ) be a probability space and let (F
n, n ∈ N) be a non-decreasing sequence of σ-algebras. For simplicity, we suppose that
σ
[
∞n=0
F
n:=
∞
_
n=0
F
n= A.
The expectation operator and the conditional expectation operator relative to a σ-algebra C are denoted by E and E
C, respectively. We briefly write L
pfor the L
p(Ω, A, P ) space with the norm (or quasinorm) kf k
p:= (E|f |
p)
1/p(0 < p ≤ ∞).
In this paper we consider sequences X = (X
n, n ∈ N) of integrable and adapted (i.e. X
nis F
n-measurable for all n ∈ N) functions. We always suppose that X
0= 0. The q-variation S
q(X) and the conditional q-variation s
q(X) (0 < q < ∞) of X are defined by
S
q(X) := X
∞n=0
|X
n|
q1/qand s
q(X) := X
∞n=0
E
Fn−1|X
n|
q1/q, respectively, while for q = ∞ we let
S
∞(X) := s
∞(X) := sup
n∈N
|X
n|.
Let us introduce the Hardy spaces h
Spqand h
spq(0 < p, q ≤ ∞) consisting of the sequences X = (X
n) of adapted functions for which
kXk
hSqp:= kS
q(X)k
p< ∞ and kXk
hsqp
:= ks
q(X)k
p< ∞,
respectively. Note that h
Spqis a subspace of the well-known space L
p(l
q) that contains sequences ξ = (ξ
n, n ∈ N) of A-measurable functions and is equipped with the norm
kξk
Lp(lq):= h E X
∞n=0
|ξ
n|
qp/qi
1/p.
Now we introduce the corresponding bmo spaces. For 1 ≤ q < ∞, bmo
qand bmo
−qconsist of all sequences X = (X
n) of adapted functions for which kXk
bmoq=
sup
n∈N
E
Fn∞
X
k=n+1
|X
k|
q1/q∞
< ∞ and
kXk
bmo− q=
sup
n∈N
E
Fn∞
X
k=n
|X
k|
q1/q∞
< ∞,
respectively. Furthermore, let bmo
∞= bmo
−∞= h
S∞∞.
Dellacherie and Meyer [3] showed that the dual of h
S12is bmo
−2. It will be proved that the dual of h
S1qis bmo
−q0and the one of h
s1qis bmo
q0(1 ≤ q < ∞, 1/q + 1/q
0= 1).
3. Duality results. We shall embed the spaces h
Spqand h
spqin Hardy spaces of martingales, the duals of which are known. Let
D
k:= σ(r
0, . . . , r
k−1) = σ{[l2
−k, (l + 1)2
−k) : 0 ≤ l < 2
k}
be the dyadic σ-algebras (see Weisz [11]), where r
kis the Rademacher func- tion on [0, 1), i.e.
r
k(x) := 1 if x ∈ [2l/2
k+1, (2l + 1)/2
k+1) (0 ≤ l < 2
k),
−1 if x ∈ [(2l + 1)/2
k+1, (2l + 2)/r2
k+1) (0 ≤ l < 2
k).
Set
A
n:= σ(F
n× D
n) (n ∈ N).
Consider the probability space (Ω ×[0, 1), σ(A×B), P ×λ) and the stochastic basis (A
n, n ∈ N), where B denotes the Borel measurable sets and λ is Lebesgue measure.
We investigate the martingales relative to (A
n), i.e. the sequences f = (d
nf, n ∈ N) of adapted functions relative to (A
n) for which E
An−1d
nf = 0 (n ∈ N). The q-variation and conditional q-variation (0 < q < ∞) of a martingale f is given by
S
q(f ) := X
∞n=0
|d
nf |
q1/qand s
q(f ) := X
∞n=0
E
An−1|d
nf |
q1/q, respectively, and for q = ∞ we let
S
∞(f ) := s
∞(f ) := sup
k∈N
|d
kf |.
The martingale Hardy spaces H
pSqand H
psq(0 < p, q ≤ ∞) containing martingales relative to (A
n) are defined with the help of the norms
kf k
HpSq:=
\
Ω 1
\
0
S
q(f )
pdP dλ
1/pand kf k
Hsqp
:=
\
Ω 1
\
0
s
q(f )
pdP dλ
1/p, respectively. The corresponding dual spaces are equipped with the norms
kf k
BMOq= sup
n∈N
E
An∞
X
k=n+1
|d
kf |
q1/q∞
and
kf k
BMO− q=
sup
n∈N
E
An∞
X
k=n
|d
kf |
q1/q∞
.
Set
BMO
∞= BMO
−∞= H
∞S∞. It is easy to see that the operator
(2) X 7→ f
X:= (X
nr
n−1, n ∈ N)
maps h
Spqin H
pSqisometrically (0 < p ≤ ∞, 1 ≤ q ≤ ∞). Indeed, the function d
nf
X:= X
nr
n−1is A
n-measurable and integrable, of course. On the other hand,
E
An−1(d
nf
X) = E
Fn−1(X
n)E
Dn−1(r
n−1) = 0
because the Rademacher functions are independent. Since S
q(f
X) = S
q(X), our statement is proved. As E
An−1|d
nf
X|
q= E
Fn−1|X
n|
q, we have s
q(f
X) = s
q(X), and so (2) is isometric from h
spqto H
psq(0 < p ≤ ∞, 1 ≤ q ≤ ∞).
Similarly, we can show that (2) is an isometry from bmo
qto BMO
qand from bmo
−qto BMO
−q(1 ≤ q ≤ ∞).
We can prove in the same way as Theorem 14 in Weisz [12] that the dual of h
Spqis h
Sp0q0, where 1 < p, q < ∞ or 1 = q ≤ p < ∞ and 1/p + 1/p
0= 1/q + 1/q
0= 1. The following result, due to Dellacherie and Meyer [3] for q = 2, extends this result to p = 1.
Theorem 1. The dual of h
S1qis bmo
−q0whenever 1 ≤ q < ∞ and 1/q + 1/q
0= 1.
P r o o f. Since the proof is similar to that of Theorem 1 in Weisz [10], we sketch it only. For Y ∈ bmo
−q0consider the functional
l
Y(X) := E
X
∞n=0
X
nY
n(X ∈ h
Sqq).
Notice that h
Sqqis dense in h
S1q. We verified in [12] that the dual of H
1Sqis BMO
−q0with the same assumption on q and q
0as in the theorem. Using this we conclude that
|l
Y(X)| =
\
Ω 1
\
0
∞
X
n=0
d
nf
Xd
nf
YdP dλ
≤ Ckf
Xk
H1Sq
kf
Yk
BMO− q0= CkXk
hSq1
kY k
bmo− q0, which yields that l
Yis bounded on h
S1q.
Conversely, if l is in the dual of h
S1qthen it is also in the dual of h
Sqq. Consequently, there exists Y ∈ h
Sq0q0such that
(3) l(X) = E
X
∞n=0
X
nY
n(X ∈ h
Sqq).
On the other hand, l can be extended preserving its norm onto H
1Sq. There- fore there exists g ∈ BMO
−q0such that
(4) l(X) = l(f
X) =
\
Ω 1
\
0
∞
X
n=0
X
nr
n−1d
ng dP dλ (X ∈ h
Sqq) and
kgk
BMO− q0≤ Cklk.
It follows from (3) and (4) that Y
n(ω) =
1
\
0
r
n−1(x)d
ng(ω, x) dλ(x).
Applying this we obtain kY k
bmo−q0
= sup
n∈N
E
Fn∞
X
k=n
|Y
k|
q01/q0∞
≤ sup
n∈N
E
An∞
X
k=n
|d
kg|
q01/q0∞
= kgk
BMO− q0and the theorem is proved.
The following theorem can be proved similarly.
Theorem 2. The dual of h
spqis h
spq00, where 1 < p ≤ q < ∞ or p ≥ q ≥ 2 and 1/p+1/p
0= 1/q +1/q
0= 1. Moreover , the dual of h
s1qis bmo
q0provided that 1 ≤ q < ∞ and 1/q + 1/q
0= 1.
It is interesting to note that the duals of bmo
q0and bmo
−q0are not h
s1qand h
S1q, respectively. However, a kind of special subspaces of bmo
q0and bmo
−q0can be defined, having duals h
s1qand h
S1q, respectively.
Let vmo
q(resp. vmo
−q) contain all elements X ∈ bmo
q(resp. X ∈ bmo
−q) for which
n→∞
lim
E
Fn∞
X
k=n+1
|X
k|
q1/q∞
= 0
resp. lim
n→∞
E
Fn∞
X
k=n
|X
k|
q1/q∞
= 0 .
With the method used in Weisz [12] one can show that if every σ-algebra
F
nis generated by finitely many atoms then the dual of vmo
q0is h
s1qand
the dual of vmo
−q0is h
S1qwhenever 1 < q
0< ∞ and 1/q + 1/q
0= 1.
4. Inequalities. It follows from the convexity and concavity lemma (see Garsia [4], pp. 113–114) that
(5)
X
∞n=0
E
Fn−1|X
n|
q1/qp
≤ C
pX
∞n=0
|X
n|
q1/qp
(q ≤ p < ∞), and
X
∞n=0
|X
n|
q1/qp
≤ C
pX
∞n=0
E
Fn−1|X
n|
q1/qp
(0 < p ≤ q), where (X
n) is a sequence of A-measurable functions. Note that by H¨ older’s inequality (1) follows from (5) for q ≤ p < ∞.
In case there exists a constant R > 0 such that for all f ∈ L
1one has E
Fn|f | ≤ RE
Fn−1|f | (n ∈ N), the stochastic basis (F
n) is said to be regular . Since the sequence of dyadic σ-algebras is regular, it can easily be seen that whenever (F
n) is regular, so is (A
n). It is proved in [12] that in this case the spaces H
psqand H
pSqare equivalent (0 < p < ∞, 1 ≤ q < ∞). Hence h
spqand h
Spqare also equivalent. This means, amongst other things, that if (F
n) is regular then (5) also holds for 0 < p < ∞ and 1 ≤ q < ∞ when (X
n) is an adapted function sequence. Consequently, under these conditions we obtain (1) for the parameters 0 < p < ∞ and 1 ≤ q < ∞.
If (F
n) is not regular then (1) is not true for p = 1 (see Lepingle [8]).
However, if we take again adapted sequences then it holds for p = 1, too.
The case q = 2 can also be found in Lepingle [8].
Theorem 3. If (X
n, n ∈ N) is a sequence of adapted functions and 1 ≤ q < ∞ then
X
∞n=0
|E
Fn−1X
n|
q1/q1
≤ C
X
∞n=0
|X
n|
q1/q1
.
P r o o f. Since the dual of L
1(l
q) is L
∞(l
q0) (1 ≤ q < ∞, 1/q + 1/q
0= 1) we have
E X
∞n=0
|E
Fn−1X
n|
q1/q= sup
Y ∈L∞(lq0) kY kL∞(lq0 )≤1
E h X
∞n=0
(E
Fn−1X
n)Y
ni .
By Theorem 1,
E
h X
∞n=0
(E
Fn−1X
n)Y
ni
≤ CkXk
hSq1
k(E
Fn−1Y
n, n ∈ N)k
bmo−q0
.
The inequality kY k
L∞(lq0)≤ 1 implies E
Fn∞
X
k=n
|E
Fk−1Y
k|
q0≤ |E
Fn−1Y
n|
q0+ E
Fn∞
X
k=n+1
|Y
k|
q0≤ 2, which shows that
k(E
Fn−1Y
n, n ∈ N)k
bmo−q0