We show that any discrete-time filtration can be embedded into a dyadic filtration, realized via a non-classical measure-preserving map. This embedding makes it possible to reduce certain problems involving vector-valued martingales to the case of Haar martingales. Some applications of this result will also be discussed.
Introduction
A sequence of σ-algebras {Fn}n≥1 in a probability space (X,A,μ) is said to be a filtration if Fn⊂Fn+1. A partition of a probability space (X,A,μ) is a finite or infinite family of pairwise disjoint measurable sets {Ek}⊂A such that μ(Ek)>0, ⋃kEk=X. For partitions A and B we write A≺B, if each element of A is a union of some elements of B. A sequence of partitions {Fn}n≥1 is said to be a partitioning if
Fn≺Fn+1,n≥1,n→∞lima∈Fnmaxμ(a)=0.
In particular, {Fn}n≥1 can be a partitioning on the probability space [0,1) equipped with the Lebesgue measure. We refer to these as an interval partitioning if each Fn consists of intervals of the form [a,b). If all those intervals are dyadic, then we call such a partitioning a dyadic partitioning. Given partitioning {Fn}n≥1 generates a filtration {σ(Fn)}n≥1, where σ(Fn) denotes the σ-algebra generated by the elements of the partition Fn. A key result of the paper is an embedding theorem that allows a general filtration to be reflected into a dyadic one. Such a embedding is realized by a non-standard measure-preserving map between two probability spaces. Such a measure-preserving map is not a pointwise correspondence; rather, it acts on the measurable sets of the probability spaces. Namely,
Research was supported by the Higher Education and Science Committee of RA, in the frames of the research project 21AG-1A045.
Definition 1.1. Let (X,A,μ) and (Y,B,ν) be probability spaces. We say a function Θ:A→B is measure-preserving set map (MPS-map), if
1)μ(E)=ν(Θ(E)),
2)μ(E∪F)=ν(Θ(E)∪Θ(F)),(1)
for any E,F∈A. If the probability spaces X and Y coincide, then Θ is called a measure-preserving transformation (MPS-transformation).
We will examine in § 2 the properties of such a measure-preserving map and explore its relationship to the classical notion of a measure-preserving transformation in the next section. In particular, we show that any MPS-map admits a natural extension to a map between the spaces of measurable functions on X and Y that preserves joint cumulative distribution functions (see Definition 4.2). Specifically, for a simple function f=∑jcj1aj, its Θ-image is defined by Θ(f)=∑jcj1Θ(aj). For an arbitrary measurable function f on X, the function Θ(f) is defined as the almost everywhere limit of Θ(sn), where sn is a sequence of simple functions converging to f almost everywhere (see Lemmas 2.1 and 2.2 for details).
Introduce some preliminary notations. Define the distance of two partitions A and B as
dist(A,B)=a∈A∖B∑μ(a)=b∈B∖A∑μ(b).
Note that if δ is the distance between two partitions, then they share common elements whose union has measure 1−δ. In particular, if the distance is zero, the two partitions coincide. For a partition F we denote by σ(F) the σ-algebra generated by the elements of F. Then the conditional expectation of a function f∈L1(X) with respect to the σ-algebra σ(F) will be denoted by
EFf(x)=a∈F∑μ(a)1∫af⋅1a(x).
where 1a is the indicator function of the set a. Given partitioning {Fn}n≥1 denote by F∞ the σ-algebra generated by all the elements of partitions Fn, i.e. F∞=σ(⋃n≥1Fn).
Definition 1.2. Let {An}n≥1 and {Bn}n≥1 be partitionings on probability spaces (X,A,μ) and (Y,B,ν) respectively. The partitioning {An}n≥1 is said to be embedded in {Bn}n≥1 if there exists a MPS-map Θ:A∞→B∞ such that
Θ(EAn(1E))=EBn(1Θ(E))for all E∈j≥1⋃Aj and n≥1.(2)
The embedding of the partitioning {An}n≥1 into {Bn}n≥1 will be denoted by the symbol {An}n≥1↪{Bn}n≥1.
The following is the main result of the paper.
Theorem 1.1. Let εn>0, n≥1 and suppose {An}n≥1 and {Bn}n≥1 be partitionings on a probability space X. Then there exist sequence of integers kn and a partitioning {Dn}n≥1 such that
Dn⊂k≥1⋃Bk,(3)
dist(Dn,Bkn)<εn,(4)
{An}n≥1↪{Dn}n≥1.
In § 4 we will apply Theorem 1.1 to provide a framework for interpreting an arbitrary martingale (or martingale difference sequence) in terms of the Haar system (Theorem 4.1). In § 5 this equivalence principle is used to establish a logarithmic Lp-norm estimate for the maximal function of N different transformations of a martingale. Finally, in § 6 we deduce from Theorem 1.1 its multiple version.
Measure-preserving transformations
Let (X,A,μ) be a probability space. By the classical definition, a measure-preserving transformation (MP-transformation) is a function ϕ:X→X such that ϕ−1(E)∈A for every E∈A and μ(ϕ−1(E))=μ(E). Observe that any MP-transformation ϕ:X→X generates a MPS-transformation Θ:A→A defined by Θ(E)=ϕ−1(E). Nevertheless, not all MPS-transformations can be represented by a MP-transformation. An example of such an MPS-transformation can be seen in the proof of Theorem 1.1.
Lemma 2.1. If Θ is a MPS-map from a probability space (X,A,μ) to (Y,B,ν), then for any measurable sets E,F∈A, Ek∈A, k=1,2,…,
This implies (7) for two sets and so for a finite collection of measurable sets. Then passing to a limit we will get (7) for any countable collection of sets. Similarly (1) implies (6).
Throughout the paper, B denotes a Banach space. Let (X,A,μ) be a probability space. We denote by LB0(X) the space of B-valued measurable functions on X. For an arbitrary family of measurable set F⊂A, we denote by LB0(X,F) the space of σ(F)-measurable B-valued functions on X. Similarly, the notation LBp(X,F), 1≤p≤∞, denotes the Bochner-Lebesgue spaces of σ(F)-measurable B-valued functions on X.
Let Θ be a MPS-map from a probability space (X,A,μ) to (Y,B,ν). Then Θ induces a map that takes any indicator function 1E, E∈A, to 1Θ(E). The following lemma gives a suitable extension of Θ to entire space LB0(X) of B-valued measurable functions.
Lemma 2.2. If Θ is a MPS-map from a probability space (X,A,μ) to a probability space (Y,B,ν), then it can be uniquely extended as an operator from LB0(X) into LB0(Y) such that every sequence {fk}k≥1⊂LB0(X) is probabilistically equivalent to its image sequence {Θ(fk)}k≥1.
Proof. Let s(x) be a simple function, i.e.
s(x)=i=1∑mαi1Ei(x),
where Ei∈A, i=1,2,…,m, are pairwise disjoint measurable sets and αi∈B. Then we define
Θs(x)=i=1∑mαi1Θ(Ei)(x).
For an arbitrary f∈LB0(X), one can find a sequence of simple functions sn(x), such that
f(x)=n→∞limsn(x)a.e.(8)
Using Lemma 2.1 and a standard argument one can observe that the sequences sn(x) and Θsn(x) are probabilistically equivalent. Consequently, Θsn(x) converges almost everywhere and the limit doesn’t depend on the particular choice of simple functions sn in (8). We therefore define this limit as Θf(x). It is also clear that any sequence {fk}k≥1⊂LR0(X) is probabilistically equivalent to its image sequence {Θ(fk)}k≥1.
Lemma 2.3. Let {Fn}n≥1 be a filtration on a probability space (X,F∞,μ) and suppose that Θ:⋃nFn→F∞ is a map, which restriction to each σ-algebra Fn is a MPS-map from Fn into F∞. Then Θ can be extended as a MPS-transformation on F∞.
Proof. It is well known in measure theory that any set E∈F∞ can be well-approximated by sets from ⋃nFn. In particular, for any set E∈F∞ there exists a sequence Fn∈⋃nFn such that 1Fn→1E a.e.. Since the sequence Θ(Fn) is probabilistically equivalent to the sequence Fn, we can say that 1Θ(Fn) converges a.e. to an indicator function 1F, F∈F∞, and F doesn’t depend on the choice of the sequence Fn. Then we define Θ(E)=F. In view of Lemma 2.1, one can check that with this definition Θ becomes a MPS-transformation on F∞.
Proof of Theorem 1.1
Lemma 3.1. Let {An}n≥1 and {Bn}n≥1 be partitionings on probability spaces (X,A,μ) and (Y,B,ν) respectively such that {An}n≥1↪{Bn}n≥1 and suppose that this embedding is realized by the MPS-transformation Θ. Then,
1)f∈LR0(X,An)⇒Θ(f)∈LR0(X,Bn),(9)
2)Θ(EAn(f))=EBn(Θ(f))for all f∈LR1(X,A∞) and n≥1.(10)
Proof. Observe that equation (2) in the definition of embedding implies Θ(σ(An))⊂σ(Bn), which is the same as (9) when f=1E is an indicator function of a measurable set E. We also note that equation (2) is a special case of (10) when f=1E. The proofs of both relations (9) and (10) are based on a standard approximation argument using simple functions.
The following lemma follows directly from the definition of the embedding property of partitionings.
Lemma 3.2. If partitionings {An}n≥1, {Bn}n≥1 and {Cn}n≥1 satisfy
{An}n≥1↪{Bn}n≥1,{Bn}n≥1↪{Cn}n≥1
realized by MPS-maps ϕ and ψ respectively, then it holds the embedding
{An}n≥1↪{Cn}n≥1
realized by ϕ∘ψ.
Let a⊂[0,1) be an interval, and ak be a (finite or infinite) partition of a into pairwise disjoint subintervals,
a=k⋃ak,ak∩aj=∅ as k=j.(11)
To this partition, we associate a Markov partition transformation ϕ:a→a, which maps each interval ak on a linearly. Namely,
ϕ(x)=k∑lak→a(x)⋅1ak(x),(12)
where lak→a denotes the increasing linear function that maps ak onto a. Clearly, ϕ is a MP-transformation of the interval a.
Remark 1. Observe that from equation (2) it follows that Θ(An)≺Bn. Thus to check equation (2) we will need to show that
∣A∣∣Θ(E)∩Θ(A)∣=∣B∣∣Θ(E)∩B∣(13)
for every choice of sets
A∈An,B∈Bn,B⊂Θ(A),E∈Ak.(14)
Lemma 3.3. Let {An}n≥1 and {Cn}n≥1 be interval partitionings on [0,1). Then for any ε>0 and for any fixed integer m≥1 there exist an interval partitioning {Bn}n≥1 such that
Bm⊂n≥1⋃Cn,(15)
dist(Bm,Ckm)≤ε for an integer km.(16)
Besides, we have embedding
{An}n≥1↪{Bn}n≥1(17)
realized by a MPS-transformation θm, satisfying
θm(A)=A∈Bk for all A∈Ak,k<m.(18)
Remark 2. One can check that from (18) it follows that θm(Ak)=Ak≺Bk for all k<m.
Proof. For every interval a∈Am we associate an arbitrarily chosen partition {aj} consisting of the elements of ⋃n≥1Cn. Moreover, for sufficiently large km, most of the elements {aj} can be chosen from Ckm. More precisely,
aj∈Ckm∑∣aj∣>∣a∣(1−ε)(19)
and the integer km is common for all elements a∈Am. Let ϕm,a be the MP-transformation of the interval a, corresponding to its partition {aj} (see (12)). Denote by ϕm, the MP-transformation of [0,1), such that ϕm(x)=ϕm,a(x), x∈a, for all a∈Am. Define the MPS-transformation θm=ϕm−1. Clearly, if {aj} is the defined partition of an element a∈Am, then for any measurable set E⊂a we have
∣θm(E)∩aj∣=∣a∣∣aj∣⋅∣E∣.(20)
Set
Bn=Anif n<m,
while for n≥m, Bn is defined as follows. For each b∈An there exists a unique a∈Am such that b⊆a. Moreover, θm(b) can be written as a union of the component-intervals bj=aj∩θm(b) (see (11)). We define the elements of Bn to be all such component-intervals bj, where b∈An. Hence the σ-algebras Bn for all n≥1 are defined and one can check that those form an interval partitioning. If n=m, then in the above notations we will have b=a and bj=aj. Since {aj}⊂⋃n≥1Cn, we get (15). Moreover, condition (19) implies (16). Observe that for any a∈Am we have θm(a)=a∈Am. Since each b∈An with 1≤n<m is a union of some elements of Am, we conclude that θm(E)=E for any E∈An, n<m. This implies (18). To check embedding (17), we need to show equation (13) (with Θ=θm) for every choice of the sets (14). Consider the following cases of integers k and n in (14):
Case 1, n<m. Then from (18) (see also Remark 2) it follows that θm(A)=A=B and (13) holds trivially.
Case 2, n≥m, k<m. Again by (18) we have θm(E)=E. If A∩E=∅, then θm(A)∩θm(E)=∅. Since B⊂θm(A) ((14)), thus we get B∩θm(E)=∅ and (13) is trivial. If A∩E=∅, then E⊃A so θm(E)⊃θm(A)⊃B and we get (13).
Case 3, n≥m, k≥m: There exists a unique a∈Am such that A⊂a. Since θm(a)=a (see the definition of θm), we get B⊂θm(A)⊂a. If E∩a=∅ so θm(E)∩a=∅, then equation (13) is trivial. Hence we can assume that E⊂a. Let {aj} be the dyadic partition of a considered in the definition of θm. By the definition of Bn we have B=aj0∩θm(A) for some j0. Therefore,
Proof of Theorem 1.1. Without loss of generality we can suppose that our partitioning {An}n≥1 is an interval partitioning on [0,1). Applying Lemma 3.3 consecutively for m=1,2,…, we will get interval partitionings {An(m)}n≥1, m=1,2,… on [0,1), satisfying
Am(m)⊂n≥1⋃Bn,m≥1,(21)
dist(Am(m),Bkm)<εm,m≥1,(22)
{An(m−1)}n≥1↪{An(m)}n≥1,m≥1,(23)
θm(A)=A∈An(m) for all A∈An(m−1),1≤n<m,(24)
where θm is the MPS-transformation realizing the embedding (23) and An(0)=An. Define Dn=An(n), n≥1. From (24) (see also Remark 2) it follows that
Thus we obtain (25). Then for any set E∈σ(An) we can define
Θ(E)=k→∞limΘk(E)=Θn(E).(26)
Hence Θ is defined for any E∈⋃n≥1σ(An) and it gives a MPS-transformation induced on each σ(An). Applying Lemma 2.3 we can extend Θ up to a MPS-transformation on A∞. Let us prove that Θ realizes the embedding {An}n≥1↪{Dn}n≥1. Using Lemma 3.2, we can say that Θm realizes the embedding
{An}n≥1↪{An(m)}n≥1.
In view of Remark 1 it means that
∣A∣∣Θm(E)∩Θm(A)∣=∣B∣∣Θm(E)∩B∣.(27)
whenever
A∈An,B∈An(m),B⊂Θm(A),E∈Ak.(28)
If m≥n, then from (24) it follows that An(m)=An(n)=Dn and we can replace An(m) by Dn in (28). Then letting m→∞ and taking into account (26), we can also replace Θm by Θ in (27) and (28). In such a new setting relation (27) gives {An}n≥1↪{Dn}n≥1, completing the proof of theorem. □
An equivalence principle with Haar martingales
Let B be a Banach space. A martingale adapted to a partitioning {Fn}n≥1 is a sequence of functions fn:X→B such that fn is σ(Fn)-measurable and for the conditional expectation we have EFn−1(fn)=fn−1, n≥2. Given martingale {fn}n≥1 defines a martingale difference sequence dn=fn−fn−1 (here f0=0) so that fn=∑k=1ndk, n≥1. Thus, the defining property of a martingale difference sequence is EFn−1(dn)=0, n≥2. The simplest example of a martingale difference sequence is given by the Haar system hn in the form cnhn, where each coefficient cn∈B. Recall the definition of Haar system on [0,1). We have h1≡1. For an integer n≥2, with the representation n=2k+j, 1≤j≤2k, k=0,1,…, denote
Δn=[2kj−1,2kj)
and let Δn+ and Δn− be the left and the right halves of the interval Δn. Then n'th Haar function for n≥2 is defined as
hn=2k/2(1Δn+−1Δn−).
Definition 4.1. A sequence of pairwise disjoint sets of positive integers Gn⊂N, n=1,2,…, is said to be locally increasing if for any integers j and k with
j∈Gn and k∈Gn+1,
we have either Δk⊂Δj or Δk∩Δj=∅. If maxGn<minGn+1, then we say that the sequence Gn is increasing.
Observe that if Gn⊂N is locally increasing sequence then for any coefficients bj∈B the functions
fn=j∈Gn∑bjhj(29)
form a martingale difference sequence. We also refer to such sequences of Haar sums as locally increasing.
Definition 4.2. Let B be a Banach space and that (X,A,μ) and (Y,B,ν) are probability spaces. Sequences of measurable functions fk:X→B and gk:Y→B, k=1,2,…, are said to be probabilistically equivalent if they share the same joint cumulative distribution function, i.e. we have
μ(k=1⋂nfk−1(Ek))=ν(k=1⋂ngk−1(Ek))
for every choice of Borel sets Ek⊂B, k=1,2,…,n. Similarly we say that the sequences of measurable sets {Ak}⊂A and {Bk}⊂B are probabilistically equivalent if the corresponding sequences of indicator functions {1Ak} and {1Bk} are probabilistically equivalent.
Theorem 4.1. Let B be a separable Banach space and εn→0. Then any B-valued atomic martingale difference sequence {dn}n≥1 is probabilistically equivalent to a sequence of B-valued Haar sums (4.1), where the sequence
Gn is locally increasing and bj∈B. Moreover, there are subsets Gn′⊂Gn such that Gn′ is increasing and
⎩⎨⎧j∈Gn∑bjhj=j∈Gn′∑bjhj⎭⎬⎫<εn.(30)
For f∈LB1[0,1] we consider the corresponding vector-valued Haar series
n=1∑∞(f,hn)hn(x),where (f,hn)=∫01f⋅hn.(31)
The following standard properties of (31) can be derived by applying the same argument as in the classical case, where B=R.
Proposition 4.1. Let B be an arbitrary Banach space. Then for any f∈CB[0,1] the series (31) uniformly converges to f.
Proposition 4.2. If B is a separable Banach space, then for any f∈LBp[0,1], 1≤p<∞, the series (31) converges in Lp-norm and almost everywhere.
Definition 4.3. Let F be partition of a probability space (X,F,μ) and denote
m(F)=a∈Fmaxμ(a)
We say that an element a∈F is maximal, if μ(a)=m(F). A partition A is said to be ε-uniform, ε>0, if
a∈F,μ(a)<m(F)∑μ(a)<ε.
Lemma 4.1. Let {Fn}n≥1 be a dyadic partitioning on [0,1). Then any B-valued martingale difference {fn}n≥1 adapted to {Fn}n≥1 admits a Haar representation
fn=j∈Gn∑(fn,hj)hj,(32)
where the sets Gn⊂N, n=1,2,…, form a locally increasing sequence of positive integers. Moreover, if in addition Fn is εn-uniform, εn>0, then there exists an increasing sequence Gn′ such that Gn′⊂Gn and
suppj∈Gn∖Gn′∑(fn,hj)hj<εn.(33)
Proof. Denote
G1={k∈N:Δk⊆I for any I∈F1},
Gn={k∈N:Δk⊆I for any I∈Fn and Δk⊆I for some I∈Fn−1},n≥2.
One can check that the sequence {Gn}n≥1 is locally increasing. Then, in view of Proposition 4.2, it remains to show that
(fn,hk)=0 for all k∈/Gn.
Indeed, if k∈/Gn, then one of the following conditions hold:
Δk⊆I for some I∈Fn,
Δk⊆I for every I∈Fn−1.
Suppose we have 1), i.e. there is I∈Fn such that Δk⊆I. Since fn is constant on I, and therefore on Δk, it follows that (fn,hk)=0. If a Δk satisfies 2), then we get that both Δk+ and Δk− are unions of some elements of Fn−1, since each interval I∈Fn−1 itself is dyadic. On the other hand, by the martingale-difference property of fn we have (fn,1I)=0 for any I∈Fn−1. This implies (fn,1Δk±)=0, and therefore, (fn,hk)=0. If Fn is εn-uniform, then we have
a∈Fn,μ(a)<m(Fn)∑μ(a)<εn.(34)
Denote
Gn′={k∈Gn:there exists a maximal element I∈Fn such that Δk⊋I}.(35)
Observe that
Δk⊆a∈Fn,μ(a)<m(Fn)⋃aif k∈Gn∖Gn′.
Combining this with (34), we obtain (33). Now consider integers k∈Gn−1′ and j∈Gn′. By the definitions of Gn−1′ ((35)) we have ∣Δk∣>m(Fn−1). Since j∈Gn′⊂Gn, by the definition of Gn there exists an atom I∈Fn−1 such that Δj⊆I and so ∣Δj∣≤m(Fn−1). Thus we get ∣Δj∣<∣Δk∣, which clearly implies j>k. Hence {Gn′} is an increasing sequence of integer sets. □
Proof of Theorem 4.1. Let {dn}n≥1 be a B-valued martingale difference sequence adapted to a partitioning {Fn}n≥1. Applying Theorem 1.1, we find a dyadic partitioning {Dn}n≥1 such that each Dn is εn-uniform and {Fn}n≥1↪{Dn}n≥1. Let Θ be the MPS-map realizing this embedding. Using the property Θ(Fn)⊂Dn (see Remark 1) and applying Lemma 2.2, we conclude that the sequence {fn=Θ(dn)}n≥1 is probabilistically equivalent to {dn}n≥1 and is adapted to the partitioning {Dn}n≥1. On the other hand according to Lemma 3.1 we have
EDn−1(Θ(dn))=Θ(EFn−1(dn)).
Since {fn}n≥1 is a martingale difference, we have Θ(EFn−1(dn))=0. Therefore,
EDn−1(fn)=EDn−1(Θ(dn))=0,
i.e. {fn}n≥1 is also a martingale difference adapted to the dyadic partitioning {Dn}n≥1. Then applying Lemma 4.1, we get the representation (32), where {Gn} is a locally increasing sequence of integer sets. Using also (33), we get (30). This completes the proof of theorem. □
Martingale transforms
Vector-valued Haar martingales represent the simplest models of general vector-valued martingales. Owing to Theorem 4.1, the verification of various martingale inequalities in the vector-valued setting can be reduced to the dyadic Haar case, from which the corresponding results for arbitrary vector-valued martingales follow. Important examples of such inequalities are those arising in the study of martingale transforms. It is of interest Banach spaces B for which classical theorems on real-valued martingales can be extended for B-valued martingales. A Banach space B is said to have a UMD-property if there exists a positive constant cp such that the inequality
∥ε1d1+⋯+εndn∥p≤cp∥d1+⋯+dn∥p,1<p<∞,(36)
holds for all B-valued martingale difference sequences {dj}j≥1, for all so called predictable sequences εj=±1 and all n≥1. A geometric characterization of UMD-spaces was given by Burkholder [1]. It is well known that R is a UMD-space. For Haar scalar-martingales this was proved by Paley [9]. The inequality for the general scalar-martingales was established by Burkholder [2]. Moreover, it was found in [2] the best constant cp that can be in (36), where {dk} are scalar-valued martingales. In view of Theorem 4.1, problems of this type may, without loss of generality, be restricted to Haar martingales, taking dk=ckhk with ck∈B. In particular, if one seeks the best constant in (36) for a given Banach space B, then the same constant is obtained by considering only B-valued Haar martingale difference sequences instead of general ones. Indeed, by approximation, the optimal constant cp remains unchanged when passing from arbitrary martingale differences to atomic ones. Then, applying Theorem 4.1, for atomic martingale differences in (36) we find a sequence
fk=j∈Gk∑bjhj,k=1,2,…,n,(37)
which is probabilistically equivalent to the d1,…,dn. Then the inequality (36) can be equivalently written as
∥ε1f1+⋯+εnfn∥p≤cp∥f1+⋯+fn∥p.
On the other hand, both f1+⋯+fn and ε1f1+⋯+εnfn become Haar sums after substitutions (37), where the latter is simply the martingale transformation of the former. Thus we obtain the following result:
Theorem 5.1. Let B be a UMD-space. If (36) holds for all B-valued Haar martingales and all choices εk=±1, then it holds for all B-valued martingale difference sequences, with the same constant cp.
The inequalities in (36) may also be considered with εk taking values in [0,1] or in {0,1}. The statement of Theorem 5.1 remains valid for such choices of predictable εk. In particular, the case of εk∈{0,1} was studied by [4], who determined the optimal constant in (36) for scalar-valued martingale difference sequences.
Now let {Fn}n≥1 be an arbitrary filtration in a probability space (X,F∞,μ) and suppose that f={fn}n≥1 is a scalar-valued martingale adapted to {Fn}n≥1 with assumption that f0=0. For 1≤p≤∞ denote
∥f∥p=n≥1sup∥fn∥p,f∗=n≥1sup∣fn∣.
The martingale Lp-space consists of all martingales with finite norm ∥f∥p<∞. It is well known that for 1<p<∞, the martingale Lp-space is isomorphic to Lp(X,F∞). Moreover, if f={fn}n≥1 belongs to a martingale Lp-space, then fn converges both almost everywhere and in the Lp-norm.
In this case, by convention, when writing a martingale f={fn}n≥1 we identify f with the limit function of fn. If Θ={θn}n≥1 is a sequence of real valued functions with ∣θn∣≤1, then we say Θ is predictable if θn is Fn−1-measurable. The sequence g={gn}n≥1 defined by
gn=k=1∑nθkdk,dk=fk−fk−1,
is also a martingale and it is called the martingale transform of f by Θ. Let denote by TΘ the operator assigning any martingale f to its Θ-transformation g. From the inequality (36) it follows that TΘ is a bounded operator on Lp, 1<p<∞. Moreover, we have the inequality
∥TΘ∗(f)∥p≤cp∥f∥p,
which is the combination of Burkholder’s [1] and Doob’s [5] inequalities. For a finite family of predictables Θk={θn(k)}n≥1, k=1,2,…,N, consider the maximal operator
T(f)=1≤k≤NmaxTΘk∗(f).(38)
Applying Theorem 4.1, we establish the following inequality, extending the p=2 case proved in [6]. Moreover, in [6], martingale transforms are considered in the case where the predictable sequences consist of constant functions.
Theorem 5.2. Let {Fn}n≥1 be a filtration in a probability space (X,F∞,μ). Then operator (38) satisfies
∥T(f)∥p≤cplog(N+1)⋅∥f∥p,1<p<∞,
for any f∈Lp(X,F∞).
Proof. By Theorem 4.1 it is enough to prove the theorem for Haar martingales. Let ∑k=1∞akhk be the Fourier-Haar series of a function f∈Lp(0,1). Recall the definitions of the maximal function and the square function
It is well known the boundedness of both operators on Lp, 1<p<∞ (see for example [8], ch. 3, Theorems 4 and 9). A key point in the proof of Theorem 1.1 is the following good-λ inequality due to Chang-Wilson-Wolff (see [3], Corollary 3.1):
which implies the theorem for Haar martingales, and therefore for arbitrary martingales as well.
Remark on multiple martingales
We consider the probability space (Qd,Ld,md) of Lebesgue measure on d-dimensional unit cube Qd=[0,1)d. Let {An(k),n∈N}, k=1,2,…,d, be partitionings of [0,1). The Cartesian product
determines a multiple partitioning of the unite cube Qd. For a given sequence of integers N={νi∈N:ν1<ν2<…} consider the multiple sequence of partitioning
An=Aνn1,…,νnd,n=(n1,…,nd)∈Nd.(42)
Partitioning (42) is said to be a sub-partitioning of (41). The following theorem is an improved version of the analogous result from [7]. It is obtained by a direct application of Theorem 1.1.
Theorem 6.1. Let {An,n∈Nd} and {Bn,n∈Nd} be arbitrary multiple interval partitionings of the unit cube Qd. Then for any sequence of numbers εk>0, k=1,2,…, there exist a multiple partitioning Dn, a MPS-transformation Θ, and a sub-partitioning {Bn} of {Bn}, such that
Dn⊂k≥1⋃Bk,(43)
dist(Dn,Bn)<εm,1≤i≤dminni≥m,(44)
(Θ∘EAn)f(x)=(EDn∘Θ)f(x),n∈Nd,(45)
for all f∈L1(Qd).
If T is a bounded linear operator on L1[0,1), then we denote by Tk, 1≤k≤d, the operator
Lemma 6.1 ([6]). If T is a bounded linear operator on L1[0,1), then Tk determines a bounded operator on L1(Qd).
Lemma 6.2 ([6]). If T and P are bounded linear operators on L1[0,1), then
Pn∘Tm=Tm∘Pn,n=m,1≤n,m≤d.
Proof of Theorem 6.1. Applying Theorem 1.1 for each k=1,2,…,d we find a partitioning {Dn(k)}n≥1 of the interval [0,1) such that
Dn(k)⊂m≥1⋃Bm(k),(46)
dist(Dm(k),Bνm(k))<εm/d,(47)
{An(k)}n≥1⟶{Dn(k)}n≥1.(48)
Clearly, (46)⇒(43), and (47)⇒(44). Let θk be the MPS-transformation of [0,1) realizing (48). By Lemma 2.2 it can be extended as a linear operator on L1[0,1). Then by Lemma 3.1 we can write
(θk∘EFn(k))(f)=(EFn(k)∘θk)(f), for all f∈L1[0,1).
Applying θk along kth coordinate, by Lemma 6.1 we can consider θk as a linear bounded operator on L1(Qd). Thus Θ=θ1∘θ2∘⋯∘θd is also a linear bounded operator on L1(Qd). We also have
[1]D. L. Burkholder, A geometrical characterization of Banach spaces in which martingale difference sequences are unconditional, Ann. Probab. 9 (1981), no. 6, 997–1011. MR632972DOI
[2]———, Boundary value problems and sharp inequalities for martingale transforms, Ann. Probab. 12 (1984), no. 3, 647–702. MR744226DOI
[3]S.-Y. A. Chang, J. M. Wilson, and T. H. Wolff, Some weighted norm inequalities concerning the Schrödinger operators, Comment. Math. Helv. 60 (1985), no. 2, 217–246, DOI 10.1007/BF02567411. MR800004DOI
[4]K. P. Choi, A sharp inequality for martingale transforms and the unconditional basis constant of a monotone basis in L^p(0,1), Trans. Amer. Math. Soc. 330 (1992), no. 2, 509–529, DOI 10.2307/2153920. MR1034661DOI
[5]J. L. Doob, Stochastic processes, Wiley Classics Library, John Wiley & Sons, Inc., New York, 1990. Reprint of the 1953 original; A Wiley-Interscience Publication. MR1038526
[6]G. A. Karagulyan, On systems of non-overlapping Haar polynomials, Ark. Mat. 58 (2020), no. 1, 121–131, DOI 10.4310/arkiv.2020.v58.n1.a8. MR4094642DOI
[7]———, Equivalence of martingales and some related problems, Izv. Nats. Akad. Nauk Armenii Mat. 48 (2013), no. 2, 15–38, DOI 10.3103/S1068362313020015 (Russian, with English and Russian summaries); English transl., J. Contemp. Math. Anal. 48 (2013), no. 2, 51–65. MR3112686
[8]B. S. Kashin and A. A. Saakyan, Orthogonal series, Translations of Mathematical Monographs, vol. 75, American Mathematical Society, Providence, RI, 1989. Translated from the Russian by Ralph P. Boas; Translation edited by Ben Silver. MR1007141
[9]R. E. A. C. Paley, A Remarkable Series of Orthogonal Functions (I), Proc. London Math. Soc. (2) 34 (1932), no. 4, 241–264, DOI 10.1112/plms/s2-34.1.241. MR1576148DOI