betaFeedback

WorksMathematics

An equivalence principle for martingales

Abstract

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\sigma-algebras {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} in a probability space (X,A,μ)(X,\mathcal{A},\mu) is said to be a filtration if Fn⊂Fn+1\mathcal{F}_n \subset\mathcal{F}_{n+1}. A partition of a probability space (X,A,μ)(X,\mathcal{A},\mu) is a finite or infinite family of pairwise disjoint measurable sets {Ek}⊂A\{E_k\}\subset\mathcal{A} such that μ(Ek)>0\mu(E_k)>0, ⋃kEk=X\bigcup_k E_k=X. For partitions A\mathcal{A} and B\mathcal{B} we write A≺B\mathcal{A}\prec\mathcal{B}, if each element of A\mathcal{A} is a union of some elements of B\mathcal{B}. A sequence of partitions {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} is said to be a partitioning if

Fn≺Fn+1,n≥1,lim⁡n→∞max⁡a∈Fnμ(a)=0.\mathcal{F}_n\prec\mathcal{F}_{n+1},\qquad n\geq1,\qquad\lim_{n\to\infty}\max_{a\in\mathcal{F}_n}\mu(a)=0.

In particular, {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} can be a partitioning on the probability space [0,1)[0,1) equipped with the Lebesgue measure. We refer to these as an interval partitioning if each Fn\mathcal{F}_n consists of intervals of the form [a,b)[a,b). If all those intervals are dyadic, then we call such a partitioning a dyadic partitioning. Given partitioning {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} generates a filtration {σ(Fn)}n≥1\{\sigma(\mathcal{F}_n)\}_{n\geq1}, where σ(Fn)\sigma(\mathcal{F}_n) denotes the σ\sigma-algebra generated by the elements of the partition Fn\mathcal{F}_n. 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,

2010 Mathematics Subject Classification. 60G42, 42C05, 42C10, 28D05.

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)(X,\mathcal{A},\mu) and (Y,B,ν)(Y,\mathcal{B},\nu) be probability spaces. We say a function Θ:A→B\Theta:\mathcal{A}\to\mathcal{B} is measure-preserving set map (MPS-map), if

1) μ(E)=ν(Θ(E)),1)\ \mu(E)=\nu(\Theta(E)),
2) μ(E∪F)=ν(Θ(E)∪Θ(F)),(1)2)\ \mu(E\cup F)=\nu(\Theta(E)\cup\Theta(F)), \tag*{(1)}

for any E,F∈AE,F\in\mathcal{A}. If the probability spaces XX and YY coincide, then Θ\Theta 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 XX and YY that preserves joint cumulative distribution functions (see Definition 4.2). Specifically, for a simple function f=∑jcj1ajf=\sum_j c_j1_{a_j}, its Θ\Theta-image is defined by Θ(f)=∑jcj1Θ(aj)\Theta(f)=\sum_j c_j1_{\Theta(a_j)}. For an arbitrary measurable function ff on XX, the function Θ(f)\Theta(f) is defined as the almost everywhere limit of Θ(sn)\Theta(s_n), where sns_n is a sequence of simple functions converging to ff almost everywhere (see Lemmas 2.1 and 2.2 for details).

Introduce some preliminary notations. Define the distance of two partitions A\mathcal{A} and B\mathcal{B} as

dist⁡(A,B)=∑a∈A∖Bμ(a)=∑b∈B∖Aμ(b).\operatorname{dist}(\mathcal{A},\mathcal{B})=\sum_{a\in\mathcal{A}\setminus\mathcal{B}}\mu(a)=\sum_{b\in\mathcal{B}\setminus\mathcal{A}}\mu(b).

Note that if δ\delta is the distance between two partitions, then they share common elements whose union has measure 1−δ1-\delta. In particular, if the distance is zero, the two partitions coincide. For a partition F\mathcal{F} we denote by σ(F)\sigma(\mathcal{F}) the σ\sigma-algebra generated by the elements of F\mathcal{F}. Then the conditional expectation of a function f∈L1(X)f\in L^1(X) with respect to the σ\sigma-algebra σ(F)\sigma(\mathcal{F}) will be denoted by

EFf(x)=∑a∈F1μ(a)∫af⋅1a(x).\mathbb{E}^{\mathcal{F}}f(x)=\sum_{a\in\mathcal{F}}\frac{1}{\mu(a)}\int_a f\cdot1_a(x).

where 1a1_a is the indicator function of the set aa. Given partitioning {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} denote by F∞\mathcal{F}_\infty the σ\sigma-algebra generated by all the elements of partitions Fn\mathcal{F}_n, i.e. F∞=σ(⋃n≥1Fn)\mathcal{F}_\infty=\sigma\left(\bigcup_{n\geq1}\mathcal{F}_n\right).

Definition 1.2. Let {An}n≥1\{\mathcal{A}_n\}_{n\geq1} and {Bn}n≥1\{\mathcal{B}_n\}_{n\geq1} be partitionings on probability spaces (X,A,μ)(X,\mathcal{A},\mu) and (Y,B,ν)(Y,\mathcal{B},\nu) respectively. The partitioning {An}n≥1\{\mathcal{A}_n\}_{n\geq1} is said to be embedded in {Bn}n≥1\{\mathcal{B}_n\}_{n\geq1} if there exists a MPS-map Θ:A∞→B∞\Theta:\mathcal{A}_\infty\to\mathcal{B}_\infty such that

Θ(EAn(1E))=EBn(1Θ(E))for all E∈⋃j≥1Aj and n≥1.(2)\Theta\left(\mathbb{E}^{\mathcal{A}_n}(1_E)\right)=\mathbb{E}^{\mathcal{B}_n}\left(1_{\Theta(E)}\right)\quad\text{for all }E\in\bigcup_{j\geq1}\mathcal{A}_j\text{ and }n\geq1. \tag*{(2)}

The embedding of the partitioning {An}n≥1\{\mathcal{A}_n\}_{n\geq1} into {Bn}n≥1\{\mathcal{B}_n\}_{n\geq1} will be denoted by the symbol {An}n≥1↪{Bn}n≥1\{\mathcal{A}_n\}_{n\geq1} \hookrightarrow\{\mathcal{B}_n\}_{n\geq1}.

The following is the main result of the paper.

Theorem 1.1. Let εn>0\varepsilon_n > 0, n≥1n \geq1 and suppose {An}n≥1\{\mathcal{A}_n\}_{n\geq1} and {Bn}n≥1\{\mathcal{B}_n\}_{n\geq1} be partitionings on a probability space XX. Then there exist sequence of integers knk_n and a partitioning {Dn}n≥1\{\mathcal{D}_n\}_{n\geq1} such that

Dn⊂⋃k≥1Bk,(3)\mathcal{D}_n \subset\bigcup_{k\geq1}\mathcal{B}_k, \tag*{(3)}
dist⁡(Dn,Bkn)<εn,(4)\operatorname{dist}(\mathcal{D}_n,\mathcal{B}_{k_n}) < \varepsilon_n, \tag*{(4)}

{An}n≥1↪{Dn}n≥1\{\mathcal{A}_n\}_{n\geq1} \hookrightarrow\{\mathcal{D}_n\}_{n\geq1}.

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 LpL^p-norm estimate for the maximal function of NN different transformations of a martingale. Finally, in § 6 we deduce from Theorem 1.1 its multiple version.

Measure-preserving transformations

Let (X,A,μ)(X,\mathcal{A},\mu) be a probability space. By the classical definition, a measure-preserving transformation (MP-transformation) is a function ϕ:X→X\phi: X \to X such that ϕ−1(E)∈A\phi^{-1}(E) \in\mathcal{A} for every E∈AE \in\mathcal{A} and μ(ϕ−1(E))=μ(E)\mu(\phi^{-1}(E)) = \mu(E). Observe that any MP-transformation ϕ:X→X\phi: X \to X generates a MPS-transformation Θ:A→A\Theta: \mathcal{A} \to\mathcal{A} defined by Θ(E)=ϕ−1(E)\Theta(E) = \phi^{-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\Theta is a MPS-map from a probability space (X,A,μ)(X,\mathcal{A},\mu) to (Y,B,ν)(Y,\mathcal{B},\nu), then for any measurable sets E,F∈AE,F\in\mathcal{A}, Ek∈AE_k\in\mathcal{A}, k=1,2,…k=1,2,\ldots,

μ(E∖F)=ν(Θ(E)∖Θ(F)),(5)\mu(E\setminus F)=\nu\bigl(\Theta(E)\setminus\Theta(F)\bigr), \tag*{(5)}
μ(⋃kEk)=ν(⋃kΘ(Ek)),(6)\mu\left(\bigcup_k E_k\right)=\nu\left(\bigcup_k\Theta(E_k)\right), \tag*{(6)}
μ(⋂kEk)=ν(⋂kΘ(Ek)).(7)\mu\left(\bigcap_k E_k\right)=\nu\left(\bigcap_k\Theta(E_k)\right). \tag*{(7)}

Proof. Using the definition of MPS-map, we obtain

μ(F)+μ(E∖F)=μ(F∪E)=ν(Θ(F)∪Θ(E))=ν(Θ(F))+ν(Θ(E)∖Θ(F))=μ(F)+ν(Θ(E)∖Θ(F)).\mu(F) + \mu(E \setminus F) = \mu(F \cup E) = \nu(\Theta(F) \cup\Theta(E)) = \nu(\Theta(F)) + \nu(\Theta(E) \setminus\Theta(F)) = \mu(F) + \nu(\Theta(E) \setminus\Theta(F)).

Thus (5) follows. Similarly,

μ(E)+μ(F)−μ(E∩F)=μ(E∪F)=ν(Θ(E)∪Θ(F))=ν(Θ(E))+ν(Θ(F))−ν(Θ(E)∩Θ(F))=μ(E)+μ(F)−ν(Θ(E)∩Θ(F)).\mu(E) + \mu(F) - \mu(E \cap F) = \mu(E \cup F) = \nu(\Theta(E) \cup\Theta(F)) = \nu(\Theta(E)) + \nu(\Theta(F)) - \nu(\Theta(E) \cap\Theta(F)) = \mu(E) + \mu(F) - \nu(\Theta(E) \cap\Theta(F)).

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\mathbb{B} denotes a Banach space. Let (X,A,μ)(X,\mathcal{A},\mu) be a probability space. We denote by LB0(X)L^0_{\mathbb{B}}(X) the space of B\mathbb{B}-valued measurable functions on XX. For an arbitrary family of measurable set F⊂A\mathcal{F} \subset\mathcal{A}, we denote by LB0(X,F)L^0_{\mathbb{B}}(X,\mathcal{F}) the space of σ(F)\sigma(\mathcal{F})-measurable B\mathbb{B}-valued functions on XX. Similarly, the notation LBp(X,F)L^p_{\mathbb{B}}(X,\mathcal{F}), 1≤p≤∞1 \le p \le\infty, denotes the Bochner-Lebesgue spaces of σ(F)\sigma(\mathcal{F})-measurable B\mathbb{B}-valued functions on XX.

Let Θ\Theta be a MPS-map from a probability space (X,A,μ)(X,\mathcal{A},\mu) to (Y,B,ν)(Y,\mathcal{B},\nu). Then Θ\Theta induces a map that takes any indicator function 1E\mathbf{1}_E, E∈AE \in\mathcal{A}, to 1Θ(E)\mathbf{1}_{\Theta(E)}. The following lemma gives a suitable extension of Θ\Theta to entire space LB0(X)L^0_{\mathbb{B}}(X) of B\mathbb{B}-valued measurable functions.

Lemma 2.2. If Θ\Theta is a MPS-map from a probability space (X,A,μ)(X,\mathcal{A},\mu) to a probability space (Y,B,ν)(Y,\mathcal{B},\nu), then it can be uniquely extended as an operator from LB0(X)L^0_{\mathbb{B}}(X) into LB0(Y)L^0_{\mathbb{B}}(Y) such that every sequence {fk}k≥1⊂LB0(X)\{f_k\}_{k\geq1} \subset L^0_{\mathbb{B}}(X) is probabilistically equivalent to its image sequence {Θ(fk)}k≥1\{\Theta(f_k)\}_{k\geq1}.

Proof. Let s(x)s(x) be a simple function, i.e.

s(x)=∑i=1mαi1Ei(x),s(x) = \sum_{i=1}^{m} \alpha_i \mathbf{1}_{E_i}(x),

where Ei∈AE_i \in\mathcal{A}, i=1,2,…,mi = 1,2,\ldots,m, are pairwise disjoint measurable sets and αi∈B\alpha_i \in\mathbb{B}. Then we define

Θs(x)=∑i=1mαi1Θ(Ei)(x).\Theta s(x) = \sum_{i=1}^{m} \alpha_i \mathbf{1}_{\Theta(E_i)}(x).

For an arbitrary f∈LB0(X)f \in L^0_{\mathbb{B}}(X), one can find a sequence of simple functions sn(x)s_n(x), such that

f(x)=lim⁡n→∞sn(x)a.e.(8)f(x) = \lim_{n \to\infty} s_n(x) \quad\text{a.e.} \tag*{(8)}

Using Lemma 2.1 and a standard argument one can observe that the sequences sn(x)s_n(x) and Θsn(x)\Theta s_n(x) are probabilistically equivalent. Consequently, Θsn(x)\Theta s_n(x) converges almost everywhere and the limit doesn’t depend on the particular choice of simple functions sns_n in (8). We therefore define this limit as Θf(x)\Theta f(x). It is also clear that any sequence {fk}k≥1⊂LR0(X)\{f_k\}_{k\geq1} \subset L_{\mathbb{R}}^0(X) is probabilistically equivalent to its image sequence {Θ(fk)}k≥1\{\Theta(f_k)\}_{k\geq1}.

Lemma 2.3. Let {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} be a filtration on a probability space (X,F∞,μ)(X,\mathcal{F}_\infty,\mu) and suppose that Θ:⋃nFn→F∞\Theta:\bigcup_n\mathcal{F}_n\to\mathcal{F}_\infty is a map, which restriction to each σ\sigma-algebra Fn\mathcal{F}_n is a MPS-map from Fn\mathcal{F}_n into F∞\mathcal{F}_\infty. Then Θ\Theta can be extended as a MPS-transformation on F∞\mathcal{F}_\infty.

Proof. It is well known in measure theory that any set E∈F∞E\in\mathcal{F}_\infty can be well-approximated by sets from ⋃nFn\bigcup_n\mathcal{F}_n. In particular, for any set E∈F∞E\in\mathcal{F}_\infty there exists a sequence Fn∈⋃nFnF_n\in\bigcup_n\mathcal{F}_n such that 1Fn→1E\mathbf{1}_{F_n}\to\mathbf{1}_E a.e.. Since the sequence Θ(Fn)\Theta(F_n) is probabilistically equivalent to the sequence FnF_n, we can say that 1Θ(Fn)\mathbf{1}_{\Theta(F_n)} converges a.e. to an indicator function 1F\mathbf{1}_F, F∈F∞F\in\mathcal{F}_\infty, and FF doesn’t depend on the choice of the sequence FnF_n. Then we define Θ(E)=F\Theta(E)=F. In view of Lemma 2.1, one can check that with this definition Θ\Theta becomes a MPS-transformation on F∞\mathcal{F}_\infty.

Proof of Theorem 1.1

Lemma 3.1. Let {An}n≥1\{\mathcal{A}_n\}_{n\geq1} and {Bn}n≥1\{\mathcal{B}_n\}_{n\geq1} be partitionings on probability spaces (X,A,μ)(X,\mathcal{A},\mu) and (Y,B,ν)(Y,\mathcal{B},\nu) respectively such that {An}n≥1↪{Bn}n≥1\{\mathcal{A}_n\}_{n\geq1}\hookrightarrow\{\mathcal{B}_n\}_{n\geq1} and suppose that this embedding is realized by the MPS-transformation Θ\Theta. Then,

1) f∈LR0(X,An)⇒Θ(f)∈LR0(X,Bn),(9)1)\ f\in L_{\mathbb{R}}^0(X,\mathcal{A}_n)\Rightarrow\Theta(f)\in L_{\mathbb{R}}^0(X,\mathcal{B}_n), \tag*{(9)}
2) Θ(EAn(f))=EBn(Θ(f))for all f∈LR1(X,A∞) and n≥1.(10)2)\ \Theta\left(\mathbb{E}^{\mathcal{A}_n}(f)\right)=\mathbb{E}^{\mathcal{B}_n}\left(\Theta(f)\right)\quad\text{for all }f\in L_{\mathbb{R}}^1(X,\mathcal{A}_\infty)\text{ and }n\geq1. \tag*{(10)}

Proof. Observe that equation (2) in the definition of embedding implies Θ(σ(An))⊂σ(Bn)\Theta(\sigma(\mathcal{A}_n))\subset\sigma(\mathcal{B}_n), which is the same as (9) when f=1Ef=\mathbf{1}_E is an indicator function of a measurable set EE. We also note that equation (2) is a special case of (10) when f=1Ef=\mathbf{1}_E. 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\{\mathcal{A}_n\}_{n\geq1}, {Bn}n≥1\{\mathcal{B}_n\}_{n\geq1} and {Cn}n≥1\{\mathcal{C}_n\}_{n\geq1} satisfy

{An}n≥1↪{Bn}n≥1,{Bn}n≥1↪{Cn}n≥1\{\mathcal{A}_n\}_{n\geq1}\hookrightarrow\{\mathcal{B}_n\}_{n\geq1},\qquad\{\mathcal{B}_n\}_{n\geq1}\hookrightarrow\{\mathcal{C}_n\}_{n\geq1}

realized by MPS-maps ϕ\phi and ψ\psi respectively, then it holds the embedding

{An}n≥1↪{Cn}n≥1\{\mathcal{A}_n\}_{n\geq1}\hookrightarrow\{\mathcal{C}_n\}_{n\geq1}

realized by ϕ∘ψ\phi\circ\psi.

Let a⊂[0,1)a \subset[0,1) be an interval, and aka_k be a (finite or infinite) partition of aa into pairwise disjoint subintervals,

a=⋃kak,ak∩aj=∅ as k≠j.(11)a = \bigcup_k a_k,\qquad a_k \cap a_j = \varnothing\text{ as } k \ne j. \tag*{(11)}

To this partition, we associate a Markov partition transformation ϕ:a→a\phi: a \to a, which maps each interval aka_k on aa linearly. Namely,

ϕ(x)=∑klak→a(x)⋅1ak(x),(12)\phi(x) = \sum_k l_{a_k\to a}(x) \cdot\mathbf{1}_{a_k}(x), \tag*{(12)}

where lak→al_{a_k\to a} denotes the increasing linear function that maps aka_k onto aa. Clearly, ϕ\phi is a MP-transformation of the interval aa.

Remark 1. Observe that from equation (2) it follows that Θ(An)≺Bn\Theta(\mathcal{A}_n) \prec\mathcal{B}_n. Thus to check equation (2) we will need to show that

∣Θ(E)∩Θ(A)∣∣A∣=∣Θ(E)∩B∣∣B∣(13)\frac{|\Theta(E) \cap\Theta(A)|}{|A|} = \frac{|\Theta(E) \cap B|}{|B|} \tag*{(13)}

for every choice of sets

A∈An,B∈Bn,B⊂Θ(A),E∈Ak.(14)A \in\mathcal{A}_n,\qquad B \in\mathcal{B}_n,\qquad B \subset\Theta(A),\qquad E \in\mathcal{A}_k. \tag*{(14)}

Lemma 3.3. Let {An}n≥1\{\mathcal{A}_n\}_{n\ge1} and {Cn}n≥1\{\mathcal{C}_n\}_{n\ge1} be interval partitionings on [0,1)[0,1). Then for any ε>0\varepsilon> 0 and for any fixed integer m≥1m \ge1 there exist an interval partitioning {Bn}n≥1\{\mathcal{B}_n\}_{n\ge1} such that

Bm⊂⋃n≥1Cn,(15)\mathcal{B}_m \subset\bigcup_{n\ge1}\mathcal{C}_n, \tag*{(15)}
dist⁡(Bm,Ckm)≤ε for an integer km.(16)\operatorname{dist}(\mathcal{B}_m,\mathcal{C}_{k_m}) \le\varepsilon\text{ for an integer } k_m. \tag*{(16)}

Besides, we have embedding

{An}n≥1↪{Bn}n≥1(17)\{\mathcal{A}_n\}_{n\ge1} \hookrightarrow\{\mathcal{B}_n\}_{n\ge1} \tag*{(17)}

realized by a MPS-transformation θm\theta_m, satisfying

θm(A)=A∈Bk for all A∈Ak, k<m.(18)\theta_m(A) = A \in\mathcal{B}_k \text{ for all } A \in\mathcal{A}_k,\ k < m. \tag*{(18)}

Remark 2. One can check that from (18) it follows that θm(Ak)=Ak≺Bk\theta_m(\mathcal{A}_k) = \mathcal{A}_k \prec\mathcal{B}_k for all k<mk < m.

Proof. For every interval a∈Ama \in\mathcal{A}_m we associate an arbitrarily chosen partition {aj}\{a_j\} consisting of the elements of ⋃n≥1Cn\bigcup_{n\ge1}\mathcal{C}_n. Moreover, for sufficiently large kmk_m, most of the elements {aj}\{a_j\} can be chosen from Ckm\mathcal{C}_{k_m}. More precisely,

∑aj∈Ckm∣aj∣>∣a∣(1−ε)(19)\sum_{a_j\in\mathcal{C}_{k_m}} |a_j| > |a|(1-\varepsilon) \tag*{(19)}

and the integer kmk_m is common for all elements a∈Ama \in\mathcal{A}_m. Let ϕm,a\phi_{m,a} be the MP-transformation of the interval aa, corresponding to its partition {aj}\{a_j\} (see (12)). Denote by ϕm\phi_m, the MP-transformation of [0,1)[0,1), such that ϕm(x)=ϕm,a(x)\phi_m(x)=\phi_{m,a}(x), x∈ax \in a, for all a∈Ama \in\mathcal{A}_m. Define the MPS-transformation θm=ϕm−1\theta_m=\phi_m^{-1}. Clearly, if {aj}\{a_j\} is the defined partition of an element a∈Ama \in\mathcal{A}_m, then for any measurable set E⊂aE \subset a we have

∣θm(E)∩aj∣=∣aj∣∣a∣⋅∣E∣.(20)|\theta_m(E)\cap a_j|=\frac{|a_j|}{|a|}\cdot|E|. \tag*{(20)}

Set

Bn=An if n<m,\mathcal{B}_n=\mathcal{A}_n\ \text{if }n<m,

while for n≥mn \ge m, Bn\mathcal{B}_n is defined as follows. For each b∈Anb \in\mathcal{A}_n there exists a unique a∈Ama \in\mathcal{A}_m such that b⊆ab \subseteq a. Moreover, θm(b)\theta_m(b) can be written as a union of the component-intervals bj=aj∩θm(b)b_j=a_j\cap\theta_m(b) (see (11)). We define the elements of Bn\mathcal{B}_n to be all such component-intervals bjb_j, where b∈Anb \in\mathcal{A}_n. Hence the σ\sigma-algebras Bn\mathcal{B}_n for all n≥1n \ge1 are defined and one can check that those form an interval partitioning. If n=mn=m, then in the above notations we will have b=ab=a and bj=ajb_j=a_j. Since {aj}⊂⋃n≥1Cn\{a_j\}\subset\bigcup_{n\ge1}\mathcal{C}_n, we get (15). Moreover, condition (19) implies (16). Observe that for any a∈Ama \in\mathcal{A}_m we have θm(a)=a∈Am\theta_m(a)=a\in\mathcal{A}_m. Since each b∈Anb\in\mathcal{A}_n with 1≤n<m1\le n<m is a union of some elements of Am\mathcal{A}_m, we conclude that θm(E)=E\theta_m(E)=E for any E∈AnE\in\mathcal{A}_n, n<mn<m. This implies (18). To check embedding (17), we need to show equation (13) (with Θ=θm\Theta=\theta_m) for every choice of the sets (14). Consider the following cases of integers kk and nn in (14):

Case 1, n<mn<m. Then from (18) (see also Remark 2) it follows that θm(A)=A=B\theta_m(A)=A=B and (13) holds trivially.

Case 2, n≥mn\ge m, k<mk<m. Again by (18) we have θm(E)=E\theta_m(E)=E. If A∩E=∅A\cap E=\varnothing, then θm(A)∩θm(E)=∅\theta_m(A)\cap\theta_m(E)=\varnothing. Since B⊂θm(A)B\subset\theta_m(A) ((14)), thus we get B∩θm(E)=∅B\cap\theta_m(E)=\varnothing and (13) is trivial. If A∩E≠∅A\cap E\ne\varnothing, then E⊃AE\supset A so θm(E)⊃θm(A)⊃B\theta_m(E)\supset\theta_m(A)\supset B and we get (13).

Case 3, n≥mn\ge m, k≥mk\ge m: There exists a unique a∈Ama\in\mathcal{A}_m such that A⊂aA\subset a. Since θm(a)=a\theta_m(a)=a (see the definition of θm\theta_m), we get B⊂θm(A)⊂aB\subset\theta_m(A)\subset a. If E∩a=∅E\cap a=\varnothing so θm(E)∩a=∅\theta_m(E)\cap a=\varnothing, then equation (13) is trivial. Hence we can assume that E⊂aE\subset a. Let {aj}\{a_j\} be the dyadic partition of aa considered in the definition of θm\theta_m. By the definition of Bn\mathcal{B}_n we have B=aj0∩θm(A)B=a_{j_0}\cap\theta_m(A) for some j0j_0. Therefore,

B∩θm(E)=aj0∩θm(A)∩θm(E)=aj0∩θm(A∩E).B\cap\theta_m(E)=a_{j_0}\cap\theta_m(A)\cap\theta_m(E)=a_{j_0}\cap\theta_m(A\cap E).

On the other hand, by (20) we have

∣B∣=∣aj0∩θm(A)∣=∣aj0∣∣a∣⋅∣A∣,|B| = |a_{j_0} \cap\theta_m(A)| = \frac{|a_{j_0}|}{|a|} \cdot|A|,
∣B∩θm(E)∣=∣aj0∩θm(A∩E)∣=∣aj0∣∣a∣⋅∣A∩E∣.|B \cap\theta_m(E)| = |a_{j_0} \cap\theta_m(A \cap E)| = \frac{|a_{j_0}|}{|a|} \cdot|A \cap E|.

Thus we obtain

∣θm(E)∩B∣∣B∣=∣aj0∣∣a∣⋅∣B∣⋅∣A∩E∣=∣A∩E∣∣A∣=∣θm(A)∩θm(E)∣∣A∣\frac{|\theta_m(E) \cap B|}{|B|} = \frac{|a_{j_0}|}{|a| \cdot|B|} \cdot|A \cap E| = \frac{|A \cap E|}{|A|} = \frac{|\theta_m(A) \cap\theta_m(E)|}{|A|}

and so (13). This completes the proof of lemma. □

Proof of Theorem 1.1. Without loss of generality we can suppose that our partitioning {An}n≥1\{\mathcal{A}_n\}_{n \ge1} is an interval partitioning on [0,1)[0,1). Applying Lemma 3.3 consecutively for m=1,2,…m=1,2,\ldots, we will get interval partitionings {An(m)}n≥1\{\mathcal{A}^{(m)}_n\}_{n \ge1}, m=1,2,…m=1,2,\ldots on [0,1)[0,1), satisfying

Am(m)⊂⋃n≥1Bn,m≥1,(21)\mathcal{A}^{(m)}_m \subset\bigcup_{n \ge1}\mathcal{B}_n,\quad m \ge1, \tag*{(21)}
dist⁡(Am(m),Bkm)<εm,m≥1,(22)\operatorname{dist}(\mathcal{A}^{(m)}_m,\mathcal{B}_{k_m}) < \varepsilon_m,\quad m \ge1, \tag*{(22)}
{An(m−1)}n≥1↪{An(m)}n≥1,m≥1,(23)\{\mathcal{A}^{(m-1)}_n\}_{n \ge1} \hookrightarrow\{\mathcal{A}^{(m)}_n\}_{n \ge1},\quad m \ge1, \tag*{(23)}
θm(A)=A∈An(m) for all A∈An(m−1),1≤n<m,(24)\theta_m(A)=A\in\mathcal{A}^{(m)}_n\text{ for all }A\in\mathcal{A}^{(m-1)}_n,\quad1\le n<m, \tag*{(24)}

where θm\theta_m is the MPS-transformation realizing the embedding (23) and An(0)=An\mathcal{A}^{(0)}_n=\mathcal{A}_n. Define Dn=An(n)\mathcal{D}_n=\mathcal{A}^{(n)}_n, n≥1n\ge1. From (24) (see also Remark 2) it follows that

Dm−1=Am−1(m−1)=θm(Am−1(m−1))=Am−1(m)≺Am(m)=Dm.\mathcal{D}_{m-1}=\mathcal{A}^{(m-1)}_{m-1}=\theta_m(\mathcal{A}^{(m-1)}_{m-1})=\mathcal{A}^{(m)}_{m-1}\prec\mathcal{A}^{(m)}_m=\mathcal{D}_m.

Thus {Dm}m≥1\{\mathcal{D}_m\}_{m\ge1} provides a partitioning, and we obtain (21)⇒\Rightarrow(1.3), (22)⇒\Rightarrow(1.4). Consider the MPS-transformation

Θn=(θn∘⋯∘θ2∘θ1).\Theta_n=(\theta_n\circ\cdots\circ\theta_2\circ\theta_1).

Observe that for any set E∈σ(An)E\in\sigma(\mathcal{A}_n) we have

Dn=An(n)∋Θn(E)=Θn+1(E)=Θn+2(E)=⋯(25)\mathcal{D}_n=\mathcal{A}^{(n)}_n\ni\Theta_n(E)=\Theta_{n+1}(E)=\Theta_{n+2}(E)=\cdots \tag*{(25)}

Indeed, using (23) and Remark 1, we can write

E⊂An=An(0)⇒θ1(E)∈An(1)⇒(θ2∘θ1)(E)∈An(2)⇒⋮(θn∘…∘θ2∘θ1)(E)∈An(n)⇒Θn(E)∈An(n)=Dn.\begin{aligned} E &\subset\mathcal{A}_n=\mathcal{A}_n^{(0)} \Rightarrow\\ \theta_1(E)&\in\mathcal{A}_n^{(1)} \Rightarrow\\ (\theta_2\circ\theta_1)(E)&\in\mathcal{A}_n^{(2)} \Rightarrow\\ &\vdots\\ (\theta_n\circ\ldots\circ\theta_2\circ\theta_1)(E)&\in\mathcal{A}_n^{(n)} \Rightarrow\\ \Theta_n(E)&\in\mathcal{A}_n^{(n)}=\mathcal{D}_n. \end{aligned}

Then from (24) it follows that

Θn(E)=Θn+1(E)=Θn+2(E)=….\Theta_n(E)=\Theta_{n+1}(E)=\Theta_{n+2}(E)=\ldots.

Thus we obtain (25). Then for any set E∈σ(An)E\in\sigma(\mathcal{A}_n) we can define

Θ(E)=lim⁡k→∞Θk(E)=Θn(E).(26)\Theta(E)=\lim_{k\to\infty}\Theta_k(E)=\Theta_n(E). \tag*{(26)}

Hence Θ\Theta is defined for any E∈⋃n≥1σ(An)E\in\bigcup_{n\geq1}\sigma(\mathcal{A}_n) and it gives a MPS-transformation induced on each σ(An)\sigma(\mathcal{A}_n). Applying Lemma 2.3 we can extend Θ\Theta up to a MPS-transformation on A∞\mathcal{A}_\infty. Let us prove that Θ\Theta realizes the embedding {An}n≥1↪{Dn}n≥1\{\mathcal{A}_n\}_{n\geq1}\hookrightarrow\{\mathcal{D}_n\}_{n\geq1}. Using Lemma 3.2, we can say that Θm\Theta_m realizes the embedding

{An}n≥1↪{An(m)}n≥1.\{\mathcal{A}_n\}_{n\geq1}\hookrightarrow\{\mathcal{A}_n^{(m)}\}_{n\geq1}.

In view of Remark 1 it means that

∣Θm(E)∩Θm(A)∣∣A∣=∣Θm(E)∩B∣∣B∣.(27)\frac{|\Theta_m(E)\cap\Theta_m(A)|}{|A|} = \frac{|\Theta_m(E)\cap B|}{|B|}. \tag*{(27)}

whenever

A∈An,B∈An(m),B⊂Θm(A),E∈Ak.(28)A\in\mathcal{A}_n,\qquad B\in\mathcal{A}_n^{(m)},\qquad B\subset\Theta_m(A),\qquad E\in\mathcal{A}_k. \tag*{(28)}

If m≥nm\geq n, then from (24) it follows that An(m)=An(n)=Dn\mathcal{A}_n^{(m)}=\mathcal{A}_n^{(n)}=\mathcal{D}_n and we can replace An(m)\mathcal{A}_n^{(m)} by Dn\mathcal{D}_n in (28). Then letting m→∞m\to\infty and taking into account (26), we can also replace Θm\Theta_m by Θ\Theta in (27) and (28). In such a new setting relation (27) gives {An}n≥1↪{Dn}n≥1\{\mathcal{A}_n\}_{n\geq1}\hookrightarrow\{\mathcal{D}_n\}_{n\geq1}, completing the proof of theorem. □

An equivalence principle with Haar martingales

Let B\mathbb{B} be a Banach space. A martingale adapted to a partitioning {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} is a sequence of functions fn:X→Bf_n:X\to\mathbb{B} such that fnf_n is σ(Fn)\sigma(\mathcal{F}_n)-measurable and for the conditional expectation we have EFn−1(fn)=fn−1\mathbb{E}_{\mathcal{F}_{n-1}}(f_n)=f_{n-1}, n≥2n\geq2. Given martingale {fn}n≥1\{f_n\}_{n\geq1} defines a martingale difference sequence dn=fn−fn−1d_n=f_n-f_{n-1} (here f0=0f_0=0) so that fn=∑k=1ndkf_n=\sum_{k=1}^{n}d_k, n≥1n \ge1. Thus, the defining property of a martingale difference sequence is EFn−1(dn)=0\mathbb{E}^{\mathcal{F}_{n-1}}(d_n)=0, n≥2n \ge2. The simplest example of a martingale difference sequence is given by the Haar system hnh_n in the form cnhnc_nh_n, where each coefficient cn∈Bc_n \in\mathbb{B}. Recall the definition of Haar system on [0,1)[0,1). We have h1≡1h_1 \equiv1. For an integer n≥2n \ge2, with the representation n=2k+jn=2^k+j, 1≤j≤2k1 \le j \le2^k, k=0,1,…k=0,1,\ldots, denote

Δn=[j−12k,j2k)\Delta_n=\left[\frac{j-1}{2^k},\frac{j}{2^k}\right)

and let Δn+\Delta_n^+ and Δn−\Delta_n^- be the left and the right halves of the interval Δn\Delta_n. Then nn'th Haar function for n≥2n \ge2 is defined as

hn=2k/2(1Δn+−1Δn−).h_n=2^{k/2}\left(1_{\Delta_n^+}-1_{\Delta_n^-}\right).

Definition 4.1. A sequence of pairwise disjoint sets of positive integers Gn⊂NG_n \subset\mathbb{N}, n=1,2,…n=1,2,\ldots, is said to be locally increasing if for any integers jj and kk with

j∈Gn and k∈Gn+1,j \in G_n \text{ and } k \in G_{n+1},

we have either Δk⊂Δj\Delta_k \subset\Delta_j or Δk∩Δj=∅\Delta_k \cap\Delta_j=\varnothing. If max⁡Gn<min⁡Gn+1\max G_n < \min G_{n+1}, then we say that the sequence GnG_n is increasing.

Observe that if Gn⊂NG_n \subset\mathbb{N} is locally increasing sequence then for any coefficients bj∈Bb_j \in\mathbb{B} the functions

fn=∑j∈Gnbjhj(29)f_n=\sum_{j\in G_n}b_jh_j \tag*{(29)}

form a martingale difference sequence. We also refer to such sequences of Haar sums as locally increasing.

Definition 4.2. Let B\mathbb{B} be a Banach space and that (X,A,μ)(X,\mathcal{A},\mu) and (Y,B,ν)(Y,\mathcal{B},\nu) are probability spaces. Sequences of measurable functions fk:X→Bf_k:X\to\mathbb{B} and gk:Y→Bg_k:Y\to\mathbb{B}, k=1,2,…k=1,2,\ldots, are said to be probabilistically equivalent if they share the same joint cumulative distribution function, i.e. we have

μ(⋂k=1nfk−1(Ek))=ν(⋂k=1ngk−1(Ek))\mu\left(\bigcap_{k=1}^{n}f_k^{-1}(E_k)\right)=\nu\left(\bigcap_{k=1}^{n}g_k^{-1}(E_k)\right)

for every choice of Borel sets Ek⊂BE_k \subset\mathbb{B}, k=1,2,…,nk=1,2,\ldots,n. Similarly we say that the sequences of measurable sets {Ak}⊂A\{A_k\}\subset\mathcal{A} and {Bk}⊂B\{B_k\}\subset\mathcal{B} are probabilistically equivalent if the corresponding sequences of indicator functions {1Ak}\{1_{A_k}\} and {1Bk}\{1_{B_k}\} are probabilistically equivalent.

Theorem 4.1. Let B\mathbb{B} be a separable Banach space and εn→0\varepsilon_n\to0. Then any B\mathbb{B}-valued atomic martingale difference sequence {dn}n≥1\{d_n\}_{n\ge1} is probabilistically equivalent to a sequence of B\mathbb{B}-valued Haar sums (4.1), where the sequence

GnG_n is locally increasing and bj∈Bb_j \in\mathbb{B}. Moreover, there are subsets Gn′⊂GnG'_n \subset G_n such that Gn′G'_n is increasing and

∣{∑j∈Gnbjhj≠∑j∈Gn′bjhj}∣<εn.(30)\left|\left\{\sum_{j\in G_n} b_jh_j \ne\sum_{j\in G'_n} b_jh_j\right\}\right| < \varepsilon_n. \tag*{(30)}

For f∈LB1[0,1]f \in L^1_{\mathbb{B}}[0,1] we consider the corresponding vector-valued Haar series

∑n=1∞(f,hn)hn(x),where (f,hn)=∫01f⋅hn.(31)\sum_{n=1}^{\infty}(f,h_n)h_n(x),\quad\text{where } (f,h_n)=\int_0^1 f\cdot h_n. \tag*{(31)}

The following standard properties of (31) can be derived by applying the same argument as in the classical case, where B=R\mathbb{B}=\mathbb{R}.

Proposition 4.1. Let B\mathbb{B} be an arbitrary Banach space. Then for any f∈CB[0,1]f \in C_{\mathbb{B}}[0,1] the series (31) uniformly converges to ff.

Proposition 4.2. If B\mathbb{B} is a separable Banach space, then for any f∈LBp[0,1]f \in L^p_{\mathbb{B}}[0,1], 1≤p<∞1 \le p < \infty, the series (31) converges in LpL^p-norm and almost everywhere.

Definition 4.3. Let F\mathcal{F} be partition of a probability space (X,F,μ)(X,\mathcal{F},\mu) and denote

m(F)=max⁡a∈Fμ(a)m(\mathcal{F})=\max_{a\in\mathcal{F}}\mu(a)

We say that an element a∈Fa\in\mathcal{F} is maximal, if μ(a)=m(F)\mu(a)=m(\mathcal{F}). A partition A\mathcal{A} is said to be ε\varepsilon-uniform, ε>0\varepsilon>0, if

∑a∈F,μ(a)<m(F)μ(a)<ε.\sum_{\substack{a\in\mathcal{F},\,\mu(a)<m(\mathcal{F})}} \mu(a)<\varepsilon.

Lemma 4.1. Let {Fn}n≥1\{\mathcal{F}_n\}_{n\ge1} be a dyadic partitioning on [0,1)[0,1). Then any B\mathbb{B}-valued martingale difference {fn}n≥1\{f_n\}_{n\ge1} adapted to {Fn}n≥1\{\mathcal{F}_n\}_{n\ge1} admits a Haar representation

fn=∑j∈Gn(fn,hj)hj,(32)f_n=\sum_{j\in G_n}(f_n,h_j)h_j, \tag*{(32)}

where the sets Gn⊂NG_n\subset\mathbb{N}, n=1,2,…n=1,2,\ldots, form a locally increasing sequence of positive integers. Moreover, if in addition Fn\mathcal{F}_n is εn\varepsilon_n-uniform, εn>0\varepsilon_n>0, then there exists an increasing sequence Gn′G'_n such that Gn′⊂GnG'_n\subset G_n and

∣supp⁡(∑j∈Gn∖Gn′(fn,hj)hj)∣<εn.(33)\left|\operatorname{supp}\left(\sum_{j\in G_n\setminus G'_n}(f_n,h_j)h_j\right)\right|<\varepsilon_n. \tag*{(33)}

Proof. Denote

G1={k∈N:Δk⊈I for any I∈F1},G_1 = \{k \in\mathbb{N} : \Delta_k \not\subseteq I \text{ for any } I \in\mathcal{F}_1\},
Gn={k∈N:Δk⊈I for any I∈Fn and Δk⊆I for some I∈Fn−1},n≥2.G_n = \{k \in\mathbb{N} : \Delta_k \not\subseteq I \text{ for any } I \in\mathcal{F}_n \text{ and } \Delta_k \subseteq I \text{ for some } I \in\mathcal{F}_{n-1}\}, \quad n \ge2.

One can check that the sequence {Gn}n≥1\{G_n\}_{n\ge1} is locally increasing. Then, in view of Proposition 4.2, it remains to show that

(fn,hk)=0 for all k∉Gn.(f_n,h_k)=0 \text{ for all } k \notin G_n.

Indeed, if k∉Gnk \notin G_n, then one of the following conditions hold:

  1. Δk⊆I\Delta_k \subseteq I for some I∈FnI \in\mathcal{F}_n,

  1. Δk⊈I\Delta_k \not\subseteq I for every I∈Fn−1I \in\mathcal{F}_{n-1}.

Suppose we have 1), i.e. there is I∈FnI \in\mathcal{F}_n such that Δk⊆I\Delta_k \subseteq I. Since fnf_n is constant on II, and therefore on Δk\Delta_k, it follows that (fn,hk)=0(f_n,h_k)=0. If a Δk\Delta_k satisfies 2), then we get that both Δk+\Delta_k^+ and Δk−\Delta_k^- are unions of some elements of Fn−1\mathcal{F}_{n-1}, since each interval I∈Fn−1I \in\mathcal{F}_{n-1} itself is dyadic. On the other hand, by the martingale-difference property of fnf_n we have (fn,1I)=0(f_n,1_I)=0 for any I∈Fn−1I \in\mathcal{F}_{n-1}. This implies (fn,1Δk±)=0(f_n,1_{\Delta_k^\pm})=0, and therefore, (fn,hk)=0(f_n,h_k)=0. If Fn\mathcal{F}_n is εn\varepsilon_n-uniform, then we have

∑a∈Fn,μ(a)<m(Fn)μ(a)<εn.(34)\sum_{\substack{a\in\mathcal{F}_n,\,\mu(a)<m(\mathcal{F}_n)}} \mu(a)<\varepsilon_n. \tag*{(34)}

Denote

Gn′={k∈Gn:there exists a maximal element I∈Fn such that Δk⊋I}.(35)G'_n=\{k\in G_n:\text{there exists a maximal element }I\in\mathcal{F}_n \text{ such that }\Delta_k\supsetneq I\}. \tag*{(35)}

Observe that

Δk⊆⋃a∈Fn,μ(a)<m(Fn)aif k∈Gn∖Gn′.\Delta_k\subseteq\bigcup_{\substack{a\in\mathcal{F}_n,\,\mu(a)<m(\mathcal{F}_n)}}a \quad\text{if }k\in G_n\setminus G'_n.

Combining this with (34), we obtain (33). Now consider integers k∈Gn−1′k\in G'_{n-1} and j∈Gn′j\in G'_n. By the definitions of Gn−1′G'_{n-1} ((35)) we have ∣Δk∣>m(Fn−1)|\Delta_k|>m(\mathcal{F}_{n-1}). Since j∈Gn′⊂Gnj\in G'_n\subset G_n, by the definition of GnG_n there exists an atom I∈Fn−1I\in\mathcal{F}_{n-1} such that Δj⊆I\Delta_j\subseteq I and so ∣Δj∣≤m(Fn−1)|\Delta_j|\le m(\mathcal{F}_{n-1}). Thus we get ∣Δj∣<∣Δk∣|\Delta_j|<|\Delta_k|, which clearly implies j>kj>k. Hence {Gn′}\{G'_n\} is an increasing sequence of integer sets. □

Proof of Theorem 4.1. Let {dn}n≥1\{d_n\}_{n\ge1} be a B\mathbb{B}-valued martingale difference sequence adapted to a partitioning {Fn}n≥1\{\mathcal{F}_n\}_{n\ge1}. Applying Theorem 1.1, we find a dyadic partitioning {Dn}n≥1\{\mathcal{D}_n\}_{n\ge1} such that each Dn\mathcal{D}_n is εn\varepsilon_n-uniform and {Fn}n≥1↪{Dn}n≥1\{\mathcal{F}_n\}_{n\ge1}\hookrightarrow\{\mathcal{D}_n\}_{n\ge1}. Let Θ\Theta be the MPS-map realizing this embedding. Using the property Θ(Fn)⊂Dn\Theta(\mathcal{F}_n) \subset\mathcal{D}_n (see Remark 1) and applying Lemma 2.2, we conclude that the sequence {fn=Θ(dn)}n≥1\{f_n=\Theta(d_n)\}_{n\geq1} is probabilistically equivalent to {dn}n≥1\{d_n\}_{n\geq1} and is adapted to the partitioning {Dn}n≥1\{\mathcal{D}_n\}_{n\geq1}. On the other hand according to Lemma 3.1 we have

EDn−1(Θ(dn))=Θ(EFn−1(dn)).\mathbb{E}^{\mathcal{D}_{n-1}}(\Theta(d_n))=\Theta(\mathbb{E}^{\mathcal{F}_{n-1}}(d_n)).

Since {fn}n≥1\{f_n\}_{n\geq1} is a martingale difference, we have Θ(EFn−1(dn))=0\Theta(\mathbb{E}^{\mathcal{F}_{n-1}}(d_n))=0. Therefore,

EDn−1(fn)=EDn−1(Θ(dn))=0,\mathbb{E}^{\mathcal{D}_{n-1}}(f_n)=\mathbb{E}^{\mathcal{D}_{n-1}}(\Theta(d_n))=0,

i.e. {fn}n≥1\{f_n\}_{n\geq1} is also a martingale difference adapted to the dyadic partitioning {Dn}n≥1\{\mathcal{D}_n\}_{n\geq1}. Then applying Lemma 4.1, we get the representation (32), where {Gn}\{G_n\} 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\mathbb{B} for which classical theorems on real-valued martingales can be extended for B\mathbb{B}-valued martingales. A Banach space B\mathbb{B} is said to have a UMD-property if there exists a positive constant cpc_p such that the inequality

∥ε1d1+⋯+εndn∥p≤cp∥d1+⋯+dn∥p,1<p<∞,(36)\lVert\varepsilon_1d_1+\cdots+\varepsilon_nd_n\rVert_p\leq c_p\lVert d_1+\cdots+d_n\rVert_p,\qquad1<p<\infty, \tag*{(36)}

holds for all B\mathbb{B}-valued martingale difference sequences {dj}j≥1\{d_j\}_{j\geq1}, for all so called predictable sequences εj=±1\varepsilon_j=\pm1 and all n≥1n\geq1. A geometric characterization of UMD-spaces was given by Burkholder [1]. It is well known that R\mathbb{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 cpc_p that can be in (36), where {dk}\{d_k\} 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=ckhkd_k=c_kh_k with ck∈Bc_k\in\mathbb{B}. In particular, if one seeks the best constant in (36) for a given Banach space B\mathbb{B}, then the same constant is obtained by considering only B\mathbb{B}-valued Haar martingale difference sequences instead of general ones. Indeed, by approximation, the optimal constant cpc_p 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∈Gkbjhj,k=1,2,…,n,(37)f_k = \sum_{j\in G_k} b_j h_j,\qquad k=1,2,\ldots,n, \tag*{(37)}

which is probabilistically equivalent to the d1,…,dnd_1,\ldots,d_n. Then the inequality (36) can be equivalently written as

∥ε1f1+⋯+εnfn∥p≤cp∥f1+⋯+fn∥p.\|\varepsilon_1 f_1+\cdots+\varepsilon_n f_n\|_p\leq c_p\|f_1+\cdots+f_n\|_p.

On the other hand, both f1+⋯+fnf_1+\cdots+f_n and ε1f1+⋯+εnfn\varepsilon_1f_1+\cdots+\varepsilon_nf_n 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\mathbb{B} be a UMD-space. If (36) holds for all B\mathbb{B}-valued Haar martingales and all choices εk=±1\varepsilon_k=\pm1, then it holds for all B\mathbb{B}-valued martingale difference sequences, with the same constant cpc_p.

The inequalities in (36) may also be considered with εk\varepsilon_k taking values in [0,1][0,1] or in {0,1}\{0,1\}. The statement of Theorem 5.1 remains valid for such choices of predictable εk\varepsilon_k. In particular, the case of εk∈{0,1}\varepsilon_k\in\{0,1\} was studied by [4], who determined the optimal constant in (36) for scalar-valued martingale difference sequences.

Now let {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} be an arbitrary filtration in a probability space (X,F∞,μ)(X,\mathcal{F}_\infty,\mu) and suppose that f={fn}n≥1f=\{f_n\}_{n\geq1} is a scalar-valued martingale adapted to {Fn}n≥1\{\mathcal{F}_n\}_{n\geq1} with assumption that f0=0f_0=0. For 1≤p≤∞1\leq p\leq\infty denote

∥f∥p=sup⁡n≥1∥fn∥p,f∗=sup⁡n≥1∣fn∣.\|f\|_p=\sup_{n\geq1}\|f_n\|_p,\qquad f^*=\sup_{n\geq1}|f_n|.

The martingale LpL^p-space consists of all martingales with finite norm ∥f∥p<∞\|f\|_p<\infty. It is well known that for 1<p<∞1<p<\infty, the martingale LpL^p-space is isomorphic to Lp(X,F∞)L^p(X,\mathcal{F}_\infty). Moreover, if f={fn}n≥1f=\{f_n\}_{n\geq1} belongs to a martingale LpL^p-space, then fnf_n converges both almost everywhere and in the LpL^p-norm.

In this case, by convention, when writing a martingale f={fn}n≥1f=\{f_n\}_{n\geq1} we identify ff with the limit function of fnf_n. If Θ={θn}n≥1\Theta=\{\theta_n\}_{n\geq1} is a sequence of real valued functions with ∣θn∣≤1|\theta_n|\leq1, then we say Θ\Theta is predictable if θn\theta_n is Fn−1\mathcal{F}_{n-1}-measurable. The sequence g={gn}n≥1g=\{g_n\}_{n\geq1} defined by

gn=∑k=1nθkdk,dk=fk−fk−1,g_n=\sum_{k=1}^{n}\theta_kd_k,\qquad d_k=f_k-f_{k-1},

is also a martingale and it is called the martingale transform of ff by Θ\Theta. Let denote by TΘT_\Theta the operator assigning any martingale ff to its Θ\Theta-transformation gg. From the inequality (36) it follows that TΘT_\Theta is a bounded operator on LpL^p, 1<p<∞1 < p < \infty. Moreover, we have the inequality

∥TΘ∗(f)∥p≤cp∥f∥p,\|T_\Theta^*(f)\|_p \leq c_p\|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\Theta_k = \{\theta_n^{(k)}\}_{n\geq1}, k=1,2,…,Nk = 1, 2, \ldots, N, consider the maximal operator

T(f)=max⁡1≤k≤NTΘk∗(f).(38)T(f)=\max_{1\leq k\leq N}T_{\Theta_k}^*(f). \tag*{(38)}

Applying Theorem 4.1, we establish the following inequality, extending the p=2p=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\{\mathcal{F}_n\}_{n\geq1} be a filtration in a probability space (X,F∞,μ)(X,\mathcal{F}_\infty,\mu). Then operator (38) satisfies

∥T(f)∥p≤cplog⁡(N+1)⋅∥f∥p,1<p<∞,\|T(f)\|_p \leq c_p\sqrt{\log(N+1)}\cdot\|f\|_p,\qquad1<p<\infty,

for any f∈Lp(X,F∞)f\in L^p(X,\mathcal{F}_\infty).

Proof. By Theorem 4.1 it is enough to prove the theorem for Haar martingales. Let ∑k=1∞akhk\sum_{k=1}^{\infty}a_kh_k be the Fourier-Haar series of a function f∈Lp(0,1)f\in L^p(0,1). Recall the definitions of the maximal function and the square function

f∗(x)=sup⁡n≥1∣∑k=1nakhk(x)∣,Sf(x)=(∑k=1∞ak2hk2(x))1/2.f^*(x)=\sup_{n\geq1}\left|\sum_{k=1}^{n}a_kh_k(x)\right|,\qquad Sf(x)=\left(\sum_{k=1}^{\infty}a_k^2h_k^2(x)\right)^{1/2}.

It is well known the boundedness of both operators on LpL^p, 1<p<∞1<p<\infty (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\lambda inequality due to Chang-Wilson-Wolff (see [3], Corollary 3.1):

∣{x∈[0,1):f∗(x)>λ, Sf(x)<ελ}∣≲exp⁡(−cε2)∣{f∗(x)>λ/2}∣,λ>0, 0<ε<1.(39)|\{x\in[0,1):f^*(x)>\lambda,\ Sf(x)<\varepsilon\lambda\}|\lesssim\exp\left(-\frac{c}{\varepsilon^2}\right)|\{f^*(x)>\lambda/2\}|,\qquad\lambda>0,\ 0<\varepsilon<1. \tag*{(39)}

Denote

fk=TΘk(f)=∑j=1∞θj(k)ajhj,k=1,2,…,N.f_k=T_{\Theta_k}(f)=\sum_{j=1}^{\infty}\theta_j^{(k)}a_jh_j,\qquad k=1,2,\ldots,N.

We have Sfk(x)≤Sf(x)Sf_k(x)\leq Sf(x), k=1,2,…,Nk=1,2,\ldots,N. Thus, applying inequality (39) with εN=(c/ln⁡N)1/2\varepsilon_N=(c/\ln N)^{1/2}, we obtain

∣{fk∗>λ, S(f)≤εNλ}∣≲exp⁡(−cεN2)∣{fk∗>λ/2}∣.(40)|\{f_k^*>\lambda,\ S(f)\leq\varepsilon_N\lambda\}|\lesssim\exp\left(-\frac{c}{\varepsilon_N^2}\right)|\{f_k^*>\lambda/2\}|. \tag*{(40)}

For

T(f)=max⁡1≤k≤NTΘk∗(f)=max⁡1≤k≤Nfk∗T(f)=\max_{1\leq k\leq N}T_{\Theta_k}^*(f)=\max_{1\leq k\leq N}f_k^*

we obviously have

{T(f)>λ}⊂{T(f)>λ, S(f)≤εNλ}∪{S(f)>εNλ}=A(λ)∪B(λ),\{T(f)>\lambda\}\subset\{T(f)>\lambda,\ S(f)\leq\varepsilon_N\lambda\}\cup\{S(f)>\varepsilon_N\lambda\}=A(\lambda)\cup B(\lambda),

and thus

∥T(f)∥pp≤p∫0∞λp−1∣A(λ)∣ dλ+p∫0∞λp−1∣B(λ)∣ dλ.\|T(f)\|_p^p\leq p\int_0^\infty\lambda^{p-1}|A(\lambda)|\,d\lambda+p\int_0^\infty\lambda^{p-1}|B(\lambda)|\,d\lambda.

From (40) it follows that

p∫0∞λp−1∣A(λ)∣ dλ≤p∑k=1N∫0∞λp−1∣{fk∗>λ, S(f)≤εNλ}∣ dλ≤exp⁡(−cεN2)∑k=1Np∫0∞λp−1∣{fk∗>λ/2}∣ dλ\begin{aligned} p\int_0^\infty\lambda^{p-1}|A(\lambda)|\,d\lambda\leq p\sum_{k=1}^{N}\int_0^\infty\lambda^{p-1}|\{f_k^*>\lambda,\ S(f)\leq\varepsilon_N\lambda\}|\,d\lambda\\ &\leq\exp\left(-\frac{c}{\varepsilon_N^2}\right)\sum_{k=1}^{N}p\int_0^\infty\lambda^{p-1}|\{f_k^*>\lambda/2\}|\,d\lambda \end{aligned}
≲1N∑k=1N∥fk∗∥pp\lesssim\frac{1}{N}\sum_{k=1}^{N}\|f_k^*\|_p^p
≲1N∑k=1N∥fk∥pp\lesssim\frac{1}{N}\sum_{k=1}^{N}\|f_k\|_p^p
≲∥f∥pp.\lesssim\|f\|_p^p.

Combining this and

p∫0∞λp−1∣B(λ)∣ dλ=εN−p∥S(f)∥pp≲(log⁡N)p/2⋅∥f∥pp,p\int_0^\infty\lambda^{p-1}|B(\lambda)|\,d\lambda=\varepsilon_N^{-p}\|S(f)\|_p^p\lesssim(\log N)^{p/2}\cdot\|f\|_p^p,

we get

∥T(f)∥p≲log⁡N⋅∥f∥p\|T(f)\|_p\lesssim\sqrt{\log N}\cdot\|f\|_p

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)(Q_d,\mathcal{L}_d,m_d) of Lebesgue measure on dd-dimensional unit cube Qd=[0,1)dQ_d=[0,1)^d. Let {An(k), n∈N}\{\mathcal{A}_n^{(k)},\,n\in\mathbb{N}\}, k=1,2,…,dk=1,2,\ldots,d, be partitionings of [0,1)[0,1). The Cartesian product

An=An1,…,nd=⨂k=1dAnk(k),n=(n1,…,nd)∈Nd.(41)\mathcal{A}_{\mathbf{n}}=\mathcal{A}_{n_1,\ldots,n_d}=\bigotimes_{k=1}^{d}\mathcal{A}^{(k)}_{n_k},\qquad\mathbf{n}=(n_1,\ldots,n_d)\in\mathbb{N}^{d}. \tag*{(41)}

determines a multiple partitioning of the unite cube QdQ_d. For a given sequence of integers N={νi∈N:ν1<ν2<…}\mathcal{N}=\{\nu_i\in\mathbb{N}:\nu_1<\nu_2<\ldots\} consider the multiple sequence of partitioning

A~n=Aνn1,…,νnd,n=(n1,…,nd)∈Nd.(42)\widetilde{\mathcal{A}}_{\mathbf{n}}=\mathcal{A}_{\nu_{n_1},\ldots,\nu_{n_d}},\qquad\mathbf{n}=(n_1,\ldots,n_d)\in\mathbb{N}^{d}. \tag*{(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}\{\mathcal{A}_{\mathbf{n}},\mathbf{n}\in\mathbb{N}^{d}\} and {Bn,n∈Nd}\{\mathcal{B}_{\mathbf{n}},\mathbf{n}\in\mathbb{N}^{d}\} be arbitrary multiple interval partitionings of the unit cube QdQ_d. Then for any sequence of numbers εk>0\varepsilon_k>0, k=1,2,…k=1,2,\ldots, there exist a multiple partitioning Dn\mathcal{D}_{\mathbf{n}}, a MPS-transformation Θ\Theta, and a sub-partitioning {B~n}\{\widetilde{\mathcal{B}}_{\mathbf{n}}\} of {Bn}\{\mathcal{B}_{\mathbf{n}}\}, such that

Dn⊂⋃k≥1Bk,(43)\mathcal{D}_{\mathbf{n}}\subset\bigcup_{k\geq1}\mathcal{B}_k, \tag*{(43)}
dist⁡(Dn,B~n)<εm,min⁡1≤i≤dni≥m,(44)\operatorname{dist}(\mathcal{D}_{\mathbf{n}},\widetilde{\mathcal{B}}_{\mathbf{n}})<\varepsilon_m,\qquad\min_{1\leq i\leq d}n_i\geq m, \tag*{(44)}
(Θ∘EAn)f(x)=(EDn∘Θ)f(x),n∈Nd,(45)(\Theta\circ\mathbb{E}^{\mathcal{A}_{\mathbf{n}}})f(x)=(\mathbb{E}^{\mathcal{D}_{\mathbf{n}}}\circ\Theta)f(x),\qquad\mathbf{n}\in\mathbb{N}^{d}, \tag*{(45)}

for all f∈L1(Qd)f\in L^1(Q_d).

If TT is a bounded linear operator on L1[0,1)L^1[0,1), then we denote by TkT_k, 1≤k≤d1\leq k\leq d, the operator

Tk:L1(Qd)→L1(Qd)T_k:L^1(Q_d)\to L^1(Q_d)

defined by

Tkf(x1,…,xd)=Tf(x1,…,xk−1,⋅,xk+1,…,xd).T_kf(x_1,\ldots,x_d)=Tf(x_1,\ldots,x_{k-1},\cdot,x_{k+1},\ldots,x_d).

The following lemmas are proved in [6].

Lemma 6.1 ([6]). If TT is a bounded linear operator on L1[0,1)L^1[0,1), then TkT_k determines a bounded operator on L1(Qd)L^1(Q_d).

Lemma 6.2 ([6]). If TT and PP are bounded linear operators on L1[0,1)L^1[0,1), then

Pn∘Tm=Tm∘Pn,n≠m,1≤n,m≤d.P_n\circ T_m=T_m\circ P_n,\qquad n\ne m,\quad1\leq n,m\leq d.

Proof of Theorem 6.1. Applying Theorem 1.1 for each k=1,2,…,dk=1,2,\ldots,d we find a partitioning {Dn(k)}n≥1\{\mathcal{D}^{(k)}_n\}_{n\geq1} of the interval [0,1)[0,1) such that

Dn(k)⊂⋃m≥1Bm(k),(46)\mathcal{D}^{(k)}_n\subset\bigcup_{m\geq1}\mathcal{B}^{(k)}_m, \tag*{(46)}
dist⁡(Dm(k),Bνm(k))<εm/d,(47)\operatorname{dist}(\mathcal{D}^{(k)}_m,\mathcal{B}^{(k)}_{\nu_m})<\varepsilon_m/d, \tag*{(47)}
{An(k)}n≥1⟶{Dn(k)}n≥1.(48)\{\mathcal{A}^{(k)}_n\}_{n\geq1}\longrightarrow\{\mathcal{D}^{(k)}_n\}_{n\geq1}. \tag*{(48)}

Clearly, (46) ⇒\Rightarrow (43), and (47) ⇒\Rightarrow (44). Let θk\theta_k be the MPS-transformation of [0,1)[0,1) realizing (48). By Lemma 2.2 it can be extended as a linear operator on L1[0,1)L^1[0,1). Then by Lemma 3.1 we can write

(θk∘EFn(k))(f)=(EFn(k)∘θk)(f), for all f∈L1[0,1).\left(\theta_k \circ\mathbb{E}_{\mathcal{F}_{n}^{(k)}}\right)(f)=\left(\mathbb{E}_{\mathcal{F}_{n}^{(k)}}\circ\theta_k\right)(f),\text{ for all }f\in L^1[0,1).

Applying θk\theta_k along kkth coordinate, by Lemma 6.1 we can consider θk\theta_k as a linear bounded operator on L1(Qd)L^1(\mathbb{Q}_d). Thus Θ=θ1∘θ2∘⋯∘θd\Theta=\theta_1\circ\theta_2\circ\cdots\circ\theta_d is also a linear bounded operator on L1(Qd)L^1(\mathbb{Q}_d). We also have

EFn=EFn1(1)∘⋯∘EFnd(d),EAn=EFn1(1)∘⋯∘EFnd(d).\mathbb{E}^{\mathcal{F}_n}=\mathbb{E}^{\mathcal{F}_{n_1}^{(1)}}\circ\cdots\circ\mathbb{E}^{\mathcal{F}_{n_d}^{(d)}},\qquad\mathbb{E}^{\mathcal{A}_n}=\mathbb{E}^{\mathcal{F}_{n_1}^{(1)}}\circ\cdots\circ\mathbb{E}^{\mathcal{F}_{n_d}^{(d)}}.

Applying Lemma 6.2, we can write

Θ∘EAn=θ1∘⋯∘θd∘EFn1(1)∘⋯∘EFnd(d)=(θ1∘EFn1(1))∘⋯∘(θd∘EFnd(d))=(EFn1(1)∘θ1)∘⋯∘(EFnd(d)∘θd)=EFn1(1)∘⋯∘EFnd(d)∘θ1∘⋯∘θd=EFn∘Θ,\begin{aligned} \Theta\circ\mathbb{E}^{\mathcal{A}_n} &=\theta_1\circ\cdots\circ\theta_d\circ\mathbb{E}^{\mathcal{F}_{n_1}^{(1)}}\circ\cdots\circ\mathbb{E}^{\mathcal{F}_{n_d}^{(d)}}\\ &=\left(\theta_1\circ\mathbb{E}^{\mathcal{F}_{n_1}^{(1)}}\right)\circ\cdots\circ\left(\theta_d\circ\mathbb{E}^{\mathcal{F}_{n_d}^{(d)}}\right)\\ &=\left(\mathbb{E}^{\mathcal{F}_{n_1}^{(1)}}\circ\theta_1\right)\circ\cdots\circ\left(\mathbb{E}^{\mathcal{F}_{n_d}^{(d)}}\circ\theta_d\right)\\ &=\mathbb{E}^{\mathcal{F}_{n_1}^{(1)}}\circ\cdots\circ\mathbb{E}^{\mathcal{F}_{n_d}^{(d)}}\circ\theta_1\circ\cdots\circ\theta_d\\ &=\mathbb{E}^{\mathcal{F}_n}\circ\Theta, \end{aligned}

which gives (45).

Acknowledgments

Research was supported by the Higher Education and Science Committee of RA, in the frames of the research project 21AG-1A045

Conflicts of Interest: The authors declare no conflicts of interest.

Data Availability Statement: Data sharing is not applicable to this article as no data sets were generated or analyzed during the current study.

References

Institute of Mathematics of the National Academy of Sciences of the Republic of Armenia, Marshal Baghramian ave., 24/5, Yerevan, 0019, Armenia

Email address: [email protected]

Institute of Mathematics of the National Academy of Sciences of the Republic of Armenia, Marshal Baghramian ave., 24/5, Yerevan, 0019, Armenia

Email address: [email protected]

Institute of Mathematics of the National Academy of Sciences of the Republic of Armenia, Marshal Baghramian ave., 24/5, Yerevan, 0019, Armenia

Email address: khvn [email protected]

References

  1. [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. [2]———, Boundary value problems and sharp inequalities for martingale transforms, Ann. Probab. 12 (1984), no. 3, 647–702. MR744226DOI
  3. [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. [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. [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. [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. [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. [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. [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

Paper details

vlnc:W_5wnrepph
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents