In~\cite{BFR} we constructed processes Wn, built from a family of independent standard Poisson processes, that converge almost surely to the Brownian sheet uniformly on [0,1]2, and in~\cite{Rov} we proved that P(max[0,1]2∣Wn−W∣>αn−β)=o(n−q) for every β<λ/2, where λ∈(0,51) is the parameter of the construction; this gives rates n−β with β<101. In this paper, we show that the same construction, with λ≥1, satisfies P(max(s,t)∈[0,1]2∣Wn(s,t)−W(s,t)∣>αn−1/2(logn)3)=o(n−q)for every q>0, which is the one-parameter rate of Gorostiza and Griego~\cite{GG} up to a factor (logn)1/2. In particular the restriction λ<51 is not needed.
Introduction
Let W={W(s,t):(s,t)∈[0,1]2} be a Brownian sheet. Griego, Heath and Ruiz-Moncayo [5] constructed realizations of uniform transport processes converging almost surely to Brownian motion, uniformly on [0,1]; the rate of convergence was obtained by Gorostiza and Griego [4] and by Csörgő and Horváth [3]. In [2] we extended the almost sure convergence to the Brownian sheet: they cut [0,1]2 into ⌊nλ⌋ vertical strips, observe that the sheet integrated over each strip is, after rescaling, a standard Brownian motion, approximate each of these independent Brownian motions by a uniform transport process as in [5], and sum over the strips. They obtained almost sure uniform convergence for λ∈(0,51) and left the rate of convergence as an open problem. In [7] we obtained the rate
P([0,1]2max∣Wn−W∣>αn−β)=o(n−q),β<2λ,λ∈(0,51),
and observed [7 Remark 2.3] that it is worse than the one-parameter rate n−1/2(logn)5/2 of [4], because the latter, valid for each of the ⌊nλ⌋ Brownian motions, could not be kept for their sum.
The purpose of this note is to show that the one-parameter rate is kept for the sum, up to a factor (logn)1/2. In [2, 7] the errors of the different strips are added through the triangle inequality, but these errors are independent and centered, so their partial sums behave like a random walk rather than like a sum of absolute values. On the other hand, using the Borell–TIS inequality, it no longer constrains λ from above, so we can take λ≥1.
Section 2 recalls the construction and states precisely, as a hypothesis, the properties of the one-parameter approximation that are used. Section 3 contains the results and Section 4 the proof. The probabilistic tools used in the proof are collected in Section 5.
The construction and the one-parameter input
We follow [2, 7]. Fix λ>0 and write L=⌊nλ⌋, assuming n large enough that L≥1. It is convenient to assume that W is the restriction to [0,1]2 of a Brownian sheet defined on [0,2]×[0,1]; this is no loss of generality, since the statements are about the existence of suitable realizations. For k=1,…,L put
The processes W1,…,WL are independent standard Brownian motions on [0,1] (they are centered, Gaussian, continuous, and E[Wk(t)Wk′(t′)]=(t∧t′)1{k=k′}), and W(l/nλ,t)=∑k≤lWk(t).
For each k, a uniform transport process Wk(n) is constructed from Wk by the Skorohod embedding procedure of [5]. Throughout the paper we write
rn=n−1/2(logn)5/2,n≥2,
for the one-parameter rate of Gorostiza and Griego [4]. In the proofs we will use that
rn≥n−1for all n≥3.(1)
We use the normalization of [4], in which the transport process moves with speed n. The transport process of the k-th strip has the law of
where N is a standard Poisson process and A an independent Bernoulli(21) variable: the process moves at speed n and reverses its direction at the jump times of a Poisson process of rate n2. It is realized on the same probability space as Wk as in [5, 2]. This is also the process constructed in the proof of [2 Theorem 2.1]: there the pieces have slope ±n, durations exponential with parameter 2n2 and i.i.d. symmetric signs, and since each new sign differs from the previous one with probability 21, independently, the direction is reversed at the jump times of a Poisson process of rate n2.
From the construction of the approximations we get the following properties:
(H1) For every n the pairs (Wk,Wk(n)), k≤L, are independent and identically distributed. In particular the error processes Dk=Wk(n)−Wk, k≤L, are independent and identically distributed.
(H2) For every n, k and t∈[0,1], EWk(n)(t)=0; and almost surely, for all t,t′∈[0,1],
∣Wk(n)(t)∣≤n,∣Wk(n)(t)−Wk(n)(t′)∣≤n∣t−t′∣.
(H3) For every q>0 there are αq>0 and nq≥3, depending on q but not on n or k, such that, with rn=n−1/2(logn)5/2,
P(0≤t≤1max∣Wk(n)(t)−Wk(t)∣>αqrn)≤n−qfor all n≥nq and all k≤L.(2)
The approximating process is defined at the strip boundaries by
and max[0,1]2∣Wn−W∣=O(n−1/2(logn)3) almost surely. In particular the almost sure convergence of Wn to W holds.
Proof. Recall rn=n−1/2(logn)5/2 and apply Theorem 3.1. Choose any λ≥1. Then n−λ/2≤n−1/2≤rn, so the bound is at most 2Crn(logn)1/2.
Using q=2: with An={max[0,1]2∣Wn−W∣>αn−1/2(logn)3} we have
n≥1∑P(An)≤n0+n≥n0∑n−2<∞.
By the Borel–Cantelli lemma, almost surely An occurs for finitely many n only, i.e. max[0,1]2∣Wn−W∣≤αn−1/2(logn)3 for all n large enough. □
Remark 3.1. The rates of [7] are n−β with β<λ/2<101. Theorem 3.2 gives n−1/2(logn)3, the rate n−1/2(logn)5/2 of [4] for one Brownian motion up to a factor (logn)1/2. This answers the question raised in [7 Remark 2.3].
Remark 3.2. (d-parameter Wiener processes.) The same method applies to the d-parameter Wiener process, d≥2, approximated as in [2 Section 4]: the strips are replaced by the ⌊nλ⌋d−1 boxes ∏i<d((ki−1)/nλ,ki/nλ]×[0,1], each carrying an independent transport process, the normalization n−λ/2 becomes n−λ(d−1)/2, and Wn is defined by (d−1)-linear interpolation.
Proof of Theorem 3.1
Fix q>0. All constants below may depend on q, λ and the constants of (H3), but not on n, and n is assumed large enough.
Proof. If s∈[l/nλ,(l+1)/nλ) with l≤L−1, write θ=nλs−l∈[0,1). Then
Wn(s,t)−W(s,t)=(1−θ)Δn(l,t)+θΔn(l+1,t)
+(1−θ)[W(nλl,t)−W(s,t)]+θ[W(nλl+1,t)−W(s,t)].
The first bracket is bounded by Zl and the second, by the triangle inequality through W(l/nλ,t), by 2Zl. If s≥L/nλ, then Wn(s,t)−W(s,t)=Δn(L,t)+W(L/nλ,t)−W(s,t), and the last difference is bounded by ZL.
Thus it suffices to prove that for n large enough,
is a Brownian sheet, and Zl=n−λ/2sup[0,1]2∣Bl∣. Let m=Esup[0,1]2∣Bl∣, which does not depend on l. By Lemma 5.1 (maximal variance 1) with x=2(q+λ+1)logn and a union bound over the L+1≤nλ+1 strips,
We begin studying Dk. Let q′=q+λ+7, an=αq′rn, with αq′ given by (H3), and put
Fk={t∈[0,1]sup∣Dk(t)∣≤an},Gn=k=1⋂LFk.
By (H3) and (H1), πn:=P(Fkc) does not depend on k and πn≤n−q′ for n≥nq′; hence
P(Gnc)≤Lπn≤nλ−q′≤n−q−1.
We apply Lemma 5.3 to each Dk. By (H2), EDk(t)=0 and supt∣Dk(t)∣≤Yk:=n+supt∣Wk(t)∣, so that EYk2≤2n2+8≤10n2 by Doob’s inequality. By (H1) the function mn(t)=E[Dk(t)1Fk] does not depend on k, and
using l≤nλ and (5). Moreover, for fixed t, the variables D1′(t),…,DL′(t) are independent (by (H1)) and centered, and by Lemma 5.3
∣Dk′(t)∣≤an+∣mn(t)∣≤2ansurely, for every k≤L and t∈[0,1],(7)
because ∣mn(t)∣≤10n−5/2≤αq′n−1≤an for n large, by (5) and rn≥n−1.
Partial sums on a grid
Let M=⌈nλ+3⌉, ti=i/M, i=0,…,M. For fixed l≤L and i≤M, the variables Xk=Dk′(ti), k≤l, are independent, centered and, by (7), bounded by the deterministic constant c=2an, uniformly in k, l and i. Lemma 5.2 therefore applies with c=2an and gives, since l≤nλ,
By (H2), the first term is at most n−λ/2⋅L⋅n⋅M−1≤nλ/2+1n−(λ+3)≤n−2, deterministically. For the second, for each i<M the field (s,ρ)↦M1/2W((0,s]×(ti,ti+ρ/M]) on [0,1]2 is a Brownian sheet, and the second term is bounded by M−1/2 times its supremum. By Lemma 5.1 with x=2(q+λ+4)logn and a union bound over the M≤2nλ+3 intervals,
Since M−1/2≤n−1≤rn, combining with (8) we obtain the bound (4).
Remark 4.1.The construction of Section 2 makes sense for every λ>0; nothing in the definition of the strips, of the strip Brownian motions Wk, of the transport processes Wk(n) or of the interpolated process Wn requires λ<51. The restriction λ∈(0,51) appears only in the proof of almost sure convergence in [2 pp. 324–325], where Markov’s inequality of order M leads to an exponent that can be made smaller than −1 only if λ<51. The present paper uses from [2] only the construction (hypotheses (H1)–(H2)) and from [4] only the one-parameter rate of each strip (hypothesis (H3)); none of these depends on λ. Therefore Theorem 3.1 and Theorem 3.2 hold for every λ>0, and in particular for λ≥1, which is the range giving the rate n−1/2(logn)3.
Probabilistic tools
We collect here the probabilistic inequalities used in § 4. We begin with a Borell–TIS inequality.
Lemma 5.1. Let Γ be a centered Gaussian field with continuous paths on a compact set T, let σ2=supx∈TEΓ(x)2 and m=EsupT∣Γ∣. Then m<∞ and, for every x>0,
P(Tsup∣Γ∣>m+x)≤2e−x2/(2σ2).
Proof. Since Γ has continuous paths on the compact set T, it is a.s. bounded: for almost every ω the map t↦Γ(t,ω) is continuous on a compact set, hence bounded. The Borell–TIS inequality [1 Theorem 2.1.1], applied to Γ and to −Γ, then gives
Since EsupT(−Γ)=−EinfTΓ≤EsupT∣Γ∣=m, combining the two inequalities yields the stated bound.
We will also use the following Hoeffding inequality (see [6]).
Lemma 5.2. If X1,…,Xl are independent, centered and ∣Xk∣≤c almost surely for every k, then for every x>0
P(k=1∑lXk>x)≤2exp(−2lc2x2).
The last technical result is a truncation lemma.
Lemma 5.3. Let D=(D(t))t∈[0,1] be a process with ED(t)=0 for every t and supt∣D(t)∣≤Y with EY2<∞. For a>0 let F={supt∣D(t)∣≤a} and
m(t)=E[D(t)1F],D′(t)=D(t)1F−m(t).
Then ED′(t)=0, ∣m(t)∣≤(EY2)1/2P(Fc)1/2 and, surely, ∣D′(t)∣≤a+∣m(t)∣ for every t.
Proof. Since ED(t)=0, we have m(t)=−E[D(t)1Fc], and Cauchy–Schwarz gives the bound on m(t). The bound on D′(t) holds on every sample point, because ∣D(t)1F∣≤a everywhere.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process.
During the preparation of this work the author used Claude (Anthropic) to assist with LaTeX typesetting of formulas, translation and language editing, and the verification of proofs and calculations. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the content of the published article.
References
[1]R. J. Adler and J. E. Taylor. Random Fields and Geometry. Springer Monographs in Mathematics. Springer, New York, 2007.DOI
[2]X. Bardina, M. Ferrante and C. Rovira. Strong approximations of Brownian sheet by uniform transport processes. Collect. Math. 71 (2020), 319–329.
[3]M. Csörgő and L. Horváth. Rate of convergence of transport processes with an application to stochastic differential equations. Probab. Theory Related Fields 78 (1988), no. 3, 379–387.DOI
[4]L. G. Gorostiza and R. J. Griego. Rate of convergence of uniform transport processes to Brownian motion and application to stochastic integrals. Stochastics 3 (1980), 291–303.DOI
[5]R. J. Griego, D. Heath and A. Ruiz-Moncayo. Almost sure convergence of uniform transport processes to Brownian motion. Ann. Math. Statist. 42 (1971), no. 3, 1129–1131.DOI
[6]W. Hoeffding. Probability inequalities for sums of bounded random variables. J. Amer. Statist. Assoc. 58 (1963), 13–30.
[7]C. Rovira. Rate of convergence of uniform transport processes to a Brownian sheet. Open Math. 18 (2020), 731–737.arxiv.org/abs/1909.01981