Merryfield's inequality for multiparameter martingales

Guillermo Rey

REVIEW 3 major objections 5 minor 11 references

Merryfield's inequality for multiparameter martingales

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proves that the square function is controlled by the maximal function for discrete multiparameter martingales on regular filtrations satisfying the Cairoli-Walsh condition, for every $0<p<\infty$.

desk verdict A credible multiparameter extension of Brossard's L^p comparison, built on a new one-parameter inequality, but Section 4's proof of Theorem C has a load-bearing gap that needs a separation lemma. read the letter →

arxiv 2506.02974 v1 pith:NSE5H4GO submitted 2025-06-03 math.PR math.CA

Guillermo Rey This is my paper

classification math.PRmath.CA MSC 60G4260G4842B25

keywords martingalemultiparametersquarefunctionmaximalBurkholder-Davis-GundyinequalityMerryfieldregularfiltrationCairoli-Walshcondition

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proves a multiparameter analogue of the Burkholder-Davis-Gundy inequality for discrete martingales. The main theorem states that for a $k$-parameter martingale $f$ on a regular filtration satisfying the Cairoli-Walsh (F4) condition, $E[(Sf)^p] \lesssim E[(f^*)^p]$ for every $0

What carries the argument

The machinery is the one-parameter "Merryfield inequality" (Theorem A): an $L^2$ summation-by-parts estimate driven by the conditional variance identity $E[(\Delta f_m)^2 \mid \mathcal{F}_{m-1}] = E[\Delta(f^2)_m \mid \mathcal{F}_{m-1}]$. The iteration uses (F4) to build one-parameter filtrations $\mathcal{G}_j = \sigma(\bigcup_m \mathcal{F}_{m,j})$ in which the slices of a multiparameter martingale are martingales, so Theorem A applies along each coordinate. The residual terms are organized by the boundary slices $\partial_I(M) = \{m \le M : m_i = M_i \text{ for } i \notin I\}$. Regularity enters through the enlargement operator $\mathrm{Enl}(E) = \{1^*_E > R^{-k-1}\}$, which lets a stopping set be detected one step earlier without multiplying its probability by more than a fixed constant; this is what turns the boundary sum into the Brossard distributional bound.

What would settle it

Construct a two-parameter martingale on a regular filtration satisfying (F4) with $E[Sf]=+\infty$ but $E[f^*]<+\infty$; Theorem C at $p=1$ predicts $E[Sf] \lesssim E[f^*]$, so any such example refutes it. The product dyadic filtration is the natural place to compute both sides explicitly.

Watch

Extended reading notes

Core claim

On the paper's own terms the discovery is Theorem C: for any $k$-parameter martingale $f$ with respect to a regular filtration satisfying (F4), $E[(Sf)^p] \lesssim E[(f^*)^p]$ for all $0<p<\infty$. The proof is in two layers. Theorem A is a one-parameter inequality of Merryfield type, proved by a conditional variance identity and summation by parts, giving $E[\sum_{m=1}^M (\Delta f_m)^2 a_{m-1}^2 \mid \mathcal{F}_0] \le 20E[f_M^2 a_M^2 + \sum_{m=1}^M (f_m^2 + f_{m-1}^2)(\Delta a_m)^2 \mid \mathcal{F}_0]$. Theorem B iterates that scalar inequality through the $k$ coordinates; the (F4) condition makes one-parameter slices of a multiparameter martingale into martingales on the enlarged one-parameter filtrations, and the iteration leaves a sum over boundary slices $\partial_I(M)$ for all $I \subseteq [k]$. Assuming regularity, Section 4 converts this boundary sum into Brossard's distributional inequality $P(Sf>\lambda) \lesssim P(f^*>\lambda) + \lambda^{-2}E[(f^*)^2; f^* \le \lambda]$, and the layer-cake formula turns it into the $L^p$ comparison.

Load-bearing premise

The load-bearing premise is that the filtration is regular, meaning that relaxing one coordinate of the conditioning information can enlarge the probability of any set by at most a fixed factor; the final distributional estimate depends on that finite bound, and without it even the two-parameter case is noted to be open.

Editorial extensions

If this is right

For every $0<p<\infty$, the norms $\|Sf\|_p$ and $\|f^*\|_p$ are comparable for $k$-parameter martingales on regular (F4) filtrations, so the square-function and maximal-function definitions of multiparameter martingale Hardy spaces agree.
The distributional inequality $P(Sf>\lambda) \lesssim P(f^*>\lambda)+\lambda^{-2}E[(f^*)^2; f^* \le \lambda]$ holds in any number of parameters, extending Brossard's biparameter theorem.
The endpoint estimate $P(Sf>\lambda) \lesssim E[\frac{|f|}{\lambda}\log(e+\frac{|f|}{\lambda})^{k-1}]$ follows from the layer-cake formula, as the paper notes.
Theorem B provides a weighted square-function bound in which the weights are differences $\Delta^I a$ on boundary slices; this is a reusable tool for future multiparameter martingale estimates.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Tracking the constants in Theorem A and in the enlargement bound would give an explicit dependence of the comparison constant on the regularity constant $R$ and on $k$, a quantitative version the paper does not state.
The same one-parameter inequality plus (F4) slicing is a natural starting point for weighted or sparse forms of the multiparameter square-function estimate, with the boundary differences $\Delta^I a$ of Theorem B playing the role of testing weights.
A natural test is whether the inequality remains valid when the regularity constant is allowed to grow slowly with $k$; Theorem C as stated does not address that asymptotic regime.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

One thing to know: this is a credible, genuinely multiparameter version of Brossard's L^p comparison, with a new one-parameter inequality at its base. The printed text has a few misprints and one real gap in Section 4 that should be patched before the theorem is taken as proven.

The new content: Theorem 2.1 is a one-parameter martingale inequality of Merryfield type, proved by summation by parts and conditional variance. The algebra is correct up to the stated constant. Theorem 3.1 is the F4 induction; the key step—showing the one-parameter slices are martingales—uses (F4) in the right way, and the induction is coherent. For a reader who knows Brossard's two-parameter argument, this is a clean generalization.

The soft spots: In Section 2 the filtration is stated decreasing, and in Section 3.1 the directional differences are misprinted. These are typos; any referee will see through them. The serious one is in Section 4. The proof of Theorem C uses the stopping-bump martingale a_m = E[1_{Enl(E)^c} | F_m] and asserts that on a set F^c—never defined—a_{m-1} ≳ 1. That inequality is what replaces (f^*_m)^2 by λ^2 in the terms Σ_{I≠∅} ∫ (f^*_m)^2 (Δ_I a_m)^2. Without it, Theorem B leaves f^* factors uncontrolled and the argument does not close. The regularity condition plus (F4) gives a_{N_i(m)} ≤ R a_m, but it does not by itself give the needed lower bound. One expects a separation lemma: if sup_l E[1_E | F_l] ≤ R^{-k-1}, then the conditional expectation of the enlargement is bounded away from 1. The threshold R^{-k-1} is clearly chosen for that, but the lemma is not stated or proved, and the constants must be independent of the stopping horizon. This is a load-bearing gap, not a cosmetic one.

The citation pattern is fine: Brossard, Cairoli-Walsh, Merryfield, and Weisz are the natural references, and the paper is honest about what is known in the non-regular case.

Bottom line: if the Section 4 estimate is true—and I suspect it is—this paper is a solid contribution. As written, the proof of Theorem C is incomplete. I would send it to a serious referee, and ask the referee to focus on Section 4. The result deserves scrutiny.

Referee Report

3 major / 5 minor

Summary. The paper proves a discrete martingale analogue of Merryfield's inequality. Theorem A is a one-parameter inequality comparing E[Σ(Δf_m)^2 a_{m-1}^2 | F_0] with boundary and quadratic-variation terms of the two martingales f and a. Theorem B iterates Theorem A under the Cairoli-Walsh (F4) condition to obtain a k-parameter version with sums over boundary slices ∂_I(M). Theorem C claims the L^p comparison E[(Sf)^p] ≲ E[(f*)^p] for all 0<p<∞ for regular filtrations satisfying (F4), via Brossard's stopping-bump argument and a distributional inequality. The one-parameter algebra and the multiparameter induction are largely coherent, but the proof of Theorem C relies on several unproved assertions in Section 4, including the central lower bound a_{m-1}≳1 on an undefined set F^c.

Significance. If Theorem C is established, it would extend Brossard's two-parameter result to arbitrary dimensions under regularity plus (F4), and would yield the stated endpoint estimates; this is a worthwhile contribution to multiparameter martingale theory. The paper has genuine strengths: Theorem A is proved by an explicit elementary summation-by-parts argument with no fitted constants, the induction in Theorem B has a clear structure, and the derivation does not assume the target inequality. The main weakness is that the decisive stopping-bump step in Section 4 is not actually proved: the enlargement estimates are asserted rather than demonstrated, and the limiting argument for arbitrary f is omitted. The central claim is therefore not fully established as written.

major comments (3)

[Section 4] The passage beginning 'Note that, on F^c, we have a_{m-1} ≳ 1' is the load-bearing step of the proof of Theorem C, but the set F is never defined, and the asserted lower bound is not proved. The same paragraph also states without proof that P(Enl(E)) ≲ P(E). This lower bound is exactly what allows the f*_m factors in the I≠∅ terms of Theorem B to be replaced by λ^2; without it, the application of Theorem B leaves uncontrolled maximal-function factors and the argument collapses. A separation lemma should be stated and proved: on the set where sup_l E[1_E|F_l] ≤ R^{-k-1}, one must have E[1_{Enl(E)}|F_{m-1}] ≤ 1-c for a constant c>0 independent of m and of the stopping horizon M. The constants must also be independent of M for the final estimate to hold for arbitrary f. This is the decisive missing piece.
[Section 3.1] The proof that \tilde f_{m,n} = Δ_1 f_{m,n} is a martingale with respect to G^1_n contains the unjustified equality E[f_{m,n+1}|F_{j,n}] = E[E[f_{m,n+1}|F_{m,n+1}]|F_{j,n}], which would require F_{j,n} ⊆ F_{m,n+1}; for j>m this inclusion fails in general. The desired martingale property can be derived instead from the (F4) conditional-independence structure, but the printed argument does not supply that derivation. Since this one-parameter reduction is the engine of Theorem B and is reused in the k-parameter induction in Section 3.2, the proof of Theorem B is incomplete at this point and needs to be repaired.
[Section 4] The proof of Theorem C is carried out only for a stopped f with Δf_m = 0 outside a finite box m≤M and on the coordinate axes. The paper does not explain why the constants in the stopping-bump argument are independent of M, nor how the distributional inequality for stopped f is transferred to arbitrary f by a limiting argument. This is necessary for the theorem as stated ('for any k-parameter martingale f'), and the independence of the constants from M must be verified explicitly.

minor comments (5)

[Section 2] The filtration is defined as decreasing, F_{m+1} ⊆ F_m, which is inconsistent with the increasing filtrations used in Section 3 and with the later martingale definitions; the definition should be corrected to F_m ⊆ F_{m+1}, and the formula for F_∞ should be adjusted accordingly.
[Section 3.1] The definition of Δ_2 has the sign reversed: for n≥1 it is written as f_{m,n-1} - f_{m,n}, while the displayed formula for Δ f_{m,n} and the standard convention require f_{m,n} - f_{m,n-1}. The definition of Δ_1 also contains a typo ('f_{m,n} - f_{m,n}' for m≥1). These should be corrected for consistency.
[Section 4] The notation is incomplete: the set F in the split P(Sf>λ) ≤ P(Enl^2(E)) + λ^{-2}∫_{F^c} ... is not defined, and Enl^2(E) is not explained; moreover integrals such as ∫_{F^c} Σ(Δf)^2 should be written as expectations or as integrals over Ω to avoid ambiguity.
[Theorem C] The distributional inequality established in Section 4 yields E[(Sf)^p] ≲ E[(f*)^p] for p<2 by the layer-cake formula; for p≥2 the paper must explicitly state that it is relying on the known p>1 result cited in the introduction, since the new stopping-bump argument does not by itself cover p≥2.
[Introduction] Equation (1.2) contains typographical errors: it should read P(Sf>λ) ≲ P(f*>λ) + λ^{-2} E[(f*)^2; f*≤λ], and the occurrences of 'f^*λ' and 'h f^*' should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the multiparameter comparison is proved from one-parameter summation by parts, an F4 induction, and a stopping-bump argument, none of which assumes the target inequality.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs. Theorem A is proved from the conditional-variance identity (2.2) and a summation-by-parts formula; it does not invoke the desired L^p comparison. Theorem B is obtained by induction on the number of parameters, using only the (F4) condition to verify that parameter slices are one-parameter martingales; the induction hypothesis is applied to lower-dimensional slices, and the target inequality is not assumed. Theorem C is deduced from Theorem B through the standard Brossard stopping-bump route: one defines a_m = E[1_{Enl(E)^c} | F_m] and uses regularity to control the increments, then applies the layer-cake formula; the p>1 regime is cited to an external known result and is not the claim being reduced. The only fragile spots are unproved quantitative assertions in Section 4, such as a_{m-1} ≳ 1 on F^c and P(Enl(E)) ≲ P(E), but these are missing details or correctness risks rather than circularity: they are not equivalent by construction to any fitted parameter and do not assume the target inequality, and no load-bearing premise is justified only by a self-citation. Hence no circular step can be exhibited, and the score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The proof is self-contained except for standard probability facts, the assumed F4 and regularity conditions on the filtration, the L2 integrability hypothesis used in Theorem A, and the cited one-parameter results for p>1. There are no fitted parameters or invented entities.

assumptions (5)

standard math Standard measure-theoretic probability: conditional expectations, the variance identity, and the layer-cake formula.
Used throughout, e.g., Section 2 conditional variance identity (2.2) and Section 4 layer-cake step.
domain assumption Cairoli-Walsh (F4) condition for k-parameter filtrations.
Assumed in Theorems B and C; used to prove the directional martingale property in the induction (Sections 3.1 and 3.2).
domain assumption Regularity of the filtration: a_{N_i(m)} ≤ R a_m for nonnegative martingales.
Assumed in Theorem C; used in the enlargement argument to control a_{m-1} from below and to obtain the stopping bump property (Section 4).
domain assumption L^2 integrability condition f_m a_n ∈ L^2 for all m,n.
Assumed in Theorems A and B to justify the conditional variance identity; standard truncation arguments would relax it.
standard math One-parameter Burkholder-Davis-Gundy inequality for vector-valued martingales for p>1, as cited from [10].
Used to cover the p>1 range in Theorem C; it is an external classical result, not the target inequality.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Merryfield's inequality for multiparameter martingales." pith.science (2026). https://pith.science/paper/NSE5H4GO

@misc{pith2026250602974,
  author       = {Pith},
  title        = {Pith review of: Merryfield's inequality for multiparameter martingales},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NSE5H4GO}},
  note         = {Machine review of arXiv:2506.02974}
}

abstract

We extend an inequality of Merryfield, valid in the continuous setting, to discrete multiparameter martingales. As a consequence, we obtain the $L^p$ comparison of the maximal function with the square function: \begin{align*} E[(Sf)^p] \lesssim E[(f^*)^p] \end{align*} for regular multiparameter filtrations and $0 < p < \infty$.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

A. Bernard. EspacesH 1 de martingales ` a deux indices. Dualit´ e avec les martingales de type “BMO”.Bull. Sci. Math. (2), 103(3):297–303, 1979

work page 1979
[2]

Brossard

J. Brossard. Comparison des “normes”L p du processus croissant et de la variable maximale pour les martingales r´ eguli` eres ` a deux indices. Th´ eor` eme local correspondant.Ann. Probab., 8(6):1183–1188, 1980

work page 1980
[3]

Brossard

J. Brossard. R´ egularit´ e des martingales ` a deux indices et in´ egalit´ es de normes. InTwo-index random processes (Paris, 1980), volume 863 ofLecture Notes in Math., pages 91–121. Springer, Berlin, 1981

work page 1980
[4]

D. L. Burkholder, B. J. Davis, and R. F. Gundy. Integral inequalities for convex functions of operators on martingales. InProceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory, pages 223–240. Univ. California Press, Berkeley, CA, 1972

work page 1970
[5]

Cairoli and J

R. Cairoli and J. B. Walsh. Stochastic integrals in the plane.Acta Math., 134:111–183, 1975

work page 1975
[6]

M. G. Cowling, Z. Fan, J. Li, and L. Yan.Characterizations of Product Hardy Spaces on Stratified Groups by Singular Integrals and Maximal Functions, pages 193–227. Springer Nature Switzerland, Cham, 2025

work page 2025
[7]

M. G. Cowling, M.-Y. Lee, J. Li, and J. Pipher. An endpoint estimate for product singular integral operators on stratified lie groups.Canadian Journal of Mathematics, page 1–32, 2025

work page 2025
[8]

R. F. Gundy and E. M. Stein.H p theory for the poly-disc.Proc. Nat. Acad. Sci. U.S.A., 76(3):1026–1029, 1979

work page 1979

Show all 11 references

[9]

K. G. Merryfield. On the area integral, Carleson measures andH p in the polydisc.Indiana Univ. Math. J., 34(3):663–685, 1985

1985
[10]

Treil.H 1 and dyadicH 1

S. Treil.H 1 and dyadicH 1. InLinear and complex analysis, volume 226 ofAmer. Math. Soc. Transl. Ser. 2, pages 179–193. Amer. Math. Soc., Providence, RI, 2009

2009
[11]

Weisz.Martingale Hardy spaces and their applications in Fourier analysis, volume 1568 ofLecture Notes in Mathematics

F. Weisz.Martingale Hardy spaces and their applications in Fourier analysis, volume 1568 ofLecture Notes in Mathematics. Springer-Verlag, Berlin, 1994. Universidad Aut´onoma de Madrid Email address:guillermo.rey@uam.es

1994

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

A. Bernard. EspacesH 1 de martingales ` a deux indices. Dualit´ e avec les martingales de type “BMO”.Bull. Sci. Math. (2), 103(3):297–303, 1979

work page 1979

[2] [2]

Brossard

J. Brossard. Comparison des “normes”L p du processus croissant et de la variable maximale pour les martingales r´ eguli` eres ` a deux indices. Th´ eor` eme local correspondant.Ann. Probab., 8(6):1183–1188, 1980

work page 1980

[3] [3]

Brossard

J. Brossard. R´ egularit´ e des martingales ` a deux indices et in´ egalit´ es de normes. InTwo-index random processes (Paris, 1980), volume 863 ofLecture Notes in Math., pages 91–121. Springer, Berlin, 1981

work page 1980

[4] [4]

D. L. Burkholder, B. J. Davis, and R. F. Gundy. Integral inequalities for convex functions of operators on martingales. InProceedings of the Sixth Berkeley Symposium on Mathematical Statistics and Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theory, pages 223–240. Univ. California Press, Berkeley, CA, 1972

work page 1970

[5] [5]

Cairoli and J

R. Cairoli and J. B. Walsh. Stochastic integrals in the plane.Acta Math., 134:111–183, 1975

work page 1975

[6] [6]

M. G. Cowling, Z. Fan, J. Li, and L. Yan.Characterizations of Product Hardy Spaces on Stratified Groups by Singular Integrals and Maximal Functions, pages 193–227. Springer Nature Switzerland, Cham, 2025

work page 2025

[7] [7]

M. G. Cowling, M.-Y. Lee, J. Li, and J. Pipher. An endpoint estimate for product singular integral operators on stratified lie groups.Canadian Journal of Mathematics, page 1–32, 2025

work page 2025

[8] [8]

R. F. Gundy and E. M. Stein.H p theory for the poly-disc.Proc. Nat. Acad. Sci. U.S.A., 76(3):1026–1029, 1979

work page 1979