Martingale Construction & Asymptotic Approximation

• Martingale construction is a probabilistic framework that represents stochastic processes through martingale identities and limit theorems, facilitating precise asymptotic analysis.
• It enables rigorous derivation of convergence rates, norm inequalities, quadratic form approximations, and higher-order expansions across various models.
• The approach is applied in contexts ranging from singular integrals and random walks to Pólya urns and martingale posterior inference, impacting both theory and practice.

Martingale construction encompasses a collection of probabilistic methods for representing and analyzing stochastic processes through martingale structures, primarily to derive sharp asymptotic approximations for distributions, rates of convergence, or limiting behaviors in both discrete and continuous settings. Asymptotic approximation in this context refers to precise quantitative descriptions—often including limiting distributions, rates, expansions, or tail asymptotics—attained via martingale representations and associated limit theorems or identities. The interplay between martingale constructions and asymptotic analysis is central to modern probability theory, stochastic process analysis, mathematical statistics, singular integral theory, and the probabilistic study of partial differential equations.

1. Martingale Identities and Cotlar’s Structure

A fundamental illustration of martingale construction appears in the context of singular integral operators, notably in the "Cotlar martingale transforms." The Cotlar identity, originally formulated for the Hilbert transform $H$ on $\mathbb{R}$,

$|Hf|^2 = 2 H(f Hf) + |f|^2,$

admits a nontrivial martingale analogue. For a Brownian-motion driven martingale

$M_t = M_0 + \int_0^t K_s \cdot dB_s,$

and a $d \times d$ matrix $A$ satisfying $A v \cdot v = 0 \ \forall v$, the Cotlar martingale identity reads

$$|A * M_t|^2 = 2 A * (M(A * M))_t + \int_0^t |A K_s|^2 ds - 2 \int_0^t [M_s A^2 K_s] \cdot dB_s.$$

If $A^2 = -I$ (a "Cotlar matrix"), this reduces to

$|A * M_t|^2 = 2 A * (M(A * M))_t + M_t^2,$

from which sharp $\mathbb{R}$0 operator-norm inequalities can be derived at dyadic exponents via trigonometric induction, mirroring those for the Hilbert transform. The same approach, extended to the Riesz transforms via conformal martingale transforms (notably in odd dimensions through block-diagonal Cotlar matrices), establishes both analytic structure and asymptotic norm equivalence as $\mathbb{R}$1, specifically

$\mathbb{R}$2

for the Riesz vector and the one-dimensional Hilbert transform (Bañuelos, 10 Apr 2026).

This structure reveals deep algebraic cancellations in singular integral theory and provides new avenues for longstanding open problems—such as sharp constants in the vector Riesz transforms and the Beurling–Ahlfors operator norm—while bypassing the method of rotations and Bellman function techniques.

2. Martingale Constructions in Stochastic Approximation and Convergence Rates

General martingale and supermartingale constructions are central to the quantitative analysis of stochastic processes, especially in the context of stochastic approximation, convergence rates, and explicit finite-sample bounds. Consider a nonnegative adapted process $\mathbb{R}$3 satisfying a relaxed supermartingale recursion: $\mathbb{R}$4 with summable error and fluctuation sequences. Introducing "slowing-down" functions—concave, strictly increasing, and supermultiplicative such as $\mathbb{R}$5 or $\mathbb{R}$6—enables analysis of

$\mathbb{R}$7

and, via Ville's inequality and Doob–Ville supermartingale constructions, explicit rates for the $\mathbb{R}$8 convergence and almost sure deviation probabilities: $\mathbb{R}$9 where $|Hf|^2 = 2 H(f Hf) + |f|^2,$0 are computable from the process parameters and error moduli (Neri et al., 17 Apr 2025).

Applications include:

• Robbins–Siegmund convergence rates: Explicit $|Hf|^2 = 2 H(f Hf) + |f|^2,$1 rates in mean and matching tail bounds for classic stochastic-approximation schemes when strong monotonicity or contractivity is present.
• Dvoretzky’s theorem and stochastic quasi-Fejér monotonicity: Two-stage applications and generalizations to metric spaces with deterministic or stochastic perturbations.
• Quasi-Fejér monotonicity: Generalizes the convergence paradigm to metric spaces via stochastic descent inequalities, yielding effective rates under quadratic contraction.

3. Martingale Approximations for Quadratic Forms and Dependent Data

Quadratic forms of Markov chains or time-series processes often admit martingale approximations via Poisson equation-based construction, yielding functional central limit theorems and refined asymptotic expansions. For

$|Hf|^2 = 2 H(f Hf) + |f|^2,$2

with $|Hf|^2 = 2 H(f Hf) + |f|^2,$3 symmetric, the canonical decomposition

$|Hf|^2 = 2 H(f Hf) + |f|^2,$4

isolates a martingale-difference array $|Hf|^2 = 2 H(f Hf) + |f|^2,$5 and a negligible remainder $|Hf|^2 = 2 H(f Hf) + |f|^2,$6 under suitable geometric-ergodicity conditions and moment bounds (Atchade et al., 2011).

This approach applies directly to:

• Lag-window estimators for long-run variance in time series, where the nondegenerate term is shown to be asymptotically Gaussian or driven by Brownian functionals;
• U-statistics with varying kernels, with CLTs under general dependence provided the second-order components are negligible.

4. Martingale Expansions and Higher-Order Asymptotics

Martingale expansions support higher-order (e.g., Edgeworth-type) asymptotic approximations in both normal and mixed-normal settings. For vector martingales $|Hf|^2 = 2 H(f Hf) + |f|^2,$7, under stable convergence and Malliavin calculus regularity, Yoshida’s random symbol method gives

$|Hf|^2 = 2 H(f Hf) + |f|^2,$8

with $|Hf|^2 = 2 H(f Hf) + |f|^2,$9 the normal density. The expansion utilizes the adaptive and anticipative random symbols arising from conditional variances and Malliavin derivatives of limit objects (e.g., realized volatility functionals of diffusions) (Yoshida, 2012, Yoshida, 2012).

Applications include bias correction and coverage improvements in inference about volatility, option pricing, and integrated variance in ergodic and non-ergodic models.

5. Martingale-Based Construction for Discrete Models and Random Walk Extremes

Explicit martingale functionals are essential for deriving subexponential and heavy-tail asymptotics in random walk theory. In the context of a random walk $M_t = M_0 + \int_0^t K_s \cdot dB_s,$0 with negative drift and heavy-tailed increments, consider the integrated tail functional

$M_t = M_0 + \int_0^t K_s \cdot dB_s,$1

with $M_t = M_0 + \int_0^t K_s \cdot dB_s,$2. The key is to show that $M_t = M_0 + \int_0^t K_s \cdot dB_s,$3 is a sub-/supermartingale, allowing an application of the Optional Stopping Theorem to produce tight asymptotics for the maximum: $M_t = M_0 + \int_0^t K_s \cdot dB_s,$4 and for the excursion maximum,

$M_t = M_0 + \int_0^t K_s \cdot dB_s,$5

This method circumvents renewal theory and direct large-deviation estimates, relying purely on martingale properties and the optional stopping framework (Denisov et al., 2011).

6. Martingale Construction in Combinatorial Stochastic Models

In generalized Pólya urn processes and partition-structure models (e.g., Ewens–Pitman), martingale constructions provide direct access to precise limiting theorems. For Pólya urns, after explicit normalization,

$M_t = M_0 + \int_0^t K_s \cdot dB_s,$6

is a vector-martingale capturing the principal fluctuation direction. The regime $M_t = M_0 + \int_0^t K_s \cdot dB_s,$7 governs limiting behavior, from central limit theorems and laws of large numbers (for “small urns”) to superdiffusive scaling (for “large urns”) (Laulin, 2020). In the Ewens–Pitman model with linearly-scaling diversity ($M_t = M_0 + \int_0^t K_s \cdot dB_s,$8), a suitably normalized process is a mean-one martingale, yielding LLN, CLT, and Berry–Esseen bounds in both classical and reinforced ($M_t = M_0 + \int_0^t K_s \cdot dB_s,$9) regimes (Ribeiro, 25 Mar 2025).

7. Maximal Inequalities, Martingale Random Fields, and Empirical Process Theory

Refined maximal inequalities for martingales underpin tight control of infinite-dimensional function-indexed stochastic processes. Recent advances establish that for separable martingale random fields totally bounded in $d \times d$0,

$d \times d$1

linking finite approximation with the expected supremum (Nishiyama, 31 Mar 2026). Oracle maximal inequalities (via discrete integration by parts) for finite classes extend to infinite-dimensional inequalities (generalized Lenglart), which in turn yield entropy-free Donsker and central limit theorems for empirical and martingale-indexed processes.

8. Martingale Posterior Inference and Predictive CLT

The martingale posterior framework generalizes Bayesian inference to sequential predictive settings by recursively updating predictive densities. Parametric and nonparametric martingale posteriors ensure the key property

$d \times d$2

leading to strong consistency and (by martingale CLT) Bernstein–von Mises results. Predictive CLTs enable hybrid resampling schemes: the terminal posterior can be efficiently approximated via a single Gaussian increment after a moderate number of predictive updates, achieving both computational and statistical efficiency in large-scale or online inferential settings (Fong et al., 2024, Fong et al., 2024).

---

In summary, martingale construction is a unifying technique for both the rigorous representation of stochastic processes and the extraction of sharp, quantitative asymptotic approximations. Its scope encompasses analytic identities (e.g., Cotlar’s trick in harmonic analysis), discrete and continuous time probabilistic models (urn schemes, Markov chains), high-frequency data asymptotics (quadratic forms, realized volatility), infinite-dimensional random field theory, and contemporary algorithmic statistics (martingale posteriors, sequential inference). Explicit constructions and corresponding asymptotic results continue to drive advances in the precise quantitative understanding of complex stochastic systems.