From oracle maximal inequalities to martingale random fields via finite approximation from below

Abstract: A novel approach is proposed to establish a sharp upper bound on the expected supremum of a separable martingale random field, serving as an alternative to classical universal chaining-based methods. The proposed approach begins by deriving a new "oracle maximal inequality" for a finite class of submartingales. This is achieved via integration by parts rather than a simplistic application of the triangle inequality. Consequently, we obtain a generalization of Lenglart's inequality for discrete-time martingales, extending it from the one-dimensional case to finite-dimensional settings, and further to certain infinite-dimensional cases through a "finite approximation device". The primary applications include several weak convergence theorems for sequences of separable martingale random fields under the uniform topology. In particular, new results are established for i.i.d. sequences, including a necessary and sufficient condition for a countable class of functions to possess the Donsker property. Additionally, we provide new moment bounds for the supremum of empirical processes indexed by classes of sets or functions.

• The paper introduces oracle maximal inequalities to provide sharp bounds on the suprema of martingale random fields.
• It employs a novel finite approximation approach that bypasses traditional entropy methods via integration by parts.
• The methodology yields new Donsker-type results and advances empirical process theory without parasitic log factors.

Oracle Maximal Inequalities and Martingale Random Fields via Finite Approximation from Below

Introduction and Context

The paper "From oracle maximal inequalities to martingale random fields via finite approximation from below" (2603.29739) introduces a methodology for analyzing the supremum of separable martingale random fields without relying on entropy-based arguments typical in classical empirical process theory. The core contributions are twofold: (i) the derivation of oracle maximal inequalities (OMIs) for finite classes of submartingales and their quadratic forms, and (ii) the establishment of a finite approximation device which allows asymptotic results for infinite-dimensional martingale random fields to be built from finite-dimensional cases in a rigorous way. This approach bypasses the standard chaining and entropy integrability techniques, offering sharper or entropy-free sufficient conditions for weak convergence and Donsker-type results.

Oracle Maximal Inequalities: Structure and Consequences

The oracle maximal inequalities (OMIs) constitute a set of sharp maximal bounds for the supremum of finite classes of martingales and related processes. Notably, the derivation eschews the classical triangle inequality approach, instead using integration by parts and martingale transforms that preserve structure and avoid log-type factors introduced by entropy bounds.

A principal result is the OMI for the squares of martingales. For a finite class $\{ \xi^i \}_{i \in I}$ of one-dimensional martingale difference sequences, the quadratic maximal process can be controlled as follows (see Lemma: OMI for squares of martingales): \begin{align*} \max_{i \in I}\left( \sum_{k=1}^n \xi_k^i \right)^2 &\leq 4 \sum_{k=1}^n \max_{i \in I} \mathbb{E}[ (\xi_k^i)^2 | \mathcal{F}{k-1} ] + 4 \max{1 \leq k \leq n} \max_{i \in I} \mathbb{E}[ (\xi_k^i)^2 | \mathcal{F}_{k-1} ] + \text{(martingale terms)}, \end{align*} almost surely. This form parallels but improves on classical maximal inequalities (e.g., Doob's, Lenglart's), since it scales via maxima of conditional variances instead of sums or entropy bounds. The result is robust under finite approximation and is the cornerstone for the subsequent infinite-dimensional extension.

The Finite Approximation Device

To bridge the finite and infinite cases, the author introduces a precise formulation of supremum expectation operations: for a random field $X = \{ X(\theta) : \theta \in \Theta \}$ and $p \geq 1$,

$$\widetilde{\mathbb{E}}\left[ \sup_{\theta \in \Theta} |X(\theta)|^p \right] := \sup_{F \subset \Theta, |F|<\infty} \mathbb{E}\left[ \max_{\theta \in F} |X(\theta)|^p \right].$$

Rigorous analysis shows that for separable random fields over totally bounded index sets, the supremum of the expectations aligns closely (within constants) with the expectation of the true supremum (Lemma: finite approximation device). This observation justifies carrying supremum inequalities from finite to infinite classes, provided separability and total boundedness are present.

Donsker Theorems without Entropy Conditions

A central theoretical outcome is the establishment of necessary and sufficient conditions for a class $\mathcal{H} \subset L^2(P)$ to be $P$-Donsker, formulated purely in geometric terms of total boundedness (covering numbers) with respect to the intrinsic $L^2$ pseudometric, and the (mild) separability assumption for the empirical process. Formally, for any countable $\mathcal{H}$ with envelope $H\in L^2(P)$, it is $P$-Donsker if and only if

$X = \{ X(\theta) : \theta \in \Theta \}$0

where $X = \{ X(\theta) : \theta \in \Theta \}$1 is the $X = \{ X(\theta) : \theta \in \Theta \}$2-covering number and $X = \{ X(\theta) : \theta \in \Theta \}$3 is the natural $X = \{ X(\theta) : \theta \in \Theta \}$4 pseudometric. Notably, no entropy integral or local bracketing conditions are needed (Theorem: new Donsker theorem), in contrast with classical entropy methods. Moreover, the weak convergence results (CLT) extend to empirical processes and random fields indexed by general metric or pseudometric spaces when total boundedness and separability hold.

Implications for Empirical Process Theory and High-Dimensional Statistics

The theoretical developments have immediate applications to empirical process theory, stochastic process limit theorems, and high-dimensional statistics:

• The OMI-driven approach provides new moment bounds for the supremum of empirical processes, crucial for rates of convergence of $X = \{ X(\theta) : \theta \in \Theta \}$5-estimators and for theoretical guarantees in high-dimensional inference (Theorem: supremal inequalities).
• For classes $X = \{ X(\theta) : \theta \in \Theta \}$6 not totally bounded (e.g., certain indicator function classes), the analysis distinguishes between finite supremum expectations $X = \{ X(\theta) : \theta \in \Theta \}$7 and actual supremum expectations $X = \{ X(\theta) : \theta \in \Theta \}$8, revealing subtle failures of naive monotone convergence, and motivating care in empirical process approximations (see Example: Suzuki's counterexample).
• The entropy-free Donsker theorems enable new weak convergence and CLT results in stochastic process settings, as in martingale difference arrays and stochastic process models for statistical inference.

Further Theoretical and Practical Implications

The analysis highlights that, for separable and totally bounded random fields, one may avoid the often rough entropy inequalities and obtain sharp constants and sharper asymptotic results (no "parasitic logarithmic factor" as in Talagrand). This has theoretical consequences for the study of Gaussian random field regularity (Talagrand-Fernique majorizing measures phenomenon) and for the sharpness of tail and moment inequalities.

On the practical side, these advances have potential impacts on the statistical theory of high-dimensional inference (e.g., LASSO, Dantzig selector for stochastic processes), where controlling suprema of empirical processes is central. The new methodology could yield better rates and more refined probabilistic bounds, especially for stochastic models where the natural geometry is induced by data.

Conclusion

This work presents a significant, technically sophisticated approach for controlling the supremum of martingale random fields and empirical processes, replacing entropy-based arguments by OMIs and a robust finite approximation device. The resulting characterization of the Donsker property and the associated weak convergence theory for empirical processes indexed by general classes set a new standard in empirical process theory, with both theoretical and practical ramifications for stochastic analysis and high-dimensional statistics. The techniques presented highlight the nuanced geometric nature of infinite-dimensional probability, stressing that separability and total boundedness—rather than entropy integrability alone—can suffice to guarantee limit theorems and sharp supremum bounds.

Further research may consider generalizations beyond martingale-dominated processes, extensions to continuous-time semimartingales, or the adaptation of OMI machinery to dependent structures and non-identically distributed models. Integrating these tools with model selection, regularization, and other statistical tasks remains a promising avenue.