Stochastic Variant via Supermartingale Theory

• The topic defines stochastic variants via supermartingale theory, where adapted processes maintain non-increasing conditional expectations to secure strong convergence and risk bounds.
• It applies these principles to optimal transport, model checking, and safety verification, enabling explicit quantitative guarantees in complex stochastic systems.
• The methodology underpins accelerated optimization and financial models by leveraging explicit descent recursions and probabilistic certificates for robust algorithmic performance.

A stochastic variant via supermartingale theory refers broadly to stochastic process, optimization, or verification frameworks that exploit the supermartingale property—adapted processes where the conditional expectation of the next value is not greater than the present—to obtain strong probabilistic guarantees, convergence rates, and quantitative risk or safety certificates. This methodology appears across stochastic approximation, stochastic optimal transport, probabilistic model checking, verification of safety specifications, and accelerated optimization, with quantitative control over sample-path behavior, rates, and rare-event probabilities. Below, key developments and applications are organized to detail the mathematical principles, representative results, and algorithmic methodologies for stochastic variants via supermartingale theory.

1. Supermartingale Foundations in Stochastic Analysis

The supermartingale property for an adapted process $(X_n, \mathcal{F}_n)$ is the core structure: $$\mathbb{E}[ X_{n+1} \mid \mathcal{F}_n ] \leq X_n \quad \text{a.s.}$$ This sub-expectation property underpins a wide range of stochastic process convergence theorems (Doob's supermartingale convergence), stopping-time inequalities, and optional sampling results. Relaxed or perturbed forms, for instance

$$\mathbb{E}[ X_{n+1} \mid \mathcal{F}_n ] \leq (1 + A_n) X_n + C_n,$$

enable fine-grained, non-asymptotic control on stochastic iterative schemes, covering error accumulation and drift (Neri et al., 17 Apr 2025).

Such structure leads to key results:

• Almost sure convergence with explicit (often uniform) rates depending only on bounds for $(A_n)$ and summability of $(C_n)$.
• Quantitative versions of classical stochastic approximation theorems (e.g., Robbins–Siegmund, quasi-Fejér convergence, Dvoretzky's theorem) with explicit, constructive rates for iterates and function values (Neri et al., 17 Apr 2025).
• Extension to coupled processes and delayed updates, crucial for modern accelerated methods (Ming-Kun, 2024).

2. Stochastic Variants in Optimal Transport and Coupling Theory

Supermartingale coupling generalizes classical optimal transport by imposing the constraint $\mathbb{E}[Y \mid X] \leq X$ (second-order convex–decreasing stochastic order). The stochastic variant of canonical couplings (Hoeffding–Fréchet, antitone) is constructed using the shadow measure framework, resulting in two canonical supermartingale couplings: increasing and decreasing supermartingale transports (Nutz et al., 2016).

Notable characteristics:

• Supermartingale transports are optimizers for constrained Monge–Kantorovich problems with inequality constraints, captured via duality with “slope” variables $h(x) \ge 0$ and supporting dual potentials $(\varphi, \psi)$ (Nutz et al., 2016).
• Decomposition into martingale and classical transport regions, determined by where the constraint binds, yields structured no-crossing geometric properties for support (Nutz et al., 2016).
• These couplings interpolate between classical (unconstrained) and martingale transports, forming the stochastic (supermartingale) variant of the classical transport theory.

The continuous-time supermartingale Brenier theorem (Bayraktar et al., 2022) extends these ideas, constructing pure-jump processes as limits of Markovian iterations of one-step supermartingale couplings. Limit points are identified as weak solutions to local Lévy-type SDEs, optimal for specific path-dependent cost functionals, tying supermartingale optimal transport to stochastic process laws in the sense of Carr–Geman–Madan–Yor.

3. Quantitative Verification, Certificates, and Model Checking

Supermartingale-based stochastic invariants enable rigorous quantitative probability bounds for invariance-type properties in stochastic systems. For a shift-invariant specification $\varphi$, one constructs a measurable function $V: S \to [0,\infty)$, the stochastic invariant, such that: $$P V(s) \le V(s) \quad \forall s \in I, \qquad V(s) \ge 1 \quad \forall s \notin I,$$ yielding the probability bound $P_\mu[\text{exit } I] \leq \mu V$ (Abate et al., 7 Apr 2025). This approach generalizes classical (almost-sure) supermartingale certificates to precise, quantitative estimates for $\omega$-regular, LTL, or reach-avoid properties (Henzinger et al., 24 May 2025).

Key points:

• For arbitrary shift-invariant properties, completeness is ensured up to any desired $\varepsilon > 0$ in the general case and exactly in finite-state settings (Abate et al., 7 Apr 2025).
• Supermartingale certificates extend to the quantitative regime by coupling stochastic invariants with almost-sure arguments, lifting qualitative safety proofs to quantitative guarantees.
• Algorithmic synthesis uses sum-of-squares, SMT, or quantified entailment to generate polynomial stochastic invariants, automata embeddings, and optimal controllers (Henzinger et al., 24 May 2025, Abate et al., 7 Apr 2025).

4. Safety Certification for Stochastic Systems

Discrete- and continuous-time stochastic dynamical systems admit finite-horizon probabilistic safety certificates via barrier functions constructed so that their composition with system dynamics forms a supermartingale (Mathiesen et al., 2022, Neustroev et al., 2024). For instance, a barrier function $B(x)$ satisfying: $$\mathbb{E}[ B(X_{k+1}) \mid X_k = x ] \leq B(x) + \beta \quad \forall x\in S,$$ bounds the probability of unsafe events by $\gamma + \beta H$ over horizon $H$ (Mathiesen et al., 2022). Neural network parameterizations, bound-propagation training, and piecewise linear relaxations extend this framework to highly nonlinear/nonconvex domains, outperforming sum-of-squares approaches.

For SDEs, a neural supermartingale certificate $V(x;\theta)$ with generator inequality $\mathcal{L}_x V(x;\theta) \leq -\lambda(x)$ quantifies reach-avoid probabilities, with formal verification using interval-bound propagation in the state space (Neustroev et al., 2024).

Empirical studies show tight quantitative lower bounds that closely track empirical Monte Carlo estimates, and outperform discrete-time or polynomial-baseline certificates on classical control benchmarks (Mathiesen et al., 2022, Neustroev et al., 2024).

5. Stochastic Optimization: Algorithms and Convergence via Supermartingale Arguments

Stochastic optimization methods, including splitting proximal point, subgradient, and accelerated schemes with Nesterov momentum, utilize supermartingale-type inequalities to establish almost sure convergence and explicit rates (Brito et al., 11 Jan 2026, Ming-Kun, 2024, Neri et al., 17 Apr 2025). For a stochastic variant of the splitting proximal point algorithm, the key descent recursion: $$\mathbb{E}[ \|x_{k+1} - x^* \|^2 \mid \mathcal{F}_k ] \leq \|x_k - x^* \|^2 - \alpha_k [f(x_k) - f(x^*)] + W_k,$$ enables application of the supermartingale convergence theorem, showing $x_k \to x^*$ almost surely (Brito et al., 11 Jan 2026).

Innovations:

• Delayed supermartingale lemmas, featuring companion-matrix formalisms, allow treatment of momentum and delayed-feedback algorithms, guaranteeing almost sure convergence under classical stepsize conditions (Ming-Kun, 2024).
• Quantitative rates are uniform, depending only on summation bounds for error and drift terms, easily computed for practical algorithms (Neri et al., 17 Apr 2025).
• These results seamlessly extend to composite, proximal, and block-coordinate stochastic approximation schemes, underpinning the convergence of a wide array of modern iterative algorithms.

6. Supermartingale Methods in Financial and Statistical Models

Supermartingale frameworks capture risk, growth, and ruin behavior in stochastic games or investment models, as illustrated by the martingale/supermartingale classification of the Kelly criterion (Miller, 24 Feb 2025). For wealth process $\{W_n\}$, betting fractions above a regime-specific threshold $F_*$ ensure the process is a supermartingale, leading to almost sure ruin, while sub-threshold strategies yield submartingale behavior and exponential growth at rate $U(F, p) > 0$.

Mathematically,

$$\mathbb{E}[W_{n+1} \mid \mathcal{F}_n] \leq W_n \quad \text{if and only if } U(F, p) < 0,$$

where $U(F, p)$ is the average log-utility, underpinning rigorous exponential rates of growth or decay based on betting strategy.

7. Extensions: Strong Supermartingale Envelopes and Variational Inequalities

The construction of strong supermartingale envelopes (Mertens/Snell envelopes) and their Doob–Meyer decompositions undergird probabilistic representation and solution of reflected backward stochastic differential equations (BSDEs) with lower barriers (Aazizi et al., 2011). The minimal supermartingale dominating an obstacle provides a pathwise solution to stochastic variational inequalities, characterized by Skorokhod conditions, and ensures both existence and uniqueness under flexible regularity of the barrier.

This theory provides a foundation for pathwise reflection mechanisms in stochastic control, finance, and PDE theory.

---

Stochastic variants via supermartingale theory unify a wide array of quantitative methods in stochastic analysis, optimization, verification, and control. The underlying mathematical principles—supermartingale inequalities, envelope constructions, and coupling—enable construction of explicit, rigorous, and often optimal probabilistic guarantees across domains, from algorithmic rates in high-dimensional optimization to safety certificates and robust stochastic transport couplings.