Neural-Brownian Motion (NBM) Models

• Neural-Brownian Motion (NBM) is a stochastic framework that integrates learned neural operators with Brownian dynamics, enabling adaptive uncertainty and physical system modeling.
• NBM employs neural-parameterized BSDEs, nonlocal directional derivatives, and implicit volatility constraints to boost generalization and data efficiency in deep learning.
• While providing theoretical guarantees and empirical performance, NBM faces ongoing challenges in extending its applications to underdamped regimes and multi-dimensional dynamics.

Neural-Brownian Motion (NBM) encompasses a rigorous family of stochastic models in which Brownian motion-like structures are synthesized or parameterized using neural architectures, with notable applications in uncertainty modeling, physically inspired neural differential equations, and deep learning with stochastic activation functions. The essential feature is the infusion of learnable, data-driven, or nonlocal aspects—typically via neural networks, backward stochastic differential equations (BSDEs), or advanced nonlocal calculus—into the generative laws of Brownian-like noise and dynamics. NBM frameworks offer an interface between machine learning, stochastic analysis, and physical simulation, supporting both theoretical advances in neural modeling of stochasticity and empirical gains in generalization and data efficiency (Nagaraj et al., 2024, Qi, 19 Jul 2025, Bishnoi et al., 2023).

1. Foundational Concepts and Axiomatic Formulations

NBM models depart from classical Brownian motion $W_t$ defined as a linear-expectation martingale, $E[W_t|\mathcal{F}_s]=W_s$, by generalizing the martingale structure through learned, non-linear expectations. The principal axiom of NBM posits a neural expectation operator $\mathbb{E}^\theta$, constructed via a neural-network-driven BSDE: $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$ where $f_\theta$ is parameterized by neural weights $\theta$ and $Y_t = \mathbb{E}^\theta[\xi | \mathcal{F}_t]$ (Qi, 19 Jul 2025). A continuous Itô process $M$ is called a Neural-Brownian Motion if it is a $\mathbb{E}^\theta$-martingale, i.e., $M_s = \mathbb{E}^\theta[M_t|\mathcal{F}_s]$ for $E[W_t|\mathcal{F}_s]=W_s$0, with $E[W_t|\mathcal{F}_s]=W_s$1 and continuous paths.

In another stream (Nagaraj et al., 2024), NBM is realized through the embedding of Brownian sample paths directly into neural architectures via nonlocal directional derivatives (NDDs), providing a rigorous extension of differentiation to highly irregular functions and enabling stochastic neural activation mechanisms.

2. Stochastic Differential Equations and Implicit Volatility

A key structural result of NBM is the representation theorem: a canonical NBM is the unique strong solution to an SDE of the form

$E[W_t|\mathcal{F}_s]=W_s$2

where the volatility function $E[W_t|\mathcal{F}_s]=W_s$3 is not directly specified but uniquely determined through the algebraic constraint $E[W_t|\mathcal{F}_s]=W_s$4, with $E[W_t|\mathcal{F}_s]=W_s$5 derived from the BSDE driver. The function $E[W_t|\mathcal{F}_s]=W_s$6 is $E[W_t|\mathcal{F}_s]=W_s$7 and at most linear in $E[W_t|\mathcal{F}_s]=W_s$8, guaranteed by the differentiability and regularity of the driver under an implicit function theorem (Assumption: Implicit Volatility) (Qi, 19 Jul 2025).

In applied settings, NBM formalism is instantiated by BroGNet (Bishnoi et al., 2023), which learns both drift and diffusion coefficients in overdamped Langevin SDEs from data, with diffusion amplitude and friction parameters $E[W_t|\mathcal{F}_s]=W_s$9, $\mathbb{E}^\theta$0 per particle being outputs of neural networks. Both the algebraic and architecture-driven NBM models provide highly flexible, data-adaptive stochastic processes beyond fixed Brownian dynamics.

3. Nonlocal Calculus and Brownian Neurons

In NBM systems incorporating nonlocal calculus (Nagaraj et al., 2024), the nonlocal directional derivative (NDD) is given for $\mathbb{E}^\theta$1 by

$\mathbb{E}^\theta$2

where $\mathbb{E}^\theta$3 is a kernel sequence concentrating at $\mathbb{E}^\theta$4. This operator generalizes ordinary differentiation and supports first-order nonlocal Taylor expansions, enabling gradient-based optimization even when activations are only Hölder continuous or nowhere classically differentiable—a property satisfied by Brownian sample paths.

Sample paths of Brownian motion $\mathbb{E}^\theta$5 are almost surely nowhere differentiable but are $\mathbb{E}^\theta$6-Hölder continuous for $\mathbb{E}^\theta$7. NDDs of such sample paths are themselves Gaussian processes with explicit mean and variance.

This nonlocal machinery is leveraged in Brownian-ReLU activations: $\mathbb{E}^\theta$8 with $\mathbb{E}^\theta$9 and $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$0. The nonlocal derivative used in backpropagation replaces the standard pseudo-gradient $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$1 with $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$2.

4. Stochastic Calculus, Girsanov-Type Theorems, and Learned Attitudes

For NBM processes $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$3, stochastic calculus tools apply via Itô’s formula, with infinitesimal generator

$$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$4

Beyond standard calculus, a Girsanov-type theorem is established for canonical NBM with quadratic driver: $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$5 implying constant volatility $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$6. A change of measure with the Doléans-Dade exponential introduces a drift that depends on the learned parameter $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$7. $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$8 (convex driver) produces "learned pessimism" (positive drift under the new measure); $$-\,dY_s = f_\theta(s, X_s, Y_s, Z_s)ds - Z_s dW_s,\quad Y_T = \xi$$9 yields "learned optimism" (negative drift), making ambiguity attitude an endogenous output of the neural parameterization (Qi, 19 Jul 2025).

5. Empirical Realizations: Generalization and Physical Systems

In deep learning architectures, NBM-style stochasticity improves generalization, particularly in low-data regimes. Experimental results support that replacing a ReLU layer with Brownian-ReLU in a 3-layer MLP on E-MNIST with $f_\theta$0 of data increases Top-1 accuracy by $f_\theta$1 (from $f_\theta$2 to $f_\theta$3). In adversarially deep MLPs, Brownian–ReLU lifts Top-1 accuracy from $f_\theta$4 to $f_\theta$5 at $f_\theta$6 data. On GLUE benchmarks for text, inserting a single Brownian layer into LoRA-finetuned transformers yields an average $f_\theta$7 enhancement over deterministic counterparts (Nagaraj et al., 2024).

BroGNet demonstrates NBM in modeling Brownian dynamics of multi-particle physical systems:

• Incorporates message-passing GNNs to model drift and diffusion;
• Satisfies exact linear momentum conservation by anti-symmetric force summation;
• Outperforms non-GNN and GNN baselines (by $f_\theta$8–$f_\theta$9) across metrics like trajectory KL divergence, Brownian error, and position error;
• Admits zero-shot generalization to much larger system sizes and different temperatures without retraining (Bishnoi et al., 2023).

6. Theoretical Guarantees and Sample Complexity

For NBM architectures using nonlocal gradients, sample complexity bounds are established. Suppose the objective $\theta$0 satisfies the Polyak-Łojasiewicz condition and nonlocal gradient bias $\theta$1, the expected difference to optimum after $\theta$2 iterations is

$\theta$3

with $\theta$4 (Nagaraj et al., 2024). These guarantees rely on the nonlocal gradients acting as $\theta$5-subgradients with controlled bias and variance, ensuring the viability of stochastic optimization in highly irregular or nonconvex settings.

7. Limitations and Open Directions

NBM frameworks, while yielding empirical and theoretical advances, exhibit limitations:

• Overdamped SDEs are primarily modeled; extensions to full underdamped/Langevin regimes remain open (Bishnoi et al., 2023).
• Euler–Maruyama is standard for integration; higher-order SDE solvers could improve fidelity.
• Interpretability of learned interaction laws and further incorporation of physical symmetries (e.g., rotational equivariance) are ongoing areas (Bishnoi et al., 2023).
• Extensions to multi-dimensional and vector-valued NBM require additional care for BSDE solvability and uniqueness (Qi, 19 Jul 2025).

NBM unifies approaches where Brownian stochasticity is itself neural, adaptive, or nonlocally regularized, endowing models with data-driven stochastic dynamics while retaining control over regularity and computational tractability. Empirical studies indicate that such neural stochasticity systematically regularizes learning, especially when traditional deterministic models are prone to overfit due to limited data.