Suppose you want to use a neural network to emulate a stochastic partial differential equation (SPDE)—say, a physical field evolving under random noise. Traditional neural operators (e.g., the Fourier Neural Operator, FNO) train quickly and produce a prediction in a single forward pass.
But there's a problem: if your PDE is stochastic—tiny fluctuations in initial conditions cause huge differences in outcomes—a neural operator only outputs the conditional mean. It tells you "on average, the temperature at this location is 23.5 degrees." It does not tell you "there's a 10% chance the temperature exceeds 40 degrees." No variance, no tail risk. For uncertainty quantification, that's nearly the same as saying nothing.
The usual fix is to attach a generative model—a diffusion model, a flow model—to learn the conditional distribution. But that sacrifices the most attractive property of neural operators: single-forward-pass efficiency and resolution invariance.
Hidajat—the same author who recently published a grokking paper on arXiv—recently proposed a new architecture (arXiv:2605.15806) that solves this in a surprising way: encode a probability theorem into the neural network's architecture.
The Doob-Meyer Decomposition
The Doob-Meyer theorem is a core result of martingale theory developed in the 1960s. It states: any semimartingale (a sufficiently general class of stochastic processes) can be uniquely decomposed into a predictable drift term and an unpredictable zero-expectation martingale.
> "Semimartingale" sounds intimidating, but it just means "a stochastic process with a regular drift part and a completely random noise part." Stock prices are an example—price movements decompose into a "fundamentals-driven trend" (drift) and "unpredictable market noise" (the martingale part).
What Hidajat did is simple: treat this decomposition as an architecture design principle. The neural operator backbone predicts the drift term—the conditional mean. A lightweight low-rank branch predicts the martingale term—the conditional covariance.
The key word is "lightweight." The standard approach trains a separate generative model to predict the full conditional distribution. MNO instead has the network output mean and covariance simultaneously, with the covariance parameterized via a low-rank factor B_φ, so that B_φ^T B_φ is positive semidefinite by construction—no extra constraints needed.
Striking Results
They tested on three task families: 1D stochastic PDEs, rough volatility (a stochastic process from finance), and 2D neural operator tasks.
Results: on the φ⁴ field theory standard benchmark, MNO's Wasserstein distance is 120x better than a conditional diffusion baseline. On the stochastic Burgers equation, 68x better. And it runs roughly 3x faster than the conditional diffusion baseline—under the same training budget.
On 2D tasks, MNO matches standard FNO on zero-shot resolution transfer and turbulence simulation.
One interesting failure mode: "quasi-deterministic systems" like the Gray-Scott reaction-diffusion equation. In these systems the process has very little stochasticity and is dominated by deterministic dynamics. MNO's martingale decomposition offers little advantage here—because there's barely any randomness to learn.
What Remains Unclear
First, the Doob-Meyer decomposition is defined for temporally-indexed stochastic processes. But in many SPDE emulation tasks, the "randomness" may live in initial conditions, boundary conditions, or driving noise. MNO applies the Doob-Meyer architecture to the map "from initial conditions to terminal distribution"—effectively treating initial conditions as the drift and packing all uncertainty into the martingale. Does this map always preserve the Doob-Meyer decomposition structure mathematically? I'm not sure, since the path between initial conditions and terminal distribution can be highly nonlinear.
Second, the low-rank covariance assumption. B_φ^T B_φ is a low-rank matrix—meaning it assumes the stochastic uncertainty is concentrated in a low-dimensional subspace. For many physical systems this may hold: the dominant uncertainty modes are indeed few. But which systems violate it? Systems where uncertainty energy is spread uniformly across all modes—fully developed turbulence, for instance—may not be well captured by a low-rank covariance.
Third, the paper uses a "Gaussian residual instantiation"—assuming conditional Gaussianity of the terminal distribution. Many physical systems have non-Gaussian conditional distributions: bimodal, skewed, heavy-tailed. The Gaussian assumption simplifies covariance parameterization but limits the ability to capture non-Gaussian structure.
Still, overall this is an elegant architectural idea: rather than treating uncertainty quantification as a bolt-on step after training, encode it into the network's architecture itself. A 1960s theorem from financial mathematics, put to work in 2026 accelerating SPDE simulation.
---
References
1. Hidajat, K. (2026). *Martingale Neural Operators: Learning Stochastic Marginals via Doob-Meyer Factorization*. arXiv:2605.15806 [cs.LG]. https://arxiv.org/abs/2605.15806 2. Doob, J. L. (1953). *Stochastic Processes*. Wiley. 3. Li, Z., et al. (2021). *Fourier Neural Operator for Parametric Partial Differential Equations*. ICLR 2021. 4. Kovachki, N., et al. (2023). *Neural Operator: Learning Maps Between Function Spaces*. JMLR, 24(1), 1-97. 5. Meyer, P. A. (1966). *Probability and Potentials*. Blaisdell.