Neural Proposals for Exact Markov Chain Monte Carlo

This direction is merged into another direction: local mode entrapment. Its last version stays readable and citable here.

Version 1 of 19 papers16 conceptsfrozen 2026-10-04

Sampling from high-dimensional, multi-modal probability distributions is a fundamental challenge in fields ranging from condensed matter physics to Bayesian inference. Traditional Markov Chain Monte Carlo (MCMC) methods, while mathematically rigorous, often suffer from "slow mixing"—they get stuck in local energy minima and take an eternity to explore the full landscape. Conversely, modern generative AI models like Normalizing Flows and Diffusion models are incredibly fast at generating samples but often suffer from mode collapse or subtle biases that make them unsuitable for high-precision scientific work.

A growing body of research has converged on a powerful hybrid approach: using neural networks as proposal engines within a Metropolis-Hastings framework. This combination offers the best of both worlds—the global exploration capabilities of deep learning and the asymptotic exactness of classical statistics.

The Ergodicity Trap and Global Proposals

In traditional MCMC, such as the Metropolis-Hastings algorithm, the sampler typically moves by making small, local perturbations to the current state. While this ensures the sampler eventually visits all regions of the distribution, it leads to high "autocorrelation"—the samples are so similar to one another that you need millions of them to get a representative picture of the system.

Kim Andrea Nicoli, Pan Kessel, and their colleagues addressed this in Asymptotically unbiased estimation of physical observables with neural samplers. They proposed using neural networks to suggest entire system configurations in a single step. Because these "global proposals" can jump across energy barriers that local updates would never cross, they drastically reduce the integrated autocorrelation time.

The brilliance of this setup is the Metropolis-Hastings correction step. Even if the neural network is imperfectly trained or biased, the MH step acts as a filter, accepting or rejecting the neural proposals to ensure the final sequence of samples follows the exact target distribution. This transforms a "black-box" generative model into a rigorous tool for Asymptotically unbiased estimation.

Bridging Distributions with Neural Transports

One of the most robust ways to sample complex distributions is Parallel Tempering (PT), where multiple versions of a system are simulated at different "temperatures." High-temperature replicas explore the space freely, while low-temperature replicas settle into the peaks. The efficiency of PT depends on how easily these replicas can swap states.

Leo Zhang, Arnaud Doucet, and their team introduced Accelerated Parallel Tempering (APT) to supercharge this process. In Accelerated Parallel Tempering via Neural Transports, they use neural networks—specifically normalizing flows and diffusion models—to learn "transports" that map one distribution to another.

Instead of hoping a high-temperature state happens to look like a low-temperature one, the neural network actively transforms the state to match the target temperature's geometry. This "neural transport" makes swaps much more likely to be accepted, bridging the gap between distributions and achieving orders-of-magnitude better efficiency than classical PT.

Symmetry and Physical Exactness

For these samplers to work in physics, they must often respect the fundamental symmetries of the universe, such as rotation or translation invariance. In Simulating the Hubbard Model with Equivariant Normalizing Flows, D. Schuh and Kim Andrea Nicoli demonstrate how baking these symmetries directly into the neural architecture (equivariance) allows the model to learn the physics of materials like graphene more effectively.

However, even with equivariant architectures, a model might still have slight artifacts. The researchers emphasize that the hybrid MCMC-neural approach is essential here: the neural network provides the "physical intuition" to propose smart moves, while the MCMC layer provides the "mathematical proof" that the results are statistically exact.

Why It Works: The Synergy of Two Worlds

The core intuition behind this research direction is a division of labor:

  1. Neural Networks handle the "Where": They learn the high-probability regions of the space and how to navigate between them, providing high-quality, global proposals.
  2. MCMC handles the "How Much": The Metropolis-Hastings rule uses the known (but perhaps hard-to-sample) energy function of the system to correct for any inaccuracies in the neural network's learned density.

As Kim Andrea Nicoli and colleagues noted in their work on saVANt, this framework allows us to use biased, variational tools to produce unbiased, exact results. By treating deep learning as a proposal engine rather than a direct generator, researchers have unlocked a way to simulate complex physical systems with both the speed of AI and the rigor of traditional science.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_proposals_for_exact_markov_chain__81733b8f,
  title = {Neural Proposals for Exact Markov Chain Monte Carlo},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-proposals-for-exact-markov-chain-monte-carlo/rdr_81733b8f0ed5448fa8f0fcb3aa44323f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_neural_proposals_for_exact_markov_chain__81733b8f_v1,
  title = {Neural Proposals for Exact Markov Chain Monte Carlo},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-proposals-for-exact-markov-chain-monte-carlo/rdr_81733b8f0ed5448fa8f0fcb3aa44323f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}