SRMC: A Breakthrough in Constant-Memory Non-Markovian Samplers

The 'History Forgetting' Dilemma of Monte Carlo Sampling

Monte Carlo methods are among the most essential computational tools in statistical inference and machine learning, appearing virtually everywhere from Bayesian inference to generative model training. However, traditional Markov Chain Monte Carlo (MCMC) methods suffer from a fundamental efficiency bottleneck — the Markov property requires the sampler to be 'memoryless,' relying solely on the current state to determine the next step. This means the sampler may repeatedly visit regions that have already been thoroughly explored, resulting in substantial redundant computation and significantly increased long-term variance.

A recent paper published on arXiv (arXiv:2604.22948) introduces a novel framework called "Score-Repellent Monte Carlo" (SRMC), offering an elegant solution to this classic problem. The method compresses trajectory history information through sliding averages of score functions, achieving efficient non-Markovian sampling with only constant memory requirements, and is directly applicable to general continuous state spaces.

Core Idea: Using Score Functions to 'Repel' Visited Regions

The central concept of history-dependent sampling is not new — by recording the sampler's historical trajectory, it actively 'repels' visited regions, thereby encouraging the sampler to more thoroughly explore the support of the target distribution and reducing long-term Monte Carlo variance.

However, the bottleneck of existing methods lies in how to encode history. Traditional approaches typically rely on maintaining empirical measures over finite state spaces — that is, recording the visit frequency of each state. This approach becomes untenable in two types of scenarios:

• High-dimensional discrete spaces: The number of states grows exponentially, making it infeasible to store empirical measures in memory;
• Continuous state spaces: Empirical measures themselves are ill-posed and cannot be directly applied.

The key innovation of the SRMC framework is that it no longer attempts to record complete empirical measures, but instead compresses trajectory history into sliding averages of score function evaluations. The score function (i.e., the gradient of the log-density) is a vector residing in $\mathbb{R}^d$, so regardless of how complex the original state space may be, the history summary is always a fixed-dimensional vector requiring only constant-level memory.

The sampler uses this history summary to construct a 'repulsive force': when the sliding average indicates that the sampler is repeatedly lingering in a certain region, the repulsion mechanism guides the sampler toward regions that have not yet been sufficiently explored. This mechanism breaks the Markov property, endowing the sampler with global exploration capabilities while maintaining extremely low computational and storage overhead.

Technical Advantages and Theoretical Significance

Constant memory complexity is SRMC's most prominent engineering advantage. Traditional non-Markovian samplers (such as adaptive MCMC, the Wang-Landau algorithm, etc.) often need to store historical data that grows linearly with time, or maintain data structures proportional to the size of the state space. SRMC compresses historical information into a fixed-size statistic, allowing the algorithm to run indefinitely without increasing memory burden — a significant benefit for large-scale computational scenarios.

Applicability to general state spaces is another major highlight. By shifting history encoding from the state space to the score space ($\mathbb{R}^d$), SRMC is naturally suited for continuous distribution sampling without requiring discretization or binning preprocessing steps. This allows it to be directly embedded into modern probabilistic programming frameworks and deep generative model training pipelines.

From a theoretical perspective, SRMC provides a new paradigm for designing non-Markovian samplers. It demonstrates that score functions are not only core tools for generative modeling methods such as diffusion models and score matching, but can also serve as efficient compressed representations of sampling history, playing a critical role in the MCMC domain.

Connections to Current Research Trends

Notably, SRMC's theoretical foundations are closely aligned with score matching and diffusion models, which have achieved remarkable success in the AI field in recent years. As an implicit representation of probability distributions, the score function has already proven its powerful expressive capacity in areas such as image generation and molecular simulation. The introduction of SRMC further expands the application landscape of score functions, extending them from generative modeling to Monte Carlo inference — a more fundamental computational level.

Furthermore, with the growing demand for Bayesian deep learning and large-scale probabilistic inference, research on efficient MCMC methods is regaining attention. SRMC's approach of 'compressing history with score functions' may inspire more work that brings modern deep learning tools into classical statistical computation.

Outlook: Toward Smarter Samplers

The SRMC framework is currently still in the theoretical research stage, and its performance on large-scale practical problems awaits further experimental validation. However, its core idea — achieving effective history-dependent sampling under constant memory constraints — points to a promising direction for the development of Monte Carlo methods.

In the future, SRMC is expected to be combined with adaptive step-size strategies, parallel sampling techniques, and neural network-parameterized transition kernels to build more intelligent and efficient sampling systems. In fields such as AI model training, scientific computing, and uncertainty quantification, breakthroughs in such methods could yield significant efficiency improvements. For researchers and engineers pushing the limits of computational efficiency, this paper is well worth close attention.