Diffusion Path Samplers via Sequential Monte Carlo

James Matthew Young, Paula Cordero-Encinar, Sebastian Reich, Andrew Duncan, O. Deniz Akyildiz

TL;DR

Diffusion Path Sequential Monte Carlo (DPSMC) is a new sampling algorithm that brings the power of generative diffusion models to the world of unnormalized probability distributions (like Bayesian posteriors). By using Sequential Monte Carlo (SMC) to estimate score functions on the fly and introducing mathematically optimal "control variate" schedules to minimize variance, DPSMC achieves state-of-the-art sampling performance with massive potential for parallelization on modern hardware.

What Problem Does This Paper Address?

Sampling from complex, high-dimensional distributions $\pi \propto e^{-V_\pi}$ is a cornerstone of Bayesian inference and statistical physics. Traditional methods often use "annealing," which creates a geometric path from a simple distribution (like a Gaussian) to the complex target. However, these geometric paths are notorious for creating "spurious modes"—artificial peaks and valleys—that trap samplers and lead to poor results.

Recently, "diffusion paths" (the backbone of AI image generators) have emerged as a smoother alternative. The catch? To follow a diffusion path, you need to know the score function (the gradient of the log-density) at every step. In generative modeling, we learn this score using a neural network trained on millions of images. But if you're just trying to sample from a specific Bayesian posterior, you don't have a training set. You need to estimate that score accurately and efficiently in real-time.

Methods

The authors introduce DPSMC, which combines three clever components:

  1. SMC Score Estimation: Instead of a neural network, the algorithm maintains a small "cloud" of auxiliary particles for every main sample. These auxiliary particles use Sequential Monte Carlo to track the conditional distribution of the path, providing a principled, unbiased-in-limit estimate of the score.
  2. Optimal Control Variates: There are two classic ways to calculate a score: the Denoising Score Identity (DSI) and the Target Score Identity (TSI). DSI is stable at the start of the path; TSI is stable at the end. The authors derive the optimal matrix-valued weighting to combine these two, minimizing the variance of the score estimate across the entire path.
  3. Path Design: They use a "cosine schedule" for the diffusion process. By applying the calculus of variations, they prove this schedule minimizes the "action" (essentially the total effort or path length) required to reach the target, making the sampler more efficient.

Datasets

The authors evaluated DPSMC against several strong baselines (like AIS, SLIPS, and RDMC) on:

  • Synthetic "Toy" Targets:
    • GMM40: A mixture of 40 Gaussians in 2D and 50D to test multimodality.
  • Rings: Four concentric rings in 2D to test complex geometry.
  • Funnel: A 10D "narrowing" distribution that is famously difficult for MCMC.
  • Real-World Bayesian Logistic Regression:
    • Ionosphere (35D) and Sonar (61D) datasets, evaluating the predictive likelihood of the resulting samples.

Key Insights and Ideas

  • The "Aha" Moment on Variance: Most researchers use a simple heuristic to switch between score identities. This paper shows that the "Mixed Score Identity" (MSI) is actually just a special case of their optimal scalar schedule. By moving to a matrix-valued schedule, they can handle "anisotropic" targets (distributions that are much wider in one direction than another) far better than previous methods.
  • Non-Equilibrium Efficiency: Unlike MCMC-based samplers that require "equilibration" (waiting for the chain to settle) at every single time step, DPSMC evolves its auxiliary variables in a non-equilibrium manner. This removes a major sequential bottleneck.
  • Hardware Alignment: Because the auxiliary SMC steps for each sample are independent, the algorithm is "embarrassingly parallel," making it much faster on GPUs than traditional sequential samplers.

Key Results

  • Superior Accuracy: On the 50-dimensional GMM40 task, DPSMC achieved significantly lower Wasserstein-2 error (28.80) compared to classic AIS (115.13) and SMC (111.81).
  • Wall-Clock Speedup: DPSMC required only 1,000 batched energy evaluations, whereas traditional AIS and SMC required 131,000. This represents a massive improvement in practical runtime for hardware that can handle large batches.
  • Robustness: The "tempered" version of the algorithm (which smooths the path even further) was particularly effective at discovering all modes in the GMM40 benchmark, where other samplers often missed parts of the distribution.

Limitations and Future Work

  • Hyperparameter Sensitivity: The algorithm relies on a hyperparameter $T$ (the total simulation time). While the authors provide a "crude approximation" formula for $T$, it still requires some tuning or grid searching to find the sweet spot between speed and accuracy.
  • Terminal Instability: As the path approaches the very end (the pure target), the auxiliary posterior can contract into a single point (a Dirac delta), causing numerical issues. The authors currently use a heuristic to halt the SMC and switch to the target score to handle this.
  • Future Directions: The authors suggest that combining DPSMC with more advanced "underdamped" Langevin dynamics or better initialization schemes (like coarsely discretized AIS) could further improve performance on even more "pathological" distributions.

Summary updated

Abstract

We develop a diffusion-based sampler for target distributions known up to a normalising constant. To this end, we rely on the well-known diffusion path that smoothly interpolates between a (simple) base distribution and the target distribution, widely used in diffusion models. Our approach is based on a practical implementation of diffusion-annealed Langevin Monte Carlo, which approximates the diffusion path with convergence guarantees. We tackle the score estimation problem by developing an efficient sequential Monte Carlo sampler that evolves auxiliary variables from conditional distributions along the path, which provides principled score estimates for time-varying distributions. We further develop novel control variate schedules that minimise the variance of these score estimates. Finally, we provide theoretical guarantees and empirically demonstrate the effectiveness of our method on several synthetic and real-world datasets.
BibTeX
@misc{young2026diffusion,
  author = {James Matthew Young and Paula Cordero-Encinar and Sebastian Reich and Andrew Duncan and O. Deniz Akyildiz},
  title = {{Diffusion Path Samplers via Sequential Monte Carlo}},
  year = {2026},
  eprint = {2601.21951},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.21951},
}
Figure