Abstract:Langevin Dynamics is a Stochastic Differential Equation (SDE) central to sampling and generative modeling and is implemented via time discretization. Langevin Monte Carlo (LMC), based on the Euler-Maruyama discretization, is the simplest and most studied algorithm. LMC can suffer from slow convergence - requiring a large number of steps of small step-size to obtain good quality samples. This becomes stark in the case of diffusion models where a large number of steps gives the best samples, but the quality degrades rapidly with smaller number of steps. Randomized Midpoint Method has been recently proposed as a better discretization of Langevin dynamics for sampling from strongly log-concave distributions. However, important applications such as diffusion models involve non-log concave densities and contain time varying drift. We propose its variant, the Poisson Midpoint Method, which approximates a small step-size LMC with large step-sizes. We prove that this can obtain a quadratic speed up of LMC under very weak assumptions. We apply our method to diffusion models for image generation and show that it maintains the quality of DDPM with 1000 neural network calls with just 50-80 neural network calls and outperforms ODE based methods with similar compute.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper attempts to solve the problems of slow convergence speed and the need for a large number of small - step - length iterations in the time - discretization process of Langevin Dynamics (LD) in the standard Euler - Maruyama method. Specifically: 1. **Convergence problems of Langevin Monte Carlo (LMC)**: - LMC is the simplest and most widely studied algorithm based on Euler - Maruyama discretization, but its convergence speed is slow, and usually a large number of small - step - length steps are required to obtain high - quality samples. - In diffusion models, this problem is particularly prominent. In order to obtain the best sample quality, usually 1,000 small - step - length steps are required, and when the step length is larger or the number of steps is smaller, the sample quality will decline rapidly. 2. **Applications of non - strongly log - concave distributions**: - The Randomized Midpoint Method (RLMC) has been proposed for sampling from strongly log - concave distributions, but in many important applications (such as diffusion models), the target distribution is non - log - concave and contains time - varying drift terms. - Therefore, the applicability of the existing RLMC methods in these application scenarios is limited. ### Proposed methods To solve the above problems, the paper proposes the Poisson Midpoint Method (PMLC), which is an improved Langevin dynamics discretization method. The main features are as follows: 1. **Principles of the Poisson Midpoint Method**: - PMLC approximates multiple small - step - length Euler - Maruyama discretization steps into a single large - step - length step through stochastic approximation. - This method can achieve quadratic acceleration of LMC under very weak assumptions without additional assumptions (such as isoperimetry or strong log - concavity). 2. **Theoretical analysis**: - The author proves the strong error bounds of PMLC under general conditions (including non - log - concave distributions and time - varying drift terms). - For the target distributions that satisfy the Logarithmic Sobolev Inequality (LSI), PMLC can achieve quadratic acceleration in both overdamped Langevin Monte Carlo (OLMC) and underdamped Langevin Monte Carlo (ULMC) cases. 3. **Experimental verification**: - The author applies PMLC to the diffusion model for image generation, and the results show that PMLC can maintain the same quality as DDPM with 1,000 steps while only using 50 - 80 neural network calls. - Compared with ODE - based methods (such as DPM - Solver and DDIM), PMLC shows higher sample quality under similar computational amounts. ### Summary By introducing the Poisson Midpoint Method, the paper solves the problem of slow convergence in Langevin dynamics discretization and demonstrates its effectiveness under non - log - concave distributions and time - varying drift terms. The experimental results show that PMLC significantly improves efficiency and sample quality in image generation tasks. Future research can further explore methods to improve performance with fewer neural network calls and consider the convergence of algorithms with time - dependent drift functions.

The Poisson Midpoint Method for Langevin Dynamics: Provably Efficient Discretization for Diffusion Models

Langevin Monte Carlo for strongly log-concave distributions: Randomized midpoint revisited

Parallelized Midpoint Randomization for Langevin Monte Carlo

Pseudo Numerical Methods for Diffusion Models on Manifolds

Stochastic Runge-Kutta Methods: Provable Acceleration of Diffusion Models

Faster Diffusion Sampling with Randomized Midpoints: Sequential and Parallel

Sequential Controlled Langevin Diffusions

Optimizing the diffusion coefficient of overdamped Langevin dynamics

Reweighted Interacting Langevin Diffusions: an Accelerated Sampling Methodfor Optimization

Faster Sampling without Isoperimetry via Diffusion-based Monte Carlo

Kinetic Interacting Particle Langevin Monte Carlo

Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity

Accelerating Convergence of Score-Based Diffusion Models, Provably

Sampling and estimation on manifolds using the Langevin diffusion

Penalized Overdamped and Underdamped Langevin Monte Carlo Algorithms for Constrained Sampling

Mean-Square Analysis with An Application to Optimal Dimension Dependence of Langevin Monte Carlo

A Langevin-like Sampler for Discrete Distributions

Fast Diffusion Probabilistic Model Sampling through the lens of Backward Error Analysis

Accelerated convergence to equilibrium and reduced asymptotic variance for Langevin dynamics using Stratonovich perturbations

Improved Convergence Rate for Diffusion Probabilistic Models

Parallel simulation for sampling under isoperimetry and score-based diffusion models