Abstract:In this paper we propose and analyze a novel multilevel version of Stein variational gradient descent (SVGD). SVGD is a recent particle based variational inference method. For Bayesian inverse problems with computationally expensive likelihood evaluations, the method can become prohibitive as it requires to evolve a discrete dynamical system over many time steps, each of which requires likelihood evaluations at all particle locations. To address this, we introduce a multilevel variant that involves running several interacting particle dynamics in parallel corresponding to different approximation levels of the likelihood. By carefully tuning the number of particles at each level, we prove that a significant reduction in computational complexity can be achieved. As an application we provide a numerical experiment for a PDE driven inverse problem, which confirms the speed up suggested by our theoretical results.

What problem does this paper attempt to address?

This paper aims to solve the computational efficiency problems encountered by the Stein Variational Gradient Descent (SVGD) method when dealing with Bayesian inverse problems. Specifically, when the evaluation of the likelihood function is computationally costly, the traditional SVGD method can become very time - consuming because it is necessary to evolve a discrete dynamical system in multiple time steps, and each time step requires the evaluation of the likelihood function for all particle positions. To solve this problem, the author introduced a multi - layer variant of the SVGD method (ML - SVGD), which significantly reduces the computational complexity by running multiple interacting particle dynamics corresponding to different approximation levels of the likelihood function in parallel. ### Main Contributions 1. **Proposed a new multi - layer SVGD method**: This method is based on carefully combining multiple finite interacting particle systems with different sample sizes and different levels of precision. 2. **Analyzed the behavior of the multi - layer SVGD method**: Studied the changes in the solution of the mean - field system when the target probability distribution function (pdf) $\pi$ changes, and proved that small changes will lead to small changes in the Wasserstein - 2 distance. 3. **Provided a complete error analysis**: Analyzed the error of the proposed multi - layer estimator on the expected value of the mean - field solution, and reduced the overall computational cost of the algorithm by carefully adjusting the number of samples required for each level. ### Technical Details - **Mean - Field Limit**: SVGD can be regarded as the Euler approximation of the Wasserstein gradient flow in the Reproducing Kernel Hilbert Space (RKHS) for minimizing the Kullback - Leibler (KL) divergence between the reference distribution and the target distribution. - **Multi - Level Monte Carlo (MLMC) Method**: Reduces computational complexity by combining likelihood function approximations of different levels of precision through a multi - layer particle method. - **Stability Analysis**: Proved the stability of the mean - field equation when the likelihood function changes, which is crucial for understanding the performance of SVGD when facing noisy or uncertain data. ### Numerical Experiments The author carried out numerical experiments through an inverse problem driven by a partial differential equation to verify the validity of the theoretical results. The experimental results show that the multi - layer SVGD method can indeed significantly accelerate the execution speed of the algorithm while maintaining high accuracy. ### Conclusion This paper proposed a new multi - layer SVGD method and conducted a detailed analysis of its convergence and computational complexity. This method is particularly suitable for Bayesian inverse problems with high - cost likelihood function evaluations. Through the ingenious combination of multi - layer particle systems, it significantly improves computational efficiency. Future research can further explore the possibility of combining this method with other incremental precision - enhancing techniques.

On the mean-field limit for Stein variational gradient descent: stability and multilevel approximation

Augmented Message Passing Stein Variational Gradient Descent

Regularized Stein Variational Gradient Flow

Stein Variational Gradient Descent with Learned Direction

Provably Fast Finite Particle Variants of SVGD via Virtual Particle Stochastic Approximation

Stein Variational Gradient Descent on Infinite-Dimensional Space and Applications to Statistical Inverse Problems

Stein Variational Gradient Descent: A General Purpose Bayesian Inference Algorithm

Long-time asymptotics of noisy SVGD outside the population limit

Improved Finite-Particle Convergence Rates for Stein Variational Gradient Descent

Stein Variational Evolution Strategies

Riemannian Stein Variational Gradient Descent for Bayesian Inference.

Improving the performance of Stein variational inference through extreme sparsification of physically-constrained neural network models

Stein Variational Inference for Discrete Distributions

Accelerating Convergence of Stein Variational Gradient Descent via Deep Unfolding

A Stein Gradient Descent Approach for Doubly Intractable Distributions

A Stochastic Version of Stein Variational Gradient Descent for Efficient Sampling

ELBOing Stein: Variational Bayes with Stein Mixture Inference

Stochastic Gradient Descent as Approximate Bayesian Inference

Annealed Stein Variational Gradient Descent for Improved Uncertainty Estimation in Full-Waveform Inversion

Stein transport for Bayesian inference

Stochastic viscosity approximations of Hamilton-Jacobi equations and variance reduction