Abstract:Solving inverse problems without the use of derivatives or adjoints of the forward model is highly desirable in many applications arising in science and engineering. In this paper, we propose a new version of such a methodology, a framework for its analysis, and numerical evidence of the practicality of the method proposed. Our starting point is an ensemble of over-damped Langevin diffusions which interact through a single preconditioner computed as the empirical ensemble covariance. We demonstrate that the nonlinear Fokker-Planck equation arising from the mean-field limit of the associated stochastic differential equation (SDE) has a novel gradient flow structure, built on the Wasserstein metric and the covariance matrix of the noisy flow. Using this structure, we investigate large time properties of the Fokker-Planck equation, showing that its invariant measure coincides with that of a single Langevin diffusion, and demonstrating exponential convergence to the invariant measure in a number of settings. We introduce a new noisy variant on ensemble Kalman inversion (EKI) algorithms found from the original SDE by replacing exact gradients with ensemble differences; this defines the ensemble Kalman sampler (EKS). Numerical results are presented which demonstrate its efficacy as a derivative-free approximate sampler for the Bayesian posterior arising from inverse problems.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to solve inverse problems without using the derivatives of the forward model or the adjoint operator. In many applications in science and engineering, this need is very urgent. The paper proposes a new methodological framework - Ensemble Kalman Sampler (EKS) to analyze the effectiveness of this method and provides numerical evidence to prove the practical feasibility of this method. Specifically, the paper focuses on sampling from the Bayesian posterior distribution by approximating the Langevin - type stochastic dynamical system through a set of over - damped Langevin diffusion processes. Ensemble Kalman Inversion (EKI) and its variants play an important role in solving large - scale science and engineering problems because these methods avoid calculating the derivatives and adjoint operators defined by the forward mapping \(G\), which is a difficult point in many practical applications. The main contributions of the paper include: 1. **Introduction of new noise perturbations**: A new noise - perturbed version of the continuous - time Ensemble Kalman Inversion algorithm is proposed, forming an interacting particle system in the form of a stochastic differential equation (SDE), namely Ensemble Kalman Sampler (EKS). 2. **Research on related SDEs**: A related SDE is studied, in which the ensemble differences are approximated by gradients; for linear inverse problems, this approximation is exact. The mean - field limit of this related SDE is studied and a new Kalman - Wasserstein gradient - flow structure in the associated nonlinear Fokker - Planck equation is shown. 3. **Steady - state characteristics**: Using the Kalman - Wasserstein structure, the steady - state of the nonlinear Fokker - Planck equation is characterized and it is shown that one of the steady - states is the posterior density \(\pi(u)\). 4. **Global attractor**: By explicitly solving the nonlinear Fokker - Planck equation in the case of linear \(G\), it is proved that the posterior density is the global attractor for all densities with finite initial energy and not Dirac measures. 5. **Numerical experiments**: Numerical examples are provided to prove that the EKS algorithm can provide good approximate samples of the posterior distribution for simple low - dimensional test problems and PDE inverse problems in Darcy flow. Through these contributions, the paper provides a new and effective method for generating approximate samples from the Bayesian posterior distribution without using derivatives.

Interacting Langevin Diffusions: Gradient Structure And Ensemble Kalman Sampler

Less interaction with forward models in Langevin dynamics

Gradient flow structure and convergence analysis of the ensemble Kalman inversion for nonlinear forward models

Large-Scale Wasserstein Gradient Flows

The Ensemble Kalman Filter for Dynamic Inverse Problems

Ensemble Kalman Diffusion Guidance: A Derivative-free Method for Inverse Problems

Efficient, Multimodal, and Derivative-Free Bayesian Inference With Fisher-Rao Gradient Flows

A Langevinized Ensemble Kalman Filter for Large-Scale Static and Dynamic Learning

Flow Annealed Kalman Inversion for Gradient-Free Inference in Bayesian Inverse Problems

A two-stage ensemble Kalman filter based on multiscale model reduction for inverse problems in time fractional diffusion-wave equations

Ensemble Kalman inversion: mean-field limit and convergence analysis

Enhancing Score-Based Sampling Methods with Ensembles

Solving Fredholm Integral Equations of the First Kind via Wasserstein Gradient Flows

Analysis of the Ensemble Kalman Filter for Inverse Problems

Subsampling in ensemble Kalman inversion

Sampling as optimization in the space of measures: The Langevin dynamics as a composite optimization problem

Kinetic Interacting Particle Langevin Monte Carlo

Variable-Separation Based Iterative Ensemble Smoother For Bayesian Inverse Problems In Anomalous Diffusion Reaction Models

HJ-sampler: A Bayesian sampler for inverse problems of a stochastic process by leveraging Hamilton-Jacobi PDEs and score-based generative models

Nonreversible Langevin Samplers: Splitting Schemes, Analysis and Implementation