Neural Sampling with Reweighted Normalizing Flows via the Wasserstein--Fisher--Rao JKO Scheme

Chenguang Duan; Johannes Hertrich; Gabriele Steidl

Home/Archive/arXiv

arXiv · 2610.10278

Neural Sampling with Reweighted Normalizing Flows via the Wasserstein--Fisher--Rao JKO Scheme

Chenguang Duan·Johannes Hertrich·Gabriele Steidl

View original source ↗Read source document ↗Export metadata ↓

Abstract

We propose a neural algorithm for sampling from distributions specified by unnormalized Boltzmann densities. Our approach is based on the Jordan--Kinderlehrer--Otto scheme for the Kullback--Leibler divergence in the Wasserstein--Fisher--Rao geometry (WFR JKO scheme). Our contributions are twofold. First, we prove that, for any fixed step size, the exact WFR JKO iterates converge exponentially fast to the target as the number of iterations tends to infinity. Notably, this result requires no structural assumptions on the target, such as log-concavity or a logarithmic Sobolev inequality. Second, we develop a neural implementation of the WFR JKO scheme that parametrizes its transport and reaction components using reweighted normalizing flows. Numerical experiments on challenging multimodal targets demonstrate the promising performance of the proposed method.

Explore related subjects

math.NA cs.LG cs.NA math.PR stat.ML

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Chenguang Duan, Johannes Hertrich, Gabriele Steidl. 2026-10-07. Neural Sampling with Reweighted Normalizing Flows via the Wasserstein--Fisher--Rao JKO Scheme. https://arxiv.org/abs/2610.10278

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Robust boundary integral equations for the solution of elastic scattering problems via Helmholtz decompositions

Helmholtz decompositions of the elastic fields open up new avenues for the solution of linear elastic scattering problems via boundary integral equations (BIE) [Dong, Lai, Li, Mathematics of Computation,2021]. The main appeal of this approach is that the ensuing systems of BIE feature only integral operators associated with the Helmholtz equation. However, these BIE involve non standard boundary integral operators that do not result after the application of either the Dirichlet or the Neumann trace to Helmholtz single and double layer potentials. Rather, the Helmholtz decomposition approach leads to BIE formulations of elastic scattering problems with Neumann boundary conditions that involve boundary traces of the Hessians of Helmholtz layer potential. As a consequence, the classical combined field approach applied in the framework of the Helmholtz decompositions leads to BIE formulations which, although robust, are not of the second kind. Following the regularizing methodology introduced in [Boubendir, Dominguez, Levadoux, Turc, SIAM Journal on Applied Mathematics 2015] we design and analyze novel robust Helmholtz decomposition BIE for the solution of elastic scattering that are of the second kind in the case of smooth scatterers in two dimensions. We present a variety of numerical results based on Nystrom discretizations that illustrate the good performance of the second kind regularized formulations in connections to iterative solvers.

math.NA↗

arXiv2026

A Moment-Based Shallow Water Model for Coupled Bedload-Suspended Load Transport with Variable Density: Hyperbolicity and Numerical Validation

In this paper, we develop a Hyperbolic Shallow Water Exner Moment model with Erosion and Deposition (HSWEMED), which extends the recent shallow water moment framework to capture coupled morphodynamics with erosion and deposition. Extending existing moment models, HSWEMED introduces a sediment concentration equation for suspended load, couples a concentration-dependent sediment-water mixture density with the momentum equation and higher-order moments, and adds source terms arising from erosion and deposition. Starting from the incompressible Navier-Stokes equations for a water-sediment mixture, we derive a coupled system that integrates (i) the shallow water equations, (ii) moment equations for polynomial velocity coefficients, (iii) a depth-averaged suspended-sediment equation, and (iv) an Exner equation for bedload transport with erosion-deposition coupling. The reduced model evolves only the depth-averaged concentration, while the vertical structure is retained for the horizontal velocity through the moment expansion. For the regularized model, we derive the characteristic polynomial and identify a real-root condition for the sediment-coupled part. The analysis shows that strict hyperbolicity can hold for even moment orders, while odd moment orders lead to weak hyperbolicity because of a repeated eigenvalue. We also derive dissipative energy balance relations for the lower-order models. Numerical results are obtained with a second-order path-conservative MUSCL-Rusanov finite-volume scheme with a well-balanced correction for the water-at-rest equilibrium. We verify the well-balanced property, perform a reference-convergence study, and compare dam-break simulations with laboratory experiments from the literature. The results illustrate how the moment approximation and the erosion-deposition coupling affect the free surface, bed evolution, and suspended concentration.

math.NA↗

arXiv2026

Asymptotic error distributions for diffusion-implicit schemes via appurtenant methods

We develop an appurtenant-method framework for analyzing asymptotic error distributions and implement it in two diffusion-implicit settings for stochastic differential equations. The argument rests on an explicit appurtenant method which has the same strong order as the implicit method and approximates it with a higher strong order. A local-to-global strong convergence result for two Markov chains turns the one-step expansion of the implicit equation into the required appurtenant method; the limiting distribution can then be obtained from an adapted continuization of the appurtenant method. We first illustrate the framework with the fully implicit stochastic midpoint method for a multidimensional Stratonovich equation, without imposing commutativity on the diffusion fields. This method has strong order $1/2$, and its normalized error is governed by the Lévy areas omitted by the numerical method. We then turn to the main application, namely, a class of fully implicit stochastic Runge--Kutta (SRK) methods of strong order $1$ for Stratonovich equations with scalar multiplicative noise. We derive their asymptotic error distributions and identify a parameter, depending only on the coefficients of the Butcher tableau, that reflects the growth rate of the root-mean-square error of the SRK method.

math.NA↗