This work constructs a cube-to-target map by composing a Gaussian base transformation (the component-wise inverse Gaussian CDF) with an Euler-discretized probability flow ODE, and establishes conditions for diffusion probability-flow transport under mild bounded-derivative assumptions on the learned vector field.
Abstract
We study high-dimensional numerical integration with respect to complex target measures using diffusion-based transport maps and randomized quasi-Monte Carlo (RQMC). Score-based diffusion models induce a deterministic probability flow ODE that transports a simple prior to the target, suggesting a principled way to transform low-discrepancy points on the unit cube into informative samples. We construct a cube-to-target map by composing a Gaussian base transformation (the component-wise inverse Gaussian CDF) with an Euler-discretized probability flow ODE. To retain unbiasedness under transport approximation, we formulate integration as importance sampling (IS) on the cube. Our main result provides verifiable conditions under which the resulting IS integrand satisfies the boundary growth condition, implying an $O(N^{-1+\epsilon})$ RMSE for scrambled nets. We then establish these conditions for diffusion probability-flow transport under mild bounded-derivative assumptions on the learned vector field, explicitly controlling the boundary singularities introduced by the inverse Gaussian CDF. Experiments range from a 2D mixture to 784D images and a 40,960D conditional vorticity-assimilation task; in the latter, blocked scrambled Sobol'sampling reduces the randomization standard deviation of nonlinear accuracy metrics at essentially unchanged online denoising cost. Together, these results give a theoretical and empirical foundation for combining diffusion generative modeling with high-precision RQMC integration.
This analysis transfers reverse-time discretization errors to the forward corruption law and treats two numerical schemes within a common framework, which identifies a local error, accumulates it through the forward evolution, and inserts the result into a common KL decomposition.
We introduce Ordered Diffusion Kernels (ODKs), a novel class of local kernels that can approximate the infinitesimal generator of an arbitrary It\^o Stochastic Differential Equation (SDE). ODKs are designed to be applied to data sampled from dynamical systems where little dynamical information is available a priori. The Laplacian of classical diffusion kernels approximates the Laplace-Beltrami operator on the underlying manifold; adjusting the normalisation introduces an advection term that depends on the sampling density; recently, TMDmap generalised this normalisation to target an arbitrary measure, but at the cost of coupling advection to diffusion. More general local kernels can learn arbitrary second-order elliptic operators but are formulated in terms of known velocity fields --- making the first step of any analysis a potentially ill-posed inference problem. To formulate ODK, we first relax the problem of potential estimation to the more tractable task of inferring an ordering of the data, which we represent through an ordering function. We prove ODK's Laplacian converges to the infinitesimal generator of a gradient-flow SDE with state-dependent isotropic diffusion, without coupling advection and diffusion. We provide various extensions of ODK to: arbitrary drifts via local ordering functions; anisotropic diffusions via a Strang splitting scheme; multiple ordering functions; and self-tuning bandwidths. In addition, we introduce two loss functions which exploit the structure of ODKs to solve a non-parametric inference problem. We validate this framework on synthetic data from deterministic and stochastic systems, demonstrating accurate recovery of operators, velocity fields, extrinsic curvature, and spatially dependent drift and diffusion coefficients.
Jack H. Soulsby, A. C. S. Jørgensen, Atiyo Ghosh et al.· 0 citations
The probability-flow ordinary differential equation (PF-ODE) associated with a diffusion process is widely used in score-based generative modeling as a deterministic sampler that reproduces the marginal distributions of the diffusion. The validity of this marginal-matching property depends on the well-posedness of an ordinary differential equation whose velocity field is constructed from the score function of the diffusion. We examine the precise mathematical relation between a diffusion process, the Fokker-Planck equation and the associated PF-ODE under weak regularity assumptions on the drift and score. We establish existence and uniqueness of the marginal density flow as a solution of the Fokker--Planck equation under minimal regularity assumptions. We then study the corresponding Lagrangian problem using the DiPerna-Lions-Ambrosio theory of regular Lagrangian flows. We prove existence, uniqueness and stability of the flow, and show that it transports the initial distribution onto the diffusion marginals, under Sobolev or bounded-variation regularity of the score together with one-sided bounds on the divergence of the probability-flow velocity. We identify sufficient conditions for the required regularity in diffusion models relevant for applications. Our analysis underlines a fundamental distinction between Eulerian and Lagrangian descriptions. We construct a counterexample in which the Fokker--Planck equation has a unique density flow while the associated PF-ODE fails to admit a regular Lagrangian flow from the initial time, demonstrating that uniqueness of the density evolution does not in general imply the existence of a deterministic probability-flow representation. Finally, we derive stability estimates for probability-flow trajectories under learned score approximations. Our findings have implications for the training and deployment of score-based diffusion models.
We introduce weighted Laplacian flow (WLF), a deterministic particle-flow framework for sampling from a target distribution known only up to normalization. The proposed method evolves the logarithmic density ratio through a transport equation and determines the particle velocity from a target-weighted Poisson problem, resulting in a nonlocal and kernel-free mechanism for redistributing mass. For bounded smooth domains, we establish global well-posedness of the flow for Lipschitz initial data and prove that the flow contracts the discrepancy between the evolving and target densities at an exact exponential rate in terms of the oscillation of the log-density ratio, as well as in the associated projective metric. The relative entropy also satisfies a precise dissipation relation involving the two directional Kullback--Leibler divergences. In addition, we identify a variational interpretation of WLF. The method can be viewed as a gradient flow of the reverse Kullback--Leibler divergence under a target-anchored metric. Consequently, the method achieves an explicit relaxation scale without requiring log-concavity or spectral-gap assumptions on the target distribution. Numerical experiments on multimodal, heavy-tailed, and high-dimensional targets demonstrate the effectiveness of the proposed approach in capturing long-range mass transport.
Moving probability distributions in the context of stochastic processes efficiently is central to modern generative modeling, uncertainty quantification, and control of large engineered systems. Most current methods generate temporal trajectories independently, leaving open whether the trajectories can cooperate through shared population information to reduce transport cost. Here we show that mean-field path-integral diffusion turns this question into a self-consistent stochastic control problem in which each trajectory responds to the evolving population. In a broad linear-quadratic setting, the problem reduces to a finite system of ordinary differential equations. For quadratic interactions with no background drift, we prove that the optimal population guidance is exactly the straight-line interpolation between the initial and target means, for arbitrary endpoint distributions with finite means. This yields an explicit construction for mixture targets. In demand-response control of multi-zone buildings, the method reduces control energy by 19–24% while maintaining the desired final distribution. These results suggest a practical route to coordinated generative transport for energy and other large-agent systems. Efficiently moving probability distributions in the context of stochastic processes is important in artificial intelligence, physics, and smart-building control. Here, the authors show that temporal trajectories can act like a coordinated population, reducing control energy by 19–24% in demand response while still reaching the desired final distribution.
M. Chertkov· Communications Physics· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
