Radon–Wasserstein Gradient Flows

• Radon–Wasserstein gradient flows are defined via one-dimensional projections using the Radon transform to compute KL divergence gradients over probability measures.
• They enable scalable interacting-particle approximations that scale linearly in both sample size and dimension, ensuring robust convergence for high-dimensional sampling.
• Regularized variants incorporate kernel smoothing and entropic regularization to stabilize density estimates and guarantee well-posed long-time convergence.

Radon–Wasserstein gradient flows constitute a new class of gradient flows of the Kullback–Leibler (KL) divergence, formulated over novel transportation-based Riemannian geometries on spaces of probability measures. These geometries are defined using the Radon transform so that the gradient-flow velocities depend solely on one-dimensional projections of the evolving measure. The resulting flows—specifically, the Radon–Wasserstein (RW) and Regularized Radon–Wasserstein (RRW) flows—not only admit particle approximations that scale linearly in both the number of particles and the ambient dimension but also demonstrate robust convergence and stability properties, making them effective for high-dimensional sampling tasks (Hess-Childs et al., 5 Feb 2026).

1. Radon–Wasserstein Geometry

The foundational component of RW geometry is the Radon transform. Given $S^{d-1}$, the unit sphere in $\mathbb{R}^d$, and $\theta \in S^{d-1}$, the projection $\Pi_\theta(x) = x \cdot \theta$ defines the one-dimensional projection of $x$ onto $\theta$. For a probability density $\rho \in \mathcal P_2(\mathbb{R}^d)$, the Radon transform is defined as

$$(R^\theta \rho)(p) = \int_{\{x : x \cdot \theta = p\}} \rho(x) \, d\mathcal{H}^{d-1}(x), \quad p \in \mathbb{R}.$$

The full Radon transform $R\rho$ is the measure on $$D^d = (S^{d-1} \times \mathbb{R})/\{(\theta,p)\sim(-\theta,-p)\}$$ given by $\mathbb{R}^d$0. Its dual, the back-projection $\mathbb{R}^d$1, is

$\mathbb{R}^d$2

The RW metric is defined via spherical averages of one-dimensional motions. The RW metric tensor at $\mathbb{R}^d$3 is: $\mathbb{R}^d$4 where the infimum is over all $\mathbb{R}^d$5 such that $\mathbb{R}^d$6. This restriction to one-dimensional projections maintains low complexity even as $\mathbb{R}^d$7 (Hess-Childs et al., 5 Feb 2026).

2. KL Gradient Flows in Radon–Wasserstein Geometry

Let the target measure be $\mathbb{R}^d$8. The KL divergence relative to $\mathbb{R}^d$9 is

$\theta \in S^{d-1}$0

The RW gradient flow of $\theta \in S^{d-1}$1 seeks a velocity $\theta \in S^{d-1}$2 so that

$\theta \in S^{d-1}$3

The associated PDE is the continuity equation

$\theta \in S^{d-1}$4

with

$\theta \in S^{d-1}$5

Equivalently,

$\theta \in S^{d-1}$6

The velocity is thus determined by the family of one-dimensional projections and related convolution operations.

3. Regularized Radon–Wasserstein Flows

The RRW metric introduces kernel smoothing and an entropic regularization parameter. Given a smooth even kernel $\theta \in S^{d-1}$7 and $\theta \in S^{d-1}$8,

$\theta \in S^{d-1}$9

This ensures strictly positive densities along each projection, avoiding density-estimation artifacts in empirical approximations.

The RRW gradient flow PDE is

$\Pi_\theta(x) = x \cdot \theta$0

In practice, final smoothing may be omitted for the congruent particle scheme.

4. Interacting-Particle Approximations and Algorithms

The interaction terms in the RW and RRW flows depend only on one-dimensional projections $\Pi_\theta(x) = x \cdot \theta$1. Particle systems evolve by repeatedly projecting all particles along a sampled random direction, using one-dimensional convolution-based updates for velocities:

• At each time step, sample $\Pi_\theta(x) = x \cdot \theta$2.
• Compute projections $\Pi_\theta(x) = x \cdot \theta$3 and projected score terms $\Pi_\theta(x) = x \cdot \theta$4.
• Calculate velocities using either a direct convolution or FFT-accelerated scheme.

For the kernel-density Radon–Wasserstein (KDRW) update,

$\Pi_\theta(x) = x \cdot \theta$5

with naive $\Pi_\theta(x) = x \cdot \theta$6 cost per projection. The FFT-accelerated variant bins projected positions, performs convolutions via FFTs (cost $\Pi_\theta(x) = x \cdot \theta$7 given grid-size $\Pi_\theta(x) = x \cdot \theta$8), and is overall $\Pi_\theta(x) = x \cdot \theta$9, linear in both $x$0 and $x$1 (Hess-Childs et al., 5 Feb 2026).

5. Theoretical Results and Guarantees

The primary results are:

• Well-posedness: For smooth kernels $x$2, positive $x$3, and $x$4 Lipschitz with quadratic confinement, RRW and KDRW flows admit unique global solutions with controlled second moments.
• Particle system well-posedness: The finite $x$5 ODEs have unique global solutions, and their empirical measures maintain the moment bound.
• Stability and mean-field convergence: The PDE and particle flows are stable with respect to initial measure, and the particle approximation converges to the continuum solution as $x$6.
• Stochastic-gradient convergence: The single-direction-per-step Euler discretization achieves mean-squared error $x$7, uniformly in $x$8.
• Long-time convergence: Under nondegeneracy of the kernel’s Fourier transform, RRW flows converge weakly to the target measure as $x$9.

These results establish the RW and especially the RRW frameworks as robust methods for interacting-particle sampling and transport, with strong guarantees on well-posedness, stability, and convergence that are dimension-independent in critical respects (Hess-Childs et al., 5 Feb 2026).

6. Numerical Performance and Comparative Assessment

The performance of KDRW and RRW methods was assessed on Gaussian and Rosenbrock “banana” target distributions in various dimensions. Metrics included maximum mean discrepancy (MMD$\theta$0) and mean error. Algorithmic comparisons included i.i.d. sampling, Stein variational gradient descent (SVGD), and Langevin Monte Carlo (LMC).

Key empirical findings:

• Bandwidth sensitivity: MMD$\theta$1 profiles versus kernel bandwidth show a U-shape; in low $\theta$2, optimal bandwidth is around 0.2, while in high $\theta$3 performance is relatively insensitive to bandwidth.
• Per-step cost: FFT-accelerated RW methods achieve per-step times under 100 ms for $\theta$4, with wall-time scaling linear in both $\theta$5 and $\theta$6.
• Convergence: RW methods decrease MMD$\theta$7 rapidly, surpassing the i.i.d. baseline in few steps. SVGD performs comparably for $\theta$8 but suffers from variance collapse when $\theta$9.
• Error scaling with sample size:

| $\rho \in \mathcal P_2(\mathbb{R}^d)$0 | i.i.d. Slope | KDRW_fft | RRW_fft | |-----------|--------------|----------|---------| | 2 | –1.10 | –1.49 | –1.65 | | 32 | –1.02 | –1.23 | –1.20 | | 256 | –0.99 | –1.06 | –1.05 | | 2048 | –1.00 | –1.02 | –0.99 |

• Comparison to SVGD: RW methods outperform SVGD particularly in regimes with $\rho \in \mathcal P_2(\mathbb{R}^d)$1, where SVGD collapses variance.

These results indicate consistent improvement in quantization error, with super-Monte-Carlo rates in low dimensions and competitive scaling in high dimensions, highlighting the scalability and robustness of the RW and RRW approaches (Hess-Childs et al., 5 Feb 2026).