Barycentric Stochastic Interpolant

• Barycentric stochastic interpolant is a framework that defines continuous interpolation between multiple probability distributions using barycentric coordinates on a simplex.
• It leverages neural network parameterizations and stochastic differential equations to model complex, multimarginal relationships efficiently.
• The approach unifies optimal transport, generative modeling, and scientific simulation, offering improved computational speed and statistical accuracy.

A barycentric stochastic interpolant is a mathematical and algorithmic framework for constructing stochastic processes or generative models that perform continuous, barycenter-based interpolation between multiple probability distributions, datasets, or fields. This approach leverages barycentric coordinates on a simplex to blend or transport between marginals, generalizing the classical two-marginal (source-to-target) case to arbitrary multimarginal and multimodal settings. Barycentric stochastic interpolants underpin exact sampling schemes for barycentrically weighted surfaces in geometry, high-dimensional generative modeling with controlled multimarginal correspondences, and domain-bridging simulation models in scientific applications.

1. Mathematical Foundations

At its core, the barycentric stochastic interpolant constructs an interpolation (deterministic or stochastic) between $K+1$ probability densities $\rho_0, \rho_1, \ldots, \rho_K$ via barycentric coordinates $\alpha = (\alpha_0, \ldots, \alpha_K)$ on the $K$-simplex: $$\Delta^K = \left\{ \alpha \in \mathbb{R}^{K+1} \mid \sum_{k=0}^{K} \alpha_k = 1,\, \alpha_k \geq 0 \right\}$$ Given realizations $(x_0,\ldots,x_K)$ from these marginals or their coupling, a sample at location $\alpha$ is $x(\alpha) = \sum_{k=0}^K \alpha_k x_k$. The law of $x(\alpha)$ yields an interpolated (barycentric) distribution whose statistical properties interpolate among the marginals in a convex fashion (Albergo et al., 2023).

The associated probability density $\rho(\alpha,x)$ evolves under a system of continuity equations with vector fields $g_k(\alpha,x) = \mathbb{E}[x_k \mid x(\alpha)=x]$: $$\partial_{\alpha_k}\rho(\alpha,x) + \nabla_x \cdot (g_k(\alpha,x) \rho(\alpha,x)) = 0$$ For paths in the simplex, $\alpha(t)$, this induces a transport equation with drift $b(t,x) = \sum_k \dot{\alpha}_k(t) g_k(\alpha(t),x)$, generalizing dynamical optimal transport to the multimarginal case (Albergo et al., 2023).

The stochastic interpolant can be further generalized to stochastic processes, where a Brownian or Wiener term injects diffusion, yielding an SDE along paths in simplex space with score fields determined by the log-gradients of the interpolated density.

2. Algorithmic Implementation

Barycentric stochastic interpolants are practically realized by parameterizing the vector fields $g_k(\alpha,x)$ (and optionally, the drift $b$ and score $s$ for SDEs) via neural networks, often with architectures such as U-Nets. Training is accomplished by simple quadratic objectives: $$L_k(\hat{g}_k) = \int_{\Delta^K} \mathbb{E}\left[ |\hat{g}_k(\alpha,x(\alpha))|^2 - 2x_k\cdot\hat{g}_k(\alpha,x(\alpha)) \right] d\alpha$$ The score-matching variant minimizes $$J_s(s) = \int_{\Delta^K} \mathbb{E}[| s(\alpha, x(\alpha)) - \nabla_x \log \rho(\alpha, x(\alpha))|^2 ] d\alpha$$ (Albergo et al., 2023).

Sampling from the trained model involves solving characteristic ODEs (for deterministic interpolations) or SDEs (for score-based diffusion) along simplex paths from source to target barycentric coordinates. Interior points in the simplex yield true barycentric blends or barycenters among the marginals, while vertices recover exact marginals.

In high-dimensional scenarios, such as connecting particle mesh (PM) dark matter simulations to baryonic field predictions in cosmological models, 3D U-Net variants are used, conditioning the drift on additional physical or astrophysical parameters (Horowitz et al., 22 Oct 2025).

3. Key Applications

Barycentric stochastic interpolants have emerged as unifying tools in a wide array of domains:

• Geometry Processing: Efficient barycentric sampling of meshes, where the surface density is specified by barycentric interpolation of vertex weights. Techniques leverage analytical inversion using barycentric coordinates $(u,v,w)$ under constraints $u+v+w=1$ to sample points according to a nonuniform, linearly varying surface measure. Newton–Raphson iterations are used for inversion of marginal CDFs, yielding unbiased, statistically validated samples in constant time (Portsmouth, 2017).
• Generative Modeling: Multimarginal generative modeling, e.g., image-to-image translation across multiple datasets or styles, exploiting the simplex structure to enable smooth, multi-way correspondences and barycentric interpolations. The learned joint distribution builds multi-way correspondences and dynamic couplings between domains (Albergo et al., 2023).
• Scientific Simulation/Super-resolution: Fast simulation of complex physical outcomes (e.g., baryonic fields from dark matter in cosmology) using stochastic interpolants that map fast, approximate simulations to accurate, physically rich outputs. These pipelines realize dramatic computational speed-ups over traditional solvers, while matching physical statistics (e.g., Ly-α forest flux power spectra) to sub-5% accuracy at relevant spatial scales (Horowitz et al., 22 Oct 2025).

4. Theoretical Properties and Connections

The barycentric stochastic interpolant framework generalizes classical time-parameterized (two-marginal) transport, connecting to optimal-transport barycenters, Benamou–Brenier kinetic cost, and Monge maps. For $K=1$, the classical source-to-target transport is recovered (Albergo et al., 2023). For $K\geq 2$, the approach allows both sampling explicit barycenters and extracting multi-way correspondence maps.

A notable feature is that learning the conditional-expectation vector fields $g_k$ via least-squares regression suffices to recover the multimarginal coupling, and the approach does not require inner optimal transport solves or adversarial losses. In Monge-type settings (deterministic couplings), explicit invertible mappings between all pairs of marginals can be inferred.

Optimization over paths in the simplex (i.e., non-linear schedules for $\alpha(t)$) can reduce the kinetic transport cost and the number of integration steps required for high-fidelity generative modeling (Albergo et al., 2023).

5. Statistical Validation and Computational Performance

Rigorous statistical testing ensures the unbiasedness and correctness of barycentric stochastic interpolant-based schemes. In mesh sampling, empirical CDFs of generated samples closely match theoretical predictions (Kolmogorov–Smirnov $D<0.005$ for $N\geq 10^6$), as do chi-squared tests in uv-bin grids (Portsmouth, 2017). In generative modeling, the approach yields cross-correlations exceeding 0.9 and power spectra errors below 10% up to high spatial frequencies (Horowitz et al., 22 Oct 2025).

Computationally, the inversion-based barycentric sampling runs in expected constant time per sample (dominated by Newton iteration, square roots, and one CDF lookup). For high-dimensional applications with U-Net parameterizations, O(N³) scaling with grid size enables tractable simulation at moderate resolutions (e.g., 256³), requiring only minutes on modern GPUs per realization, several orders of magnitude faster than traditional hydrodynamical simulations (Horowitz et al., 22 Oct 2025).

Application Domain	Reference	Characteristic Algorithmic Feature
Mesh point sampling	(Portsmouth, 2017)	Analytical inversion in barycentric coordinates
Multimarginal modeling	(Albergo et al., 2023)	U-Net regression of vector fields over simplex
Fast simulation mapping	(Horowitz et al., 22 Oct 2025)	Conditional 3D U-Net drift in stochastic interpolant SDEs

6. Relation to Optimal Transport and Barycenters

Barycentric stochastic interpolants provide a practical computational pathway to sample or interpolate among optimal-transport barycenters by dynamic blending of multiple data distributions. Unlike traditional OT barycenters, which require solving multi-marginal OT problems that scale poorly with dimension and number of marginals, this framework sidesteps combinatorial complexity through learnable flows or score fields parameterized over the simplex (Albergo et al., 2023).

For path optimization, the kinetic cost induced by a drift $b(t,x)$ driven by curves in the simplex can be minimized to approximate the Wasserstein barycenter according to the Benamou–Brenier formulation, but with considerably improved computational tractability in practice.

7. Significance and Impact

The barycentric stochastic interpolant unites diverse lines of research in stochastic dynamics, optimal mass transport, generative modeling, and scientific computation. It enables exact and unbiased sampling in mesh geometry, efficient multimarginal translation in machine learning, and high-fidelity surrogate modeling in scientific domains. Empirical results and theoretical guarantees confirm that it generalizes two-way interpolation to arbitrarily many sources, maintains statistical accuracy, and achieves significant acceleration over legacy algorithms, facilitating novel applications in style transfer, algorithmic fairness, simulation-based inference, and mesh processing (Portsmouth, 2017, Albergo et al., 2023, Horowitz et al., 22 Oct 2025).