Stochastic Gradient Variational Inference with Price's Gradient Estimator from Bures-Wasserstein to Parameter Space

Kyurae Kim, Qiang Fu, Yian Ma, Jacob R. Gardner, Trevor Campbell

Introduction

Variational Inference (VI) is a cornerstone of modern Bayesian machine learning, widely used to approximate complex target distributions when only an unnormalized log-density is available. For years, researchers have debated the optimal space in which to perform this optimization. The two primary paths are:

  1. Black-Box VI (BBVI): Performing gradient descent in the Euclidean space of parameters (e.g., optimizing the mean and covariance of a Gaussian).
  2. Wasserstein VI (WVI): Performing gradient descent directly in the space of probability measures, utilizing the Bures–Wasserstein metric.

Recent theoretical work suggested that WVI was fundamentally superior, offering faster convergence rates than BBVI. This led to a prevailing belief that the Wasserstein geometry provided a unique mathematical advantage. However, this paper challenges that assumption, asking a critical question: Is the geometry actually better, or are we comparing apples to oranges because WVI happens to use a more sophisticated gradient estimator?

The Core Problem: Geometry vs. Estimator

By conducting a rigorous theoretical and empirical comparison, the authors settle this long-standing debate. They demonstrate that the perceived superiority of Wasserstein-based methods over standard black-box approaches is not due to the "geometry" of the space. Instead, it stems entirely from the choice of gradient estimator.

Standard BBVI typically relies on the "reparametrization trick," which requires only first-order gradients of the log-density. Conversely, WVI implementations often employ an estimator derived from Price’s theorem, which leverages second-order information (Hessians). By applying Price’s gradient estimator to both frameworks, the authors achieve identical state-of-the-art convergence guarantees, proving that the estimator—not the manifold—is the true driver of performance.

Methodology: Unifying the Theory

The researchers compared two specific algorithms: Stochastic Proximal Gradient Descent (SPGD) for parameter space and Stochastic Proximal Bures–Wasserstein Gradient Descent (SPBWGD) for measure space.

Their key technical contributions include:

  • Theoretical Unification: They proved that when both BBVI and WVI use Price’s gradient, their iteration complexity is identical.
  • Refined Analysis: By utilizing Bregman divergences, they tightened the convergence bounds, successfully removing unnecessary factors related to the condition number ($\kappa$) of the target distribution.
  • Cross-Pollination: They demonstrated how to apply first-order reparametrization gradients within WVI and second-order Price gradients within BBVI, allowing for a true 1-to-1 comparison.

Key Insights and Results

The authors evaluated their claims using the PosteriorDB benchmark suite, covering a diverse set of Bayesian inference problems ranging from low-dimensional logistic regression ($d=3$) to high-dimensional Poisson mixed effects models ($d=237$).

1. Theoretical Parity

When equipped with Price’s gradient, both SPGD and SPBWGD achieve an identical iteration complexity of: $$O\left(d\kappa\epsilon^{-1} + \sqrt{d}\kappa^{3/2}\log(\kappa\Delta^2)\epsilon^{-1/2} + \kappa^2\log\epsilon^{-1}\right)$$ This breakthrough closes the theoretical gap that previously favored WVI.

2. Empirical Confirmation

Across almost all PosteriorDB benchmarks, the choice of gradient estimator was the dominant factor in success. Methods using Price’s gradient converged significantly faster and to lower free energy levels than those using the first-order reparametrization gradient, regardless of whether they operated in Bures-Wasserstein or parameter space.

3. The Cost of Curvature

Price’s gradient is highly effective because its variance vanishes when the target distribution is Gaussian. However, this statistical efficiency comes at a computational cost: it requires $O(d^3)$ operations per iteration due to matrix-matrix products, compared to $O(d^2)$ for first-order methods.

4. Numerical Robustness

Interestingly, the authors noted that WVI requires numerically sensitive operations, such as matrix square roots and Cholesky decompositions of the full covariance matrix. In practice, this makes WVI potentially less robust than parameter-space BBVI. In challenging models (like the "Rats" dataset), SPGD in parameter space remained stable over a wider range of step sizes than its Wasserstein counterpart.

Limitations and Future Work

While this paper provides a definitive answer to the BBVI vs. WVI debate for Gaussian families, several challenges remain:

  • Computational Scalability: Because Price’s gradient requires computing and storing Hessians, it scales poorly to very high-dimensional problems ($d \gg 1000$).
  • Beyond Gaussians: The current theoretical analysis is restricted to the Gaussian variational family. Extending these insights to more complex distributions, such as Normalizing Flows, is an open challenge.
  • Natural Gradients: The study focuses on Wasserstein versus Euclidean geometry but does not deeply explore Natural Gradient VI (NGVI), which uses the Fisher information metric. The authors suggest NGVI might still offer unique benefits for location parameters that Price’s gradient does not address.

Summary updated

Abstract

For approximating a target distribution given only its unnormalized log-density, stochastic gradient-based variational inference (VI) algorithms are a popular approach. For example, Wasserstein VI (WVI) and black-box VI (BBVI) perform gradient descent in measure space (Bures-Wasserstein space) and parameter space, respectively. Previously, for the Gaussian variational family, convergence guarantees for WVI have shown superiority over existing results for black-box VI with the reparametrization gradient, suggesting the measure space approach might provide some unique benefits. In this work, however, we close this gap by obtaining identical state-of-the-art iteration complexity guarantees for both. In particular, we identify that WVI's superiority stems from the specific gradient estimator it uses, which BBVI can also leverage with minor modifications. The estimator in question is usually associated with Price's theorem and utilizes second-order information (Hessians) of the target log-density. We will refer to this as Price's gradient. On the flip side, WVI can be made more widely applicable by using the reparametrization gradient, which requires only gradients of the log-density. We empirically demonstrate that the use of Price's gradient is the major source of performance improvement.
BibTeX
@inproceedings{kim2026stochastic,
  author = {Kyurae Kim and Qiang Fu and Yian Ma and Jacob R. Gardner and Trevor Campbell},
  title = {{Stochastic Gradient Variational Inference with Price's Gradient Estimator from Bures-Wasserstein to Parameter Space}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=4HquuQN4qK},
}
Figure