Aligning Theoretical Gradient Flows with Practical Loss Functions

19 concepts16 papers1 proposal

Imagine a cloud of particles representing a probability distribution. To solve tasks like image generation, Bayesian sampling, or reinforcement learning, we must reshape this cloud to match a target distribution. Mathematically, the most elegant way to describe this evolution is a Wasserstein gradient flow—a continuous-time trajectory that drives a distribution along the steepest geometric descent path to minimize an energy functional.

As Espen Bernton demonstrates in Langevin Monte Carlo and JKO splitting, treating sampling algorithms as discrete approximations of these continuous geometric flows provides powerful tools to prove convergence rates. Yet, a profound friction exists at the heart of modern machine learning: the geometric metrics used to prove that these flows work are rarely the same as the loss functions we use to train them.

This theory-practice gap—often called the optimization metric mismatch—means that models with excellent theoretical guarantees can fail to optimize in practice, while models that succeed empirically often do so by "secretly" satisfying the math in unexpected ways.


The Metric Mismatch: Geometry vs. Information

Why does this mismatch happen? Theoretical guarantees—such as universal approximation, mode preservation, or operator contraction—are easiest to establish using geometric distances like the Wasserstein distance or Total Variation (TV) distance. These metrics are highly sensitive to the underlying spatial geometry of the data.

In contrast, practical optimization almost exclusively relies on maximum likelihood estimation (MLE) or adversarial objectives, which minimize information-theoretic divergences like the Kullback-Leibler (KL) or Jensen-Shannon (JS) divergence. These metrics focus on point-by-point probability ratios rather than spatial transport.

This discrepancy creates severe theoretical blind spots across several domains:


The Bias of Practical Gradient Approximations

Even when we explicitly design an algorithm to follow a Wasserstein gradient flow, we cannot compute the exact continuous-time velocity fields for an arbitrary distribution. Instead, particle-based methods like Stein Variational Gradient Descent (SVGD) approximate these fields using kernels.

This approximation introduces a structural bias. In On the Convergence of SVGD in KL divergence via Approximate gradient flow, Fujisawa et al. formalize this via an $(\epsilon, \delta)$-approximate gradient flow framework. They model practical particle updates as biased versions of the ideal Wasserstein flow, explicitly quantifying how kernel approximations introduce a mathematical bias that prevents perfect KL minimization.

Ideal Wasserstein Flow (Exact Gradients) ───► Perfect KL Minimization
                                                     ▲
                                                     │ (Kernel Bias / Gap)
Practical Particle Flow (Kernel Approximations) ─────┘

Because of this kernel-induced bias, deterministic particle methods often achieve only a constant-order approximation of the true target distribution, a limitation explored by He et al. in Regularized Stein Variational Gradient Flow. Furthermore, Korba et al. note in Statistical and Geometrical properties of the Kernel Kullback-Leibler divergence that proving global convergence for these kernelized gradient flows remains an open mathematical challenge.

This gap extends to non-particle methods as well: Franceschi et al. argue in A Neural Tangent Kernel Perspective of GANs that establishing a formal connection between theoretical gradient flows and non-Integral Probability Metric losses remains unresolved. Similarly, Rehman et al. show in Efficient Regression-based Training of Normalizing Flows for Boltzmann Generators that regression-based flow training bypasses MLE instabilities but loses the theoretical guarantee of a perfect likelihood match. Finally, Zhang et al. point out in Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning that non-monotonic sampling schedules escape local modes but lack integration into standard Wasserstein flow theory.


Secret Alignments and Path Corrections

If the metric mismatch is so pervasive, why do modern generative models like diffusion and flow matching work so remarkably well? Recent insights suggest that successful objectives "secretly" optimize the missing metrics.

In Score-based Generative Modeling Secretly Minimizes the Wasserstein Distance, Kwon et al. resolve a core paradox of diffusion models. While score-based models are trained via a score-matching loss (closely tied to KL divergence), the authors prove they are simultaneously minimizing the Wasserstein distance. This dual optimization ensures that even when the training loss does not explicitly bound the target distribution, the generated samples still converge in a geometrically meaningful way.

Score-Matching Loss (KL-like) ───► [Secretly Minimizes] ───► Wasserstein Distance
                                                                  │
                                                                  ▼
                                                      Geometrically Sound Samples

To explicitly close these gaps without relying on accidental alignments, researchers are designing new objectives that directly couple flow dynamics with target metrics:

By explicitly identifying where our loss functions deviate from our geometric theories, this research direction transforms distribution matching from an empirical art into a mathematically aligned science.

Key Foundational Work

A Distributional Perspective on Reinforcement Learning (2017) — contributed 1 concept elements to this direction.
Langevin Monte Carlo and JKO splitting (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_optimization_metric_mismatch_2afc5892,
  title = {optimization metric mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-metric-mismatch/rdr_2afc5892fcc74e568c1f11bfc8af06a5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_optimization_metric_mismatch_2afc5892_v1,
  title = {optimization metric mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-metric-mismatch/rdr_2afc5892fcc74e568c1f11bfc8af06a5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake