When do explicit penalties improve flow matching?
Abstract
Flow matching can learn structure directly from data, raising the question of when additional constraints improve generation. We study penalties that encode such constraints, with physics-informed flow matching as a motivating example. Satisfying a constraint more closely need not make the generated distribution more accurate. We therefore compare penalised and unpenalised generative modelling through their expected squared $2$-Wasserstein error. Our analysis covers both vector-valued and functional data in a Hilbert-space setting. We first derive geometric criteria that account for constraint violations and the distribution of samples after projection onto the constraint set. We then solve the statistical comparison explicitly for specific velocity-field classes and constraints via flow matching. In this model, penalisation improves accuracy exactly when the reduction in estimation variance outweighs the increase in squared bias. The resulting formula shows how this trade-off depends on sample size, the model class, the noise schedule, and the penalty weight.
Related
- Beyond Reachability: When Edit Order Preserves Couplings in Graph Flow Matching
- Flowz: When Does Learning Forecast Errors Add Value? Conditional Flow Matching for Financial Price Forecasts
- Average Energy Dissipation Rates of Implicit-Explicit Runge-Kutta Methods for Gradient Flow Problems
- Scale or speed? When do path signatures improve volatility-regime forecasts?
- When do data mixtures improve scaling laws? Insights from high-dimensional regression
BibTeX
@misc{kunkel2026when,
title = {{When do explicit penalties improve flow matching?}},
author = {Lea Kunkel and Alberto González-Sanz and Sophie Langer},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.15689},
eprint = {2610.15689},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.15689},
}
RIS
TY - UNPB
TI - When do explicit penalties improve flow matching?
AU - Lea Kunkel
AU - Alberto González-Sanz
AU - Sophie Langer
PY - 2026
DA - 2026/10/11
PB - arXiv
DO - 10.48550/arxiv.2610.15689
UR - https://arxiv.org/abs/2610.15689
AB - Flow matching can learn structure directly from data, raising the question of when additional constraints improve generation. We study penalties that encode such constraints, with physics-informed flow matching as a motivating example. Satisfying a constraint more closely need not make the generated distribution more accurate. We therefore compare penalised and unpenalised generative modelling through their expected squared $2$-Wasserstein error. Our analysis covers both vector-valued and functional data in a Hilbert-space setting. We first derive geometric criteria that account for constraint violations and the distribution of samples after projection onto the constraint set. We then solve the statistical comparison explicitly for specific velocity-field classes and constraints via flow matching. In this model, penalisation improves accuracy exactly when the reduction in estimation variance outweighs the increase in squared bias. The resulting formula shows how this trade-off depends on sample size, the model class, the noise schedule, and the penalty weight.
ER -