← 研踪 PaperTrailarXiv:2611.18394 · v1
Lipschitz-Constrained Discriminators in Generative Modeling: Asymptotic Behavior and Support Mismatch
首次提交 2026-11-14 00:38
版本更新 2026-11-14 00:38
雷达观测 2026-11-16 11:41
摘要
We study the asymptotic behavior of optimal discriminators associated to Lipschitz regularized KL-divergences of two probability distributions and that interpolate between the rescaled Wasserstein metric by some parameter and the KL divergence. The parameter plays the role of a Lipschitz constraint and is viewed as a regularization parameter: small values of enforce a more geometry-aware discriminator, while as increases the geometric regularization is progressively relaxed. It is a known result [3] that such divergences converge to the KL divergence as . In this paper, we characterize the asymptotic behavior of the optimal discriminators as in different regimes: Firstly, under and the Radon-Nikodym derivative is bounded from above and bounded away from zero. Secondly, when and are discrete distributions satisfying , and finally when and are non absolutely continuous discrete distributions. This result motivates training strategies which involve discriminators, such as adversarial generative models. In practice, the data distribution is often represented by a discrete empirical measure, while, especially at the early stages of the training, the model distribution may have little or no overlap with the empirical data distribution.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-14 00:38 · 最新观测版本
Lipschitz-Constrained Discriminators in Generative Modeling: Asymptotic Behavior and Support Mismatch
We study the asymptotic behavior of optimal discriminators associated to Lipschitz regularized KL-divergences of two probability distributions and that interpolate between the rescaled Wasserstein metric by some parameter and the KL divergence. The parameter plays the role of a Lipschitz constraint and is viewed as a regularization parameter: small values of enforce a more geometry-aware discriminator, while as increases the geometric regularization is progressively relaxed. It is a known result [3] that such divergences converge to the KL divergence as . In this paper, we characterize the asymptotic behavior of the optimal discriminators as in different regimes: Firstly, under and the Radon-Nikodym derivative is bounded from above and bounded away from zero. Secondly, when and are discrete distributions satisfying , and finally when and are non absolutely continuous discrete distributions. This result motivates training strategies which involve discriminators, such as adversarial generative models. In practice, the data distribution is often represented by a discrete empirical measure, while, especially at the early stages of the training, the model distribution may have little or no overlap with the empirical data distribution.
查看此版本解读 →查看此版本原文 ↗