References & Citations
Computer Science > Information Theory
Title: A generalization of the $α$-divergences based on comparable and distinct weighted means
(Submitted on 27 Jan 2020 (v1), revised 5 Feb 2020 (this version, v3), latest version 17 Feb 2020 (v4))
Abstract: We generalize the renown family of $\alpha$-divergences in information geometry using comparable and distinct weighted means. After stating several general definitions, we report the explicit closed-form formula for the quasi-arithmetic $\alpha$-divergences and instantiate it for its subfamily of power $\alpha$-divergences, including in their limit cases of $\alpha=0$ and $\alpha=1$, generalizations of the Kullback-Leibler divergence ($1$-divergence) and the reverse Kullback-Leibler divergence ($0$-divergence), respectively. We show that these generalized Kullback-Leibler divergences can be rewritten as conformal representational Bregman divergences.
Submission history
From: Frank Nielsen [view email][v1] Mon, 27 Jan 2020 10:09:21 GMT (12kb)
[v2] Thu, 30 Jan 2020 03:00:16 GMT (15kb)
[v3] Wed, 5 Feb 2020 02:46:06 GMT (17kb)
[v4] Mon, 17 Feb 2020 09:23:43 GMT (20kb)
Link back to: arXiv, form interface, contact.