We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Mathematics > Probability

Title: Logarithmic divergences from optimal transport and Rényi geometry

Abstract: Divergences, also known as contrast functions, are distance-like quantities defined on manifolds of non-negative or probability measures. Using the duality in optimal transport, we introduce and study the one-parameter family of $L^{(\pm \alpha)}$-divergences. It includes the Bregman divergence corresponding to the Euclidean quadratic cost, and the $L$-divergence introduced by Pal and the author in connection with portfolio theory and a logarithmic cost function. They admit natural generalizations of exponential family that are closely related to the $\alpha$-family and $q$-exponential family. In particular, the $L^{(\pm \alpha)}$-divergences of the corresponding potential functions are R\'{e}nyi divergences. Using this unified framework we prove that the induced geometries are dually projectively flat with constant sectional curvatures, and a generalized Pythagorean theorem holds true. Conversely, we show that if a statistical manifold is dually projectively flat with constant curvature $\pm \alpha$ with $\alpha > 0$, then it is locally induced by an $L^{(\mp \alpha)}$-divergence. We define in this context a canonical divergence which extends the one for dually flat manifolds.
Comments: 39 pages. Revised
Subjects: Probability (math.PR); Information Theory (cs.IT); Statistics Theory (math.ST)
Cite as: arXiv:1712.03610 [math.PR]
  (or arXiv:1712.03610v3 [math.PR] for this version)

Submission history

From: Ting-Kam Leonard Wong [view email]
[v1] Sun, 10 Dec 2017 23:42:29 GMT (27kb)
[v2] Wed, 27 Dec 2017 23:35:27 GMT (32kb)
[v3] Tue, 4 Sep 2018 01:52:46 GMT (32kb)

Link back to: arXiv, form interface, contact.