Confliction of the Convexity and Metric Properties in f-Divergences

Authors:
Mohammadali Khosravifard;Dariush Fooladivanda;T. Aaron Gulliver
Affiliations:
-;-;-
Venue:
IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences
Year:
2007

Citing 2
Cited 1

Elements of information theory

Elements of information theory
A new metric for probability distributions

IEEE Transactions on Information Theory

Information, Divergence and Risk for Binary Experiments

The Journal of Machine Learning Research

Quantified Score

Hi-index	0.00

Visualization

Abstract

We prove that the Variational distance (and its positive multiples) is the only f-divergence that satisfies both the identity of indiscernibles and the triangle inequality. Therefore it is the unique f-divergence which serves as a metric. This point is interpreted as a fundamental confliction of the convexity for f(x) with the metric properties for its associated f-divergence. Therefore, we relax the convexity of f(x) and replace it with other constraints to create new metrics.