Answers are grounded in this paper and cite their sources.
The directed divergences of two probability densities <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">q</tex> are given by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">∫ p(x) log (p(x)/q(x))dx</tex> and by the same expression with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">q</tex> interchanged; the divergence is the sum of the directed divergences. These quantities have applications in information theory and to the problem of assigning prior probabilities subject to constraints. It is shown that the directed divergences and their positive linear combinations, including the divegeoce, are characterized by axioms of positivity, additivity, and finiteness, which are fundamental in work on prior probabilities. In the course of the proof, the latter two are shown to imply yet another axiom: linear invariance.
No takes yet. Share an insight, caveat, or question.
Ryan F. Johnson (1979) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: