Two diagonal Gaussian reference measures give the same measures of finite relative entropy and of finite relative Fisher information; the relative entropies differ by half the integral of the difference of the weighted squares plus a constant, and the relative scores differ by the difference of the scaling maps.
In the settings of The Real Numbers: Standing Notation and Background and The Intrinsic Calculus on the Wasserstein Space: Standing Notation, let and be variance vectors, with scaling maps and , weighted squares and () and normalizing constants and , and let and be the diagonal Gaussian measures with variances and . Finite relative entropy and are those of that definition on ; finite Fisher information relative to a diagonal Gaussian measure and the relative scores are those of that definition, and for the classes are as there; integrable is as in Probability Measures on Euclidean Space and Random Vectors: Standing Notation §measures. Let .
1. (Equal domains) has finite relative entropy with respect to if and only if it has finite relative entropy with respect to , and has finite Fisher information relative to if and only if it has finite Fisher information relative to .
2. (The shift of the relative entropy) The function is integrable with respect to , and if has finite relative entropy with respect to , then
3. (The shift of the relative score) If has finite Fisher information relative to , so that is defined by clause 1, then in .
Loading…
No relations recorded yet.