Conditional mutual information for continuous random variables

492 Views Asked by At

Cover and Thomas provides definition of Conditional Mutual Information (CMI) for discrete random variables but doesn't say anything about continuous variables. Wikipedia has a section about a "more general definition" of CMI that is applicable to continuous case, which looks a bit involved given my current knowledge.

My question is, can't we extend the definition of CMI from discrete case to continuous case as follows :

$$I(X;Y \mid Z) = \iiint f_{X,Y,Z}(x,y,z)\log \frac{f_z(z)\ \ f_{X,Y,Z}(x,y,z)}{f_{X,Z}(x,z)\ \ f_{Y,Z}(y,z)} \, dx \, dy \, dz$$

where the $\ f$ s are density functions.

Is this definition wrong ? If yes, what is wrong here conceptually and what is the simplest correct definition of CMI for continuous variables ?