Conditional mutual information

For random variables ${\displaystyle X}$, ${\displaystyle Y}$, and ${\displaystyle Z}$ with support sets ${\displaystyle {\mathcal {X}}}$, ${\displaystyle {\mathcal {Y}}}$ and ${\displaystyle {\mathcal {Z}}}$, we define the conditional mutual information as

This may be written in terms of the expectation operator: ${\displaystyle I(X;Y|Z)=\mathbb {E} _{Z}[D_{\mathrm {KL} }(P_{(X,Y)|Z}\|P_{X|Z}\otimes P_{Y|Z})]}$.

Thus ${\displaystyle I(X;Y|Z)}$ is the expected (with respect to ${\displaystyle Z}$) Kullback–Leibler divergence from the conditional joint distribution ${\displaystyle P_{(X,Y)|Z}}$ to the product of the conditional marginals ${\displaystyle P_{X|Z}}$ and ${\displaystyle P_{Y|Z}}$. Compare with the definition of mutual information.

For discrete random variables ${\displaystyle X}$, ${\displaystyle Y}$, and ${\displaystyle Z}$ with support sets ${\displaystyle {\mathcal {X}}}$, ${\displaystyle {\mathcal {Y}}}$ and ${\displaystyle {\mathcal {Z}}}$, the conditional mutual information ${\displaystyle I(X;Y|Z)}$ is as follows

${\displaystyle I(X;Y|Z)=\sum _{z\in {\mathcal {Z}}}p_{Z}(z)\sum _{y\in {\mathcal {Y}}}\sum _{x\in {\mathcal {X}}}p_{X,Y|Z}(x,y|z)\log {\frac {p_{X,Y|Z}(x,y|z)}{p_{X|Z}(x|z)p_{Y|Z}(y|z)}}}$

where the marginal, joint, and/or conditional probability mass functions are denoted by ${\displaystyle p}$ with the appropriate subscript. This can be simplified as

For (absolutely) continuous random variables ${\displaystyle X}$, ${\displaystyle Y}$, and ${\displaystyle Z}$ with support sets ${\displaystyle {\mathcal {X}}}$, ${\displaystyle {\mathcal {Y}}}$ and ${\displaystyle {\mathcal {Z}}}$, the conditional mutual information ${\displaystyle I(X;Y|Z)}$ is as follows

${\displaystyle I(X;Y|Z)=\int _{\mathcal {Z}}{\bigg (}\int _{\mathcal {Y}}\int _{\mathcal {X}}\log \left({\frac {p_{X,Y|Z}(x,y|z)}{p_{X|Z}(x|z)p_{Y|Z}(y|z)}}\right)p_{X,Y|Z}(x,y|z)dxdy{\bigg )}p_{Z}(z)dz}$

where the marginal, joint, and/or conditional probability density functions are denoted by ${\displaystyle p}$ with the appropriate subscript. This can be simplified as

Alternatively, we may write in terms of joint and conditional entropies as[3]

${\displaystyle I(X;Y|Z)=H(X,Z)+H(Y,Z)-H(X,Y,Z)-H(Z)=H(X|Z)-H(X|Y,Z)=H(X|Z)+H(Y|Z)-H(X,Y|Z).}$

This can be rewritten to show its relationship to mutual information

${\displaystyle I(X;Y|Z)=I(X;Y,Z)-I(X;Z)}$

usually rearranged as the chain rule for mutual information

${\displaystyle I(X;Y,Z)=I(X;Z)+I(X;Y|Z)}$

Another equivalent form of the above is[4]

${\displaystyle I(X;Y|Z)=H(Z|X)+H(X)+H(Z|Y)+H(Y)-H(Z|X,Y)-H(X,Y)-H(Z)=I(X;Y)+H(Z|X)+H(Z|Y)-H(Z|X,Y)-H(Z)}$

Like mutual information, conditional mutual information can be expressed as a Kullback–Leibler divergence:

${\displaystyle I(X;Y|Z)=D_{\mathrm {KL} }[p(X,Y,Z)\|p(X|Z)p(Y|Z)p(Z)].}$

Or as an expected value of simpler Kullback–Leibler divergences:

${\displaystyle I(X;Y|Z)=\sum _{z\in {\mathcal {Z}}}p(Z=z)D_{\mathrm {KL} }[p(X,Y|z)\|p(X|z)p(Y|z)]}$,

${\displaystyle I(X;Y|Z)=\sum _{y\in {\mathcal {Y}}}p(Y=y)D_{\mathrm {KL} }[p(X,Z|y)\|p(X|Z)p(Z|y)]}$.

A more general definition of conditional mutual information, applicable to random variables with continuous or other arbitrary distributions, will depend on the concept of regular conditional probability. (See also.[5][6])

Let ${\displaystyle (\Omega ,{\mathcal {F}},{\mathfrak {P}})}$ be a probability space, and let the random variables ${\displaystyle X}$, ${\displaystyle Y}$, and ${\displaystyle Z}$ each be defined as a Borel-measurable function from ${\displaystyle \Omega }$ to some state space endowed with a topological structure.

Consider the Borel measure (on the σ-algebra generated by the open sets) in the state space of each random variable defined by assigning each Borel set the ${\displaystyle {\mathfrak {P}}}$-measure of its preimage in ${\displaystyle {\mathcal {F}}}$. This is called the pushforward measure ${\displaystyle X_{*}{\mathfrak {P}}={\mathfrak {P}}{\big (}X^{-1}(\cdot ){\big )}.}$ The support of a random variable is defined to be the topological support of this measure, i.e. ${\displaystyle \mathrm {supp} \,X=\mathrm {supp} \,X_{*}{\mathfrak {P}}.}$

Now we can formally define the conditional probability measure given the value of one (or, via the product topology, more) of the random variables. Let ${\displaystyle M}$ be a measurable subset of ${\displaystyle \Omega ,}$ (i.e. ${\displaystyle M\in {\mathcal {F}},}$) and let ${\displaystyle x\in \mathrm {supp} \,X.}$ Then, using the disintegration theorem:

${\displaystyle {\mathfrak {P}}(M|X=x)=\lim _{U\ni x}{\frac {{\mathfrak {P}}(M\cap \{X\in U\})}{{\mathfrak {P}}(\{X\in U\})}}\qquad {\textrm {and}}\qquad {\mathfrak {P}}(M|X)=\int _{M}d{\mathfrak {P}}{\big (}\omega |X=X(\omega ){\big )},}$

where the limit is taken over the open neighborhoods ${\displaystyle U}$ of ${\displaystyle x}$, as they are allowed to become arbitrarily smaller with respect to set inclusion.

Finally we can define the conditional mutual information via Lebesgue integration:

${\displaystyle I(X;Y|Z)=\int _{\Omega }\log {\Bigl (}{\frac {d{\mathfrak {P}}(\omega |X,Z)\,d{\mathfrak {P}}(\omega |Y,Z)}{d{\mathfrak {P}}(\omega |Z)\,d{\mathfrak {P}}(\omega |X,Y,Z)}}{\Bigr )}d{\mathfrak {P}}(\omega ),}$

where the integrand is the logarithm of a Radon–Nikodym derivative involving some of the conditional probability measures we have just defined.

In an expression such as ${\displaystyle I(A;B|C),}$ ${\displaystyle A,}$ ${\displaystyle B,}$ and ${\displaystyle C}$ need not necessarily be restricted to representing individual random variables, but could also represent the joint distribution of any collection of random variables defined on the same probability space. As is common in probability theory, we may use the comma to denote such a joint distribution, e.g. ${\displaystyle I(A_{0},A_{1};B_{1},B_{2},B_{3}|C_{0},C_{1}).}$ Hence the use of the semicolon (or occasionally a colon or even a wedge ${\displaystyle \wedge }$) to separate the principal arguments of the mutual information symbol. (No such distinction is necessary in the symbol for joint entropy, since the joint entropy of any number of random variables is the same as the entropy of their joint distribution.)

The conditional mutual information can be used to inductively define a multivariate mutual information in a set- or measure-theoretic sense in the context of information diagrams. In this sense we define the multivariate mutual information as follows:

${\displaystyle I(X_{1};\ldots ;X_{n+1})=I(X_{1};\ldots ;X_{n})-I(X_{1};\ldots ;X_{n}|X_{n+1}),}$

where

${\displaystyle I(X_{1};\ldots ;X_{n}|X_{n+1})=\mathbb {E} _{X_{n+1}}[D_{\mathrm {KL} }(P_{(X_{1},\ldots ,X_{n})|X_{n+1}}\|P_{X_{1}|X_{n+1}}\otimes \cdots \otimes P_{X_{n}|X_{n+1}})].}$

This definition is identical to that of interaction information except for a change in sign in the case of an odd number of random variables. A complication is that this multivariate mutual information (as well as the interaction information) can be positive, negative, or zero, which makes this quantity difficult to interpret intuitively. In fact, for ${\displaystyle n}$ random variables, there are ${\displaystyle 2^{n}-1}$ degrees of freedom for how they might be correlated in an information-theoretic sense, corresponding to each non-empty subset of these variables. These degrees of freedom are bounded by various Shannon- and non-Shannon-type inequalities in information theory.