Mutual information

Let ${\displaystyle (X,Y)}$  be a pair of random variables with values over the space ${\displaystyle {\mathcal {X}}\times {\mathcal {Y}}}$ . If their joint distribution is ${\displaystyle P_{(X,Y)}}$  and the marginal distributions are ${\displaystyle P_{X}}$  and ${\displaystyle P_{Y}}$ , the mutual information is defined as

where ${\displaystyle D_{\mathrm {KL} }}$  is the Kullback–Leibler divergence, and ${\displaystyle P_{X}\otimes P_{Y}}$  is the outer product distribution which assigns probability ${\displaystyle P_{X}($x)\cdot P_{Y}(y)}   to each ${\displaystyle (x,y)}$ .

Notice, as per property of the Kullback–Leibler divergence, that ${\displaystyle I(X;Y)}$  is equal to zero precisely when the joint distribution coincides with the product of the marginals, i.e. when ${\displaystyle X}$  and ${\displaystyle Y}$  are independent (and hence observing ${\displaystyle Y}$  tells you nothing about ${\displaystyle X}$ ). ${\displaystyle I(X;Y)}$  is non-negative, it is a measure of the price for encoding ${\displaystyle (X,Y)}$  as a pair of independent random variables when in reality they are not.

If the natural logarithm is used, the unit of mutual information is the nat. If the log base 2 is used, the unit of mutual information is the shannon, also known as the bit. If the log base 10 is used, the unit of mutual information is the hartley, also known as the ban or the dit.

The mutual information of two jointly discrete random variables ${\displaystyle X}$  and ${\displaystyle Y}$  is calculated as a double sum:^[3]: 20

where ${\displaystyle P_{(X,Y)}}$  is the joint probability mass functionof${\displaystyle X}$  and ${\displaystyle Y}$ , and ${\displaystyle P_{X}}$  and ${\displaystyle P_{Y}}$  are the marginal probability mass functions of ${\displaystyle X}$  and ${\displaystyle Y}$  respectively.

In the case of jointly continuous random variables, the double sum is replaced by a double integral:^[3]: 251

where ${\displaystyle P_{(X,Y)}}$  is now the joint probability density function of ${\displaystyle X}$  and ${\displaystyle Y}$ , and ${\displaystyle P_{X}}$  and ${\displaystyle P_{Y}}$  are the marginal probability density functions of ${\displaystyle X}$  and ${\displaystyle Y}$  respectively.

Intuitively, mutual information measures the information that ${\displaystyle X}$  and ${\displaystyle Y}$  share: It measures how much knowing one of these variables reduces uncertainty about the other. For example, if ${\displaystyle X}$  and ${\displaystyle Y}$  are independent, then knowing ${\displaystyle X}$  does not give any information about ${\displaystyle Y}$  and vice versa, so their mutual information is zero. At the other extreme, if ${\displaystyle X}$  is a deterministic function of ${\displaystyle Y}$  and ${\displaystyle Y}$  is a deterministic function of ${\displaystyle X}$  then all information conveyed by ${\displaystyle X}$  is shared with ${\displaystyle Y}$ : knowing ${\displaystyle X}$  determines the value of ${\displaystyle Y}$  and vice versa. As a result, the mutual information is the same as the uncertainty contained in ${\displaystyle Y}$  (or${\displaystyle X}$ ) alone, namely the entropyof${\displaystyle Y}$  (or${\displaystyle X}$ ). A very special case of this is when ${\displaystyle X}$  and ${\displaystyle Y}$  are the same random variable.

Mutual information is a measure of the inherent dependence expressed in the joint distributionof${\displaystyle X}$  and ${\displaystyle Y}$  relative to the marginal distribution of ${\displaystyle X}$  and ${\displaystyle Y}$  under the assumption of independence. Mutual information therefore measures dependence in the following sense: ${\displaystyle \operatorname {I} (X;Y)=0}$  if and only if ${\displaystyle X}$  and ${\displaystyle Y}$  are independent random variables. This is easy to see in one direction: if ${\displaystyle X}$  and ${\displaystyle Y}$  are independent, then ${\displaystyle p_{(X,Y)}(x,y)=p_{X}($x)\cdot p_{Y}(y)}  , and therefore:

${\displaystyle \log {\left({\frac {p_{(X,Y)}(x,y)}{p_{X}($x)\,p_{Y}(y)}}\right)}=\log 1=0.}  

Moreover, mutual information is nonnegative (i.e. ${\displaystyle \operatorname {I} (X;Y)\geq 0}$  see below) and symmetric (i.e. ${\displaystyle \operatorname {I} (X;Y)=\operatorname {I} (Y;X)}$  see below).

Using Jensen's inequality on the definition of mutual information we can show that ${\displaystyle \operatorname {I} (X;Y)}$  is non-negative, i.e.^[3]: 28

${\displaystyle \operatorname {I} (X;Y)\geq 0}$ 

${\displaystyle \operatorname {I} (X;Y)=\operatorname {I} (Y;X)}$ 

The proof is given considering the relationship with entropy, as shown below.

If${\displaystyle C}$  is independent of ${\displaystyle (A,B)}$ , then

${\displaystyle \operatorname {I} (Y;A,B,C)-\operatorname {I} (Y;A,B)\geq \operatorname {I} (Y;A,C)-\operatorname {I} (Y;A)}$ .^[4]

Mutual information can be equivalently expressed as:

${\displaystyle {\begin{aligned}\operatorname {I} (X;Y)&{}\equiv \mathrm {H} ($X)-\mathrm {H} (X\mid Y)\\&{}\equiv \mathrm {H} (Y)-\mathrm {H} (Y\mid X)\\&{}\equiv \mathrm {H} (X)+\mathrm {H} (Y)-\mathrm {H} (X,Y)\\&{}\equiv \mathrm {H} (X,Y)-\mathrm {H} (X\mid Y)-\mathrm {H} (Y\mid X)\end{aligned}}}  

where ${\displaystyle \mathrm {H} ($X)}   and ${\displaystyle \mathrm {H} ($Y)}   are the marginal entropies, ${\displaystyle \mathrm {H} (X\mid Y)}$  and ${\displaystyle \mathrm {H} (Y\mid X)}$  are the conditional entropies, and ${\displaystyle \mathrm {H} (X,Y)}$  is the joint entropyof${\displaystyle X}$  and ${\displaystyle Y}$ .

Notice the analogy to the union, difference, and intersection of two sets: in this respect, all the formulas given above are apparent from the Venn diagram reported at the beginning of the article.

In terms of a communication channel in which the output ${\displaystyle Y}$  is a noisy version of the input ${\displaystyle X}$ , these relations are summarised in the figure:

Because ${\displaystyle \operatorname {I} (X;Y)}$  is non-negative, consequently, ${\displaystyle \mathrm {H} ($X)\geq \mathrm {H} (X\mid Y)}  . Here we give the detailed deduction of ${\displaystyle \operatorname {I} (X;Y)=\mathrm {H} ($Y)-\mathrm {H} (Y\mid X)}   for the case of jointly discrete random variables:

${\displaystyle {\begin{aligned}\operatorname {I} (X;Y)&{}=\sum _{x\in {\mathcal {X}},y\in {\mathcal {Y}}}p_{(X,Y)}(x,y)\log {\frac {p_{(X,Y)}(x,y)}{p_{X}($x)p_{Y}(y)}}\\&{}=\sum _{x\in {\mathcal {X}},y\in {\mathcal {Y}}}p_{(X,Y)}(x,y)\log {\frac {p_{(X,Y)}(x,y)}{p_{X}(x)}}-\sum _{x\in {\mathcal {X}},y\in {\mathcal {Y}}}p_{(X,Y)}(x,y)\log p_{Y}(y)\\&{}=\sum _{x\in {\mathcal {X}},y\in {\mathcal {Y}}}p_{X}(x)p_{Y\mid X=x}(y)\log p_{Y\mid X=x}(y)-\sum _{x\in {\mathcal {X}},y\in {\mathcal {Y}}}p_{(X,Y)}(x,y)\log p_{Y}(y)\\&{}=\sum _{x\in {\mathcal {X}}}p_{X}(x)\left(\sum _{y\in {\mathcal {Y}}}p_{Y\mid X=x}(y)\log p_{Y\mid X=x}(y)\right)-\sum _{y\in {\mathcal {Y}}}\left(\sum _{x\in {\mathcal {X}}}p_{(X,Y)}(x,y)\right)\log p_{Y}(y)\\&{}=-\sum _{x\in {\mathcal {X}}}p_{X}(x)\mathrm {H} (Y\mid X=x)-\sum _{y\in {\mathcal {Y}}}p_{Y}(y)\log p_{Y}(y)\\&{}=-\mathrm {H} (Y\mid X)+\mathrm {H} (Y)\\&{}=\mathrm {H} (Y)-\mathrm {H} (Y\mid X).\\\end{aligned}}}  

The proofs of the other identities above are similar. The proof of the general case (not just discrete) is similar, with integrals replacing sums.

Intuitively, if entropy ${\displaystyle \mathrm {H} ($Y)}   is regarded as a measure of uncertainty about a random variable, then ${\displaystyle \mathrm {H} (Y\mid X)}$  is a measure of what ${\displaystyle X}$  does not say about ${\displaystyle Y}$ . This is "the amount of uncertainty remaining about ${\displaystyle Y}$  after ${\displaystyle X}$  is known", and thus the right side of the second of these equalities can be read as "the amount of uncertainty in ${\displaystyle Y}$ , minus the amount of uncertainty in ${\displaystyle Y}$  which remains after ${\displaystyle X}$  is known", which is equivalent to "the amount of uncertainty in ${\displaystyle Y}$  which is removed by knowing ${\displaystyle X}$ ". This corroborates the intuitive meaning of mutual information as the amount of information (that is, reduction in uncertainty) that knowing either variable provides about the other.

Note that in the discrete case ${\displaystyle \mathrm {H} (Y\mid Y)=0}$  and therefore ${\displaystyle \mathrm {H} ($Y)=\operatorname {I} (Y;Y)}  . Thus ${\displaystyle \operatorname {I} (Y;Y)\geq \operatorname {I} (X;Y)}$ , and one can formulate the basic principle that a variable contains at least as much information about itself as any other variable can provide.

For jointly discrete or jointly continuous pairs ${\displaystyle (X,Y)}$ , mutual information is the Kullback–Leibler divergence from the product of the marginal distributions, ${\displaystyle p_{X}\cdot p_{Y}}$ , of the joint distribution ${\displaystyle p_{(X,Y)}}$ , that is,

Furthermore, let ${\displaystyle p_{(X,Y)}(x,y)=p_{X\mid Y=y}($x)*p_{Y}(y)}   be the conditional mass or density function. Then, we have the identity

The proof for jointly discrete random variables is as follows:

${\displaystyle {\begin{aligned}\operatorname {I} (X;Y)&=\sum _{y\in {\mathcal {Y}}}\sum _{x\in {\mathcal {X}}}{p_{(X,Y)}(x,y)\log \left({\frac {p_{(X,Y)}(x,y)}{p_{X}($x)\,p_{Y}(y)}}\right)}\\&=\sum _{y\in {\mathcal {Y}}}\sum _{x\in {\mathcal {X}}}p_{X\mid Y=y}(x)p_{Y}(y)\log {\frac {p_{X\mid Y=y}(x)p_{Y}(y)}{p_{X}(x)p_{Y}(y)}}\\&=\sum _{y\in {\mathcal {Y}}}p_{Y}(y)\sum _{x\in {\mathcal {X}}}p_{X\mid Y=y}(x)\log {\frac {p_{X\mid Y=y}(x)}{p_{X}(x)}}\\&=\sum _{y\in {\mathcal {Y}}}p_{Y}(y)\;D_{\text{KL}}\!\left(p_{X\mid Y=y}\parallel p_{X}\right)\\&=\mathbb {E} _{Y}\left[D_{\text{KL}}\!\left(p_{X\mid Y}\parallel p_{X}\right)\right].\end{aligned}}}  

Similarly this identity can be established for jointly continuous random variables.

Note that here the Kullback–Leibler divergence involves integration over the values of the random variable ${\displaystyle X}$  only, and the expression ${\displaystyle D_{\text{KL}}(p_{X\mid Y}\parallel p_{X})}$  still denotes a random variable because ${\displaystyle Y}$  is random. Thus mutual information can also be understood as the expectation of the Kullback–Leibler divergence of the univariate distribution ${\displaystyle p_{X}}$ of${\displaystyle X}$  from the conditional distribution ${\displaystyle p_{X\mid Y}}$ of${\displaystyle X}$  given ${\displaystyle Y}$ : the more different the distributions ${\displaystyle p_{X\mid Y}}$  and ${\displaystyle p_{X}}$  are on average, the greater the information gain.

If samples from a joint distribution are available, a Bayesian approach can be used to estimate the mutual information of that distribution. The first work to do this, which also showed how to do Bayesian estimation of many other information-theoretic properties besides mutual information, was.^[5] Subsequent researchers have rederived ^[6] and extended ^[7] this analysis. See ^[8] for a recent paper based on a prior specifically tailored to estimation of mutual information per se. Besides, recently an estimation method accounting for continuous and multivariate outputs, ${\displaystyle Y}$ , was proposed in .^[9]

The Kullback-Leibler divergence formulation of the mutual information is predicated on that one is interested in comparing ${\displaystyle p(x,y)}$  to the fully factorized outer product ${\displaystyle p($x)\cdot p(y)}  . In many problems, such as non-negative matrix factorization, one is interested in less extreme factorizations; specifically, one wishes to compare ${\displaystyle p(x,y)}$  to a low-rank matrix approximation in some unknown variable ${\displaystyle w}$ ; that is, to what degree one might have

${\displaystyle p(x,y)\approx \sum _{w}p^{\prime }(x,w)p^{\prime \prime }(w,y)}$ 

Alternately, one might be interested in knowing how much more information ${\displaystyle p(x,y)}$  carries over its factorization. In such a case, the excess information that the full distribution ${\displaystyle p(x,y)}$  carries over the matrix factorization is given by the Kullback-Leibler divergence

${\displaystyle \operatorname {I} _{LRMA}=\sum _{y\in {\mathcal {Y}}}\sum _{x\in {\mathcal {X}}}{p(x,y)\log {\left({\frac {p(x,y)}{\sum _{w}p^{\prime }(x,w)p^{\prime \prime }(w,y)}}\right)}},}$ 

The conventional definition of the mutual information is recovered in the extreme case that the process ${\displaystyle W}$  has only one value for ${\displaystyle w}$ .

Several variations on mutual information have been proposed to suit various needs. Among these are normalized variants and generalizations to more than two variables.

Many applications require a metric, that is, a distance measure between pairs of points. The quantity

${\displaystyle {\begin{aligned}d(X,Y)&=\mathrm {H} (X,Y)-\operatorname {I} (X;Y)\\&=\mathrm {H} ($X)+\mathrm {H} (Y)-2\operatorname {I} (X;Y)\\&=\mathrm {H} (X\mid Y)+\mathrm {H} (Y\mid X)\\&=2\mathrm {H} (X,Y)-\mathrm {H} (X)-\mathrm {H} (Y)\end{aligned}}}  

satisfies the properties of a metric (triangle inequality, non-negativity, indiscernability and symmetry), where equality ${\displaystyle X=Y}$  is understood to mean that ${\displaystyle X}$  can be completely determined from ${\displaystyle Y}$ .^[10]

This distance metric is also known as the variation of information.

If${\displaystyle X,Y}$  are discrete random variables then all the entropy terms are non-negative, so ${\displaystyle 0\leq d(X,Y)\leq \mathrm {H} (X,Y)}$  and one can define a normalized distance

${\displaystyle D(X,Y)={\frac {d(X,Y)}{\mathrm {H} (X,Y)}}\leq 1.}$ 

The metric ${\displaystyle D}$  is a universal metric, in that if any other distance measure places ${\displaystyle X}$  and ${\displaystyle Y}$  close by, then the ${\displaystyle D}$  will also judge them close.^{[11][dubious – discuss]}

Plugging in the definitions shows that

${\displaystyle D(X,Y)=1-{\frac {\operatorname {I} (X;Y)}{\mathrm {H} (X,Y)}}.}$ 

This is known as the Rajski Distance.^[12] In a set-theoretic interpretation of information (see the figure for Conditional entropy), this is effectively the Jaccard distance between ${\displaystyle X}$  and ${\displaystyle Y}$ .

Finally,

${\displaystyle D^{\prime }(X,Y)=1-{\frac {\operatorname {I} (X;Y)}{\max \left\{\mathrm {H} ($X),\mathrm {H} (Y)\right\}}}}  

is also a metric.

Sometimes it is useful to express the mutual information of two random variables conditioned on a third.

For jointly discrete random variables this takes the form

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

which can be simplified as

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

For jointly continuous random variables this takes the form

${\displaystyle \operatorname {I} (X;Y|Z)=\int _{\mathcal {Z}}\int _{\mathcal {Y}}\int _{\mathcal {X}}{p_{Z}($z)\,p_{X,Y|Z}(x,y|z)\log \left[{\frac {p_{X,Y|Z}(x,y|z)}{p_{X|Z}\,(x|z)p_{Y|Z}(y|z)}}\right]}dxdydz,}  

which can be simplified as

${\displaystyle \operatorname {I} (X;Y|Z)=\int _{\mathcal {Z}}\int _{\mathcal {Y}}\int _{\mathcal {X}}p_{X,Y,Z}(x,y,z)\log {\frac {p_{X,Y,Z}(x,y,z)p_{Z}($z)}{p_{X,Z}(x,z)p_{Y,Z}(y,z)}}dxdydz.}  

Conditioning on a third random variable may either increase or decrease the mutual information, but it is always true that

${\displaystyle \operatorname {I} (X;Y|Z)\geq 0}$ 

for discrete, jointly distributed random variables ${\displaystyle X,Y,Z}$ . This result has been used as a basic building block for proving other inequalities in information theory.

Several generalizations of mutual information to more than two random variables have been proposed, such as total correlation (or multi-information) and dual total correlation. The expression and study of multivariate higher-degree mutual information was achieved in two seemingly independent works: McGill (1954)^[13] who called these functions "interaction information", and Hu Kuo Ting (1962).^[14] Interaction information is defined for one variable as follows:

${\displaystyle \operatorname {I} (X_{1})=\mathrm {H} (X_{1})}$ 

and for ${\displaystyle n>1,}$ 

${\displaystyle \operatorname {I} (X_{1};\,...\,;X_{n})=\operatorname {I} (X_{1};\,...\,;X_{n-1})-\operatorname {I} (X_{1};\,...\,;X_{n-1}\mid X_{n}).}$ 

Some authors reverse the order of the terms on the right-hand side of the preceding equation, which changes the sign when the number of random variables is odd. (And in this case, the single-variable expression becomes the negative of the entropy.) Note that

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

The multivariate mutual information functions generalize the pairwise independence case that states that ${\displaystyle X_{1},X_{2}}$  if and only if ${\displaystyle I(X_{1};X_{2})=0}$ , to arbitrary numerous variable. n variables are mutually independent if and only if the ${\displaystyle 2^{n}-n-1}$  mutual information functions vanish ${\displaystyle I(X_{1};\ldots ;X_{k})=0}$  with ${\displaystyle n\geq k\geq 2}$  (theorem 2^[15]). In this sense, the ${\displaystyle I(X_{1};\ldots ;X_{k})=0}$  can be used as a refined statistical independence criterion.

For 3 variables, Brenner et al. applied multivariate mutual information to neural coding and called its negativity "synergy" ^[16] and Watkinson et al. applied it to genetic expression.^[17] For arbitrary k variables, Tapia et al. applied multivariate mutual information to gene expression.^[18][15] It can be zero, positive, or negative.^[14] The positivity corresponds to relations generalizing the pairwise correlations, nullity corresponds to a refined notion of independence, and negativity detects high dimensional "emergent" relations and clusterized datapoints ^[18]).

One high-dimensional generalization scheme which maximizes the mutual information between the joint distribution and other target variables is found to be useful in feature selection.^[19]

Mutual information is also used in the area of signal processing as a measure of similarity between two signals. For example, FMI metric^[20] is an image fusion performance measure that makes use of mutual information in order to measure the amount of information that the fused image contains about the source images. The Matlab code for this metric can be found at.^[21] A python package for computing all multivariate mutual informations, conditional mutual information, joint entropies, total correlations, information distance in a dataset of n variables is available.^[22]

Directed information, ${\displaystyle \operatorname {I} \left(X^{n}\to Y^{n}\right)}$ , measures the amount of information that flows from the process ${\displaystyle X^{n}}$ to${\displaystyle Y^{n}}$ , where ${\displaystyle X^{n}}$  denotes the vector ${\displaystyle X_{1},X_{2},...,X_{n}}$  and ${\displaystyle Y^{n}}$  denotes ${\displaystyle Y_{1},Y_{2},...,Y_{n}}$ . The term directed information was coined by James Massey and is defined as

${\displaystyle \operatorname {I} \left(X^{n}\to Y^{n}\right)=\sum _{i=1}^{n}\operatorname {I} \left(X^{i};Y_{i}\mid Y^{i-1}\right)}$ .

Note that if ${\displaystyle n=1}$ , the directed information becomes the mutual information. Directed information has many applications in problems where causality plays an important role, such as capacity of channel with feedback.^[23][24]

Normalized variants of the mutual information are provided by the coefficients of constraint,^[25] uncertainty coefficient^[26] or proficiency:^[27]

${\displaystyle C_{XY}={\frac {\operatorname {I} (X;Y)}{\mathrm {H} ($Y)}}~~~~{\mbox{and}}~~~~C_{YX}={\frac {\operatorname {I} (X;Y)}{\mathrm {H} (X)}}.}  

The two coefficients have a value ranging in [0, 1], but are not necessarily equal. This measure is not symmetric. If one desires a symmetric measure they can consider the following redundancy measure:

${\displaystyle R={\frac {\operatorname {I} (X;Y)}{\mathrm {H} ($X)+\mathrm {H} (Y)}}}  

which attains a minimum of zero when the variables are independent and a maximum value of

${\displaystyle R_{\max }={\frac {\min \left\{\mathrm {H} ($X),\mathrm {H} (Y)\right\}}{\mathrm {H} (X)+\mathrm {H} (Y)}}}  

when one variable becomes completely redundant with the knowledge of the other. See also Redundancy (information theory).

Another symmetrical measure is the symmetric uncertainty (Witten & Frank 2005), given by

${\displaystyle U(X,Y)=2R=2{\frac {\operatorname {I} (X;Y)}{\mathrm {H} ($X)+\mathrm {H} (Y)}}}  

which represents the harmonic mean of the two uncertainty coefficients ${\displaystyle C_{XY},C_{YX}}$ .^[26]

If we consider mutual information as a special case of the total correlationordual total correlation, the normalized version are respectively,

${\displaystyle {\frac {\operatorname {I} (X;Y)}{\min \left[\mathrm {H} ($X),\mathrm {H} (Y)\right]}}}   and ${\displaystyle {\frac {\operatorname {I} (X;Y)}{\mathrm {H} (X,Y)}}\;.}$ 

This normalized version also known as Information Quality Ratio (IQR) which quantifies the amount of information of a variable based on another variable against total uncertainty:^[28]

${\displaystyle IQR(X,Y)=\operatorname {E} [\operatorname {I} (X;Y)]={\frac {\operatorname {I} (X;Y)}{\mathrm {H} (X,Y)}}={\frac {\sum _{x\in X}\sum _{y\in Y}p(x,y)\log {p($x)p(y)}}{\sum _{x\in X}\sum _{y\in Y}p(x,y)\log {p(x,y)}}}-1}  

There's a normalization^[29] which derives from first thinking of mutual information as an analogue to covariance (thus Shannon entropy is analogous to variance). Then the normalized mutual information is calculated akin to the Pearson correlation coefficient,

${\displaystyle {\frac {\operatorname {I} (X;Y)}{\sqrt {\mathrm {H} ($X)\mathrm {H} (Y)}}}\;.}  

In the traditional formulation of the mutual information,

${\displaystyle \operatorname {I} (X;Y)=\sum _{y\in Y}\sum _{x\in X}p(x,y)\log {\frac {p(x,y)}{p($x)\,p(y)}},}  

each eventorobject specified by ${\displaystyle (x,y)}$  is weighted by the corresponding probability ${\displaystyle p(x,y)}$ . This assumes that all objects or events are equivalent apart from their probability of occurrence. However, in some applications it may be the case that certain objects or events are more significant than others, or that certain patterns of association are more semantically important than others.

For example, the deterministic mapping ${\displaystyle \{(1,1),(2,2),(3,3)\}}$  may be viewed as stronger than the deterministic mapping ${\displaystyle \{(1,3),(2,1),(3,2)\}}$ , although these relationships would yield the same mutual information. This is because the mutual information is not sensitive at all to any inherent ordering in the variable values (Cronbach 1954, Coombs, Dawes & Tversky 1970, Lockhead 1970), and is therefore not sensitive at all to the form of the relational mapping between the associated variables. If it is desired that the former relation—showing agreement on all variable values—be judged stronger than the later relation, then it is possible to use the following weighted mutual information (Guiasu 1977).

${\displaystyle \operatorname {I} (X;Y)=\sum _{y\in Y}\sum _{x\in X}w(x,y)p(x,y)\log {\frac {p(x,y)}{p($x)\,p(y)}},}  

which places a weight ${\displaystyle w(x,y)}$  on the probability of each variable value co-occurrence, ${\displaystyle p(x,y)}$ . This allows that certain probabilities may carry more or less significance than others, thereby allowing the quantification of relevant holisticorPrägnanz factors. In the above example, using larger relative weights for ${\displaystyle w(1,1)}$ , ${\displaystyle w(2,2)}$ , and ${\displaystyle w(3,3)}$  would have the effect of assessing greater informativeness for the relation ${\displaystyle \{(1,1),(2,2),(3,3)\}}$  than for the relation ${\displaystyle \{(1,3),(2,1),(3,2)\}}$ , which may be desirable in some cases of pattern recognition, and the like. This weighted mutual information is a form of weighted KL-Divergence, which is known to take negative values for some inputs,^[30] and there are examples where the weighted mutual information also takes negative values.^[31]

A probability distribution can be viewed as a partition of a set. One may then ask: if a set were partitioned randomly, what would the distribution of probabilities be? What would the expectation value of the mutual information be? The adjusted mutual information or AMI subtracts the expectation value of the MI, so that the AMI is zero when two different distributions are random, and one when two distributions are identical. The AMI is defined in analogy to the adjusted Rand index of two different partitions of a set.

Using the ideas of Kolmogorov complexity, one can consider the mutual information of two sequences independent of any probability distribution:

${\displaystyle \operatorname {I} _{K}(X;Y)=K($X)-K(X\mid Y).}  

To establish that this quantity is symmetric up to a logarithmic factor (${\displaystyle \operatorname {I} _{K}(X;Y)\approx \operatorname {I} _{K}(Y;X)}$ ) one requires the chain rule for Kolmogorov complexity (Li & Vitányi 1997). Approximations of this quantity via compression can be used to define a distance measure to perform a hierarchical clustering of sequences without having any domain knowledge of the sequences (Cilibrasi & Vitányi 2005).

Unlike correlation coefficients, such as the product moment correlation coefficient, mutual information contains information about all dependence—linear and nonlinear—and not just linear dependence as the correlation coefficient measures. However, in the narrow case that the joint distribution for ${\displaystyle X}$  and ${\displaystyle Y}$  is a bivariate normal distribution (implying in particular that both marginal distributions are normally distributed), there is an exact relationship between ${\displaystyle \operatorname {I} }$  and the correlation coefficient ${\displaystyle \rho }$  (Gel'fand & Yaglom 1957).

${\displaystyle \operatorname {I} =-{\frac {1}{2}}\log \left(1-\rho ^{2}\right)}$ 

The equation above can be derived as follows for a bivariate Gaussian:

${\displaystyle {\begin{aligned}{\begin{pmatrix}X_{1}\\X_{2}\end{pmatrix}}&\sim {\mathcal {N}}\left({\begin{pmatrix}\mu _{1}\\\mu _{2}\end{pmatrix}},\Sigma \right),\qquad \Sigma ={\begin{pmatrix}\sigma _{1}^{2}&\rho \sigma _{1}\sigma _{2}\\\rho \sigma _{1}\sigma _{2}&\sigma _{2}^{2}\end{pmatrix}}\\\mathrm {H} (X_{i})&={\frac {1}{2}}\log \left(2\pi e\sigma _{i}^{2}\right)={\frac {1}{2}}+{\frac {1}{2}}\log(2\pi )+\log \left(\sigma _{i}\right),\quad i\in \{1,2\}\\\mathrm {H} (X_{1},X_{2})&={\frac {1}{2}}\log \left[(2\pi e)^{2}|\Sigma |\right]=1+\log(2\pi )+\log \left(\sigma _{1}\sigma _{2}\right)+{\frac {1}{2}}\log \left(1-\rho ^{2}\right)\\\end{aligned}}}$ 

Therefore,

${\displaystyle \operatorname {I} \left(X_{1};X_{2}\right)=\mathrm {H} \left(X_{1}\right)+\mathrm {H} \left(X_{2}\right)-\mathrm {H} \left(X_{1},X_{2}\right)=-{\frac {1}{2}}\log \left(1-\rho ^{2}\right)}$ 

When ${\displaystyle X}$  and ${\displaystyle Y}$  are limited to be in a discrete number of states, observation data is summarized in a contingency table, with row variable ${\displaystyle X}$  (or${\displaystyle i}$ ) and column variable ${\displaystyle Y}$  (or${\displaystyle j}$ ). Mutual information is one of the measures of associationorcorrelation between the row and column variables.

Other measures of association include Pearson's chi-squared test statistics, G-test statistics, etc. In fact, with the same log base, mutual information will be equal to the G-test log-likelihood statistic divided by ${\displaystyle 2N}$ , where ${\displaystyle N}$  is the sample size.

In many applications, one wants to maximize mutual information (thus increasing dependencies), which is often equivalent to minimizing conditional entropy. Examples include: