Index: The Book of Statistical ProofsStatistical ModelsMultivariate normal dataGeneral linear model ▷ Log-likelihood ratio and estimated mutual information

Theorem: Consider a general linear model $m_1$ with $n \times v$ data matrix $Y$, $n \times p$ design matrix $X$ and uncorrelated observations, i.e. $V = I_n$,

\[\label{eq:m1} m_1: \; Y = X B + E_1, \; E_1 \sim \mathcal{MN}(0, I_n, \Sigma_1) \; ,\]

as well as another model $m_0$ in which $X$ has no influence on $Y$:

\[\label{eq:m0} m_0: \; Y = E_0, \; E_0 \sim \mathcal{MN}(0, I_n, \Sigma_0) \; .\]

Then, the log-likelihood ratio of $m_1$ vs. $m_0$ is equal to the estimated mutual information of $X$ and $Y$:

\[\label{eq:glm-llrmi} \ln \Lambda_{10} = \hat{I}(X,Y) \; .\]

Proof: The maximum likelihood estimates for a general linear model are

\[\label{eq:glm-mle} \begin{split} \hat{B} &= (X^\mathrm{T} V^{-1} X)^{-1} X^\mathrm{T} V^{-1} Y \\ \hat{\Sigma} &= \frac{1}{n} (Y - X\hat{B})^\mathrm{T} V^{-1} (Y - X\hat{B}) \; , \end{split}\]

such that, for the two models, the maximum likelihood estimates are:

\[\label{eq:m1-m0-mle} \begin{split} \hat{\Sigma}_1 &= \frac{1}{n} (Y - X\hat{B})^\mathrm{T} (Y - X\hat{B}) \quad \text{with} \quad \hat{B} = (X^\mathrm{T} X)^{-1} X^\mathrm{T} Y \quad \text{and} \quad \\ \hat{\Sigma}_0 &= \frac{1}{n} Y^\mathrm{T} Y \; . \end{split}\]

The log-likelihood ratio for two general linear models is

\[\label{eq:glm-llr} \ln \Lambda_{12} = - \frac{n}{2} \ln \frac{|\hat{\Sigma}_1|}{|\hat{\Sigma}_2|} \; ,\]

such that in the present case, we have:

\[\label{eq:m1-m0-llr} \ln \Lambda_{10} = - \frac{n}{2} \ln \frac{|\hat{\Sigma}_1|}{|\hat{\Sigma}_0|} \; .\]

The mutual information for the general linear model is

\[\label{eq:glm-mi} I(X,Y) = - \frac{n}{2} \ln \frac{|\Sigma_1|}{|\Sigma_0|} \; ,\]

such that with \eqref{eq:m1-m0-mle}, the estimated mutual information is:

\[\label{eq:Y-X-mi} \hat{I}(X,Y) = - \frac{n}{2} \ln \frac{|\hat{\Sigma}_1|}{|\hat{\Sigma}_0|} \; ,\]

Together, \eqref{eq:m1-m0-llr} and \eqref{eq:Y-X-mi} show that

\[\label{eq:glm-llrmi-qed} \ln \Lambda_{10} = \hat{I}(X,Y) \; .\]
Sources:

Metadata: ID: P458 | shortcut: glm-llrmi | author: JoramSoch | date: 2024-06-21, 10:27.