First, let $f: \R ^d \to \R $ be a normal density.

First, express the log likelihood of $f$ on a record $x^k$ by

\[ \log f(x^k) = - \frac{1}{2} (x^k - \mu )^{\tp} \inv{\Sigma } (x^k - \mu ) - \frac{1}{2} \log \det \Sigma - \frac{d}{2}\log 2\pi . \]

Second, use the trace to rewrite the quadratic form

\[ -\frac{1}{2}\tr{(\Sigma ^{-1} (x^k - \mu ) (x^k - \mu )^{\tp})}. \]

Third, use these two, and the linearity of trace to express the average negative log likelihood by

\[ \begin{aligned} - \frac{1}{n} \sum_{k = 1}^{n} \log f(x^k) &= \frac{1}{2} \tr{(\Sigma ^{-1}(\frac{1}{n}\sum_{i = 1}^{n} (x^k - \mu )(x^k - \mu )^{\tp}))} + \frac{1}{2}\log\det\Sigma + \frac{d}{2}\log 2\pi . \end{aligned} \]

Fourth, use matrix calculus (or the derivation in Proposition 1 of Multivariate Normal Maximum Likelihood) to see that, for a minimizer of the negative average log likelihood, the mean must be $\frac{1}{n} \sum_{i = 1}^{n} x^k$.

Fifth, recognize the empirical covariance matrix $\frac{1}{n}\sum_{k = 1}^{n} (x^k - \mu )(x^k - \mu )^{\tp}$; denote it by $S$.

Sixth, change variables with $P = \Sigma ^{-1}$ and express

\[ \log{(\det{(\Sigma )})} = \log{(\det{(\inv{P})})} = \log{(\inv{(\det P)})} = -\log{(\det{(P)})} \]

Seventh, write the likelihood in simplified form (using circulant property of trace)

\[ \frac{1}{2} \tr{(SP)} - \frac{1}{2} \log \det P - \frac{d}{2}\log 2\pi \]

Eighth, drop the constant and prefactors:

\[ \tr{(SP)} - \log \det P \]

Ninth, if $g$ is a normal with then the tree density approximation objective is the same equivalent to

\[ d(g, f) = h(g, f) - h(f) \sim h(g, f) = - \int_{\R ^d} g\log f. \]

TODO: Extra, the let $g$ be normal and $f$ be normal. The tree normal approximation problem

\[ d(g, f) \sim h(g, f) = - \int_{R^d} g \log f. \]

Since the set of optimal solutions for both optimization is contained in the set of normal densities which match on the mean we can assume that $\mu _f = \mu _g$. So the approximation objective is equivalent to

\[ \tr{P_f \Sigma _g} - \log \det P_f, \]

Thus, a solution is a maximum likelihood tree normal of a dataset if and only if it is an optimal tree approximator of the empirical normal density of the dataset.