What's the third derivative of $\log( \det X)$ with respect to $X$? If $X$ is positive definite matrix, what is the third derivative of $\log(\det X)$ with respect to $X$?
We know the first derivative of $\log(\det X)$ is $X^{-1}$
and the second derivative (hessian) is $X^{-1} \otimes X^{-1}$ but I don't know the third derivative of  $\log(\det X)$ with respect to $X$.
Thanks in advance.
 A: Let $H$ be a symmetric matrix with "small" entries. Then by writing $\Delta = X^{-1}H$ for simplicity, we may decompose $\log\det(X+H)$ as
$$ \log\det(X+H) = \log\det(X) + \log\det(I+\Delta). $$
Now by using the identities $A=\exp(\log A)$ and $\det\exp(A) = \exp\operatorname{tr}(A)$ (which surely hold for positive definite matrices), we get
\begin{align*}
\log\det(I+\Delta)
&= \log\det \exp(\log(I+\Delta)) \\
&= \log\det \exp\left(\sum_{k=1}^{\infty}\frac{(-1)^{k-1}}{k} \Delta^k \right) \\
&= \operatorname{tr}\left(\sum_{k=1}^{\infty}\frac{(-1)^{k-1}}{k} \Delta^k \right) \\
&= \sum_{k=1}^{\infty}\frac{(-1)^{k-1}}{k} \operatorname{tr}(\Delta^k).
\end{align*}

*

*For example, the first derivative of $\log\det(X)$ is the linear functional
$$ H \mapsto \operatorname{tr}(X^{-1}H), $$
which can be identified as $X^{-1}$ via the Riesz representation theorem, assuming that the space of symmetric matrices is furnished with the inner product $\langle A, B \rangle = \operatorname{tr}(A^{\mathsf{T}}B)$.


*Then the second derivative of $\log\det(X)$ is the quadratic form
$$ H \mapsto -\operatorname{tr}(X^{-1}HX^{-1}H), $$
which then induces the bilinear form
$$ (U, V) \mapsto -\operatorname{tr}(X^{-1}UX^{-1}V) $$
via polarization. (I am not sure how this can be identified with $-X^{-1}\otimes X^{-1}$, though.)


*Finally, the third derivative of $\log\det(X)$ is the cubic form
$$ H \mapsto 2\operatorname{tr}(X^{-1}HX^{-1}HX^{-1}H). $$
A: Let
$$\eqalign{
X &= X^T \\
Y &= Y^T = X^{-1} \\
f(X) &= \log\det(X) = -\log\det(Y) \\
}$$
then in index notation the first few derivatives of $f$ are
$$\eqalign{
\frac{\partial f}{\partial X_{ij}}
 &= Y_{ij} \\
\frac{\partial^2 f}{\partial X_{ij}\,\partial X_{k\ell}}
 &= ​-Y_{i\ell}Y_{jk} \\
\frac{\partial^3 f}{\partial X_{ij}\,\partial X_{k\ell}\,\partial X_{pq}}
 &​= Y_{ip}Y_{q\ell}Y_{jk} + Y_{i\ell}Y_{jp}Y_{qk} \\\\
}$$
These formula result from the repeated application of two simple rules
$$\eqalign{
dY_{i\ell}
&= -\sum_{j}\sum_{k}Y_{ij}\;dX_{jk}\,Y_{k\ell}
\qquad\qquad
\frac{\partial X_{ij}}{\partial X_{k\ell}} &= \delta_{ik}\delta_{j\ell} \\
}$$
