Jensen inequality for matrix exponential of positive definite matrix (laplacian)

384 Views Asked by At

I have a positive definite matrix $n\times n$ matrix $\mathbf{L}$. This matrix is the combinatorial Laplacian of a graph.

We can compute the expected value of the Laplacian as $\mathbb{E}[\mathbf{L}]$ as in the graph model considered, links in the graph appear independently with probability $p_{ij}$. The expected Laplacian has $\sum_k p_{ik}$ on the diagonal and $-p_{ij}$ off diagonal elements.

Now I need to compute the following quantity:

$$ \mathbb{E}\left[ \frac{e^{-\beta \mathbf{L}}}{\textrm{Tr}[e^{-\beta \mathbf{L}}]}\right] $$ where the exponential is intended as the matrix exponential and $\mathrm{Tr}$ is the trace.

I understand that the Jensen inequality can be correctly applied to the denominator $\textrm{Tr}[e^{-\beta \mathbf{L}}]$ as this is a convex scalar function and one gets $\mathbb{E}\left[\textrm{Tr}[e^{-\beta \mathbf{L}}]\right] \geq \mathrm{Tr}\left[ e^{-\beta \mathbb{E}[\mathbf{L}]}\right]$.

However I cannot find references on how to apply Jensen inequality to matrix valued functions. Is it correct to conclude that:

$$ \mathbb{E}\left[ \frac{e^{-\beta \mathbf{L}}}{\textrm{Tr}[e^{-\beta \mathbf{L}}]}\right] \geq \frac{e^{-\beta E[\mathbf{L}]}}{\textrm{Tr}[e^{-\beta E[\mathbf{L}]}]} $$

Is it possible to show this using the Taylor expansion of the exponential or there

1

There are 1 best solutions below

0
On

It is not possible to apply Jensen inequality because the matrix exponential is not an operator convex function. To see that, look at the Bhatia book "Matrix analysis".