0
$\begingroup$

How would you find the nearest (via Hilbert distance), PSD matrix (with trace = 1) to a Hermitian matrix? I found an answer to a similar question here. However, as I understand, Hingham's work only applies to real matrices.

(I am looking at reconstructing a density matrix from experimental data which needs to be a PSD with trace = 1 )

$\endgroup$
0
$\begingroup$

Isn't (as was originally asked) the closest psd matrix to $H=U^* \Delta U$ where $U$ is unitary and $\Delta$ real diagonal with diagonal entries $d_i$ given by $G=U^* \Delta_+ U$ where $\Delta_+$ is the diagonal matrix with diagonal entries $d_i'= \max(0,d_i)$? Because $X\mapsto U^* X U$ is an isometry with respect to the Hilbert distance?

Added after the question was changed. If the closest psd matrix with trace 1 to $H$ is sought, chose $\Delta_+$ to be closest to $\Delta$ subject to the trace 1 and non-negativity condition. The $U$ stuff above reduces your problem to a $\mathbb R^n$ problem from an $n\times n$ matrix problem. I think this is attained when $d_i'$ of the form $d_i'=0$ if $d_i<0$ and $d_i' = d_i+\lambda$ for $d_i\ge0$, where $\lambda$ is chosen to make $\sum d_i' = 1.$

$\endgroup$
  • $\begingroup$ I have just edited the question. What I have been doing is similar to your suggestion and then dividing it by Trace(G) to normalize the trace. However, the data suggests that it is not the best reconstruction. $\endgroup$ – Pranav M Mar 15 at 23:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.