8
$\begingroup$

This question comes from https://stats.stackexchange.com/questions/457375/recover-full-covariance-matrix-from-covariance-diagonal-and-precision-off-diagon where it have not found answers. So, let $\Sigma$ be an $N\times N$ covariance (that is, positive semidefinite) matrix, but we do only know its diagonal and the off-diagonal elements of its inverse $\Sigma^{-1}$ (known as precision matrix).

How can we find $\Sigma$ effectively?

$\endgroup$
4
  • $\begingroup$ If it is positive semi definite (and not positive definite) why does it have an inverse? $\endgroup$
    – username
    Oct 21, 2021 at 18:59
  • $\begingroup$ @username: That is just the definition of a covariance matrix, but in this case, we certainly must assume it has an inverse. $\endgroup$ Oct 21, 2021 at 19:23
  • 2
    $\begingroup$ the computation is a matter of solving $N$ equations with $N$ unknowns; there are multiple solutions, some experimentation suggests that only a single solution produces a positive definite $\Sigma$; is there a proof for that? $\endgroup$ Oct 21, 2021 at 21:26
  • 1
    $\begingroup$ I raised the issue of the uniqueness of the reconstruction in a separate thread $\endgroup$ Oct 22, 2021 at 17:12

2 Answers 2

2
$\begingroup$

Following the ideas from my answer to Carlo's question, let $M(z_1, \ldots, z_n)$ be the symmetric matrix with $M(z_1, \ldots, z_n)_{ij} = \Sigma_{ij}$ for $i \neq j$ and $M(z_1, \ldots, z_n)_{ii} = z_i$.

Use any sort of numerical optimization algorithm to maximize $$\log \det M(z_1, \ldots, z_n) - \sum_{i=1}^n \Sigma^{-1}_{ii} z_i$$ on the open set of $(z_1, \ldots, z_n)$ where $M(z_1, \ldots, z_n)$ is positive definite. The function in question is concave, so the optimum is unique and shouldn't be hard to find. This $M(z_1,\ldots, z_n)$ will have the desired property.

The equations one gets seems to have fairly high degree, so I would not attempt a symbolic solution.

$\endgroup$
1
$\begingroup$

In the linked question the OP also asks for the special case that the off-diagonal elements of $\Sigma^{-1}$ have a rank-one structure, $(\Sigma^{-1})_{ij}=u_i v_j$ for $i\neq j$. In that case the covariance matrix can be reconstructed as follows.

Denote $\Sigma^{-1}=A+uv^\top$, with $u,v$ known rank-$N$ vectors and $A={\rm diag}\,(d_1,d_2,\ldots d_N)$ an unknown diagonal matrix. According to the Sherman–Morrison formula, the covariance matrix $\Sigma$ then has diagonal elements $$\Sigma_{ii}= x_i - {x_i^2 u_iv_i \over 1 +\sum_{n=1}^N u_nv_n x_n },\;\;i=1,2,\ldots N,$$ with $\;\;x_n=1/d_n$. These are $N$ equations where the $N$ unknowns $x_n$ appear quadratically.

This method of solution is efficient, but I do not have a proof that the solution with a positive definite $\Sigma$ is unique.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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