I'm trying to prove, or disprove, that the following matrix is positive (semi-)definite and symmetric (so a covariance matrix); $$ M = C_M^{-1} + G^T C_D^{-1} G. $$ Here, $C_M$ is a covariance matrix (which is positive semi-definite and symmetric, sometimes positive definite) of dimensions $n \times n$, which is not necessarily diagonal. The matrix $C_D$, of dimensions $m \times m$, is also a covariance matrix which in my application is always diagonal, but could be extended to any arbitrary covariance matrix (so always positive semi-definite, symmetric).

G is an arbitrary matrix of dimensions $m \times n$. It does not necessarily have full rank. I think this already excludes positive definiteness of the second term, since $x^T G^T G x = (G x)^T(G x) > 0$ only if G is full rank.

Can we extend this to $G^T C_D^{-1} G$ to prove that M is positive semi-definite symmetric? I'd suggest to use the property $x^T A x \geq 0$;

$$ \begin{eqnarray} x^T G^T C_D^{-1} G x &=& (G x)^T C_D^{-1} (G x) \nonumber \\ & \overset{?}{\geq} & 0 \end{eqnarray} $$

I guess this boils down to proving that the inverse of $C_D$ (and $C_M$) is positive semi-definite, because afterwards we can again use the same definition on $(G x)^T C_D^{-1} (G x) $. And the sum of PSD matrices is also PSD (that proof is rather easy, see here).

And then a small follow up; if the second term is PSD, but the first term is PD, the resulting matrix M is also PD right?

For a little context, from the first equation matrix $M$ is used as the mass matrix in a Hamiltonian Monte Carlo sampler, and I need to Cholesky decompose it to propose momentum samples. Cholesky decomposition can (I think) only be done on a positive definite matrix. If a matrix is positive semi-definite, it is still possible to Cholesky decompose it if the diagonal entries of $L$ are allowed to be zero, but I don't know if my algorithm breaks down in this case.


If a matrix $A$ is symmetric positive definite, then $A^{-1}$ is symmetric positive definite.

Proof: Let the SVD of $A=UDU^T$, then $A^{-1}=UD^{-1}U^T$, we can see from the eigenvalues that $A^{-1}$ is positive definite.

If a matrix $A$ is positive definite and matrix $B$ is symmetric positive semidefinite, then $A+B$ is positive definite.


$$x^T(A+B)x=x^TAx+x^TBx > 0$$

  • $\begingroup$ Thanks, easy as 1, 2, 3, both proofs! The term $ (G x)^T C_D^{-1} (G x) $ then also proves PD/PSD right? $\endgroup$ – Lars Gebraad Aug 9 '17 at 8:48
  • 1
    $\begingroup$ If $C_D$ is positive definite then $(Gx)^TC_D^{-1}(Gx) \geq 0$, hence it is PSD. I don't think it is PD as $Gx$ can be zero even though $x$ is non-zero. $\endgroup$ – Siong Thye Goh Aug 9 '17 at 9:00
  • $\begingroup$ So the requirement for $M$ to be PD if $C_M$ is PSD is that $G$ has full rank, and $C_D$ is PD. $\endgroup$ – Lars Gebraad Aug 9 '17 at 9:19

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.