For a symmetric $M$-matrix $M$ and a positive diagonal matrix $D$, I'm trying to evaluate whether $$ x^TDMx = \frac{1}{2}x^T(DM+MD)x > 0 $$ for all $x \in \mathbb{R}^n$. Since $M$-matrices are $D$-stable, I believe it is true that $DM$ and $MD$ are both $M$-matrices. So, is it true that $DM + MD$ is positive definite? If not, what would a counter-example be, and what would be a sufficient condition such that this inequality holds?

What if I strengthened $M$ to be a diagonally dominant matrix, with strict dominance in at least one row?

  • $\begingroup$ If $M$ is positive definite and $D$ is positive diagonal, then both $MD$ and $DM$ are positive definite. Sum of $N$ positive definite matrices is also positive definite, so you are right. $\endgroup$ – Seyhmus Güngören Jan 15 '15 at 0:20
  • $\begingroup$ $MD$ and $DM$ are not symmetric matrices, so I cannot see how they could be positive definite in the typical sense, unless you are using some other result of which I am not aware. $\endgroup$ – John Jan 15 '15 at 0:25
  • 1
    $\begingroup$ Check this one. math.stackexchange.com/questions/4336/… what I am saying is true but not in general. $\endgroup$ – Seyhmus Güngören Jan 15 '15 at 0:47
  • $\begingroup$ So if I'm understanding things properly, if $DM$ is diagonalizable, my desired property will hold? $\endgroup$ – John Jan 15 '15 at 0:57
  • $\begingroup$ @john No, it won't. Product of two SPD matrices is always diagonalizable. $\endgroup$ – Algebraic Pavel Jan 15 '15 at 1:01

If $M$ is an M-matrix, then there is a positive definite diagonal matrix $D$ such that $DM+MD$ is positive definite (when $M$ is symmetric, that $D$ could trivially be the identity matrix). I'm afraid that this is not true for any p.d. $D$. Take, e.g., $$ M=\pmatrix{2 & -1 \\ -1 & 2}, \quad D=\pmatrix{20 & 0 \\ 0 & 1}. $$ This covers also the case of a diagonally dominant $M$.

I believe that there's no sufficient condition for this to hold true unless $M$ is diagonal. Assume for now that $n=2$. Let $$ M=\pmatrix{\alpha&-\gamma\\-\gamma&\beta}, \quad \alpha,\beta,\gamma>0. $$ Take $$ D=\pmatrix{\delta&0\\0&1}, \quad \delta>0, $$ and assume $x:=[\xi,1]^T$. We have $$ x^TDMx=\alpha\delta\xi^2 -\gamma(1+\delta)\xi + \beta. $$ This quadratic function of $\xi$ attains negative values if the discriminant $$\tag{1} d(\delta):=\gamma^2(1+\delta)^2-4\alpha\beta\delta $$ is positive for some $\delta$. Indeed, it is positive for large enough $\delta$. Hence if $\gamma>0$, there is always a diagonal matrix $D$ and a vector $x$ such that $x^TDMx<0$.

If $n>2$, you can reduce the problem to the $2\times 2$ case by looking for an $x$ having nonzeros in the positions $i$ and $j$ for which $m_{ij}<0$ ($i\neq j$) which will effectively reduce the proof the what we just did here. This gives:

For any $n\times n$ ($n\geq 2$) symmetric M-matrix $M=(m_{ij})$ such that $m_{ij}<0$ for some $i\neq j$, there is a diagonal matrix $D$ such that $DM+MD$ is not positive definite.


Let $M$ be a symmetric M-matrix such that $DM+MD$ is positive definite for any positive definite diagonal matrix $D$. Then $M$ is diagonal.

  • $\begingroup$ Thanks for the interesting counter-example. I guess the intuition is that if $$ \frac{d_{\rm max}}{d_{\rm min}} >\!\!> 1 $$ the result can fail. I wonder if one can state a sufficient condition like $$ \frac{d_{\rm max}}{d_{\rm min}} \leq \frac{\lambda_{\rm max}(M)}{\lambda_{\rm min}(M)} $$ $\endgroup$ – John Jan 17 '15 at 18:58

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.