Suppose we have a non symmetric Matrix $A$. Let also we know that $A+A^T$ is positive definite. Is it the case that A is also positive definite? If A symmetric then its easy to show that $A$ is positive definite because $A=A^T$. So if it is not the case then can we find a 2x2 matrix, $B$ which is non symmetric and $B+B^T$ is positive definite but B is not positive definite?

I have another question. Suppose $A$ is a non symmetric non positive definite matrix. is it possible its inverse is positive definite?


  • $\begingroup$ positive definite $\implies P D P^T$ where $P$ is orthogonal and $D$ is diagonal $> 0$ $\endgroup$ – reuns Jan 30 '16 at 21:59
  • 2
    $\begingroup$ What is your definition of "positive definite" for a matrix that is not symmetric???????????? $\endgroup$ – Will Jagy Jan 30 '16 at 22:10
  • $\begingroup$ Suppose the matrix [1 1; 0 1]. This matrix is not symmetric but positive definite as determinant is positive so are the eigenvalues. My question is suppose we know $A+A^T$ is positive definite. Is it mandatory for A to be positive definite? $\endgroup$ – user2104150 Jan 31 '16 at 0:27
  • $\begingroup$ @user2104150 note that your definition is definitely not the usual definition. For example, few people would consider the matrix [1 4; 0 1] to be positive definite. $\endgroup$ – Omnomnomnom Jan 31 '16 at 3:33

The answer is no, at least with your definition of positive-definite; we can show this directly in the $2 \times 2$ case:

Let the symmetric part of the matrix be $$ S = \begin{pmatrix} 2a & b \\ b & 2d \end{pmatrix}, $$ and the antisymmetric part $A$ have top-right element $-c$. It is easy to compute the eigenvalues of $S$ and $S+A$. Since $S$ is symmetric, we expect it to have real eigenvalues, and indeed it does, namely $$ \lambda_{\pm}=a+d \pm \sqrt{(a-d)^2+b^2}. $$ The square root contains a sum of real squares, so is nonnegative. We also require that $$ a+d>\sqrt{(a-d)^2+b^2} $$ in order to have $\lambda_-$ positive.

On the other hand, the eigenvalues of $S+A$ are $$ \mu_{\pm}=a+d \pm \sqrt{(a-d)^2+b^2-c^2}, $$ so we require $$ (a-d)^2+b^2-c^2>0 $$ for real eigenvalues, which is certainly not necessary, given that we can choose $c$ as large as we like. Hence we can have complex eigenvalues for $S+A$, even if $S$ has positive ones. On the other hand, supposing that $\mu_{\pm} \in \mathbb{R}$, we do automatically get the other condition: $$ a+d>\sqrt{(a-d)^2+b^2} >\sqrt{(a-d)^2+b^2-c^2}, $$ since $c^2 > 0$ and the square root is increasing.

| cite | improve this answer | |
  • $\begingroup$ So the positive definite matrix must have real eigen values? In that definition A = [1 1;-1 1] is positive definite in the sense $x^TAx > 0$ but it has complex eigenvalues? what can we say about A? is it positive definite? $\endgroup$ – user2104150 Jan 31 '16 at 14:59
  • $\begingroup$ If you use that definition of positive-definite, the antisymmetric part of $A$ makes absolutely no contribution at all: $$ (A_{ij}-A_{ji})x_i x_j = 0. $$ $\endgroup$ – Chappers Jan 31 '16 at 15:20

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.