4

Suppose there is a (non-symmetric) real square matrix $A$ with symmetric part $A+A^T$.

What are some conditions on $A$ that are sufficient for $A+A^T$ to be positive definite?

For example, if the eigenvalues of $A$ are strictly positive is $A+A^T$ positive definite? (EDIT: This part of the question is answered in the negative in the comments).

This would then give the result I actually want which is that given two positive definite matrices $C$ and $D$ it follows that the symmetric part of $CD$ is also positive definite. (EDIT: But I think it is still not clear if $CD+DC>0$ - this is (perhaps) a slightly more special case than $A+A^T$ with $A$ having positive eigenvalues.)

John U
  • 390
  • wat can you say about eigen values of $A^T$ –  Jan 11 '14 at 12:42
  • $A$ and $A^T$ have the same eigenvalues but I do not know what that says about the eigenvalues of $A+A^T$.. maybe this is obvious but I am unfamiliar – John U Jan 11 '14 at 12:58
  • 1
    oh yes.. that does not help! I am helpless!! –  Jan 11 '14 at 13:09
  • 5
    Consider $A = \begin{pmatrix}\frac14 & 1\0 & \frac14\end{pmatrix}$, the eigenvalues of $A$ are all positive but $A + A^T$ are not positive definite. – achille hui Jan 11 '14 at 13:43
  • Yes $x^T(A+A^T)x = 2x^TAx$ and $x^TAx$ may be less than zero if $A$ is non-symmetric but has positive eigenvalues.. – John U Jan 11 '14 at 13:54
  • What about other conditions on A? Does it somehow change anything if $A=CD$ for two symmetric positive definite real matrices $C$ and $D$? – John U Jan 11 '14 at 13:55
  • 2
    A matrix is sum of symmetric and anti symmetric parts. $x' A x$ is zero if $A$ is anti symmetric. So a matrix is pd, iff its symmetric part is pd. –  Jan 11 '14 at 14:12
  • 2
    Also note that eigenvalues of a non-symmetric pd matrix are not necessarily positive. –  Jan 11 '14 at 14:17

1 Answers1

3

To address the problem that the OP actually wanted solved:

Given two symmetric positive definite matrices $C$ and $D$, show that $CD+DC$ is also positive definite.

This is not true, as can be seen from the following example:

$$C = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix}, \quad D = \begin{bmatrix} 100 \\ & 1 \end{bmatrix}.$$

Then their respective spectra are:

$$\sigma(C) = \{1,3\}, \quad \sigma(D) = \{1, 100\}, \quad \sigma(CD+DC) \approx \{-20.2724, 424.272 \}.$$

The reason is simple: when computing $\det(CD+DC)$, the big factor $100$ participates in the positive part only $2$ times, but squared in the negative (counter-diagonal) part.

This might give you some insight how things behave in a more general setting. For example, I'm fairly certain that it can be shown that the following is true: if $C$ and $D$ do not commute, there exists $n \in \mathbb{N}$ such that $CD^n+D^nC$ is not positive definite. All you need to do is observe a $2 \times 2$ submatrix for which $D$ is not a multiply of $I_2$ (WLOG, you can assume that $D$ is diagonal). In other words, any such $D$ can be blown enough to make the above invalid.

This would give you that for every symmetric positive definite $C$ (such that it is not a multiple of identity), you can find a symmetric positive definite $D$ (diagonal, if you want) such that the above does not hold.

Vedran Šego
  • 11,372