1

I want to find an upper bound for $trace(XYXZ)$ in terms of $trace(X^2YZ)$. $X, Y, Z$ are Hermitian positive definite matrices.

Any idea/help is appreciated.

Update: Assume that $Y$ and $Z$ commute: $YZ=ZY$.

Mah
  • 1,251
  • There cannot be such a bound because $X^2$ may be equal to zero while your first trace may not be – lcv Jul 17 '18 at 04:21
  • @lcv but X is positive definite, so X^2 can’t be zero. – Mah Jul 17 '18 at 04:26
  • Oh I missed that. Still why should it be true? – lcv Jul 17 '18 at 04:31
  • 1
    This result comes to mind, but I'm not sure if it's directly useful. – Ben Grossmann Jul 17 '18 at 04:41
  • 4 people gave meaningful answers and comments on the question, so why it is marked as off-topic?! – Mah Jul 17 '18 at 15:33
  • @Omnomnomnom I've tried that, but I wasn't successful. – Mah Jul 17 '18 at 15:36
  • @Mah I didn't vote to close the question, but you should read the feedback; "this question is missing context or other details" etc. What are your thoughts on the problem? What have you tried? How did you come across this problem in the first place? I suspect that if you added details to this effect, your question would be reopened, in case you care about that at this point. – Ben Grossmann Jul 17 '18 at 16:18

2 Answers2

4

It is possible for $\text{trace}(X^2 Y Z) = 0$ with $X,Y,Z$ positive definite Hermitian. Try $$ X^2 = \pmatrix{2 & -1\cr -1 & 1\cr},\ Y = \pmatrix{1/10 & 0\cr 0 & 1\cr},\ Z = \pmatrix{2 & 1\cr 1 & 7/10}$$ However, this will not work with the added condition that $Y$ and $Z$ commute. If $Y$ and $Z$ commute, so do $Y$ and $Z^{1/2}$, and $YZ = Z^{1/2} Y Z^{1/2}$ is positive definite. Then $$\text{trace}(X^2 Y Z) = \text{trace}((YZ)^{1/2} X^2 (YZ)^{1/2}) > 0$$

Robert Israel
  • 470,583
3

Short answer: None of the two has a reason to be larger than the other. Some bounds in terms of the condition numbers of $Y,Z$ are $$trace(XYXZ) \leq \min(\kappa(Y), \kappa(Z)) \cdot trace(X^2YZ)$$ $$trace(X^2YZ) \leq \min(\kappa(Y), \kappa(Z)) \cdot trace(XYXZ)$$


First, we can simultaneously diagonalize $Y$ and $Z$, say $Y = PY'P^{-1}$, $Z = PZ'P^{-1}$ with $P$ unitary, then we get $$trace(X'Y'X'Z') \quad, \quad trace(X'^2Y'Z')$$ with $X' = P^{-1}XP$ still hermitian. Thus we may suppose $Y,Z$ are diagonal. Then $$\begin{align*} trace(XYXZ) &= \sum_{i,j}|X_{ij}|^2Y_{jj}Z_{ii} \\ trace(X^2YZ) &= \sum_{i,j}|X_{ij}|^2Y_{ii}Z_{ii} \end{align*}$$ The $Y_{ii},Z_{ii}>0$ are arbitrary, so all we can say is that one is bounded by $$\kappa(Y) = \max_{i \neq j}|Y_{ii}/Y_{jj}|$$ (or $\kappa(Z) = \max_{i \neq j}|Z_{ii}/Z_{jj}|$) times the other.

Bart Michels
  • 26,985
  • 6
  • 59
  • 123