3
$\begingroup$

Let $ 0 \preceq A \preceq I$ and let $B$ and $C$ be two symmetric positive definite matrices.

Can we related \begin{align} {\rm Tr}( (I-A) B+A C) \end{align}

to

\begin{align} Tr(I-A)Tr(B)+Tr(A) Tr(C) \end{align}

via some inequality?

Note, that in the case $A=aI$ for any $a\in [0,1]$ there is an equality.

Edit: Note, from a very nice answer below we have:

\begin{align} \lambda_{\min}(I-A){\rm Tr}(B)+\lambda_{\min}(A){\rm Tr}(C) \le {\rm Tr}( (I-A) B+A C) \le \lambda_{\max}(I-A){\rm Tr}(B)+\lambda_{\max}(A){\rm Tr}(C). \end{align}

$\endgroup$
  • $\begingroup$ Just an observation, though it may not be useful to you. If $0\preceq A\preceq I$, then $0\preceq I-A\preceq I$. Therefore, if $B$ and $C$ are nonnegative matrices, $0\leq\operatorname{Tr}((I-A)B)\leq\operatorname{Tr}(B)$ and $0\leq\operatorname{Tr}(AC)\leq\operatorname{Tr}(C)$. It follows that in this case, $$0\leq\operatorname{Tr}((I-A)B+AC)\leq\operatorname{Tr}(B+C)$$ and $$0\leq\operatorname{Tr}(I-A)\operatorname{Tr}(B)+\operatorname{Tr}(A)(C)\leq\operatorname{Tr}(B+C)$$ (both quantities are nonnegative and bounded above by $\operatorname{Tr}(B+C)$). $\endgroup$ – parsiad Apr 19 '17 at 13:52
3
$\begingroup$

Note that the function $$ \langle A,B \rangle = Tr(A^TB) $$ defines an inner product over the real matrices. As such, $\|A\|_2^2 = Tr(A^TA)$ is the norm induced by this inner product. We have $$ Tr((I-A)B + AC) = Tr((I-A)B) + Tr(AC) =\\ \langle I-A,B \rangle + \langle A,C \rangle \leq\\ \|I - A\|\,\|B\| + \|A\|\,\|C\| = \\ \sqrt{Tr[(I - A)^2]Tr(B^2) + Tr(A^2)Tr(C^2)} $$ We can do a bit better using the inequality $|\langle A,B \rangle| \leq \sigma_1(A)\sum \sigma_i(B)$. In the symmetric positive semidefinite case, amounts to $|\langle A,B \rangle| \leq \rho(A) Tr(B)$, where $\rho$ denotes the spectral radius. We now have $$ \langle I-A,B \rangle + \langle A,C \rangle \leq\\ \rho(I - A)Tr(B) + \rho(A) Tr(C) $$ This reduces to your result in the case that $A = aI$ for $a \in [0,1]$.


The reverse inequality works perfectly well for our purposes: note that $A \succeq \lambda_{min}(A)I$. So, $C^{1/2}AC^{1/2} \succeq \lambda_{min}(A)C^{1/2}C^{1/2}$. We therefore have $$ Tr((I-A)B + AC) = Tr((I-A)B) + Tr(AC) =\\ Tr(B^{1/2}(I-A)B^{1/2}) + Tr(C^{1/2}AC) \geq\\ Tr(\lambda_{min}(I - A)B^{1/2}B^{1/2}) + Tr( \lambda_{min}(A)C^{1/2}C^{1/2}) = \\ \lambda_{min}(I-A)Tr(B) + \lambda_{min}(A) Tr(C) $$ in retrospect, a similar inequality could have worked in reverse.

$\endgroup$
  • $\begingroup$ I don't have a reference for these inequalities of the top of my head. If someone could find such a reference, I would appreciate it. I'm using the Hölder's inequalities for Schatten p-norms. $\endgroup$ – Omnomnomnom Apr 19 '17 at 13:58
  • $\begingroup$ What is this second inequality called that you are using? $ |<A ,B>| \le \sigma_1 \sum \sigma_i(B)$. $\endgroup$ – Boby Apr 19 '17 at 14:00
  • $\begingroup$ I don't know. In Pedersen's "Analysis Now", it's lemma 3.4.10, and that's the infinite dimensional version. $\endgroup$ – Omnomnomnom Apr 19 '17 at 14:06
  • $\begingroup$ See also proposition IV.2.4 from Bhatia's Matrix Analysis $\endgroup$ – Omnomnomnom Apr 19 '17 at 14:11
  • $\begingroup$ @Boby I'm going to post my own question about that. $\endgroup$ – Omnomnomnom Apr 19 '17 at 14:13

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.