The problem:

Let $A, B$ be real matrices of dimension $n$ and $\alpha_1 \geq \dots \geq \alpha_n$ and $\beta_1 \geq \dots \geq \beta_n \geq 0$ its eigenvalues, respectively ($B$ is symmetric positive semidefinite). Also, suppose that $\sum_{i=1}^n\beta_i=1$ and that the eigenvalues are all real.

Is it true that $\mbox{trace}(A^TB)\geq \alpha_n$?

My attempt:

Let $A=PJP^T$ be the Jordan representation of $A$ and $B=QDQ^T$ the spectral decomposition of $B$, that is, $P$ and $Q$ are orthonormal matrices of eigenvector bases, $J$ is the Jordan form of $A$ and $D=diag(\beta_1,\dots,\beta_n)$.

Then $$trace(A^TB)=trace(PJ^TP^TQDQ^T)=trace(\underbrace{(Q^TP)J^T(Q^TP)^T}_{this:=S}D)=\sum_{i=1}^n S_{ii}\beta_i.$$ And I can't continue.

My second (and wrong) attempt:

My wrong attempt is similar, but the underlined step is probably wrong. I know it is in general not true, but I am not sure about this particular case. $$trace(A^TB)=\underline{trace(PJ^TP^TQDQ^T)=trace(J^tPP^TQQ^TD)}=trace(J^TD)=\sum_{i=1}^n \alpha_i\beta_i\geq \alpha_n.$$

For instance, if this is true, then the lesser eigenvalue of $A$ is the solution of a semidefinite program: $$\alpha_n=\min_B\{trace(A^TB)\mid trace(B)=1,B\succeq 0\}.$$

Thank you!


Let $$\begin{align}A&=\begin{bmatrix}0&2\\0&0\end{bmatrix}&B=\frac{1}{2}\begin{bmatrix}1&-1\\-1&1\end{bmatrix}\end{align}\text{.}$$ Under your conjecture one would have $\mathrm{Tr}\,A^{\mathsf{T}}B\geq 0$. Instead, $$\mathrm{Tr}\,A^{\mathsf{T}}B=-1\text{.}$$

  • 1
    $\begingroup$ The eigenvalues of $A$ and $B$ are $\{0,0\}$ and $\{0,1\}$, respectively... $\endgroup$ – K B Dave Mar 2 '18 at 19:56
  • $\begingroup$ Is it true when $A$ is symmetric? $\endgroup$ – Koto Mar 2 '18 at 19:59
  • $\begingroup$ Yes, because by subtracting $I\alpha_n$ we can assume without loss of generality that $A$ is psd, in which case $\mathrm{Tr}\, AB=\mathrm{Tr}\, B^{1/2}AB^{1/2}\geq 0$. $\endgroup$ – K B Dave Mar 2 '18 at 20:01
  • $\begingroup$ Sorry, I didn't get it, can you please elaborate a little more? $\endgroup$ – Koto Mar 2 '18 at 20:04
  • 3
    $\begingroup$ $A$ symmetric implies $A-\alpha_n I$ is psd, so set $A\leftarrow A-\alpha_n I$; what needs to be shown now is $\mathrm{Tr} AB\geq 0$. $B$ is psd, so its psd square root $B^{1/2}$ exists. by cyclicity of the trace, $\mathrm{Tr}AB^{1/2}B^{1/2}=\mathrm{Tr}B^{1/2}AB^{1/2}$, so the question is reduced to demonstrating the inequality $\mathrm{Tr}B^{1/2}AB^{1/2}\geq 0$. But $A$ psd implies $B^{1/2}AB^{1/2}$ psd, and traces of psd matrices are nonnegative. $\endgroup$ – K B Dave Mar 2 '18 at 20:13

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.