Given any real-valued, symmetric, positive semi-definite matrix $A, B \in \mathbb{R}^{n \times n}$, $\lambda > 0$. Does the following inequality hold for some constant $c$? $$\sigma_{\max}\Big((A+B + \lambda I_n)^{-1}A\Big) \leq c,$$ where $\sigma_{\max}(\cdot)$ denotes the maximal singular value.

The constant $c$ might be dependent on $\lambda$ and some norm of $A$ or $B$.


No. Let $A_0=\pmatrix{1&0\\ 0&0}, B_0=\pmatrix{1&1\\ 1&1}$ and $$ C_0=(A_0+B_0)^{-1}A_0=\pmatrix{1&-1\\ -1&2}\pmatrix{1&0\\ 0&0} =\pmatrix{1&0\\ -1&0}. $$ Then $\sigma_1(C_0)=\|C_0\|_2>1$ because the Euclidean norm of the first column of $C_0$ is greater than $1$.

It follows that if $A=A_0+tI$ and $B=B_0+\lambda I$ for some small $t,\lambda>0$, then $A$ is positive definite, $B\succeq\lambda I$ and $\|C\|_2>1$ for $C=(A+B)^{-1}A$.

Similarly, if $A=X_1X_1^T+\cdots+X_pX_p^T$ and $X_{p+1}X_{p+1}^T+\cdots+X_tX_t^T$ are close to $A_0$ and $B_0$ respectively and $\lambda>0$ is small, then $\|C\|_2>1$ when $C=(A+B+\lambda I)^{-1}A$.

| cite | improve this answer | |
  • $\begingroup$ Many thanks! In my problem, the matrix $B$ actually has a better property that $B \geq \lambda I_n$ for some $\lambda > 0$. So can we avoid the counterexample and will the statement hold now? $\endgroup$ – Peng Zhao May 7 at 11:05
  • 2
    $\begingroup$ @PengZhao Still no. $\sigma_1\left((A+B)^{-1}A\right)$ is continuous in the entries of $B$. If you add a small $\lambda I$ to the $B$ above, you still get a counterexample. $\endgroup$ – user1551 May 7 at 11:08
  • $\begingroup$ Could you please offer some more elaborations or some references? This would be quite helpful. $\endgroup$ – Peng Zhao May 7 at 11:11
  • $\begingroup$ @PengZhao Sorry, I don't know any reference. $\endgroup$ – user1551 May 7 at 11:13
  • $\begingroup$ Thanks for the update! I believe the counterexamples can be avoided by requiring $\lambda > 0$ and $\| X_s \| \leq S$, and the threshold will be dependent on $\lambda$ and $S$, instead of the constant $1$. $\endgroup$ – Peng Zhao May 7 at 15:55

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.