4
$\begingroup$

Assume we have the matrix product: $$A=PBD$$

where $P$ is a projection matrix (i.e., $P=P^2$, $P=P^\top$, and $\|P\|_2=1$), $B$ is a matrix whose infinite norm is equal to one ($\|B\|_\infty=1$), and $D$ is a diagonal matrix whose $\ell_2$-norm is less than one ($\|D\|_2<1$).

Is it correct to say that the spectral radius of $A$ is less than 1 ($\rho(A)< 1$). If yes, how to prove it?

$\endgroup$
2
$\begingroup$

Yes. Denote the entrywise absolute value of a complex matrix $X$ by $|X|$. In general, if $|X|\le Y$ entrywise, then $\rho(X)\le\rho(|X|)\le\rho(Y)$.

Your $P$ is an orthogonal projection. Therefore the moduli of its entries are bounded above by $1$. Hence $|A|\le|B||D|$ and $\rho(A)\le\rho(|B||D|)$. So, it suffices to prove that the latter quantity is strictly smaller than $1$.

Suppose $(\lambda,v)$ is an eigenpair of $|B||D|$. Clearly, $\|\,|B|\,\|_\infty=\|B\|_\infty=1$. As $D$ is diagonal, $\|\,|D|\,\|_\infty=\|D\|_2<1$. Therefore $|\lambda|\|v\|_\infty=\|\,|B|\,|D|\,v\|_\infty\le\|\,|B|\,\|_\infty\|\,|D|\,\|_\infty\|v\|_\infty<\|v\|_\infty$. Thus $\rho(|B||D|)<1$. $\square$

Remark. Your conjecture is true because $D$ is diagonal. Otherwise it is false in general. Here is a counterexample. Let $B=\pmatrix{-\frac12&-\frac12\\ 0&1}$. Its largest singular value is $\sigma_1=\frac14(3+\sqrt{5})>1$. Let $B=USV^T$ be its singular value decomposition. Set $P=U\pmatrix{1\\ &0}U^T$ and $D=rVU^T$ for some $r$ that is smaller than but sufficiently close to $1$. Then $A=PBD=r\sigma_1U\pmatrix{1\\ &0}U^T$, whose spectral radius is $r\sigma_1>1$.

$\endgroup$
  • $\begingroup$ Thank you for your useful response. One additional question. If I have a block diagonal matrix $D$ instead of a diagonal one where each block $D_i$ is symmetric and $\|D_i\|_2<1$. Furthermore, I know that $D$ is diagonally dominant. Is it possible to prove that $\rho(A)<1$? $\endgroup$ – user293017 Mar 31 '16 at 9:14
  • $\begingroup$ @user293017 The $D$ in the counterexample given in my answer is symmetric and diagonally dominant. Numerically $D\approx r\pmatrix{-0.9487&-0.3162\\ -0.3162&0.9487}$. So, using the $P,B$ and $D$ in the remark, $\pmatrix{P\\ &P}\pmatrix{B\\ &B}\pmatrix{D\\ &D}$ gives you a counterexample to your new conjecture.. $\endgroup$ – user1551 Mar 31 '16 at 19:17

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.