Assume that $A$ is a $n\times n$ positive matrix, whose eigenvalues are $a_1\ge a_2\ldots \ge a_n>0;$ $B$ is a $p \times p$ positive matrix, whose eigenvalues are $b_1\ge b_2\ldots \ge b_p>0.$ For $n\times p$ full rank matrix $X$ with $n\ge p,$

How to prove $$\det(X'AX+B)\ge c(X)\prod_{i=1}^p(a_{n-p+i}+b_{i}),$$

where $c(X)$ is a postive constant that depends only on $X$.

  • $\begingroup$ Maybe you have something more precise in mind. For instance, $c(X)=0$ always works. $\endgroup$ – Suvrit Jan 28 at 19:18
  • $\begingroup$ @Suvrit Thanks. $C(X)>0.$ $\endgroup$ – Xiaopai Song Jan 28 at 20:41
  • 1
    $\begingroup$ What do we get if $X=0$? $\endgroup$ – Fedor Petrov Jan 30 at 21:28
  • $\begingroup$ @Fedor Petrov Thank you again. It's full rank. $\endgroup$ – Xiaopai Song Jan 30 at 21:57
  • $\begingroup$ @Denis Serre Thanks for your edition. Are you interesting in another inequality mathoverflow.net/q/322262/134602? $\endgroup$ – Xiaopai Song Feb 2 at 15:29

Your both conjectural inequalities are equivalent to each other and false even for $n=p$, $A=B$, $X=I$ (and $a_i=\lambda^i$ for large $\lambda$, for example).

What is true that $$|X'AX+B|\geqslant c(X)\prod_{i=1}^p (a_{n-p+i}+b_i).$$

Proof. Let $\lambda_1\geqslant \lambda_{2}\geqslant \ldots \geqslant \lambda_p$ be eigenvalues of $X'AX+B$. It suffices to prove that $$\lambda_{i}\geqslant \rho(b_i+a_{n-p+i})\quad (*)$$ for all $i=1,\dots,p$, where $\rho=\rho(X)$ depends only on $X$. Consider two cases.

1) $b_i\geqslant a_{n-p+i}$. Then $(*)$ holds with $\rho=1/2$, since $X'AX+B\geqslant B$, thus $\lambda_i\geqslant b_i$.

2) $a_{n-p+i}\geqslant b_i$. Let $L$ be the image of the operator (identified with the matrix) $X$, $\dim L=p$ and $X$ is a linear isomorphism onto $L$. Denote by $4\rho^2$, $\rho>0$, the norm of the inverse map $X^{-1}:L\mapsto X$. The intersection $L_i$ of $L$ and the space generated by $u_1,\dots,u_{n-p+i}$, where $Au_i=a_i u_i$ and $u_i$ are orthogonal eigenvectors of $A$, has dimension at least $i$. We have $$((X'AX+B) x,x)\geqslant (X'AX x,x)=(AXx,Xx)\geqslant a_{n-p+i}(Xx,Xx)\geqslant 2\rho a_{n-p+i}(x,x)$$ whenever $x\in X^{-1}L_i.$ Therefore by the variational (Courant) principle $\lambda_i\geqslant 2\rho(X)a_{n-p+i}\geqslant \rho(a_{n-p+i}+b_i)$.

  • $\begingroup$ Thank you again. Your answer is right. $\endgroup$ – Xiaopai Song Jan 31 at 17:12
  • $\begingroup$ Are you interested in other similar inequalities? mathoverflow.net/q/322262/134602 and mathoverflow.net/q/322265/134602. For now, I can't prove it. $\endgroup$ – Xiaopai Song Feb 2 at 3:36
  • 1
    $\begingroup$ Again, estimate consecutive eigenvalues of all involved matrices using variational principle $\endgroup$ – Fedor Petrov Feb 2 at 9:24
  • $\begingroup$ Thanks for your constructive comment. I try to handle the inequality (2) mathoverflow.net/q/322262/134602 and makes comments in it. I am not familiar with the variational principle. So I can't follow you overall. If you have time, could you give some comments or answer it? Your comments are very important to me. $\endgroup$ – Xiaopai Song Feb 2 at 15:17
  • 1
    $\begingroup$ I mean en.wikipedia.org/wiki/Min-max_theorem In other words, in order to estimate the $i$-th eigenvalue from below you should find a subspace of dimension $i$ onto which you have a lower bound for the quadratic form $\endgroup$ – Fedor Petrov Feb 2 at 19:23

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.