3
$\begingroup$

Let $M\in\mathbb{R}^{n\times n}$ be symmetric positive definite and consider a matrix $Q\in\mathbb{R}^{n\times m}$ ($m<n$) with orthonormal columns ($Q^TQ=I$). I'm interested in finding an exact expression for $$ K\equiv\max_{x}\frac{x^Tx}{x^TQ^TM^{-1}Qx} $$ in terms of the Rayleigh quotient of $M$ (no inverse). Equivalently, $K$ can be considered as the maximum of $x^Tx/x^TM^{-1}x$ over $x\in\mathcal{R}(Q)$ (the range of $Q$).

Obviously, if $Q=I$, then $$ \max_x\frac{x^Tx}{x^TM^{-1}x}=\max_y\frac{y^TMy}{y^Ty} $$ simply by substituting $y=M^{-1/2}x$. However, that does not work for $K$ above. Of course, what I can do is $$ K=\max_{y}\frac{y^T(Q^TM^{-1}Q)^{-1}y}{y^Ty}, $$ but I can't see that to be very useful.

What I'm actually looking for is something like this: $$ K=\max_{x\in S_1}\min_{y\in S_2(x)}\frac{y^TMy}{x^Tx}, $$ where $S_1$ and $S_2(x)$ are some suitable subspaces of $\mathbb{R}^{n}$ ($S_2$ generally dependent on $x$). However, I'm not sure how to arrive to this max-min characterisation.

Thanks in advance for any suggestion.

P.S.: I have a suspicion that this might be contained in some book dealing thoroughly with variational characterisations of the eigenvalues of SPD matrices. However, I haven't found any with such a result in my library.

$\endgroup$

1 Answer 1

3
$\begingroup$

If $z=M^{-1/2}Qx$, then $$K=\max_{z\in S}\frac{z^TMz}{z^Tz},$$ where $S$ is the range of $M^{-1/2}Q$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.