I'm looking for a proof of the Kantorovich inequality, namely:

$$ \langle Ax,x\rangle \langle A^{-1}x,x\rangle \leq \frac{1}{4}\left(K(A)+\frac{1}{K(A)}\right)$$ Where $ K(A)= \lVert A\rVert_2 \lVert A^{-1}\rVert_2 $ and $A $ is an Hermitian positive definite matrix and $x$ a vector with the accurate size. Or alternatively

$$ \langle Ax,x\rangle \langle A^{-1}x,x\rangle \leq \frac{1}{4}\left(\left(\frac{\beta}{\alpha}\right)^{2}+\left(\frac{\alpha}{\beta}\right)^{2}\right)$$ where $0 < \alpha = \lambda_1 \leq \cdots \leq \lambda_n = \beta$ are the eigenvalues of $ A$.

There are a lot of proofs on internet but this is the one that I found easier to understand. Nevertheless, I'm stuck figuring out something:

They use $ f,g: [\alpha,\beta]\to \mathbb{R} $ two convex function with $ f$ positive and $f(t)\leq g(t) $ for every $t \in [\alpha, \beta] $. Then they claim that $ F=f(A)$ and $ G=g(A)$ are well defined and hermitian positive definite.

I don't even understand why they mean by $ f(A)$.

If any of you guys can suggest me alternative documentation (other proof) or help me with this one, I would be very grateful

  • $\begingroup$ $F = f(A)$ is the matrix that is obtained by expandind $f(t)$ as a Taylor series (an "infinite polynomial") and then substituting $A$ for $t$ in that expansion. $\endgroup$ – NickD Apr 4 '18 at 20:13
  • $\begingroup$ @Nick thank you but how does that implies that $ F$ and $G$ are hermitian positive definite $\endgroup$ – Natalio Apr 4 '18 at 20:16
  • $\begingroup$ If $\lambda$ is an eigenvalue of $A$ then $f(\lambda)$ is an eigenvalue of $f(A)$. I think that's enough to push through the proof, given the conditions. $\endgroup$ – NickD Apr 4 '18 at 20:39
  • $\begingroup$ @Nick, yes that is enough. Thank you very much $\endgroup$ – Natalio Apr 4 '18 at 20:46

If $A$ is diagonalizable, i.e., if $A=SDS^{-1}$, with $D= \text{diag}(\lambda_1,\dots,\lambda_n)$, and if $f:\mathbb{R} \longrightarrow \mathbb{R}$, then $$ f(A) := S f(D) S^{-1}, $$ in which $f(D):= \text{diag}(f(\lambda_1),\dots,f(\lambda_n))$.

In the problem above, since every Hermitian matrix is diagonalizable (in fact, if $A$ is Hermitian, then there is a unitary matrix $U$ such that $U^*AU$ is diagonal), it follows that $f(A)$ is well-defined and positive definite since $f$ is positive.


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.