Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

The matrices I discuss are all $N\times N$ hermitian matrices. Consider two (hermitian) matrices $A_1$ and $A_2$. For a real scalar $t$, define the following function for the matrix $A_1+t*A_2$ \begin{align} \lambda(t)=\min_{u^Hu=1}~u^H(A_1+t*A_2)u \end{align} For a given $t$, this is essentially the smallest eigenvalue of $A_1+t*A_2$. This should be a concave function. Now define the function \begin{align} f(t)=||A_1+t*A_2||_F^2 \end{align} I randomly generated two hermitian matrices in my simulation software (matlab), and plotted this for a range of $t$ The image. You can see something happening here. $f(t)$ (blue curve) appears to be a convex function. The main thing being the value of $t$ at which $f(t)$ attains its minimum and $\lambda(t)$ (red curve) attains its maximum seems to be same. Is there any explanation for this. I tried it several times (more than 15 times at least), the nature of the resulting graph is always same. Can anyone give a possible explanation for this.

UPDATE--- I checked the exact values. They are close. But not the same. But still any possible explanation for this behaviour.

share|cite|improve this question
Note that you can get proper spacing for the norm bars using \lVert and \rVert. – joriki Dec 2 '12 at 14:49
up vote 1 down vote accepted

As long as you don't quantify "close", there's nothing to explain here. It's clear that both graphs are asymptotically linear in both directions and must have a cross-over region in between. Thus if you zoom out far enough that both cross-over regions look small, their graphs will inevitably look like this.

share|cite|improve this answer
This came from a random gaussian matrix. As soon as changed the variances, the behavior started to change. So I guess there is not much of thing inside it. Thanks for the answer – dineshdileep Dec 5 '12 at 4:08

I think you plotted $g: t \mapsto \Vert A_1 + t A_2 \Vert_F$ as the upper curve, not its square $f$.

For the special case $A_2 = I$ (identity matrix), you will not get this behavior. In that case $\lambda(t) = t + \mu_{min}(A_1)$, where $\mu_{min}(A)$ is the smallest eigenvalue of the matrix $A$. More generally, the slope of $\lambda$ will be near $\mu_{max}(A_2)$ for large negative $t$ and near $\mu_{min}(A_2)$ for large positive $t$. If $A_2$ is positive (negative) definite, then $\lambda$ can be expected to increase (decrease) monotonically.

share|cite|improve this answer
Thanks for pointing out that, I was not plotting the square. $A_1$ and $A_2$ are hermitian indefinite. But I observed the behaviour you pointed when $A_2$ is positive definite (or negative). I guess that can proved from interlacing theorem. – dineshdileep Dec 4 '12 at 16:31

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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