MathOverflow is a question and answer site for professional mathematicians. 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

So $X$ and $Y$ are Hermitian matrices (or just symmetric real) of size $n$ by $n$ and suppose $Y\succeq X$, namely $Y-X$ is positive-semidefinite. Now write the eigenvalues of $Y$ as $\alpha_1\leq\ldots\leq \alpha_n$, and the eigenvalues of $X$ as $\beta_1\leq\ldots\leq \beta_n$. Is is necessarily true that $\alpha_i\geq\beta_i$ for all $i$?

I might be able to solve this myself (although with time I am less sure), but it should be be much easier for whoever already knows the answer. A quick reference would do, thanks.

share|cite|improve this question
Where does this question come from? – Igor Rivin Sep 18 '12 at 16:34
up vote 9 down vote accepted

This is true and well known. By the minimax principle, $\alpha_k$ is the minimum over all $k$-dimensional subspaces of the norm of the quadratic form $v\mapsto(v,Yv)$ restricted to the subspace. And similarly for $\beta_k$ and $(v,Xv)$. Since $(v,Yv)\ge(v,Xv)$ for every vector $v$, the same inequality holds for the norms of restrictions to subspaces and hence for eigenvalues.

I think any textbook covering the minimax principle has this corollary.

share|cite|improve this answer
Yes, you are right, it's quite simple. Too bad I missed it.. While thinking of my question above I was sidetracked to think of a stronger related question: Suppose E is a ellipsoid contained inside a bigger ellipsoid F. Can E always be moved inside F until all their axes are parallel? A positive answer also gives a positive answer to my original question, but I don't see how to obtain an implication in the other direction. Do you happen to know the solution to the latter question as well? – puzne Sep 18 '12 at 19:29

At this point I find it worth mentioning the following facts. Let $X$ and $Y$ be arbitrary square complex matrices. Let $|X|=(X^*X)^{1/2}$ be the matrix absolute value. Then, we may have the following inequalities (in decreasing order of strength, i.e., a higher one in the list implies a lower one):

  1. $|X| \succeq |Y|$
  2. $\sigma_i(X) \ge \sigma_i(Y)$, where $\sigma_i$ is the $i$th singular value
  3. $\prod_{j=1}^k \sigma_j^\downarrow(X) \ge \prod_{j=1}^k \sigma_j^\downarrow(Y)$ for $k=1,2,\ldots,n$;
  4. $\sum_{j=1}^k \sigma_j^\downarrow(X) \ge \sum_{j=1}^k \sigma_j^\downarrow(Y)$ for $k=1,2,\ldots,n$;
  5. $\| X\| \ge \|Y\|$, where $\|\cdot\|$ is any unitarily invariant norm.

Actually, in the list above, $4\Leftrightarrow 5$. The original question is a special case of $1 \implies 2$, and the textbook proof is as S. Ivanov has nicely explained above.

share|cite|improve this answer
Thanks - do you have an easy reference for the above facts? – puzne Sep 19 '12 at 16:43
Hi Puzne, these results are developed in several books; I find the book "Matrix Analysis" by R. Bhatia, particularly readable. Notice that 1,2 follows from Sergei's answer (and the fact that $\sigma(|X|)=\sigma(X)$ ); then 2 obviously implies 3. The fact that $3$ implies $4$ is a short exercise, while $4$ implies $5$ is an important theorem in matrix analysis; so basically, you'll need to refer to some book only for the last result. – Suvrit Sep 19 '12 at 18:22

No, but $$ \sum_{j=1}^{k} \alpha_j \geq \sum_{j=1}^{k} \beta_j $$ for all $k$ as EASILY follows from the mini max principle.

share|cite|improve this answer

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.