I would like to prove (or prove it is not true with a counter example) the following result:

Let $A$, $B$ be two squares matrices of size $n\times n$ with positive entries. If $A \leq B$, then $\sum_{i = 1}^n \sigma_i(A) \leq \sum_{i = 1}^n \sigma_i(B)$.

The sign $\leq$ between matrices is meant elementwise and the notation $\sigma_i(A)$ is used to denote the i-th singular value of $A$. The Perron-Frobenius theorem only provides information on the spectral radius, while the above statement involves all singular values.

  • 1
    $\begingroup$ Probably I'm missing something trivial, so let me ask. Your definition of $\leq$ implies that $\mathrm{trace}(A) \leq \mathrm{trace}(B)$, so it seems to me that what you are asking is trivially true. Where am I wrong? $\endgroup$ Feb 19 '18 at 9:42
  • 3
    $\begingroup$ @FrancescoPolizzi: The matrices $A$ and $B$ need not be symmetric, so the sums of the singular values are $\operatorname{trace}(\sqrt{A^*A})$ and $\operatorname{trace}(\sqrt{B^*B})$ rather than $\operatorname{trace}(A)$ and $\operatorname{trace}(B)$ . $\endgroup$ Feb 19 '18 at 9:46
  • $\begingroup$ @Jochen Glueck: Ah, right, I was missing that point. Thanks! $\endgroup$ Feb 19 '18 at 9:47

The inequality is not true, in general. Here is a counterexample:

Let $B = \begin{pmatrix} 1 & 1 \\ 1 & 1 \end{pmatrix} $ and let $A = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} $ (my favourite counterexample matrix).

The singular values of $B$ are $2$ and $0$. A short computation shows that the singular values of $A$ are $(\frac{3 + \sqrt{5}}{2})^{1/2}$ and $(\frac{3 - \sqrt{5}}{2})^{1/2}$; the sum of those two values is strictly larger than $2$ (approximately $2.236$, but admittedly I used a calculator to check this...)

  • 1
    $\begingroup$ Since science needs more confirmatory results and replication: I think this is entirely correct. Besides hand-calculation, e.g. the following sage-code can be used to confirm the singular vales: A = matrix(CDF, [[1,1],[1,1]]);B = matrix(CDF, [[1,1],[0,1]]); [A.SVD() , B.SVD() , N( ((3+sqrt(5))/2)^0.5 ) , N( ((3-sqrt(5))/2)^0.5 ) ]. The last statement is easily seen by squaring and rearranging to be equivalent to $3^2−(\sqrt{5})^2>1$, which is true. $\endgroup$ Feb 19 '18 at 12:28

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.