Let $A$ and $B$ be two $n \times n$ (symmetric) positive definite matrices, and denote the $k$th largest eigenvalue of a general $n \times n$ matrix by $\lambda_k(X)$, $k = 1, 2, \ldots, n$ so that $$\lambda_1(X) \geq \lambda_2(X) \geq \cdots \geq \lambda_n(X).$$ I guess the following relation holds: $$\lambda_k(A + B) > \max\{\lambda_k(A), \lambda_k(B)\}, \; k = 1, 2, \ldots, n.$$

This looks intuitive but I have difficulty to prove it, any hints?

  • $\begingroup$ How do you order eigenvalues? $\endgroup$ – Pierre-Guy Plamondon Jan 22 '16 at 18:18
  • $\begingroup$ It doesn't matter. I am trying to say any eigenvalue of the sum should be greater than the corresponding ones. To be more precise, I will edit. $\endgroup$ – Zhanxiong Jan 22 '16 at 18:20

For symmetric matrices you have the Courant-Fischer min-max Theorem: $$\lambda_k(A) = \min \{ \max \{ R_A(x) \mid x \in U \text{ and } x \neq 0 \} \mid \dim(U)=k \}$$ with $$R_A(x) = \frac{(Ax, x)}{(x,x)}.$$ Now, your assertion follows easily, since $R_{A+B}(x) > \max\{R_A(x), R_B(x)\}$.

This theorem is also helpful to prove other nice properties of the eigenvalues of symmetric matrices. For example: \begin{equation*} \lambda_k(A) + \lambda_1(B) \le \lambda_k(A+B) \le \lambda_k(A) + \lambda_n(B) \end{equation*} This shows the continuous dependence of the Eigenvalues on the entries of the matrix, and also your assertion.

  • $\begingroup$ Does this result also hold for sum of semi definite positive matrices? (changing the > for ≥ ) $\endgroup$ – Manuel Oct 29 '18 at 21:51
  • 1
    $\begingroup$ Yes and it's the same proof. $\endgroup$ – gerw Oct 30 '18 at 8:44
  • $\begingroup$ @grew : your last equation is not inline with the question. OP is considering $\lambda_n$ as the lowest eigen value and $\lambda_1$ as highest. But you seem to consider the other way. $\endgroup$ – Rajesh Dachiraju Mar 19 at 8:58

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.