Let $C = A + B$, where $A$, $B$, and $C$ are positive definite matrices. In addition, $C$ is fixed. Let $\lambda (A)$, $\lambda (B)$, and $\lambda (C)$ be smallest eigenvalues of $A$, $B$, and $C$, respectively. Is there any result about the smallest eigenvalues of $C$ in comparison with the sum of smallest eigenvalues of $A$ and $B$? Is it true that : $\lambda (A)$ + $\lambda (B)$ < $\lambda (C)$ ? Moreover, what is the smallest possible value of $\lambda (A)$ + $\lambda (B)$ given a fixed $C$, and under what condition does this happen? Many thanks!


------------------------------ Post Edit ---------------------------

Question about $\lambda_{min} (A+B) > \lambda_{min} (A) + \lambda_{min} (B) $ can be seen from Weyl's inequality.

The remaining question is about the smallest attainable value of $\lambda_{min} (A) + \lambda_{min} (B) $ given a fixed $C$?

  • $\begingroup$ Are $A,B$ and $C$ symmetric? In this case you can use Rayleigh quotient $\frac{x^tCx}{x^tx}$. $\endgroup$ – Davide Giraudo Oct 9 '12 at 20:01
  • $\begingroup$ A, B, and C are positive definite, therefore, symmetric. Could you clarify on how using Rayleigh quotient to find smallest possible $\lambda (A) + \lambda (B)$ ? $\endgroup$ – Ho Xuan Oct 9 '12 at 20:10

If $S$ is a symmetric real matrix, then the minimal eigenvalue of $S$, $\lambda_{\min}(S)$ is given by $$\lambda_{\min}(S)=\min_{x\neq 0}\frac{x^tSx}{x^tx}.$$ Indeed, it's not hard to see that when $S$ is diagonal, and in the general case diagonalize $S$ in an orthonormal basis.

By the properties of $\min$, this gives $\lambda_{\min}(A+B)\geq \lambda_{\min}(A)+\lambda_{\min}(B)$. Moreover, if we have equality $\frac{x^tSx}{x^tx}$, it's for an eigenvector for $\lambda_{\min}$, hence we have equality if $A$ and $B$ have a common eigenvector.

  • $\begingroup$ Thank you for the answer! I've just realized the first question $\lambda_{min} (A+B) > \lambda_{min} (A) + \lambda_{min} (B) $ can also be seen from Weyl's inequality. However, I am still looking for the answer about the smallest possible value of $\lambda_{min} (A) + \lambda_{min} (B) $ given a fixed $C = A + B$ . $\endgroup$ – Ho Xuan Oct 9 '12 at 20:34
  • $\begingroup$ When $C=I$, we can take $A=\pmatrix{k^{-1}&0\\0&1-k^{—1}}$ and $B=\pmatrix{1-k^{-1}&0\\0&k^{—1}}$ to see that the minimum can be arbitrary small. $\endgroup$ – Davide Giraudo Oct 9 '12 at 20:43
  • $\begingroup$ I think this is only true for certain cases of C. In general, the minimum value is dependent on C and not arbitrarily small. There should be some relationship between $v_{min} (A)$, $v_{min} (B)$, and $v_{min} (C)$ when the minimum value is attained. Here $v_{min}$ is the eigenvector associated with the smallest eigenvalue. Thanks $\endgroup$ – Ho Xuan Oct 9 '12 at 21:16
  • $\begingroup$ So, what are the the cases you are interested in? $\endgroup$ – Davide Giraudo Oct 10 '12 at 8:31
  • $\begingroup$ Hi, thanks for keep answering. I realized that the general case is difficult and there is no obvious connection among eigenvectors corresponding to smallest eigenvalues of $A$, $B$, and $C$. $\endgroup$ – Ho Xuan Oct 11 '12 at 15:55

I answer the second part of the question. For $n=1$ the answer is $\lambda_{min}(C)$. For $n > 1$, the answer is, $\lambda_{min}(A) + \lambda_{min}(B)$ can be arbitrarily close to $0$.

Since $C$ is a symmetric positive definite matrix, we consider the singular value decomposition of $C$. $$ C = \sum_{i=1}^n \lambda_i v_i v_i^T,$$ where $0 < \lambda_1 \leq \dots \leq \lambda_n$ are eigenvalues of $C$, and $v_1, \dots, v_n$ is an orthonormal basis. For sufficiently small $\varepsilon > 0$, let \begin{align*} A &= (1-\varepsilon) \lambda_1 v_1 v_1^T + \varepsilon\lambda_2 v_2 v_2^T + \frac{1}{2}\sum_{i=3}^n \lambda_i v_i v_i^T,\\ B &= \varepsilon \lambda_1 v_1 v_1^T + (1-\varepsilon)\lambda_2 v_2 v_2^T + \frac{1}{2}\sum_{i=3}^n \lambda_i v_i v_i^T. \end{align*}

Now $A+B=C$, $A$ and $B$ are positive definite, and $$\lambda_{min}(A) + \lambda_{min}(B) \leq 2\varepsilon,$$ which can be arbitrarily close to $0$.


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.