Let $ M = \left( \begin{array}{cc} A & B \\ B^T & C \end{array} \right)$ be a real symmetric matrix ($A, C$ are symmetric matrices). Let $\lambda_m$ and $\lambda_M$ denote the minimum and maximum eigenvalues of $M$, and $\mu_A$ and $\mu_C$ denote, respectively, the maximum eigenvalues of $A$ and $C$.

Show that $\lambda_m+\lambda_M \leq \mu_A + \mu_C$.

I've just got a hint from others: consider the positive semidefinite matrix $ M - \lambda_m I$. Maybe there are other methods.

  • $\begingroup$ Are you familiar with the Schur complement formula? $\endgroup$ – percusse Nov 23 '17 at 16:51

Here I follow the hint and provide an answer. Let's first assume that $ M $ is positive semi-definite and show that $\lambda_M \leq \lambda_A + \lambda_B $.

Let $z=\left[\begin{array}{c}x\\y\end{array}\right]\in \mathbb{R}^n$, $x\in \mathbb{R}^s, y\in \mathbb{R}^t$, $s, t$ are the order of matrices $A$ and $C$.

Take a unit eigenvector $z_M=(x_M, y_M)$ of $M$ associated with $\lambda_M$. Then we have $$\lambda_M=z_M^T Mz_M=x_M^T A x_M + 2x_M^T B y_M + y_M^T Cy_M.$$ Note that $A$ and $C$ are now also positive semi-definite, implying that $\lambda_A\geq 0$ and $\lambda_C\geq 0$.

If $x_M=0$, then we have $y_M^Ty_M=1$ and $\lambda_M \leq y_M^T CY_M \leq \lambda_C y_M^Ty_M = \lambda_C \leq \lambda_A+\lambda_C$.

If $y_M=0$, then we have $x_M^T x_M=1$ and $\lambda_M \leq x_M^T A x_M \leq \lambda_A x_M^Tx_M = \lambda_A \leq \lambda_A+\lambda_C$.

Now we assume that $x_M\neq 0, y_M\neq 0$. Consider $z(s,t)=\left[\begin{array}{c}sx_M\\ty_M\end{array}\right]\in \mathbb{R}^n$.

By the assumption of positive semi-definteness, we have $$0\leq s^2 x_M^T A x_M +2st x_M^TBy_M+t^2 y_M^T C y_M.$$ Combining these two yields $$\lambda_M \leq (1+s^2) x_M^T A x_M +2(st+1) x_M^TBy_M+ (1+t^2) y_M^T C y_M.$$ Take $s=\sqrt{\dfrac{y_M^Ty_M}{x_M^T x_M}}$ and $t=-\sqrt{\dfrac{x_M^T x_M}{y_M^Ty_M}}$. Thus, $st=-1$, eliminating the cross terms, $$\lambda_M \leq \frac{x_M^T A x_M}{x_M^T x_M} + \frac{y_M^T C y_M}{y_M^T y_M}\leq \lambda_A+\lambda_C.$$ In general, $M$ may not be positive semi-definite, and we consider $M'=M-\lambda_m I$. Here we have $$\lambda_M'=\lambda_M-\lambda_m,\quad\lambda_A'=\lambda_A-\lambda_m, \quad \lambda_C'=\lambda_C-\lambda_m.$$ Applying the above proven proposition to $M'$, we have $\lambda_M'\leq \lambda_A'+\lambda_C'$, that is $$\lambda_M-\lambda_m\leq (\lambda_A-\lambda_m)+(\lambda_C-\lambda_m)\Rightarrow \lambda_M +\lambda_m \leq \lambda_A +\lambda_C.$$


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.