Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.


A is an $n\times n$ complex matrix. Let $\lambda _\max(Y)$ and $\sigma _\max(Y)$ denote respectively the largest eigenvalue and the largest singular value of a square matrix $Y$. The question is to prove that: $$\lambda _\max\left(\frac{A+A^*}{2}\right)\leq \sigma _\max (A)$$

I started as follows: Since $A$ is Hermitian, then $(A+A^*)$ is Hermitian as well. Then: $$\lambda_\max\left(\frac{A+A^*}{2}\right)=\max_{\left \| x \right \|=1}x^*\left(\frac{A+A^*}{2}\right)x$$ where $x\in \mathbb{C}^n$

I need to prove the following to finish my proof: $$\max_{\left \| x \right \|=1}x^*\left(\frac{A+A^*}{2}\right)x\geq \max_{\left \| x \right \|=1}(x^*A^*Ax)^{\frac{1}{2}}\tag{1}$$

From the first sight, it looks like I have to use the fact that: $\sqrt{ab}\geq \frac{a+b}{2}$ for $a$ and $b$ positive, but this works only for numbers and not matrices.

I appreciate if someone can prove the inequality $(1)$.

share|improve this question
Are you sure you need to require that $A$ is hermitian? Because in that case, $(A+A^*)/2=A$. –  Martin Argerami Apr 23 '12 at 2:04
And you also need an absolute value in the left-hand side (inside the max); otherwise the inequality is trivially false with $A=-1$. –  Martin Argerami Apr 23 '12 at 2:12
@Martin Argerami: You're right. $A$ is not Hermitian. I will edit the original statement. –  Boyan Klo Apr 23 '12 at 2:18
If $A$ is allowed to be non-hermitian, the inequality is again false in general. For example, if $A^*=-A$, then the left-hand side is zero, while the right-hand side is not. –  Martin Argerami Apr 23 '12 at 3:03

1 Answer 1

up vote 3 down vote accepted

Your inequality is actually a well-known equality (in the case where $A$ is hermitian).

The right-hand side of your inequality is the operator norm of $A$, i.e. $$ \max_{\left \| x \right \|=1}(x^*A^*Ax)^{\frac{1}{2}}=\|A\|. $$ For the left-hand side we get, using the Cauchy-Schwarz inequality, $$ \max_{\left \| x \right \|=1}\left|x^*(\frac{A+A^*}{2})x\right|\leq\max_{\left \| x \right \|=1}\|x\|\,\left\|\frac{A+A^*}2\right\|\,\|x\|=\left\|\frac{A+A^*}2\right\|\\ \leq\frac{\|A\|+\|A^*\|}2=\frac{\|A\|+\|A\|}2=\|A\|. $$ Note that so far we don't use that $A$ is hermitian. When $A$ is hermitian, however, we can prove the reverse inequality, which requires a little more work.

Let $$ s=\max\{|x^*Ax|:\ \|x\|=1\} $$ Now take vectors $x,y$ with $\|x\|=\|y\|=1$. Then $$ (x\pm y)^*A(x\pm y)=x^*Ax+y^*Ay\pm2\text{Re}\,y^*Ax. $$ Subtracting one equation from the other we get $$ 4\text{Re}\,y^*Ax=(x+y)^*A(x+y)-(x-y)^*A(x-y) $$ (this is the real part of what is usually called the polarization identity). Now we get $$ 4\text{Re}\,y^*Ax\leq s\,\|x+y\|^2+s\,\|x-y\|^2=2s\,(\|x\|^2+\|y\|^2)=2s. $$ Fix $t$ such that $y^*Ax=e^{it}\,|y^*Ax|$. Note that we can do the above process for the vector $e^{-it}x$, and so this shows that $$ |y^*Ax|\leq s\ \ \ \ \text{ for any }x,y\text{ with }\|x\|=\|y\|=1. $$ Taking max over $y$ with $\|y\|=1$, we obtain $$ \|Ax\|\leq s \ \ \ \text{ for all } x \text{ with } \|x\|=1, $$ and thus $$ \|A\|\leq s. $$

share|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.