I am new here and this my first question, I hope I am being as clear as possible and apologize in advance for any misunderstandings.

I am researching the Arithmetic-Geometric Mean (AM-GM) inequality for matrices. Which can be seen as an "extension" of the obvious AM-GM inequality for positive numbers $a,b$ : $ \sqrt{ab} \leq \frac{1}{2}(a+b) $. One of such possible extensions is the following:

$|||A^{1/2}B^{1/2}|||\leq \frac{1}{2}|||A+B|||$ As shown in "Positive Definite Matrices" by R.Bhatia.

Where A, B are positive, hermitian matrices, and $||| \cdot |||$ is a unitarily invariant norm.

This is precisely my problem/question: all proofs and attempts to investigate the AM-GM matrix-norm-inequality use unitarily invariant norms. Does this mean that for non-unitarily invariant norms this inequality does not hold? Is there a counterexample for this? perhaps for p-norms such as $\|\cdot\|_\infty$ , $\|\cdot\|_1$? Or an explanation why a proof for non-unitarily invariant norms could not work?

I am just lost since all the texts I read regarding the AM-GM inequality, immediately jump to using unitarily invariant norms.

Thank you!!


Consider $$A = \begin{pmatrix} 46 & 7 & 8 \\ 7 & 36 & 16 \\ 8 & 16 & 30\end{pmatrix}\qquad \text{ and }\qquad B=\begin{pmatrix} 34 & 7 & 12 \\ 7 & 48 & 10 \\ 12 & 10 & 44\end{pmatrix}.$$ Then $$\frac{1}{2}\|A+B\|_1 =\frac{1}{2}\|A+B\|_{\infty}=62$$ however $$ \|A^{1/2}B^{1/2}\|_1\approx 62.1527 \qquad \text{ and }\qquad \|A^{1/2}B^{1/2}\|_{\infty}\approx 62.1037.$$

Here is the Matlab code to generate such counter examples:

n = 3;  % choose the dimension of the matrices
stop = 0;  
while stop == 0  
  % Generates a pair of symmetric positive definite matrices
  A = floor(10*rand(n));  
  A = A+10*n*eye(n);  
  B = floor(10*rand(n));  
  B = B+B';  
  B = B+10*n*eye(n);  
  % Computes their square root
  AS = A^(1/2);  
  % Check if the AM-GM is not satisfied
  if norm(AS*BS,1)>norm(A+B,1)/2 && norm(AS*BS,inf)>norm(A+B,inf)/2  
    stop =1;  

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.