For a strictly positive collection of weights $\{w_{i}\}$, consider the weighted $l_{1}$ vector norm: $$ ||x||_{W} = \sum_{i}^{N} w_{i}|x_{i}| $$ What is (or more accurately, how would you compute) the matrix norm induced by this vector norm? That is, what is $||A||_{W}$? (You may assume $A$ is square so that the weights are well-defined.)

Note that, when $w_{i} = 1$ for all $i$, you get the standard result $$||A||_{1} = \max_{j} \sum_{i=1}^{N} |a_{ij}|.$$ For this reason, I feel like the correct answer should be $$||A||_{W} = \max_{j} \sum_{i=1}^{N} w_{i}|a_{ij}|.$$ But I cannot arrive at this result using any of the standard tricks.


Let $W$ be the diagonal matrix of weights. Notice then that, $$ \|x\|_W = \sum_{i=1}^{N} w_i |x_i| = \sum_{i=1}^{N} |w_ix_i| = \|Wx\|_1 $$

By definition, $$ \|A\|_W = \sup_{\|x\|_W = 1} \|Ax\|_W $$

Thus, letting $y=Wx$ (so that $x = W^{-1}y$), $$ \|A\|_W = \sup_{\|Wx\|_1 = 1} \|WAx\| = \sup_{\|y\|_1 = 1} \|WAW^{-1}y\|_1 = \| WAW^{-1} \|_1 $$

Now, $$ [WAW^{-1}]_{i,j} = (w_i/w_j) A_{i,j} $$ so, $$ \|WAW^{-1}\|_1 = \max_j \sum_{i=1}^{N} \frac{w_i}{w_j} A_{i,j} = \max_j \frac{1}{w_j} \sum_{i=1}^{N} w_i A_{i,j} $$


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.