We define weighting matrix $W$ as follows

$$W := \frac{1}{Md}\left[\sum_{m=1}^{M}x^{(m)}(x^{(m)})^{T}\right] - \frac 1d \Bbb I_d$$

where $\Bbb I_d$ is the $d \times d$ identity matrix and $x^{(m)} \in \{\pm 1\}^d$. Why is $\| W \|_2 \leq 1$?

For reference, the claim that $\| W \|_2 \leq 1$ is made on page 4 here, on the bottom left.


Consider the matrix $ A=\frac{1}{M}\sum\limits_{m=1}^{M}x^{(m)}(x^{(m)})^{T}: $ $$ |A_{i,j}| = \left|\frac{1}{M}\sum\limits_{m=1}^{M}x_i^{(m)}x^{(m)}_j\right| \leq 1. $$ Moreover, diagonal elements $A_{i,i} = \frac{1}{M}\sum\limits_{m=1}^{M}\left(x_i^{(m)}\right)^2 = 1$. Let's $B = A - \Bbb I_d$, then $$|B_{i,j}| = |(A-\Bbb I_d)_{i,j}| \leq 1$$

Now, for $W = \frac{1}{d}B$ and arbitrary vector $x$ we have $$ \|Wx\| = \frac{1}{d}\|Bx\| = \frac{1}{d}\sqrt{\sum_i\left(\sum_jB_{i,j}x_j\right)^2} \leq \frac{1}{d}\sqrt{\sum_i\left(\sum_jB_{i,j}^2\right)\left(\sum_jx_j^2\right)} \leq \\ \le \frac{1}{d}\sqrt{d^2 \|x\|^2} = \|x\|, $$ which means $\|W\| \leq 1$.


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.