1
$\begingroup$

Given a bounded operator on a Hilbert space $A \in \mathcal{B}(H)$, I am familiar with the fact that the Hilbert-Schmidt norm of an operator is lesser or equal to the trace class norm, i.e. $\|A\|_{\mathcal{HS}} \leq \|A\|_{\mathcal{TC}}$.

While working with weighted $\ell_2$ spaces, I was faced with an apparent contradiction of this fact. Let me be clear : I have total confidence in the inequality $\|A\|_{\mathcal{HS}} \leq \|A\|_{\mathcal{TC}}$. I believe instead that I've made a logical or computational mistake somewhere. However, I'm unable to spot it.

Consider a weight function $w: \mathbb{N} \rightarrow (0,\infty)$ and let $H$ be the Hilbert space $\ell_2(\mathbb{Z} / m \mathbb{Z})$ endowed with the weighted euclidean norm $$ \|f\|_w :=~ \sqrt{\sum\limits_{k =1}^m w(k)^2 |f(k)|^2}.$$

Let's denote by $\langle \cdot, \cdot \rangle_w$ the associated inner product.

If $\{e_n\}_{n \in \mathbb{N}}$ is the standard orthonormal basis of $\ell_2(\mathbb{Z} / m \mathbb{Z})$, then $\{u_n\}_{n \in \mathbb{N}}$ is an orthonormal basis of $H$ where $u_n := \frac{e_n}{\|e_n\|_w} =\frac{e_n}{w(n)}$.

Now consider $A \in \mathcal{B}(H)$ a bounded operator on $H$.

  • The Hilbert-Schmidt norm of $A$ is defined by $\sqrt{\sum\limits_{k=1}^m \big\|A(u_k)\big\|_w^2}$.
  • The Trace Class norm of $A$ is defined by $\sum\limits_{k=1}^m \Big\langle (A^*A)^{\frac{1}{2}}(u_k),u_k\Big\rangle_w$, where $A^*$ is the adjoint operator of $A$ and $(A^*A)^{\frac{1}{2}}$ is a square root of the positive operator $A^*A$.

Now consider the operator $A:= \langle \cdot, e_i\rangle_w e_j$ for some $1 \leq i, j \leq m$.

I've computed that $\|A\|_{\mathcal{TC}} = w(i)^2$ whereas $\|A\|_{\mathcal{HS}} = w(i)w(j)$ but this is impossible because for $i \neq j$ we could find values of $w(i)$ and $w(j)$ for which $\|A\|_{\mathcal{HS}} > \|A\|_{\mathcal{TC}}$.

Hence there most be something wrong with my computations of the norms of the operator $A:= \langle \cdot, e_i\rangle_w e_j$. The thing is : when I look at my computations, I can't identity when I have sinned.

Trace Class norm : Observe first that $A$ can be identified with the $m$ by $m$ matrix whose entries are all zero except for the $(j,i)$-th which is $w(i)^2$. Hence $A^*A$ can be identified with the $m$ by $m$ matrix whose entries are all zero except for the $(i,i)$-th which is $w(i)^4$. Hence $(A^*A)^{\frac{1}{2}}$ can be identified with the $m$ by $m$ matrix whose entries are all zero except for the $(i,i)$-th which is $w(i)^2$. Therefore $$(A^*A)^{\frac{1}{2}}(u_k) = \begin{cases}w(i)e_i~~~ \text{if $k = i$}\\ 0~~~~~~~ \text{otherwise} \end{cases}$$ Hence $$\|A\|_{\mathcal{TC}} = \sum\limits_{k=1}^m \Big\langle (T^*T)^{\frac{1}{2}} u_k, u_k \Big\rangle_w = \Big\langle (T^*T)^{\frac{1}{2}} u_i, u_i \Big\rangle_w = \big\langle w(i) e_i, u_i \big\rangle_w = \|e_i\|_w^2 = w(i)^2.$$

Hilbert-Schmidt norm : $$\Big\|\langle \cdot, e_i\rangle_w e_j\Big\|_{\mathcal{HS}} = \left(\sum\limits_{k=1}^m \big\|\langle u_k, e_i\rangle_w e_j\big\|_w^2\right)^{\frac{1}{2}} =\Big(w(i)^2 w(j)^2\Big)^{\frac{1}{2}} = w(i)w(j)$$ since $\langle u_k, e_i\rangle_w = \left\langle \frac{e_k}{w(k)}, e_i\right\rangle = \begin{cases}w(i)~~~ \text{if $k = i$}\\ 0~~~~~~~ \text{otherwise} \end{cases}$.

$\endgroup$

1 Answer 1

0
+50
$\begingroup$

Your matrix elements for the adjoint are not correct. You are seemingly calculating the matrix elements for $A$ in the basis $(e_i)_i$ which is not orthonormal. Using your scalar product it should then be $w_j^2$ for the adjoint in the $(i,j)$-th place. One has: $$ \langle A^* e_m , e_k \rangle = \langle e_m,A e_k\rangle = w_m^2 w_k^2 \delta_{m,j}\delta_{k,i}$$

The symmetry becomes perhaps even more apparent if you work in the basis $(u_i)_i$ of unit vectors for your norm. Write: $$ A = A_{i,j} = w_i w_j \langle \cdot, u_i \rangle u_j$$ and the adjoint is simply $A^* = A_{j,i}$. Thus $A^* A = w_i^2 w_j^2 \langle \cdot, u_i \rangle u_i$ from which you get that the two norms are identical. This is also as one should expect since $A$ has rank 1.

$\endgroup$

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.