3

Consider

$$T\colon\ell^2\to\ell^2, (s_1,s_2,\ldots)\mapsto (s_2,s_3,\ldots)$$

$$S\colon\ell^2\to\ell^2, (s_n)\mapsto (s_n/n)$$

$$R:=TS.$$

1) Show that $R$ is a compact operator.

2) Calculate the Singular Value Decomposition of $R$.

EDIT: My result concerning 2.)

The singular value decomposition $$ \left\{(\sigma_j),(u_j),(v_j)\right\} $$ of $R$ is given by $$ \sigma_j=\frac{1}{j+1},~~~~~v_j=e_j, j\geq 1,~~~~~u_j=e_j, j\geq 2, $$ where $\left\{u_j\right\}$ is an Orthonormalsystem of $\ell^2$ and $\left\{v_j\right\}$ is an Orthonormalsystem of $\ell^2$.

Is that right? Is that the SVD of $R$?

1 Answers1

1

No, this isn't correct.

In your fourth paragraph, you chose an arbitrary sequence $(s_n)$ from $\ell_2$. It need not be bounded (limited). Your argument just shows that each $s_n$ is in $\ell_2$. The constant $C$ you found depends of the particular $s_n$ that you are looking at. This error is easily remedied: you simply assume at the start that $(x_n)$ is a bounded sequence in $\ell_2$.

A more egregious error occurs later. The vector $u$ that you constructed is not a subsequence of $(R(s_n))$; it's just an element of $\ell_2$. So, it seems you tried to identify a candidate for a subsequential limit point; but, I don't think you can show that some subsequence of $(R(s_n))$ converges $u$. You need to be more careful with your diagonalization.

But, if you want to use a diagonalization argument, you can do the following:

Assume $(s_n)$ is a bounded sequence in $\ell_2$. If we can show that $(R(s_n))$ has a convergent subsequence, it will follow that $R$ is compact.

Towards this end, set $y_n=R(s_n)$. Then "diagonalize" $(y_n)$ as follows:

Choose a subsequence $(y^1_n)$ of $(y_n)$ so that its first coordinates form a convergent sequence in $\Bbb R$: $\lim\limits_{n\rightarrow\infty} y^1_n(1)=z_1$ for some number $z_1$.

Then choose a subsequence $(y^2_n)$ of $(y^1_n)$ so that its second coordinates form a convergent sequence in $\Bbb R$: $\lim\limits_{n\rightarrow\infty} y^2_n(2)=z_2$ for some number $z_2$.

Continue...

Now set $x_n=y_n^n$. Then $(x_n)$ is a subsequence of $(y_n)$ such that for each $i$ we have $\lim\limits_{n\rightarrow\infty} x_n(i)=z_i$.

Now show that $z\in \ell_2$ and that $\Vert x_n-z\Vert\rightarrow 0$. (Alternatively, you could just ignore $z$ and show that $(x_n)$ is Cauchy in $\ell_2$.)


I'll remark that the argument can be simplified a tad by noting that, since $T$ is bounded and since the composition of a bounded operator with a compact operator is compact, it suffices to show that $S$ is compact.

Showing that $S$ is compact can be done without diagonalizing explicitly by, for instance, showing that $S(B(\ell_1))$ is totally bounded. Or, by showing that $S$ is the norm limit of finite rank operators (see, this post, for example).

This post will also prove helpful.

Community
  • 1
David Mitra
  • 74,748
  • Thanks. And how can I calculate the SVD? –  Feb 02 '13 at 19:48
  • Finding the adjoint operator $R^$ and and then finding the Eigenvalues of $RR^$? (resp. the Eigensequences?) –  Feb 02 '13 at 20:14
  • @math12 You're welcome. I'm not familiar with SVD; so I'm afraid I can't help much with that. But, I will think about it. – David Mitra Feb 02 '13 at 20:19
  • Is the adjoint operator of $R$ given by $R'\colon \ell^2\to \ell^2, (x_1,x_2,...)\mapsto \left(0,\frac{x_1}{2},\frac{x_2}{3},\frac{x_3}{4},...\right)$? –  Feb 03 '13 at 11:21
  • @math12 That looks correct. You should verify that $(Rx,y)=(x,R^*y)$ for all $x,y$. – David Mitra Feb 03 '13 at 12:35
  • I calculated it in another way. Okay, the result is correct. Now I have to calculate the eigenvalues of $RR'$ to get the singular values of $R$? –  Feb 03 '13 at 13:35
  • @math12 I'm sorry, I'm not familiar with SVD's. But, if you want to find the eigenvalues of $RR^$, note that $RR^$ is just the multiplication operator $(x_1,x_2,\ldots)\rightarrow(x_1/4, x_2/9, x_3/16,\ldots)$. So the eigenvalues are $1/4,1/9,\ldots$ and the unit vectors are the corresponding eigenvalues. – David Mitra Feb 03 '13 at 13:59
  • Can anybody say me if my calculation is correct? To my calculation the SVD $\left{(\sigma_j),(u_j),(v_j)\right}$ of $R$ is given by:

    $\sigma_j=\frac{1}{j+1}$,

    $v_j=e_j, j\geq 1$ and

    $u_j=e_j, j\geq 2$. Is that correct?

     –  Feb 03 '13 at 20:06
  • @math12 You might want to add your last comment to your original post. That way, the thread will be booted to the main page and people will see it. – David Mitra Feb 03 '13 at 21:29
  • You mean I should edit my opening post and put my result in there? –  Feb 03 '13 at 21:31
  • @math12 Yes. No one is likely to see it in these comments. – David Mitra Feb 03 '13 at 21:31
  • Okay! Thank you for that advice. I will do so. –  Feb 03 '13 at 21:32