4

I am trying to work out question 8.17 from the book 'Functional Analysis: an elementary introduction' by Haase. The question is formulated as follows. Let $(e_j)_{j\geq 1}$ be an orthonormal system in a Hilbert space $H$, and let $a_j \geq 0$ be scalars with $\sum_{j=1}^\infty a_j^2 < \infty$. Show that the set \begin{align} C := \left\{\left.\sum_{j=1}^\infty\lambda_j e_j\right|\, \lambda_j \in \mathbb{K},\,|\lambda_j| \leq a_j\right\} \end{align} is compact in $H$.

This is my attempted solution: I would go with the definition of sequential compactness here. My plan was either to construct a pairwise orthogonal sequence. Then, showing that $ \sum_{j=1}^\infty \|f_n\|^2 < \infty$ would imply that $\sum_j f_n$ converges in H. I thought about using Bessel's inequality, which turns out here to be an equality, since for some $f\in C$ we have \begin{align} \langle f, e_k\rangle = \langle\sum_{j=1}^\infty \lambda_j e_j, e_k\rangle = \sum_{j=1}^\infty \lambda_j \langle e_j, e_k\rangle = \lambda_k. \end{align} Thus, we can write $f$ as follows \begin{align} f = \sum_{j=1}^\infty\lambda_j e_j = \sum_{j=1}^\infty\langle f, e_j\rangle e_j = Pf, \end{align} the standard abstract Fourier series. From this, we can deduce that \begin{align} \|f\|^2 = \sum_{j=1}^\infty |\langle f,e_j\rangle |^2 = \sum_{j=1}^\infty |\lambda_j|^2 \left(\leq \sum_{j=1}^\infty a_j^2 < \infty\right), \end{align} using Pythagoras.
Now, To show that $C$ is compact we have to show that every sequence in $C$ has a converging subsequence. Let $(f_n)_{n\geq 1}$ be a sequence in $C$. Let $(f_{n_k})_{k\geq 1}$ be a subsequence, we need to show that $(f_{n_k})_{k\geq 1}$ converges to some $f$ in $C$.
Each $f_n$ is of the form \begin{align} f_n = \sum_{j=1}^\infty \lambda_{n,j} e_{n, j}, \end{align} the coefficients, $\lambda_{n, \cdot}$, depend on $n$, and so does the 'positions', $e_{n, \cdot}$, where we place these coefficients. The sequence $(f_n)_{n\geq 1}$ is not pairwise orthogonal. \begin{align} \langle f_n, f_m\rangle = \left\langle \sum_{j=1}^\infty \lambda_{n, j}e_{n,j}, \sum_{k=1}^\infty \lambda_{m, k}e_{m,k}\right\rangle = \sum_{j=1}^\infty\sum_{k=1}^\infty \lambda_{n,j}\lambda_{m, k} \langle e_{n,j}, e_{m,k}\rangle \end{align} we can say nothing about the term $\langle e_{n,j}, e_{m,k}\rangle$. We could also try to show that this set $C$ is closed in some other compact set $D$. But then we would have to construct some set $D$, and I do not really get far with this approach either.
Any help is much appreciated!

UpzYaDead
  • 117

1 Answers1

1

Proving compactness by showing that every sequence has a convergent subsequence is a nice strategy, and I think you can make this work. Can I suggest going about it using a diagonal argument?

Suppose $(f_1, f_2, f_3, \dots)$ is a sequence in $C$, and write each $f_n$ in the form $f_n = \sum_{j = 1}^\infty \lambda_{n, j} e_j$.

  1. Using the compactness of the closed ball $\overline{B(0, a_1)} \subset \mathbb K$, one can show that there exist positive integers $n^{(1)}_1 < n^{(1)}_2 < n^{(1)}_3 < \dots$ such that $(\lambda_{n^{(1)}_1, 1}, \lambda_{n^{(1)}_2, 1}, \lambda_{n^{(1)}_3, 1}, \dots)$ is a convergent sequence.
  2. Hence, using the compactness of the closed ball $\overline{B(0, a_2)} \subset \mathbb K$, one can show that there exist positive integers $n^{(2)}_1 < n^{(2)}_2 < n^{(2)}_3 < \dots$, chosen from the set $\{ n^{(1)}_1 , n^{(1)}_2 , n^{(1)}_3 , \dots \}$, such that $(\lambda_{n^{(2)}_1, 1}, \lambda_{n^{(2)}_2, 1}, \lambda_{n^{(2)}_3, 1}, \dots)$ and $(\lambda_{n^{(2)}_1, 2}, \lambda_{n^{(2)}_2, 2}, \lambda_{n^{(2)}_3, 2}, \dots)$ are both convergent sequences.

And so on.

Having done all of this, one can conclude that $(\lambda_{n^{(1)}_1, j}, \lambda_{n^{(2)}_2, j}, \lambda_{n^{(3)}_3, j}, \dots)$ is a convergent sequence, for any $j$.


To summarise, the above argument shows that:

There exists a subsequence $(f_{n_k})_{k \in \mathbb N}$ of $(f_n)_{n\in \mathbb N}$ such that the sequence $(\lambda_{n_k, j})_{k\in\mathbb N}$ converges, for any $j$.

(Take $n_k := n^{(k)}_k$.)

The final step is to show that $(f_{n_k})_{k \in \mathbb N}$ is itself convergent (as a sequence in the Hilbert space $H$).

This is where your ideas about Pythagoras' theorem come in. Here's the rough idea:

For each $j$, define $\lambda_j := \lim_{k \to \infty} \lambda_{n_{k}, j}$. Also define $f := \sum_{j = 1}^\infty \lambda_j e_j$. (Here, you should convince yourself that this infinite series is well defined, and that this $f$ is in fact ab element of $C$.) The goal is to prove that $(f_{n_k})_{k \in \mathbb N}$ converges to $f$ as a sequence in $H$.

So pick an $\epsilon > 0$. Since $\sum_{j = 1}^\infty |a_j|^2 < \infty$, there exists a $J$ such that $\sum_{j = J + 1}^\infty |a_j|^2 < \epsilon$.

Also, for each $j$, there exists a $K_j$ such that $k \geq K_j$ implies $(\lambda_{n_k, j} - \lambda_j)^2 < \epsilon$.

Thus one can show that if $k \geq \max \{ K_1, \dots K_J \}$, then $$ \left\| f_{n_k} - f \right\|^2 = \sum_{j = 1}^\infty (\lambda_{n_k, j} - \lambda_{j})^2 < J \epsilon + 4\epsilon.$$ (Here, Pythagoras' theorem is used.)

Feel free to leave a comment if you have questions about filling in the details.

Kenny Wong
  • 33,403
  • Thank you very much! I do not fully understand the first part of your construction. I think it is similar as done here, as suggested by @J de Ro. The second part is very clear! Thanks for the quick answer :) – UpzYaDead Nov 16 '23 at 11:20
  • Just a quick question, shouldn't the term $2\varepsilon$ be $4\varepsilon$ instead? – UpzYaDead Nov 16 '23 at 11:34
  • @UpzYaDead The first part is a diagonal argument, and it's the same as in the link you posted. Have you seen diagonal arguments before? Having looked at other answers, do you understand the diagonal argument now? For the second part, I think I do mean $2\epsilon$ not $4\epsilon$. The idea is that $\sum_{j = J + 1}^\infty (\lambda_{n_k, j} - \lambda_j)^2 \leq \sum_{j = J + 1}^\infty \lambda_{n_k, j} ^2 + \sum_{j = J + 1}^\infty \lambda_j ^2 \leq 2\sum_{j = J + 1}^\infty a_j^2 < 2\epsilon$. – Kenny Wong Nov 16 '23 at 12:04
  • I haven't really seen diagonal arguments before. But I understand the concept now! The bound also makes sense now, thanks for writing it out. My $4\varepsilon$ came from: $ |l_j - \lambda^{(n_k)}_j| \leq |l_j| + |\lambda^{(n_k)}_j| \leq a_j + a_j = 2a_j \implies |l_j - \lambda^{(n_k)}_j|^2 \leq 4a_j^2$, taking the norm on the lambdas as they are complex valued. – UpzYaDead Nov 16 '23 at 22:23
  • Actually you were right about the $4$ rather than the $2$! I made a basic error! So sorry. – Kenny Wong Nov 17 '23 at 18:34