Let $ \mathcal{H} $ be a Hilbert space and $ T: \mathcal{H} \to \mathcal{H} $ a bounded linear operator. The $ n $-th singular number $ {\mu_{n}}(T) $ of $ T $ is defined as the distance from $ T $ to the space of operators of rank at most $ n $. We say that $ T $ is in the trace class if $$ \sum_{n} {\mu_{n}}(T) < \infty. $$ Show that in this case, $ T $ is compact and if $ \{ \lambda_{n} \} $ are its eigenvalues, then $$ \sum_{n} |\lambda_{n}| < \infty. $$ Also, show that, in general, the converse is not true.

I have not seen this definition of ‘trace class’ before. Can anyone give me some hints? Can I approximate $ T $ with finite-rank operators?

  • 4
    $\begingroup$ My "guess" is: because your sum is finite, you must have $\mu_n(T)\rightarrow 0$, hence I think you can use this fact to produce the desired aproximation. $\endgroup$ – Tomás Jan 3 '13 at 22:33
  • $\begingroup$ To disprove the converse, take a diagonal operator on $\ell_2$ with diagonal entries $\lambda_n$ that converge to zero, but slowly enough so that $\sum \lambda_n=\infty$. $\endgroup$ – user53153 Jan 4 '13 at 1:31
  • 1
    $\begingroup$ @PavelM I dont see this contradiction... isnt it $\sum|\lambda_n| < \infty$ in the converse statement? $\endgroup$ – Johan Jan 4 '13 at 21:41
  • $\begingroup$ Indeed, I misunderstood what was meant by the converse here. Maybe look at some form of Volterra operator which is compact and has no eigenvalues at all. $\endgroup$ – user53153 Jan 4 '13 at 22:36
  • $\begingroup$ The operator $Te_i=\frac{e_{i+1}}{i}$, where $\{e_i\}$ is a orthonormal sequence seems to work fine as a counter example? $\endgroup$ – Jonas Wallin Jan 5 '13 at 0:41

In order to distinguish the new definition of $ {\mu_{n}}(T) $ from the old one, let us call it $ {\mu^{\text{New}}_{n}}(T) $.

We shall assume throughout this discussion that $ \displaystyle \sum_{n=1}^{\infty} {\mu_{n}}(T) < \infty $.

By the Divergence Test from calculus (it’s hard to believe that something so simple can crop up here!), we have $ \displaystyle \lim_{n \to \infty} {\mu_{n}}(T) = 0 $. Hence, for any $ \epsilon > 0 $, there exists an $ n \in \mathbb{N} $ sufficiently large so that $ {\mu_{n}}(T) < \epsilon $, which means that we can find an $ F \in B(\mathcal{H}) $ of rank $ \leq n $ such that $ \| T - F \|_{B(\mathcal{H})} < \epsilon $. Therefore, $ T $ can be approximated in the operator norm by bounded operators of finite rank, making it a compact operator.

Recall that $ {\mu_{n}}(T) $ is defined as the $ n $-th term of the null sequence that is formed by listing the eigenvalues of the positive compact operator $ |T| $ in decreasing order, taking multiplicity into account. The Minimax Principle then says that $ {\mu^{\text{New}}_{n}}(T) = {\mu_{n}}(T) $ (please click here to access a set of notes on trace-class operators that contains a proof of this result; see Lemma 12). Therefore, ‘trace-class’ in the new sense is the same as ‘trace-class’ in the old sense, and so $$ \text{Tr}(|T|) \stackrel{\text{def}}{=} \sum_{n=1}^{\infty} {\mu_{n}}(T) = \sum_{n=1}^{\infty} {\mu^{\text{New}}_{n}}(T). $$

For any $ T \in K(\mathcal{H}) $, one can find orthonormal sequences $ (\mathbf{v}_{n})_{n \in \mathbb{N}} $ and $ (\mathbf{w}_{n})_{n \in \mathbb{N}} $, not necessarily complete, such that $$ T = \sum_{n=1}^{\infty} {\mu_{n}}(T) \langle \mathbf{v}_{n},\bullet \rangle_{\mathcal{H}} \cdot \mathbf{w}_{n}. $$ We thus obtain a more explicit approximation of $ T $ by bounded operators of finite rank. This is a standard result in the theory of compact operators; please refer to the Wikipedia article on Compact Operator or to Corollary 4 of the notes mentioned above.

  • $\begingroup$ Thanks! can you please expand a little on how this implies the second statement? $\endgroup$ – Johan Jan 4 '13 at 21:42
  • $\begingroup$ @Johan: I’ve done a revision of my original post. The links will explain some of the details that I haven’t managed to provide. $\endgroup$ – Haskell Curry Jan 29 '13 at 4:54
  • $\begingroup$ Great! thank you very much! $\endgroup$ – Johan Jan 29 '13 at 14:40
  • $\begingroup$ @Johan: One more thing. We already have $ \forall n \in \mathbb{N}: ~ |{\lambda_{n}}(T)| \leq {\mu_{n}}(T) $, so by the Comparison Test, the convergence of $ \displaystyle \sum_{n=1}^{\infty} {\mu_{n}}(T) $ yields the convergence of $ \displaystyle \sum_{n=1}^{\infty} |{\lambda_{n}}(T)| $. $\endgroup$ – Haskell Curry Jan 29 '13 at 22:11
  • $\begingroup$ @HaskellCurry: Why does $\mu_n(T)<\varepsilon$ imply that we can find $F\in\mathcal{B}(\mathcal{H})$ of rank $\leq n$ such that $\|T-F\|_{\mathcal{B}(\mathcal{H})}<\varepsilon$? Also why do we assume that $\sum_{n=1}^\infty\mu_n(T)<\infty$. $\endgroup$ – C-Star-W-Star Aug 15 '15 at 12:53

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.