I'm a graduate student doing research on time-frequency analysis. I am considering the existence of a certain frame-like inequality. Let $H: L^2(\mathbb{R}^d) \rightarrow L^2(\mathbb{R}^d)$ be a compact self-adjoint operator, and for each $(x,\omega)\in \mathbb{R}^{d} \times \mathbb{R}^d,$ let $\pi(x,\omega): L^2(\mathbb{R}) \rightarrow L^2(\mathbb{R}^{d})$ be the time-frequency shift operator $\pi(x,\omega)f =f(t-x) e^{2\pi i \langle \omega,t-x \rangle}$. Fix a window function $g\in L^2(\mathbb{R}^{d}),$ and suppose there exist finite points $\{(x_n,\omega_{n})\}_{n=1}^q$ such that there are $A,B>0$ where for all $f\in L^2(\mathbb{R}^{d})$ \begin{equation} A||Hf||_2^2 \leq \sum_{n=1}^q |\langle Hf, \pi(x_n,\omega_n)g \rangle|^2 \leq B||Hf||_2^2. \end{equation} You might recognize that the inequality above looks like the Gabor Frame inequality except we only have finite 'atoms' and we have a compact self-adjoint operator inside. Let us review the spectral theorem for compact operators:

Let $H$ be a Hilbert space and $T \in K(H)$ where $K(H)$ denotes the set of compact operators from $H$ to $H$. Let $T$ be normal or self adjoint depending on whether the underlying field is $\mathbb R$ or $\mathbb C$, then there exists an orthonormal set $\{e_i \mathrel| i\in I\}$ where $I$ is either $\mathbb N$ or $\{1,2,\dotsc,k\}$ and a sequence $(\lambda_i)_{i\in I} \in \mathbb K$ which converges to $0$ such that

$$\operatorname{span}{(x_i : i ∈ I})^{⊥} = \ker(T)$$

and also $\forall x \in H$ : $Tx= \sum_{i\in I} \lambda_i \langle x, e_i \rangle e_i$ with unconditional convergence.

Now denote the eigenvalue-eigenfunction pairs of $H$ by $(\lambda_i, e_i)_{i \in \mathcal{I}}$. For each $N\in \mathbb{N}$ let $V_N = \operatorname{span}\{e_1,e_2,...,e_N \} \subset L^2(\mathbb{R}^d)$, and $P_N$ be the orthogonal projection operator from $L^2(\mathbb{R}^d)$ onto $V_N.$ Since orthogonal projections are self-adjoint and idempotent, then for all $f\in L^2(\mathbb{R}^d)$ the inequality above implies: \begin{equation} A||P_NHf||_2^2 \leq \sum_{n=1}^q |\langle P_NHf, P_N\pi(x_n,\omega_n)g \rangle|^2 \leq B||P_NHf||_2^2. \end{equation} The above is also due to the fact that $H$ commutes with $P_N,$ which should not be hard to prove. From frame theory, the inequality above implies that for all $N\in\mathbb{N},$ the finite set $\{P_N\pi(x_n,\omega_n)g \}_{n=1}^q$ is a spanning set of $\operatorname{Im}P_NH,$ and therefore $\sup_{N\in \mathbb{N}}(\operatorname{dim}\operatorname{Im}P_NH) = \sup_{N\in \mathbb{N}}(\operatorname{rank}P_NH) \leq q < \infty$ It should not also be that hard to show that $V_N = \operatorname{Im}HP_N.$ So finally we have $\sup_{N\in \mathbb{N}}\operatorname{dim}V_N \leq q.$ Am I right in thinking that this should imply that $H$ should only have a finite spectrum (and hence finite rank)?

Does this mean that our starting inequality (provided it is possible to construct one) is only valid for compact self-adjoint operators whose spectrum is finite?


Yes, your starting inequality can only hold if $H$ has finite dimensional range. In fact, the dimension can be at most $q$.

To see this, let $Y$ denote the range of $H$. Your inequality implies (why?) that if $y \in Y$ with $\langle y, \pi(x_n, \omega_n) g\rangle =0$, then $y =0$. In other words, the linear map $$ Y \to \Bbb{C}^q, y \mapsto (\langle y, \pi(x_n, \omega_n) g\rangle )_{i=1,\dots,q} $$ is injective. By elementary linear algebra, this means that $Y$ has dimension at most $q$.

Note that we did not use any properties of the operator $H$, except that its range is a linear space and that it satisfies your inequality.

Finally, it is not too hard to see that a self adjoint operator with finite rank has finite spectrum.

  • $\begingroup$ Your mapping is injective since norm 'norm' equivalences are also equivalence relations (i.e we can swap the inequality above so that $A^{\prime} \sum |\langle Hf, \pi(x_n,\omega_n)g \rangle|^2 \leq ||Hf||_2^2 \leq B^{\prime} \sum |\langle Hf, \pi(x_n,\omega_n)g \rangle|^2 $). So your hypothesis would imply $Hf = 0$ from the axioms of a normed linear space. Anyway, did you use the rank-nullity theorem for your conclusion? It's kinda confusing for me since AFAIK it can only be used if we know a priori that our mapping's domain is finite dimensional, which is kind of circular. $\endgroup$ – Kurome Jun 25 '18 at 2:00
  • $\begingroup$ @Kurmo: Let's call the map from my proof $\Phi$. Since it is injective, I we denote by $Z$ the range of $\Phi$, then $\Phi : Y \to Z$ is an isomorphism. Since $Z$ is finite dimensional as a subspace of the finite dimensional space $\Bbb{C}^q$, also $Y$ is finite dimensional [if $z_1,...,z_N$ is a basis for $Z$, then $\Phi^{-1}(z_1),...,\Phi^{-1}(z_N)$ is a basis for $Y$]. $\endgroup$ – PhoemueX Jun 25 '18 at 5:30
  • $\begingroup$ I see, I think we can also simply argue that: $Y$ cannot be isomorphic to a space of finite dimension if it is of infinite dimension. Thank you, that's illuminating. $\endgroup$ – Kurome Jun 25 '18 at 6:03

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.