Coarse geometric methods for generalized wavelet approximation theory

H. Führ

Abstract: Generalized wavelet systems, as studied in this talk, arise from the translation action on a family of mother wavelets, chosen to ensure perfect reconstruction. This rather general notion comprises Gabor systems (including the multi-window variety), generalized wavelet systems arising from the action of matrix groups such as shearlets, but also curvelets and related constructions. The approximation theoretic properties of these systems are coded in the associated decomposition spaces, defined by imposing suitably weighted Lp-norms on the transform side. Understanding the dependence of the scale of decomposition spaces on properties of the wavelet system is one of the basic problems of the theory, and generally not well understood. The talk translates this question into a problem from the domain of coarse geometry, concerning quasi-equivalence of suitably defined metrics in frequency domain, and presents several applications.

Felix Voigtlaender

Coorbit dual molecules and convolution dominated operators

Abstract: Many continuous frames—including Gabor frames, wavelets, and certain types of shearlets—are induced by group representations. An important question is whether one can obtain a discrete frame by sampling such a localized continuous system. Preferably, the discrete dual frame should be well-localized as well, ensuring that the frame expansion extends to function spaces beyond the Hilbert space setting.

One of the main results of coorbit theory shows that this is indeed possible in many cases. In this talk, we present a novel proof of this fact. This proof is more accessible than the traditional one and yields a stronger conclusion, showing that the dual frame forms a family of coorbit molecules. The main proof ingredient is a novel local holomorphic calculus for convolution-dominated operators on general locally compact groups, which might be of independent interest.

In more technical detail, let G be a locally compact group and let π be an irreducible representation of G on a Hilbert space π. The voice transform of f with respect to g is Vgf(x)=⟨f, π(x)g. We assume that Vgg belongs to the Wiener amalgam space W(L, Lw1) and is normalized so that
                                                                                                        f = ∫GVgf(x) π(x)gdμ(x)  ∀ f ∈ ℋπ.

We show that for every discrete, sufficiently dense set Λ ⊂ G, the family (π(λ)g)λ ∈ Λ is a frame for π, and there exists a dual frame (hλ)λ ∈ Λ that forms a family of coorbit-molecules, in the sense that
                                                                                                       |Vghλ(x)| ≤ Θ(λ−1x)
for all λ ∈ Λ, with a suitable envelope Θ ∈ W(L, Lw1).

This molecule condition implies that the size of the coefficients (⟨f, hλ⟩)λ ∈ Λ fully reflects the size profile of the voice transform Vgf. It also ensures that the frame expansion
                                                                                                       f = ∑λ ∈ λf, hλ⟩ π(λ)g = ∑λ ∈ λf, π(λ)g⟩ hλ
extends to the associated coorbit spaces.

Finally, for very regular sets Λ (in particular for quasi-lattices), we show that the canonical dual frame forms a family of coorbit-molecules.

Dynamical sampling and approximate frame representations via iterated systems  
Ole Christensen  
Abstract: Dynamical sampling is currently a very active research area, dealing with, e.g.,  representations of frames via iterations of a bounded operator. Unfortunately,  the class of frames having such a representation is small, and only few concrete examples are known. We prove that the restrictions on the standard theory can be removed by considering approximate representations. The talk presents joint work with Dr. Marzieh Hasannasab at TU Berlin.  
 

Function Spaces for Fourier Analysis A Fresh Approach and Banach Gelfand Triples

H. Feichtinger‬

Abstract: ‎This is not a technical‎, ‎but a conceptual talk‎, ‎with the‎ ‎goal to describe a mindset that should help to better understand‎ ‎important aspects of modern Fourier and Time-Frequency Analysis‎. ‎New problems (such as Gabor expansions) require new tools‎, ‎in‎ ‎our case new function spaces‎, ‎such as Wiener amalgams or‎ ‎modulation spaces‎.

‎It turned out that these spaces‎, ‎specifically the‎ ‎Fourier invariant Segal algebra‎ ‎S0(Rd) and its dual S*0(Rd) (endowed with the w* topology)‎ ‎also provide a much more convenient‎‎ tool for the description of the Fourier transform than classical Lp-theory‎. ‎At can be developed without reference to the theory of tempered distributions developed by Laurent‎ ‎Schwartz‎, ‎but is in fact a (simplified)‎ ‎version of it‎, ‎sufficient for most engineering applications‎. ‎Therefore we call members of S*0(Rd) mild distributions.

It can be used to derive key results‎, ‎like the Shannon Sampling Theorem‎ ‎or the description of linear‎, ‎time-invariant systems as convolution operators by some impulse response (or via a transfer function on the Fourier transform‎ ‎side)‎. ‎The setting of mild distributions allows correctly to handle‎ ‎Dirac combs‎, ‎discrete periodic measures‎, ‎or Lp-spaces on an equal footing‎ ‎and shows how to (correctly) turn typical heuristic statements found in books‎ ‎on Fourier analysis into correct claims‎, ‎using distributional convergence (resp. w*-convergence) inside of S*0(Rd).

Thus the setting of the Banach Gelfand Triple (S0,L2,S*0)(Rd)‎ ‎appears to be the appropriate setting for all this‎, ‎because it‎ ‎allows to justify mathematically what engineers and physicists‎ ‎are doing‎, ‎but also because it can be used to support computational‎ ‎schemes‎. ‎Finally the setting grew out of long-standing attempts‎ ‎to develop a theory of extended Fourier transforms over LCA groups‎.

The talk comes with a number of illustrations‎, ‎showing via‎ ‎suitable pictograms for the different function spaces what the‎ ‎correct inclusion relations are between them‎. ‎At the technical‎ ‎level we only need basic functional analysis and a bit of group theory‎.
‎Please take it as a showcase for an extensive new landscape‎, ‎allowing‎ ‎you to take just a glimpse into a novel approach to what I call‎ ‎Conceptual Harmonic Analysis (more then Abstract + Numerical‎ ‎Harmonic Analysis)‎.

Image inpainting using sparse multiscale  representations  
Demetrio Labate  
Abstract: Image inpainting is concerned with the task of recovering missing blocks of data in images or videos. Among the several strategies proposed in the literature, sparse multiscale representations offer a powerful approach to analyze the image inpainting problem using a rigorous mathematical formulation. Here, we investigate inpainting in the continuous domain as a function interpolation problem in a Hilbert space, where only a masked version  of an unknown image is known. Under the assumption that the unknown image is sparse with respect to an appropriate representation, we search for the sparsest admissible solution. Since images found in many applications are dominated by edges, we consider an image model consisting of distributions supported on curvilinear singularities and prove that the theoretical performance of the recovery depends on the sparsifying and microlocal properties of the representation system, namely, exact image recovery is achieved if the size of the missing singularity is smaller than the size of the structure elements of the representation system. As a consequence of this observation, we prove that a sparsity-based image inpainting algorithm based on the shearlet representation - a multiscale anisotropic system that provides nearly optimally sparse representation of piecewise smooth images - significantly outperforms wavelets and other conventional multiscale systems.