Coorbit Spaces: Theory and Applications

• Coorbit spaces are function and distribution spaces defined via Banach, quasi-Banach, or Fréchet conditions on voice transforms from a unitary representation of a locally compact group.
• They unify classical smoothness spaces—such as modulation, Besov, and shearlet spaces—through a common representation-theoretic framework with atomic decompositions and reproducing kernels.
• Weighted and generalized coorbit spaces extend the theory to quasi-Banach and Fréchet scales, enabling precise localization, discretization, and operator kernel representations.

Coorbit spaces are function and distribution spaces defined by imposing Banach, quasi-Banach, or Fréchet function-space conditions on a representation-theoretic voice transform of the form $V_g f(x)=\langle f,\pi(x)g\rangle$. Initiated by Hans G. Feichtinger and Karlheinz Gröchenig in the late 1980s, coorbit theory starts from a locally compact group, a unitary representation, and an analyzing vector, and then constructs spaces whose regularity, decay, or sparsity are measured on the group rather than directly on the underlying domain. In this sense, modulation spaces, Besov spaces, Sobolev–Shubin spaces, and shearlet spaces appear as instances of a common representation-theoretic mechanism, with atomic decompositions and Banach frames reflecting the geometry of the underlying group (Berge, 2021).

1. Classical construction

The classical theory starts with a locally compact group $G$, a strongly continuous unitary representation $\pi:G\to \mathcal U(\mathcal H_\pi)$, and the voice transform

$\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$

For square-integrable representations, the Duflo–Moore theorem yields admissible vectors $g$ for which $\mathcal W_g:\mathcal H_\pi\to L^2(G)$ is an isometry, and the image is a reproducing kernel Hilbert space characterized by

$F=F*_G\mathcal W_g g.$

The corresponding reconstruction formula is

$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$

understood in the weak sense (Berge, 2021).

Classical coorbit theory then strengthens square integrability to integrability. For an integrable representation, one defines the test space

$$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$

its anti-dual reservoir

$\mathcal R=(\mathcal H^1)^\urcorner,$

and the coorbit spaces

$G$0

A basic structural feature is the endpoint identification

$G$1

The extended transform remains injective on the reservoir, and the coorbit scale interpolates naturally between test vectors, the Hilbert space, and distributions (Berge, 2021).

These facts encode the original Feichtinger–Gröchenig philosophy: a function space is derived from a representation rather than imposed externally. The decisive mechanism is the reproducing identity on the transform side, which converts analytic questions about functions or distributions into convolution and localization questions on the group.

2. Weighted coorbits and standard function-space realizations

A common weighted version uses weighted Lebesgue spaces $G$2, moderate weights $G$3, and control weights $G$4. In this setting one fixes a nonzero analyzing vector $G$5, defines

$G$6

lets $G$7, and sets

$G$8

The corresponding reproducing-kernel subspace is

$G$9

and the voice transform yields an isometric isomorphism

$\pi:G\to \mathcal U(\mathcal H_\pi)$0

Within this framework, independence of the analyzing vector, completeness, duality, and the correspondence principle are all available in a form specialized to weighted Lebesgue coefficients (Zimmermann, 2024).

Two canonical realizations dominate the literature. For the affine group, the wavelet representation produces homogeneous Besov spaces; in the weighted-$\pi:G\to \mathcal U(\mathcal H_\pi)$1 formulation one obtains

$\pi:G\to \mathcal U(\mathcal H_\pi)$2

For the reduced Heisenberg group, the Schrödinger representation produces the modulation spaces $\pi:G\to \mathcal U(\mathcal H_\pi)$3. In both cases, the representation-theoretic voice transform becomes a familiar continuous transform—the wavelet transform or the short-time Fourier transform—and Banach frames arise from sufficiently regular wavelet or Gabor windows (Zimmermann, 2024).

This classical layer is historically central because it shows that coorbit spaces are not merely abstract Banach spaces of coefficients. They recover established smoothness spaces and simultaneously explain why those spaces admit transform-domain discretizations.

3. Generalizations beyond irreducibility and $\pi:G\to \mathcal U(\mathcal H_\pi)$4-integrable kernels

A major extension removes two restrictive assumptions of the classical theory: irreducibility of the representation and $\pi:G\to \mathcal U(\mathcal H_\pi)$5-integrability of the reproducing kernel. In the Fréchet-target approach, one starts from data

$\pi:G\to \mathcal U(\mathcal H_\pi)$6

where $\pi:G\to \mathcal U(\mathcal H_\pi)$7 is a continuous unitary reproducing representation, $\pi:G\to \mathcal U(\mathcal H_\pi)$8 is admissible, $\pi:G\to \mathcal U(\mathcal H_\pi)$9 is a Fréchet function space continuously embedded into $\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$0, and $\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$1 is a left-invariant Banach space continuously embedded into $\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$2. The test space is defined by

$\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$3

the distribution space is $\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$4, and the generalized coorbit space is

$\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$5

The reproducing subspace is

$\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$6

and the fundamental structural statement is

$\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$7

with inverse $\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$8. The same framework recovers

$\mathcal W_g f(x)=\langle f,\pi(x)g\rangle .$9

Its model examples include $g$0, Paley–Wiener spaces, Shannon wavelets, and Schrödingerlets (Dahlke et al., 2014).

A different line of extension treats quasi-Banach coefficient spaces. For a solid quasi-Banach function space $g$1, the correct coefficient condition is not generally $g$2 itself but

$g$3

the left Wiener amalgam associated with $g$4. In this form, coorbit spaces exist on arbitrary second countable locally compact groups, including nonunimodular groups, and the theory allows projective and reducible unitary representations, weaker localization assumptions $g$5, and molecular dual frames and Riesz sequences throughout the quasi-Banach scale (Velthoven et al., 2022).

A third extension treats the coorbit spaces themselves as Fréchet spaces when the kernel is only in

$g$6

and not in $g$7. In that case one sets

$g$8

and defines

$g$9

This projective-limit viewpoint is designed for non-integrable kernels such as the sinc kernel and leads to a Fréchet-space Correspondence Principle. Closely related work shows that, without further hypotheses, classical uniform atomic decompositions and Banach frames cannot generally persist in the non-$\mathcal W_g:\mathcal H_\pi\to L^2(G)$0 setting; stronger discretization requires additional operator bounds or auxiliary kernels (Dahlke et al., 19 Dec 2025, Dahlke et al., 2018).

A recurring misconception is that coorbit theory is intrinsically tied to $\mathcal W_g:\mathcal H_\pi\to L^2(G)$1, irreducibility, or Banach coefficients. The generalized theories show that those hypotheses are classical sufficient conditions, not an exhaustive description of the subject.

4. Decomposition-space realizations and phase-space localization

For semidirect products $\mathcal W_g:\mathcal H_\pi\to L^2(G)$2, a large class of wavelet coorbit spaces can be realized as decomposition spaces. If $\mathcal W_g:\mathcal H_\pi\to L^2(G)$3 is an integrably admissible dilation group with essential frequency support $\mathcal W_g:\mathcal H_\pi\to L^2(G)$4, one has a natural smooth analyzing space

$\mathcal W_g:\mathcal H_\pi\to L^2(G)$5

and for mixed weighted Lebesgue spaces $\mathcal W_g:\mathcal H_\pi\to L^2(G)$6 the coorbit spaces are isomorphic to decomposition spaces

$\mathcal W_g:\mathcal H_\pi\to L^2(G)$7

associated with coverings of $\mathcal W_g:\mathcal H_\pi\to L^2(G)$8 induced by the dual action of $\mathcal W_g:\mathcal H_\pi\to L^2(G)$9. In particular, the anisotropic Besov spaces associated with expansive matrices coincide precisely with the coorbit spaces induced by integrably admissible one-parameter groups (Führ et al., 2019).

The same coorbit-to-decomposition-space passage is central for concrete non-Euclidean examples. On stratified Lie groups, the quasi-regular representation of $F=F*_G\mathcal W_g g.$0 identifies homogeneous Besov spaces with coorbit spaces through

$F=F*_G\mathcal W_g g.$1

up to norm equivalence, and the coorbit machinery then yields atomic decompositions and Banach frames for these Besov spaces (Christensen et al., 2011). In the shearlet setting, the decomposition-space viewpoint provides explicit Sobolev embedding criteria for three-dimensional shearlet coorbit spaces and shows that the scaling subgroup governs the embedding behavior, while the shearing subgroup does not affect these Sobolev embeddings (Führ et al., 2019).

A complementary localization principle is provided by phase-space covers. If $F=F*_G\mathcal W_g g.$2 is a partition of unity on phase space and $F=F*_G\mathcal W_g g.$3 denotes the associated phase-space multiplier, then

$F=F*_G\mathcal W_g g.$4

This characterizes coorbit norms by arbitrary, possibly irregular, phase-space covers, and extends localization-operator descriptions from time-frequency analysis to time-scale and more general coorbit settings (Romero, 2010).

These results are structurally important because they connect coorbit spaces to Littlewood–Paley decompositions, decomposition spaces, and localization operators. The representation-theoretic definition is thereby linked to frequency coverings, embedding theory, and explicit coefficient-space descriptions.

5. Kernel theorems and operator coorbit spaces

Coorbit theory also supports a general kernel theorem for operators. If $F=F*_G\mathcal W_g g.$5 and $F=F*_G\mathcal W_g g.$6 are integrable representations of $F=F*_G\mathcal W_g g.$7 and $F=F*_G\mathcal W_g g.$8, then bounded operators

$F=F*_G\mathcal W_g g.$9

are represented by unique kernels

$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$0

through the pairing

$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$1

On the transform side, every such operator becomes an integral operator whose kernel is the generalized matrix coefficient

$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$2

and Schur-type tests give mixed-norm boundedness criteria for maps between other coorbit spaces. This abstract theorem recovers Feichtinger’s kernel theorem for modulation spaces and yields analogous results for Besov spaces and cross-representation mappings (Balazs et al., 2019).

A more recent development replaces function-valued transforms by operator-valued ones. For Hilbert–Schmidt operators, the operator-valued STFT is

$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$3

and it generates vector-valued reproducing kernel Banach spaces of operators, serving as operator coorbit spaces in direct analogy with function modulation spaces (Dörfler et al., 2022). In a parallel framework, for operator Gabor $f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$4-frames one obtains coincidence between continuous operator-STFT spaces $f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$5 and localized $f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$6-frame coorbit spaces $f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$7; moreover, the associated sparsity classes satisfy

$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$8

and best-$f=\int_G \mathcal W_g f(x)\,\pi(x)g\,d\mu_L(x),$9-term approximation obeys

$$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$0

This places operator localization, sparse approximation, and operator coorbits in a common framework, including applications to Feichtinger operators and mixed states (Dörfler et al., 19 Sep 2025).

The operator theory shows that coorbit spaces are no longer confined to scalar function spaces. The same representation-theoretic logic extends to kernels, operator-valued transforms, sparse operator dictionaries, and approximation classes.

6. Canonical examples and current directions

Several examples have become standard precisely because they lie outside the narrow classical $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$1-kernel regime. For the Paley–Wiener setting on $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$2, with $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$3, the kernel is not in $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$4 when $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$5, but it is in $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$6 for every $$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$7, and the generalized theory gives

$$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$8

The Shannon wavelet on the affine group provides another natural admissible vector with

$$\mathcal H^1=\{f\in\mathcal H_\pi:\mathcal W_g f\in L^1(G)\},$$9

and Schrödingerlets supply a highly reducible example with kernels in

$\mathcal R=(\mathcal H^1)^\urcorner,$0

These examples were among the main motivations for replacing $\mathcal R=(\mathcal H^1)^\urcorner,$1 by Fréchet targets such as $\mathcal R=(\mathcal H^1)^\urcorner,$2 (Dahlke et al., 2014).

Wavelet coorbit theory over disconnected local fields gives a different non-Euclidean realization. For a local field $\mathcal R=(\mathcal H^1)^\urcorner,$3, the quasi-regular representation of

$\mathcal R=(\mathcal H^1)^\urcorner,$4

is integrable, the test-function space

$\mathcal R=(\mathcal H^1)^\urcorner,$5

consists precisely of admissible wavelets inside $\mathcal R=(\mathcal H^1)^\urcorner,$6, and homogeneous Besov spaces satisfy

$\mathcal R=(\mathcal H^1)^\urcorner,$7

with equivalent norms. In this ultrametric setting, compact open subgroups and right-invariant wavelet coefficients yield unusually explicit tight wavelet frames and unconditional bases for the associated coorbit spaces (Abhinav et al., 10 Aug 2025).

Recent work has also shifted attention from construction to classification. In dimension $\mathcal R=(\mathcal H^1)^\urcorner,$8, for irreducibly admissible matrix groups $\mathcal R=(\mathcal H^1)^\urcorner,$9, one has

$G$00

if and only if their open dual orbits coincide and either the orbit has $G$01 or $G$02 connected components, or it has $G$03 connected components and $G$04. In this formulation, the similitude case is highly symmetric, the diagonal case is orbit-rigid, and the shearlet case is fully group-rigid (Asharaf et al., 6 Aug 2025).

These directions suggest a broader current picture of coorbit spaces. The subject now includes classical integrable representations, reducible and non-integrable reproducing systems, quasi-Banach and Fréchet coefficient scales, decomposition-space realizations, kernel theorems, operator-valued transforms, local-field wavelets, and classification problems for entire coorbit scales. What remains constant across these variants is the central idea: a space is defined by the membership of its voice transform in a model space on a group or phase space, and its structure is governed by reproducing formulas, localization, and discretization.