Reproducing formulas for generalized translation invariant systems on locally compact abelian groups
aa r X i v : . [ m a t h . F A ] A p r Reproducing formulas for generalized translation invariantsystems on locally compact abelian groups
Mads Sielemann Jakobsen ∗ , Jakob Lemvig † February 27, 2015
Abstract:
In this paper we connect the well established discrete frame theoryof generalized shift invariant systems to a continuous frame theory. To do so, welet Γ j , j ∈ J , be a countable family of closed, co-compact subgroups of a sec-ond countable locally compact abelian group G and study systems of the form ∪ j ∈ J { g j,p ( · − γ ) } γ ∈ Γ j ,p ∈ P j with generators g j,p in L ( G ) and with each P j beinga countable or an uncountable index set. We refer to systems of this form asgeneralized translation invariant (GTI) systems. Many of the familiar transforms,e.g., the wavelet, shearlet and Gabor transform, both their discrete and continu-ous variants, are GTI systems. Under a technical α local integrability condition( α -LIC) we characterize when GTI systems constitute tight and dual frames thatyield reproducing formulas for L ( G ) . This generalizes results on generalized shiftinvariant systems, where each P j is assumed to be countable and each Γ j is a uni-form lattice in G , to the case of uncountably many generators and (not necessarilydiscrete) closed, co-compact subgroups. Furthermore, even in the case of uniformlattices Γ j , our characterizations improve known results since the class of GTIsystems satisfying the α -LIC is strictly larger than the class of GTI systems sat-isfying the previously used local integrability condition. As an application of ourcharacterization results, we obtain new characterizations of translation invariantcontinuous frames and Gabor frames for L ( G ) . In addition, we will see that theadmissibility conditions for the continuous and discrete wavelet and Gabor trans-form in L ( R n ) are special cases of the same general characterizing equations. In harmonic analysis one is often interested in determining conditions on generators of functionsystems, e.g., Gabor and wavelet systems, that allow for reconstruction of any function in agiven class of functions from its associated transform via a reproducing formula. The work ofHernández, Labate, and Weiss [30] and of Ron and Shen [46] on generalized shift invariant systemsin L ( R n ) presented a unified theory for many of the familiar discrete transforms, most notablythe Gabor and the wavelet transform. The generalized shift invariant systems are collections offunctions of the form ∪ j ∈ J { T γ g j } γ ∈ Γ j , where J is a countable index set, T γ denotes translationby γ , Γ j a full-rank lattice in R n , and { g j } j ∈ J a subset of L ( R n ) . Here, the word “shift” is Mathematics Subject Classification.
Primary: 42C15, 43A32, 43A70, Secondary: 43A60, 46C05.
Key words and phrases. continuous frame, dual frames, dual generators, g-frame, Gabor frame, generalizedshift invariant system, generalized translation invariant system, LCA group, Parseval frame, wavelet frame ∗ Technical University of Denmark, Department of Applied Mathematics and Computer Science, Matematik-torvet 303B, 2800 Kgs. Lyngby, Denmark, E-mail: [email protected] † Technical University of Denmark, Department of Applied Mathematics and Computer Science, Matematik-torvet 303B, 2800 Kgs. Lyngby, Denmark, E-mail: [email protected] akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups used since the translations are discrete and the word “generalized” since the shift lattices Γ j are allowed to change with the parameter j ∈ J . The main result of Hernández, Labate, andWeiss [30] is a characterization, by so-called t α -equations, of all functions g j that give rise toisometric transforms, called Parseval frames in frame theory.The goal of this work is to connect the discrete transform theory of generalized shift invari-ant systems to a continuous/integral transform theory. In doing so, the scope of the “unifiedapproach” started in [30, 46] will be vastly extended. What more is, this new theory will cover“intermediate” steps, the semi-continuous transforms, and we will do so in a very general set-ting of square integrable functions on locally compact abelian groups. In particular, we recoverthe usual characterization results for discrete and continuous Gabor and wavelet systems asspecial cases. For discrete wavelets in L ( R ) with dyadic dilation, this result was obtained in1995, independently by Gripenberg [23] and Wang [48], and it can be stated as follows. Definethe translation operator T b f ( x ) = f ( x − b ) and dilation operator D a f ( x ) = | a | − / f ( x/a ) for b ∈ R , a = 0 . The discrete wavelet system { T j k D j ψ } j,k ∈ Z generated by ψ ∈ L ( R ) is indeeda generalized shift invariant system with J = Z , Γ j = 2 j Z , and g j = D j ψ . Now, the linearoperator W d defined by W d : L ( R ) → ℓ ( Z ) , W d f ( j, k ) = h f, T j k D j ψ i is isometric if, and only if, for all α ∈ S j ∈ Z − j Z , the following t α -equations hold: t α := X j ∈ Z : α ∈ − j Z b ψ (2 j ξ ) b ψ (2 j ( ξ + α )) = δ α, for a.e. ξ ∈ b R , (1.1)where b R denotes the Fourier domain. In the language of frame theory, we say that generators ψ ∈ L ( R ) of discrete Parseval wavelet frames have been characterized by t α -equations.Calderón [6] discovered in 1964 that any function ψ ∈ L ( R ) satisfying the Calderón admis-sibility condition Z R \{ } | b ψ ( aξ ) | | a | da = 1 for a.e. ξ ∈ b R (1.2)leads to reproducing formulas for the continuous wavelet transform. To be precise, the linearoperator W c defined by W c : L ( R ) → L (cid:0) R \{ } × R , dadba (cid:1) , W c f ( a, b ) = h f, T b D a ψ i is isometric if, and only if, the Calderón admissibility condition holds. We will see that theCalderón admissibility condition is nothing but the t α -equation (there is only one!) for thecontinuous wavelet system. Similar results hold for the Gabor case; here the continuous transformis usually called the short-time Fourier transform. Actually, the theory is not only applicableto the Gabor and wavelet setting, but to a very large class of systems of functions includingshearlet and wave packet systems, which we shall call generalized translation invariant systems.We refer the reader to the classical texts [12, 14, 27] and the recent book [38] for introductions tothe specific cases of Gabor, wavelet, shearlet and wave packet analysis.In [36], Kutyniok and Labate generalized the results of Hernández, Labate, and Weiss togeneralized shift invariant systems ∪ j ∈ J { T γ g j } γ ∈ Γ j in L ( G ) , where G is a second countablelocally compact abelian group and Γ j is a family of uniform lattices (i.e., Γ j is a discrete subgroupand the quotient group G/ Γ j is compact) indexed by a countable set J . The main goal of thepresent paper is to develop the corresponding theory for semi-continuous and continuous framesin L ( G ) . In order to achieve this, we will allow non-discrete translation groups Γ j , and we date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups will allow for each translation group to have uncountable many generators, indexed by someindex set P j , j ∈ J . We say that the corresponding family ∪ j ∈ J { T γ g j,p } γ ∈ Γ j ,p ∈ P j in L ( G ) is a generalized translation invariant system . To be precise, we will, for each j ∈ J , take P j to bea σ -finite measure space with measure µ P j and Γ j to be closed, co-compact (i.e., the quotientgroup G/ Γ j is compact) subgroups. We mention that any locally compact abelian group has aco-compact subgroup, namely the group itself. On the other hand, there exist groups that do notcontain uniform lattices, e.g., the p -adic numbers. Thus, the theory of generalized translationinvariant systems is applicable to a larger class of locally compact abelian groups than the theoryof generalized shift invariant systems.The two wavelet cases described above fit our framework. The discrete wavelet system canbe written as ∪ j ∈ Z { T γ ( D j ψ ) } γ ∈ j Z , so we see that P j is a singleton and µ P j a weighted countingmeasure for each j ∈ J = Z , and that there are countably many different (discrete) Γ j . For thecontinuous wavelet system on the form { T γ ( D p ψ ) } γ ∈ R ,p ∈ R \{ } , we have that J is a singleton,e.g., { j } since there is only one translation subgroup Γ j = R . On the other hand, here P j is uncountable and µ P j a weighted Lebesgue measure. We stress that our setup can handlecountable many (distinct) Γ j and countable many P j , each being uncountable.The characterization results in [30, 36] rely on a technical condition on the generators andthe translation lattices, the so-called local integrability condition . This condition is straightfor-ward to formulate for generalized translation invariant systems, however, we will replace it by astrictly weaker condition, termed α local integrability condition . Therefore, even for generalized shift invariant systems in the euclidean setting, our work extends the characterization resultsby Hernández, Labate, and Weiss [30]. Under the α local integrability condition, we show inTheorem 3.5 that ∪ j ∈ J { T γ g j,p } γ ∈ Γ j ,p ∈ P j is a Parseval frame for L ( G ) , that is, the associatedtransform is isometric if, and only if, t α := X j ∈ J : α ∈ Γ ⊥ j Z P j ˆ g j,p ( ω )ˆ g j,p ( ω + α ) dµ P j ( p ) = δ α, a.e. ω ∈ b G for every α ∈ ∪ j ∈ J Γ ⊥ j , where Γ ⊥ j = (cid:8) ω ∈ b G : ω ( x ) = 0 for all x ∈ Γ j (cid:9) denotes the annihilator of Γ j . Now, returning to the two main examples of this introduction, the discrete and continuouswavelet transform, we see why the number of the t α -equations in (1.1) and (1.2) are so different.In the discrete case the corresponding union of the annihilators of the translation groups is ∪ j ∈ Z − j Z , while in the continuous case the annihilator of R is simply { } , which corresponds toonly one t α -equation ( α = 0 ).Finally, as Kutyniok and Labate [36] restrict their attention to Parseval frames, there arecurrently no characterization results available for dual (discrete) frames in the setting of locallycompact abelian groups. Hence, one additional objective of this paper is to prove characterizingequations for dual generalized translation invariant frames to remedy this situation.For a related study of reproducing formulas from a purely group representation theoreticalpoint of view, we refer to the work of Führ [20], and De Mari, De Vito [15], and the referencestherein.The paper is organized as follows. We recall some basic theory about locally compact abeliangroups and introduce the generalized translation invariant systems in Section 2.1 and 2.2, re-spectively. Additionally, in Section 2.3 we give a short introduction to the theory of continuousframes and g-frames. In Section 3 we present our main characterization result for dual gen-eralized translation invariant frames (Theorem 3.4) and, as corollary, then for Parseval frames(Theorem 3.5). In Section 3.2 and 3.3 we relate several conditions used in our main results. date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups Finally, we consider the special case of translation invariant systems and apply our characteri-zation results on concrete groups and to concrete examples in Sections 3.4 and 4. Specifically,we consider discrete and continuous wavelet systems in L ( R n ) , shearlets in L ( R ) , discrete,semi-continuous and continuous Gabor frames on LCA groups and GTI systems over the p -adicintegers and numbers.During the final stages of this project, we realized that Bownik and Ross [4] have completeda related investigation. As they consider and characterize the structure of translation invariantsubspaces on locally compact abelian groups, their results do not overlap with our results in anyway. However, they do consider translations along a closed, co-compact subgroup. We adopttheir terminology of translation invariance, in place of shift invariance, to emphasize the factthat Γ j need not be discrete. In the following sections we set up notation and recall some useful results from Fourier analysison locally compact abelian (LCA) groups and continuous frame theory. Furthermore, we willprove two important lemmas, Lemma 2.2 and 2.4.
Throughout this paper G will denote a second countable locally compact abelian group. We notethat the following statements are equivalent: (i) G is second countable, (ii) L ( G ) is separable,(iii) G is metrizable and σ -compact. Note that the metric on G can be chosen to be translationinvariant.To G we associate its dual group b G consisting of all characters, i.e., all continuous homomor-phisms from G into the torus T ∼ = { z ∈ C : | z | = 1 } . Under pointwise multiplication b G is also alocally compact abelian group. We will use addition and multiplication as group operation in G and b G , respectively. Note that in the introduction we used addition as group operation in b G . Bythe Pontryagin duality theorem, the dual group of b G is isomorphic to G as a topological group,i.e., bb G ∼ = G . We recall the well-known facts that if G is discrete, then b G is compact, and viceversa.We denote the Haar measure on G by µ G . The (left) Haar measure on any locally compactgroup is unique up to a positive constant. From µ G we define L ( G ) and the Hilbert space L ( G ) over the complex field in the usual way.For functions f ∈ L ( G ) we define the Fourier transform F f ( ω ) = ˆ f ( ω ) = Z G f ( x ) ω ( x ) dµ G ( x ) , ω ∈ b G. If f ∈ L ( G ) , ˆ f ∈ L ( b G ) , and the measure on G and b G are normalized so that the Planchereltheorem holds (see [32, (31.1)]), the function f can be recovered from ˆ f by the inverse Fouriertransform f ( x ) = F − ˆ f ( x ) = Z b G ˆ f ( ω ) ω ( x ) dµ b G ( ω ) , x ∈ G. From now on we always assume that the measure on a group µ G and its dual group µ b G arenormalized this way, and we refer to them as dual measures . As in the classical Fourier analysis F can be extended from L ( G ) ∩ L ( G ) to an isometric isomorphism between L ( G ) and L ( b G ) .On any locally compact abelian group G , we define the following two linear operators. For a ∈ G , the operator T a , called translation by a , is defined by T a : L ( G ) → L ( G ) , ( T a f )( x ) = f ( x − a ) , x ∈ G. date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups For χ ∈ b G , the operator E χ , called modulation by χ , is defined by E χ : L ( G ) → L ( G ) , ( E χ f )( x ) = χ ( x ) f ( x ) , x ∈ G. Together with the Fourier transform F , the two operators E χ and T a share the following com-mutator relations: T a E χ = χ ( a ) E χ T a , F T a = E a − F , and F E χ = T χ F . For a subgroup H of an LCA group G , we define its annihilator as H ⊥ = { ω ∈ b G : ω ( x ) = 1 for all x ∈ H } . The annihilator H ⊥ is a closed subgroup in b G , and if H is closed, then b H ∼ = b G/H ⊥ and [ G/H ∼ = H ⊥ .We will repeatedly use Weil’s formula; it relates integrable functions over G with integrablefunctions on the quotient space G/H when H is a closed subgroup of G . We mention thefollowing results concerning Weil’s formula [44]. Theorem 2.1.
Let H be a closed subgroup of G . Let π H : G → G/H, π H ( x ) = x + H be thecanonical map from G onto G/H . If f ∈ L ( G ) , then the following holds:(i) The function ˙ x R H f ( x + h ) dµ H ( h ) , ˙ x = π H ( x ) defined almost everywhere on G/H , isintegrable.(ii) (Weil’s formula) Let two of the Haar measures on
G, H and
G/H be given, then the thirdcan be normalized such that Z G f ( x ) dµ G ( x ) = Z G/H Z H f ( x + h ) dµ H ( h ) dµ G/H ( ˙ x ) . (2.1) (iii) If (2.1) holds, then the respective dual measures on b G, H ⊥ ∼ = [ G/H , b G/H ⊥ ∼ = b H satisfy Z b G ˆ f ( ω ) dµ b G ( ω ) = Z b G/H ⊥ Z H ⊥ ˆ f ( ωγ ) dµ H ⊥ ( γ ) dµ b G/H ⊥ ( ˙ ω ) . (2.2) Remark . Since a Haar measure and its dual are chosen so that the Plancherel theorem holdswe have the following uniqueness result: If two of the measures on
G, H, G/H, b G, H ⊥ and b G/H ⊥ are given, and these two are not dual measures, by requiring Weil’s formulas (2.1) and (2.2), allother measures are uniquely determined.For more information on harmonic analysis on locally compact abelian groups, we refer thereader to the classical books [17, 31, 32, 44].For a Borel set E ⊂ b G with µ b G ( E ) = 0 , we define: D = (cid:8) f ∈ L ( G ) : ˆ f ∈ L ∞ ( b G ) and supp ˆ f is compact in b G \ E (cid:9) . (2.3)It is not difficult to show that D is dense in L ( G ) exactly when µ b G ( E ) = 0 . We will frequentlyprove our results on D and extend by a density argument. The role of the set E is to allow for“blind spots” of transforms – a term coined by Führ [21]. We will let E be an unspecified setsatisfying µ b G ( E ) = 0 ; the specific choice of E depends on the application, e.g., in the Gabor andwavelet case [30] one would usually take E = ∅ and E = { } , respectively.The following result relies on Weil’s formula and will play an important part of the proofs inSection 3. date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups Lemma 2.2.
Let H be a closed subgroup of an LCA group G with Haar measure µ H . Supposethat f , f ∈ D and ϕ, ψ ∈ L ( G ) . Then Z H h f , T h ϕ ih T h ψ, f i dµ H ( h ) = Z b G Z H ⊥ ˆ f ( ω ) ˆ f ( ωα ) ˆ ϕ ( ω ) ˆ ψ ( ωα ) dµ H ⊥ ( α ) dµ b G ( ω ) . Proof.
Let h ∈ H . An application of the Plancherel theorem together with Weil’s formula yields h f , T h ϕ i = h ˆ f , d T h ϕ i = h ˆ f , E − h ˆ ϕ i = Z b G ˆ f ( ω ) ˆ ϕ ( ω ) ω ( h ) dµ b G ( ω )= Z b G/H ⊥ Z H ⊥ ˆ f ( ωγ ) ˆ ϕ ( ωγ ) ω ( h ) γ ( h ) dµ H ⊥ ( γ ) dµ b G/H ⊥ ( ˙ ω )= Z b H (cid:16) ω ( h ) Z H ⊥ ˆ f ( ωγ ) ˆ ϕ ( ωγ ) dµ H ⊥ ( γ ) (cid:17) dµ b H ( ω ) , where we tacitly used that b G/H ⊥ ∼ = b H . A similar calculation can be done for h T h ψ, f i . Toease notation, we define [ ˆ f , ˆ ϕ ]( ω, H ⊥ ) = R H ⊥ ˆ f ( ωγ ) ˆ ϕ ( ωγ ) dµ H ⊥ ( γ ) for f ∈ D . Again, by thePlancherel theorem and Weil’s formula we have Z H h f , T h ϕ ih T h ψ, f i dµ H ( h )= Z H (cid:18) Z b H ω ( h )[ ˆ f , ˆ ϕ ]( ω, H ⊥ ) dµ b H ( ω ) (cid:19)(cid:18) Z b H ω ( h )[ ˆ f , ˆ ψ ]( ω, H ⊥ ) dµ b H ( ω ) (cid:19) dµ H ( h )= D F − [ ˆ f , ˆ ϕ ]( · , H ⊥ ) , F − [ ˆ f , ˆ ψ ]( · , H ⊥ ) E L ( H ) = D [ ˆ f , ˆ ϕ ]( · , H ⊥ ) , [ ˆ f , ˆ ψ ]( · , H ⊥ ) E L ( b H ) = Z b G/H ⊥ (cid:16) Z H ⊥ ˆ f ( ωγ ) ˆ ϕ ( ωγ ) dµ H ⊥ ( γ ) (cid:17)(cid:16) Z H ⊥ ˆ f ( ωγ ) ˆ ψ ( ωγ ) dµ H ⊥ ( γ ) (cid:17) dµ b G/H ⊥ ( ˙ ω )= Z b G/H ⊥ h Z H ⊥ ˆ f ( ωγ ) ˆ ϕ ( ωγ ) (cid:16) Z H ⊥ ˆ f ( ωβ ) ˆ ψ ( ωβ ) dµ H ⊥ ( β ) (cid:17) dµ H ⊥ ( γ ) i dµ b G/H ⊥ ( ˙ ω )= Z b G/H ⊥ h Z H ⊥ ˆ f ( ωγ ) ˆ ϕ ( ωγ ) (cid:16) Z H ⊥ ˆ f ( ωγα ) ˆ ψ ( ωγα ) dµ H ⊥ ( α ) (cid:17) dµ H ⊥ ( γ ) i dµ b G/H ⊥ ( ˙ ω )= Z b G ˆ f ( ω ) ˆ ϕ ( ω ) (cid:16) Z H ⊥ ˆ f ( ωα ) ˆ ψ ( ωα ) dµ H ⊥ ( α ) (cid:17) dµ b G ( ω )= Z b G Z H ⊥ ˆ f ( ω ) ˆ f ( ωα ) ˆ ϕ ( ω ) ˆ ψ ( ωα ) dµ H ⊥ ( α ) dµ b G ( ω ) . Here F denotes the Fourier transform on H . Let J ⊂ Z be a countable index set. For each j ∈ J , let P j be a countable or an uncountable indexset, let g j,p ∈ L ( G ) for p ∈ P j , and let Γ j be a closed, co-compact subgroup in G . Recall that co-compact subgroups are subgroups of G for which G/ Γ j is compact. For a compact abelian group,the group is metrizable if, and only if, the character group is countable [31, (24.15)]. Hence,since G/ Γ j is compact and metrizable, the group d G/ Γ j ∼ = Γ ⊥ j is discrete and countable. Unlessstated otherwise we equip Γ ⊥ j with the counting measure and assume a fixed Haar measure µ G on G . By Remark 1 this uniquely determines the measures on Γ j , G/ Γ j , b G , and b G/ Γ ⊥ j . date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups The generalized translation invariant (GTI) system generated by { g j,p } p ∈ P j ,j ∈ J with transla-tion along closed, co-compact subgroups { Γ j } j ∈ J is the family of functions ∪ j ∈ J { T γ g j,p } γ ∈ Γ j ,p ∈ P j .To ease notation, we will suppress the dependence of j in g j,p and write the GTI system as ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j .If we take Γ = Γ j for each j ∈ J , we obtain a translation invariant (TI) system in thesense that f ∈ ∪ j ∈ J { T γ g p } γ ∈ Γ ,p ∈ P j implies T γ f ∈ ∪ j ∈ J { T γ g p } γ ∈ Γ ,p ∈ P j for all γ ∈ Γ . However,generalized translation invariant systems are more general than translation invariant systemssince we allow for a different subgroup for each set of generators { g j,p } p ∈ P j .When each P j is countable and each Γ j is a uniform lattice, i.e., a discrete , co-compactsubgroup, we recover the generalized shift invariant (GSI) systems considered in [36]. However,we note that there exist locally compact abelian groups that do not contain any uniform lattices.As an example we mention the p -adic numbers, whose only discrete subgroup is the neutralelement which is not a uniform lattice. In other cases, such as the p -adic integers, the LCAgroup will have only trivial examples of uniform lattices, e.g., the neutral element, but haveplenty non-trivial co-compact subgroups, see Example 10 in Section 4.Finally, as an alternative generalization of uniform lattices, we mention the idea of so-calledquasi-lattices, see [28,29]. In contrast to closed, co-compact subgroups, quasi-lattices are discretesubsets in G that are not necessarily groups. The central concept of this section is that of a continuous frame. The definition is as follows.
Definition 2.3.
Let H be a complex Hilbert space, and let ( M, Σ M , µ M ) be a measure space,where Σ M denotes the σ -algebra and µ M the non-negative measure. A family of vectors { f k } k ∈ M is called a continuous frame for H with respect to ( M, Σ M , µ M ) if(a) k f k is weakly measurable, i.e., for all f ∈ H , the mapping M → C , k
7→ h f, f k i ismeasurable, and(b) there exist constants A, B > such that A k f k ≤ Z M |h f, f k i| dµ M ( k ) ≤ B k f k for all f ∈ H . (2.4)The constants A and B are called frame bounds . Remark . As we will only consider separable Hilbert spaces in this paper, we can replace weakmeasurability of k f k with (strong) measurability with respect to the Borel algebra in H byPettis’ theorem.In cases where it will cause no confusion, we will simply say that { f k } k ∈ M is a frame for H .If { f k } k ∈ M is weakly measurable and the upper bound in the above inequality (2.4) holds, then { f k } k ∈ M is said to be a Bessel family with constant B . A frame { f k } k ∈ M is said to be tight ifwe can choose A = B ; if, furthermore, A = B = 1 , then { f k } k ∈ M is said to be a Parseval frame .Two Bessel families { f k } k ∈ M and { g k } k ∈ M are said to be dual frames if h f, g i = Z M h f, g k ih f k , g i dµ M ( k ) for all f, g ∈ H . (2.5)In this case we say that the following assignment f = Z M h f, g k i f k dµ M ( k ) for f ∈ H , (2.6) date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups holds in the weak sense. Equation (2.6) is often called a reproducing formula for f ∈ H . Thefollowing argument shows that two such dual frames indeed are frames, and we shall say that theframe { f k } k ∈ M is dual to { g k } k ∈ M , and vice versa. We need to show that both Bessel families { f k } k ∈ M and { g k } k ∈ M satisfy the lower frame bound. By taking f = g in (2.5) and using theCauchy-Schwarz inequality, we have k f k = Z M h f, f k ih g k , f i dµ M ( k ) ≤ (cid:16) Z M |h f, f k i| dµ M ( k ) (cid:17) / (cid:16) Z M |h f, g k i| dµ M ( k ) (cid:17) / ≤ (cid:16) Z M |h f, f k i| dµ M ( k ) (cid:17) / p B g k f k . In the last step we used that { g k } k ∈ M has an upper frame bound B g . Rearranging the terms inthe above inequality gives B g k f k ≤ Z M |h f, f k i| dµ M ( k ) . Hence, the Bessel family { f k } k ∈ M satisfies the lower frame condition and is a frame. A similarargument shows that { g k } k ∈ M satisfies the lower frame condition. This completes the argument.Moreover, by a polarization argument, it follows that two Bessel families { f k } k ∈ M and { g k } k ∈ M are dual frames if, and only if, h f, f i = Z M h f, g k ih f k , f i dµ M ( k ) for all f ∈ H . We mention that to a given frame for H one can always find at least one dual frame. For moreinformation on (continuous) frames, we refer to [1, 2, 8, 18, 22, 34].To a frame { f k } k ∈ M for H , we associate the frame transform given by H → L ( M, µ M ) , f ( k
7→ h f, f k i ) . As mentioned in the introduction, this transform is isometric if, and only if, the family { f k } k ∈ M is a Parseval frame. A similar conclusion holds for a pair of dual frames.Let ( M , Σ , µ ) and ( M , Σ , µ ) be measure spaces. We say that a family { f k } k ∈ M in theHilbert space H is unitarily equivalent to a family { g k } k ∈ M in the Hilbert space K if there isa point isomorphism ι : M → M , i.e., ι is a (measurable) bijection such that ι (Σ ) = Σ and µ ◦ ι − = µ , a unitary mapping U : K → H , and measurable mapping M → C , k c k with | c k | = 1 such that f k = c k U g ι ( k ) for all k ∈ M . This notion of unitarily equivalence generalizesa similar concept from [1]. Unitarily equivalence is important to us since it preserves many ofthe properties we are interested in, e.g., the frame property, including the frame bounds. Thefollowing lemma tells us that “pairwise” unitarily equivalence preserves the property of beingdual frames. Lemma 2.4.
Let { f k } k ∈ M and { ˜ f k } k ∈ M be families in H , and let { g k } k ∈ M and { ˜ g k } k ∈ M befamilies in K . Suppose that f k = c k U g ι ( k ) and ˜ f k = c k U ˜ g ι ( k ) for some point isomorphism ι : M → M , a unitary mapping U : K → H , and a measurablemapping M → C , k c k with | c k | = 1 for k ∈ M . Then { f k } k ∈ M and { ˜ f k } k ∈ M are dualframes with respect to ( M , Σ , µ ) if, and only if, { g k } k ∈ M and { ˜ g k } k ∈ M are dual frames withrespect to ( M , Σ , µ ) . date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups Proof.
Assume that { f k } k ∈ M and { ˜ f k } k ∈ M are a pair of dual frames. Since the composition ofmeasurable functions is again measurable, then by our assumptions it follows that { g k } k ∈ M and { ˜ g k } k ∈ M are weakly measurable. They are obviously Bessel families. For f ∈ K and g ∈ H wecompute: h f, U ∗ g i = h U f, g i = Z M h U f, ˜ f k i h f k , g i dµ ( k ) = Z M (cid:10) U f, c k U ˜ g ι ( k ) (cid:11)(cid:10) c k U g ι ( k ) , g (cid:11) dµ ( k )= Z M (cid:10) f, ˜ g ι ( k ) (cid:11)(cid:10) g ι ( k ) , U ∗ g (cid:11) dµ ( k ) = Z M h f, ˜ g k ih g k , U ∗ g i dµ ( k ) , where the last equality follows from the properties of the point isomorphism. Since U ∗ is invertibleon all of K , this implies that { g k } k ∈ M and { ˜ g k } k ∈ M are dual frames. The opposite implicationfollows by symmetry.If µ M is the counting measure and Σ M = 2 M the discrete σ -algebra, we say that { f k } k ∈ M is a discrete frame whenever (2.4) is satisfied; for this measure space, any family of vectors isobviously weakly measurable. For discrete frames, equation (2.6) holds in the usual strong sense,i.e., with (unconditional) convergence in the H norm.Lastly, we combine the notion of continuous frames with that of generalized frames, alsoknown as g-frames. Let ( M j , Σ j , µ j ) be a measure space for each j ∈ J , where J ⊂ Z is acountable index set. We will say that a union ∪ j ∈ J { f j,k } k ∈ M j is a g-frame for H , or simply aframe, with respect to { L ( M j , µ j ) : j ∈ J } if(a) k f j,k , M j → H is measurable for each j ∈ J , and(b) there exist constants A, B > such that A k f k ≤ X j ∈ J Z M j (cid:12)(cid:12) h f, f j,k i (cid:12)(cid:12) dµ M j ( k ) ≤ B k f k for all f ∈ H . (2.7)The above definition and statements about continuous frames carry over to continuous g-frames;we refer to the original paper by Sun [47] for a detailed account of g-frames. Lemma 2.4 is alsoeasily transferred to this new setup. We will repeatedly use that it is sufficient to verify thevarious frame properties on a dense subset of H . The precise statement is as follows. Lemma 2.5.
Let D be a dense subset of H , and let ( M j , µ j ) be a measure space for each j ∈ J .(i) Suppose that ∪ j ∈ J { f j,k } k ∈ M j and ∪ j ∈ J { g j,k } k ∈ M j are Bessel families in H . If, for f ∈ D , h f, f i = X j ∈ J Z M j h f, f j,k ih g j,k , f i dµ M j ( k ) , (2.8) then equation (2.8) holds for all f ∈ H , i.e., ∪ j ∈ J { f j,k } k ∈ M j and ∪ j ∈ J { g j,k } k ∈ M j are dualframes.(ii) Suppose that ( M j , µ M j ) are σ -finite and ∪ j ∈ J { f j,k } k ∈ M j weakly measurable. If, for f ∈ D , h f, f i = X j ∈ J Z M j h f, f j,k ih f j,k , f i dµ M j ( k ) , (2.9) then equation (2.9) holds for all f ∈ H , i.e., ∪ j ∈ J { f j,k } k ∈ M j is a Parseval frame. date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups Proof. (i): The first statement follows by a straightforward generalization of the proof of the sameresult for discrete frames [19, Lemma 7]. The duality of ∪ j ∈ J { f j,k } k ∈ M j and ∪ j ∈ J { g j,k } k ∈ M j follows then by polarization.(ii): Without loss of generality we can assume that the measure space ( M j , µ M j ) is bounded foreach j ∈ J . By use of Lebesgue’s bounded convergence theorem, equation (2.9) for f ∈ D impliesthat ∪ j ∈ J { f j,k } k ∈ M j is a Bessel family on all of H ; a similar argument can be found in the proofof [43, Proposition 2.5]. The result now follows from (i). In this section we will work with generalized translation invariant systems ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j ,introduced in Section 2.2, in the setting of continuous g-frames. In order to do this, we let ( P j , Σ P j , µ P j ) be a σ -finite measure space for each j ∈ J , where J ⊂ Z is a countable indexset. For a topological space T , we let B T denote the Borel algebra of T . We now consider M j := P j × Γ j , and let Σ M j := Σ P j ⊗ B Γ j and µ M j := µ P j ⊗ µ Γ j denote the product algebra andthe product measure on P j × Γ j , respectively.We will work under the following standing hypotheses on the generalized translation in-variant system ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j . For each j ∈ J :(I) ( P j , Σ P j , µ P j ) is a σ -finite measure space,(II) the mapping p g p , ( P j , Σ P j ) → ( L ( G ) , B L ( G ) ) is measurable,(III) the mapping ( p, x ) g p ( x ) , ( P j × G, Σ P j ⊗ B G ) → ( C , B C ) is measurable.Consider T γ g p as a function of ( p, γ ) ∈ P j × Γ j into L ( G ) . This function is continuousin γ and measurable in p . Such functions are sometimes called Carathéodory functions, andsince Γ j ⊂ G is a second countable metric space, it follows that any Carathéodory function, inparticular T γ g p , is jointly measurable on ( M j , Σ M j ) = ( P j × Γ j , Σ P j ⊗ B Γ j ) . Thus, the familyof functions ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is automatically weakly measurable. A generalized translationinvariant system is therefore a frame for L ( G ) if (2.7) is satisfied with respect to the measurespaces ( M j , Σ M j , µ M j ) . Similar conclusions are valid with respect to generalized translationinvariant systems being Bessel families, Parseval frames, etc. Let us here just observe that fordual frames ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j , we have the reproducing formula f = X j ∈ J Z P j Z Γ j h f, T γ g p i T γ h p dµ Γ j ( γ ) dµ P j ( p ) for f ∈ L ( G ) , where the measure on Γ j is chosen so that the measure on Γ ⊥ j is the counting measure. Remark . In Section 3 we always assume the three standing hypotheses. However, in manyspecial cases these assumptions are automatically satisfied:(a) When P j is countable for all j ∈ J , we will equip it with a scaled counting measure kµ c , k > , and the discrete σ -algebra P j . If all P j , j ∈ J , are countable, all three standinghypotheses therefore trivially hold.(b) If P j is a second countable metric space for all j ∈ J and if p g p is continuous, thenthe standing hypotheses (II) and (III) are satisfied. Hence, if P j is also a subset of G or b G equipped with their respective Haar measure, then all three standing hypotheses hold.
10 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
The main characterization results are stated in Theorem 3.4 and 3.5. These results rely onthe following technical assumption.
Definition 3.1.
We say that two generalized translation invariant systems ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j satisfy the dual α local integrability condition (dual α -LIC) if, for all f ∈ D , X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G (cid:12)(cid:12) ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ . (3.1)In case g p = h p we refer to (3.1) as the α local integrability condition ( α -LIC) for the generalizedtranslation invariant system ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j .The α -LIC should be compared to the local integrability condition for generalized shift in-variant systems introduced in [30] for L ( R n ) and in [36] for L ( G ) . For generalized translationinvariant systems ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j the local integrability conditions (LIC) becomes X j ∈ J Z P j X α ∈ Γ ⊥ j Z supp ˆ f (cid:12)(cid:12) ˆ f ( ωα )ˆ g p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ for all f ∈ D . (3.2)Since the integrands in (3.1) and (3.2) are measurable on P j × b G , we are allowed to reorder sumsand integrals in the local integrability conditions.We will see (Lemma 3.9 and Example 1) that the LIC implies the α -LIC, but not vice versa.Moreover, we mention that dual local integrability conditions have not been considered in theliterature before. The following simple observation will often be used. Lemma 3.2.
The following assertions are equivalent:(i) The systems ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j satisfy the dual α -LIC,(ii) for each compact subset K ⊆ b G \ E X j ∈ J Z P j X α ∈ Γ ⊥ j Z K ∩ α − K (cid:12)(cid:12) ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ . Proof.
To show that (i) implies (ii), let K be any compact subset in b G and define ˆ f = K . Then,by assumption, X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G (cid:12)(cid:12) ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p )= X j ∈ J Z P j X α ∈ Γ ⊥ j Z K ∩ α − K (cid:12)(cid:12) ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ . To show that (ii) implies (i), take f ∈ D and denote supp ˆ f by K . Note that ˆ f ∈ L ∞ ( b G ) . Hence,we find that X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G (cid:12)(cid:12) ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p )
11 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups ≤ k ˆ f k ∞ X j ∈ J Z P j X α ∈ Γ ⊥ j Z K ∩ α − K (cid:12)(cid:12) ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ . In a similar way, we see that ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j satisfies the local integrability condition if,and only if, for each compact subset K ⊆ b G \ E X j ∈ J Z P j X α ∈ Γ ⊥ j Z K ∩ α − K (cid:12)(cid:12) ˆ g p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ . (3.3)Inspired by the definition of the Calderón sum in wavelet theory, we will say that the term P j ∈ J R P j | ˆ g p ( ω ) | dµ P j ( p ) is the Calderón integral . The next result shows that the Calderónintegral is bounded if the generalized translation invariant system is a Bessel family. Fromthis it follows that the t α -equations (3.6) are well-defined. We remark that Proposition 3.3generalizes [36, Proposition 3.6] and [30, Proposition 4.1] from the uniform lattice setting whereeach P j is countable to the setting of generalized translation invariant systems. Proposition 3.3.
If the generalized translation invariant system ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is a Besselfamily with bound B , then X j ∈ J Z P j | ˆ g p ( ω ) | dµ P j ( p ) ≤ B for a.e. ω ∈ b G. (3.4) Proof.
We begin by noting that the Calderón integral in (3.4) is well-defined by our standinghypothesis (III). We assume without loss of generality that J = Z . From the Bessel assumptionon ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j , we have X | j |≤ M Z P j Z Γ j (cid:12)(cid:12) h f, T γ g p i (cid:12)(cid:12) dµ Γ j ( γ ) dµ P j ( p ) ≤ B k f k for every M ∈ N and all f ∈ L ( G ) . By Lemma 2.2 we then get X | j |≤ M Z P j X α ∈ Γ ⊥ j Z b G ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ g p ( ωα ) dµ b G ( ω ) dµ P j ( p ) ≤ B k f k (3.5)for every M ∈ N and all f ∈ D . Assume towards a contradiction that there exists a Borel subset N ⊂ b G of positive measure µ b G ( N ) > for which X j ∈ J Z P j | ˆ g p ( ω ) | dµ P j ( p ) > B for a.e. ω ∈ N. In [36] it is assumed that N contains an open ball, but this needs not be the case. However,since b G is σ -compact, there exists a compact set K so that µ b G ( K ∩ N ) > . Set δ M :=inf { d ( α,
1) : α ∈ Γ ⊥ j \ { } , | j | ≤ M } . For any discrete subgroup Γ there exists a δ > such that B ( x, δ ) ∩ Γ = { x } for x ∈ Γ , where B ( x, δ ) denotes the open ball of radius δ and center x . Itfollows that δ M > since δ M is the smallest of such radii about x = 1 from a finite union ofdiscrete subgroups Γ ⊥ j . Let O be an open covering of K of sets with diameter strictly less than δ M / .
12 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Since a finite subset of O covers K , there is an open set B ∈ O so that µ b G ( B ∩ K ∩ N ) > .Define f ∈ L ( G ) by ˆ f = B ∩ K ∩ N . By Remark 4 below, we can assume that E does not intersect the closure of B ∩ K ∩ N . Therefore, f ∈ D and by our assumption we have X | j |≤ M Z P j X α ∈ Γ ⊥ j Z b G ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ g p ( ωα ) dµ b G ( ω ) dµ P j ( p )= Z b G | ˆ f ( ω ) | X | j |≤ M Z P j | ˆ g p ( ω ) | dµ P j ( p ) dµ b G ( ω ) , where the change of the order of integration above is justified by an application of the Fubini-Tonelli theorem together with the Bessel assumption (3.4) and our standing hypotheses (I) and(III). By letting M tend to infinity, we see that X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ g p ( ωα ) dµ b G ( ω ) dµ P j ( p ) > B k f k , which contradicts (3.5). Remark . In case E intersects the closure of A := B ∩ K ∩ N in the proof of Proposition 3.3,one needs to approximate the function f with functions from D as defined in (2.3). As we willuse such arguments several times in the remainder of this paper, let us consider how to do sucha modification in this specific case. Define E A = E ∩ A and F n = (cid:8) ω ∈ A : inf { d ( ω, a ) : a ∈ E A } < n (cid:9) , for each n ∈ N . Define ˆ f n = A \ F n ∈ D . Since F n +1 ⊂ F n and µ b G ( F ) < ∞ , we have k ˆ f − ˆ f n k = µ b G ( F n ) → µ b G ( ∩ n ∈ N F n ) = µ b G ( E A ) = 0 as n → ∞ , where ˆ f = B ∩ K ∩ N . Finally, we use ˆ f n in place of ˆ f in the final argument of the proof above,and let n → ∞ . We are ready to prove the first of our main results, Theorem 3.4. Under the technical dual α -LIC assumption we characterize dual generalized translation invariant frames in terms of t α -equations. We stress that these GTI systems are dual frames with respect to { L ( M j , µ j ) : j ∈ J } defined in the previous section. Recall that we assume a Haar measure on G to be given, andthat we equip every Γ ⊥ j ⊂ b G with the counting measure. Theorem 3.4.
Suppose that ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j are Bessel familiessatisfying the dual α -LIC. Then the following statements are equivalent:(i) ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j are dual frames for L ( G ) ,(ii) for each α ∈ S j ∈ J Γ ⊥ j we have t α ( ω ) := X j ∈ J : α ∈ Γ ⊥ j Z P j ˆ g p ( ω )ˆ h p ( ωα ) dµ P j ( p ) = δ α, a.e. ω ∈ b G. (3.6)
13 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Proof.
Let us first show that the t α -equations are well-defined. Take B to be a common Besselbound for the two GTI families. By two applications of the Cauchy-Schwarz inequality andProposition 3.3, we find that X j ∈ J : α ∈ Γ ⊥ j Z P j | ˆ g p ( ω ) || ˆ h p ( ωα ) | dµ P j ( p ) ≤ X j ∈ J Z P j | ˆ g p ( ω ) || ˆ h p ( ωα ) | dµ P j ( p ) ≤ X j ∈ J (cid:16) Z P j | ˆ g p ( ω ) | dµ P j ( p ) (cid:17) / (cid:16) Z P j | ˆ h p ( ωα ) | dµ P j ( p ) (cid:17) / ≤ (cid:16) X j ∈ J Z P j | ˆ g p ( ω ) | dµ P j ( p ) (cid:17) / (cid:16) X j ∈ J Z P j | ˆ h p ( ωα ) | dµ P j ( p ) (cid:17) / ≤ B, for a.e. ω ∈ b G . This shows that the t α -equations are well-defined and converge absolutely.For f ∈ D , define the function w f : G → C , w f ( x ) := X j ∈ J Z P j Z Γ j h T x f, T γ g p ih T γ h p , T x f i dµ Γ j ( γ ) dµ P j ( p ) . (3.7)By Lemma 2.2 and the calculation d T x f ( ω ) d T x f ( ωα ) = α ( x ) ˆ f ( ω ) ˆ f ( ωα ) , we have w f ( x ) = X j ∈ J Z P j Z b G X α ∈ Γ ⊥ j α ( x ) ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) dµ b G ( ω ) dµ P j ( p ) . Let ϕ α,j ( p, ω ) denote the innermost summand in the right hand side expression above. Byour standing hypothesis (III), the function ϕ α,j is (Σ P j ⊗ B G ) -measurable for each α . ApplyingBeppo Levi’s theorem to the dual α local integrability condition yields that the function P α ϕ α,j belongs to L ( P j × b G ) for each j ∈ J . An application of Fubini’s theorem now gives: w f ( x ) = X j ∈ J Z b G Z P j X α ∈ S j ∈ J Γ ⊥ j Γ ⊥ j ( α ) ϕ α,j ( p, ω ) dµ P j ( p ) dµ b G ( ω ) . Lebesgue’s dominated convergence theorem then yields: w f ( x ) = X j ∈ J X α ∈ S j ∈ J Γ ⊥ j α ( x ) Z b G Z P j Γ ⊥ j ( α ) ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) dµ P j ( p ) dµ b G ( ω ) . By the dual α local integrability condition the summand belongs to ℓ ( J × ∪ j ∈ J Γ ⊥ j ) and we cantherefore interchange the order of summations. Further, by Lebesgue’s bounded convergencetheorem, we can interchange the sum over j ∈ J and the integral over supp ˆ f ⊂ b G . Hence, w f ( x ) = X α ∈ S j ∈ J Γ ⊥ j α ( x ) Z b G ˆ f ( ω ) ˆ f ( ωα ) X j ∈ J : α ∈ Γ ⊥ j Z P j ˆ g p ( ω )ˆ h p ( ωα ) dµ P j ( p ) dµ b G ( ω ) . Finally, we arrive at: w f ( x ) = X α ∈ S j ∈ J Γ ⊥ j α ( x ) ˆ w ( α ) , where ˆ w ( α ) := Z b G ˆ f ( ω ) ˆ f ( ωα ) t α ( ω ) dµ b G ( ω ) . (3.8)
14 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
From the previous calculations and the dual α -LIC, it follows that the convergence in (3.8) isabsolute. By the Weierstrass M-test, we see that w f is the uniform limit of a generalized Fourierseries and thus an almost periodic, continuous function.We start by showing the implication (ii) ⇒ (i). Inserting (3.6) into (3.8) for x = 0 yields w f (0) = X j ∈ J Z P j Z Γ j h f, T γ g p ih T γ h p , f i dµ Γ j ( γ ) dµ P j ( p )= X α ∈ S j ∈ J Γ ⊥ j α (0) Z b G ˆ f ( ω ) ˆ f ( ωα ) δ α, dµ b G ( ω ) = h f, f i , and (i) follows by Lemma 2.5(i).For the converse implication (i) ⇒ (ii), we have w f ( x ) = X j ∈ J Z P j Z Γ j h T x f, T γ g p ih T γ h p , T x f i dµ Γ j ( γ ) dµ P j ( p ) = k f k for each f ∈ D . Consider now the function z ( x ) := w f ( x ) − k f k . We have shown that w f is continuous and by construction z is identical to the zero function. Additionally, since w f equals an absolute convergent, generalized Fourier series, also z can be expressed as an absoluteconvergent generalized Fourier series z ( x ) = X α ∈ S j ∈ J Γ ⊥ j α ( x )ˆ z ( α ) , with ˆ z ( α ) = Z b G (cid:12)(cid:12) ˆ f ( ω ) (cid:12)(cid:12) t ( ω ) dµ b G ( ω ) − k f k for α = 1 , Z b G ˆ f ( ω ) ˆ f ( ωα ) t α ( ω ) dµ b G ( ω ) for α ∈ S j ∈ J Γ ⊥ j \ { } . By the uniqueness theorem for generalized Fourier series [13, Theorem 7.12], the function z ( x ) is identical to zero if, and only if, ˆ z ( α ) = 0 for all α ∈ ∪ j ∈ J Γ ⊥ j .In case α = 1 we have R b G | ˆ f ( ω ) | (cid:0) t ( ω ) − (cid:1) dµ b G ( ω ) = 0 for f ∈ D . Hence, since D is densein L ( G ) , we conclude that t ( ω ) = 1 for a.e. ω ∈ b G . For α ∈ ∪ j ∈ J Γ ⊥ j \ { } , we have Z b G ˆ f ( ω ) ˆ f ( ωα ) t α ( ω ) dµ b G ( ω ) = 0 . (3.9)Define the multiplication operator M t α : L ( b G ) → L ( b G ) by M t α ˆ f ( ω ) = t α ( ω ) ˆ f ( ω ) . This linearoperator is bounded since by Proposition 3.3 t α ( ω ) ∈ L ∞ ( b G ) . We can now rewrite the left handside of (3.9) as an inner-product: h ˆ f , M t α T α − ˆ f i L ( b G ) = 0 , where f ∈ D . Since D is dense in the complex Hilbert space L ( G ) , this implies that M t α T α − =0 . After multiplication with T α from the right, we have M t α = 0 and therefore t α = 0 .From Theorem 3.4 we easily obtain the corresponding characterization for tight frames. Westate it for Parseval frames only as it is just a matter of scaling. Theorem 3.5.
Suppose that the generalized translation invariant system ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j sa-tisfies the α local integrability condition. Then the following assertions are equivalent:
15 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups (i) ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is a Parseval frame for L ( G ) ,(ii) for each α ∈ S j ∈ J Γ ⊥ j we have t α := X j ∈ J : α ∈ Γ ⊥ j Z P j ˆ g p ( ω )ˆ g p ( ωα ) dµ P j ( p ) = δ α, a.e. ω ∈ b G. Proof.
We first remark that the integrals in (ii) indeed converge absolutely. This follows fromtwo applications of the Cauchy-Schwarz’ inequality (as in the proof of Theorem 3.4), which gives: X j ∈ J : α ∈ Γ ⊥ j Z P j | ˆ g p ( ω ) | | ˆ g p ( ωα ) | dµ P j ( p ) ≤ . In view of Theorem 3.4 we only have to argue that the assumption on the Bessel family canbe omitted. If we assume (i), then clearly ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is a Bessel family and (ii) followsfrom Theorem 3.4.Suppose that (ii) holds. Formula (3.8) is still valid, where w f is defined as in (3.7) with h p = g p . Setting x = 0 in (3.8) yields k f k = X j ∈ J Z P j Z Γ j |h f, T γ g p i| dµ Γ j ( γ ) dµ P j ( p ) for all f ∈ D .Finally, we conclude by Lemma 2.5(ii) that ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is a Parseval frame for L ( G ) .By virtue of Lemma 2.4, we have the following extension of Theorem 3.4 and 3.5. Corollary 3.6.
The characterization results in Theorem 3.4 and 3.5 extend to systems that areunitarily equivalent to generalized translation invariant systems.
Let us now turn to sufficient conditions for a generalized translation invariant system to be aBessel family or a frame. Proposition 3.7 is a generalization of the results in, e.g., [10] and [9],which state the corresponding result for GSI systems in the euclidean space and locally compactabelian groups, respectively. The result is as follows.
Proposition 3.7.
Consider the generalized translation invariant system ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j .(i) If B := ess sup ω ∈ b G X j ∈ J Z P j X α ∈ Γ ⊥ j (cid:12)(cid:12) ˆ g p ( ω )ˆ g p ( ωα ) (cid:12)(cid:12) dµ P j ( p ) < ∞ , (3.10) then ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is a Bessel family with bound B .(ii) Furthermore, if also A := ess inf ω ∈ b G (cid:16) X j ∈ J Z P j | ˆ g p ( ω ) | dµ P j ( p ) − X j ∈ J Z P j X α ∈ Γ ⊥ j \{ } | ˆ g p ( ω )ˆ g p ( ωα ) | dµ P j ( p ) (cid:17) > , then ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j is a frame for L ( G ) with bounds A and B .
16 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Proof.
With a few adaptations the result follows from the corresponding proofs in [9] and [10].We refer to (3.10) as the absolute CC-condition, or for short, CC-condition [7]. Proposition3.7 is useful in applications as a mean to verify that a given family indeed is Bessel, or evena frame. Moreover, in relation to the characterizing results in Theorem 3.4 and 3.5, the CC-condition (3.10) is sufficient for the α -LIC to hold. In contrast, we remark that, by Example 1in Section 3.3, the CC-condition does not imply the LIC. Lemma 3.8. If ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j satisfy ess sup ω ∈ b G X j ∈ J Z P j X α ∈ Γ ⊥ j (cid:12)(cid:12) ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ P j ( p ) < ∞ and ess sup ω ∈ b G X j ∈ J Z P j X α ∈ Γ ⊥ j (cid:12)(cid:12) ˆ g p ( ωα )ˆ h p ( ω ) (cid:12)(cid:12) dµ P j ( p ) < ∞ , then the dual α local integrability condition is satisfied. Furthermore, if ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j sa-tisfies the CC-condition (3.10), then the α local integrability condition is satisfied.Proof. By applications of Cauchy-Schwarz’ inequality, we find X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G | ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) | dµ b G ( ω ) dµ P j ( p ) ≤ (cid:20) X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G | ˆ f ( ω ) | | ˆ g p ( ω )ˆ h p ( ωα ) | dµ b G ( ω ) dµ P j ( p ) (cid:21) / × (cid:20) X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G | ˆ f ( ωα ) | | ˆ g p ( ω )ˆ h p ( ωα ) | dµ b G ( ω ) dµ P j ( p ) (cid:21) / = (cid:20) Z b G | ˆ f ( ω ) | X j ∈ J Z P j X α ∈ Γ ⊥ j | ˆ g p ( ω )ˆ h p ( ωα ) | dµ P j ( p ) dµ b G ( ω ) (cid:21) / × (cid:20) Z b G | ˆ f ( ω ) | X j ∈ J Z P j X α ∈ Γ ⊥ j | ˆ g p ( ωα )ˆ h p ( ω ) | dµ P j ( p ) dµ b G ( ω ) (cid:21) / < ∞ . Finally, we show that the LIC implies the (dual) α -LIC. The precise statement is as follows. Lemma 3.9.
If both ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j satisfy the local integrabilitycondition (3.2) , then ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j satisfy the dual α local inte-grability condition. In particular, if ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j satisfies the local integrability condition,then it also satisfies the α local integrability condition.Proof. By use of Cauchy-Schwarz’ inequality and | cd | ≤ | c | + | d | , we have X j ∈ J Z P j X α ∈ Γ ⊥ j Z b G (cid:12)(cid:12) ˆ f ( ω ) ˆ f ( ωα )ˆ g p ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p )
17 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups ≤ X j ∈ J Z P j X α ∈ Γ ⊥ j (cid:18) Z α − supp ˆ f (cid:12)(cid:12) ˆ f ( ω )ˆ h p ( ωα ) (cid:12)(cid:12) dµ b G ( ω ) (cid:19) / (cid:18) Z supp ˆ f (cid:12)(cid:12) ˆ f ( ωα )ˆ g p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) (cid:19) / dµ P j ( p )= X j ∈ J Z P j X α ∈ Γ ⊥ j (cid:18) Z supp ˆ f (cid:12)(cid:12) ˆ f ( ωα − )ˆ h p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) (cid:19) / (cid:18) Z supp ˆ f (cid:12)(cid:12) ˆ f ( ωα )ˆ g p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) (cid:19) / dµ P j ( p ) ≤ X j ∈ J Z P j X α ∈ Γ ⊥ j Z supp ˆ f (cid:12)(cid:12) ˆ f ( ωα − )ˆ h p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p )+ X j ∈ J Z P j X α ∈ Γ ⊥ j Z supp ˆ f (cid:12)(cid:12) ˆ f ( ωα )ˆ g p ( ω ) (cid:12)(cid:12) dµ b G ( ω ) dµ P j ( p ) < ∞ , and the statements follow.The relationships between the various conditions considered above are summarized in thediagram below. To simplify the presentation we do not consider dual frames. An arrow meansthat the assumption at the tail of the arrow implies the assumption at the head. A crossed outarrow means that one can find a counter example for that implication; clearly, implications tothe left in the top line are also not true in general.CC Bessel Calderón integral < B LIC α -LIC ( t α -eqns. ⇔ Parseval ) X X X The crossed out arrows are shown by Example 1 and Example 2 in the next section.
In this section we consider two key examples. Both examples take place in ℓ ( Z ) ; however, theycan be extended to L ( R ) , see [5]. The first example, Example 1, shows that for a GTI systemthe α local integrability condition is strictly weaker than the local integrability condition. Example 1.
Let G = Z , N ∈ N , N ≥ and consider the co-compact subgroups Γ j = N j Z , j ∈ N . Note that b G can be identified with the half-open unit interval [0 , under addition modulo one.To each Γ j we associate N j functions g j,p , for p = 0 , , . . . , N j − . Each function g j,p is definedby its Fourier transform ˆ g j,p = ( N − / N − j/ [ p/N j , ( p +1) /N j ) . The factor ( N − / is for normalization purposes and does not play a role in the calculations.The annihilator of each Γ j is given by Γ ⊥ j = N − j Z ∩ [0 , . Note that the number of elements in Γ ⊥ j is N j . We equip both G and Γ ⊥ j with the counting measure, this implies that the measureon Γ j is the counting measure multiplied by N j . For the generalized translation invariant system ∪ j ∈ N { T γ g j,p } γ ∈ Γ j ,p =0 , ,...,N j − we show the following: (i) the LIC is violated, (ii) the α -LICholds, (iii) the system is a Parseval frame for ℓ ( Z ) . It then follows from Theorem 3.5 that the t α -equations are satisfied.Ad (i). In order for the LIC to hold we need ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j Z K ∩ ( K − α ) | ˆ g j,p ( ω ) | dω < ∞
18 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups for all compact K ⊆ [0 , , see Lemma 3.2. In particular for K = b G , we find ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j Z | ˆ g j,p ( ω ) | dω = ( N − ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j N − j = ( N − ∞ X j =1 N j − X p =0 N − j = ( N − ∞ X j =1 ∞ . Therefore, the local integrability condition is not satisfied.Ad (ii). By Lemma 3.2 it suffices to show that ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j Z b G ∩ ( b G − α ) (cid:12)(cid:12) ˆ g j,p ( ω )ˆ g j,p ( ω + α ) (cid:12)(cid:12) dω < ∞ . Due to the support of ˆ g j,p we have | ˆ g j,p ( ω )ˆ g j,p ( ω + α ) | = 0 for α ∈ Γ ⊥ j \ { } . We thus find that ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j Z | ˆ g j,p ( ω )ˆ g j,p ( ω + α ) | dω = ∞ X j =1 N j − X p =0 Z | ˆ g j,p ( ω ) | dω = ( N − ∞ X j =1 N j − X p =0 N − j = ( N − ∞ X j =1 N − j = 1 . Ad (iii). Note that P N j − p =0 | ˆ g j,p ( ω ) | = ( N − N − j [0 , ( ω ) for ω ∈ [0 , and for all j ∈ N .Using the frame bound estimates from Proposition 3.7, we have B = ess sup ω ∈ [0 , ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j | ˆ g j,p ( ω )ˆ g j,p ( ω + α ) | = ess sup ω ∈ [0 , ∞ X j =1 N j − X p =0 | ˆ g j,p ( ω ) | = ess sup ω ∈ [0 , ( N − ∞ X j =1 N − j [0 , ( ω ) = 1 . In the same way, for the lower frame bound, we find A = ess inf ω ∈ [0 , (cid:16) ∞ X j =1 N j − X p =0 | ˆ g j,p ( ω ) | − ∞ X j =1 N j − X p =0 X α ∈ Γ ⊥ j \{ } | ˆ g j,p ( ω )ˆ g j,p ( ω + α ) | (cid:17) = 1 . These calculations also show that ∪ j ∈ N { T γ g j,p } γ ∈ Γ j ,p =0 , ,...,N j − is actually a union over j ∈ N oftight frames { T γ g j,p } γ ∈ Γ j ,p =0 , ,...,N j − each with frame bound N − j . Furthermore, we see that theCC-condition is satisfied, even though the LIC fails. Hence, the CC-condition does not implyLIC (however, by Lemma 3.8 it does imply the α -LIC).The following example is inspired by similar constructions in [5] and [36]. It shows two points.Firstly, the α local integrability condition cannot be removed in Theorem 3.5. Secondly, it ispossible for a GTI Parseval frame to satisfy the t α -equations even though the α local integrabilitycondition fails. We show these observations in the reversed order.
19 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Example 2.
Let G = Z and for each m ∈ Z and k ∈ N , let [ m ] k denote the residue class of m modulo k . Then, for τ j = 2 j − − , j ∈ N , Z = [ j ∈ N [ τ j ] j = [0] ∪ [1] ∪ [3] ∪ [7] ∪ [15] . . . , where the union is disjoint. Now set g j = N − j/ τ j and Γ j = N j Z for N = 2 . The GTI system ∪ j ∈ N { T γ g j } γ ∈ Γ j is essentially a reordering of the standard orthonormal basis { e k } k ∈ Z for ℓ ( Z ) .The factor N − j/ in the definition of g j is due to the fact that we equip Γ ⊥ j with the countingmeasure. This implies that the measure on Γ j becomes N j times the counting measure. One cannow show that this GTI system does not satisfy the α -LIC. However, the system does indeedsatisfy the t α -equations. For α = 0 : ∞ X j =1 | ˆ g j ( ω ) | = ∞ X j =1 − j (cid:12)(cid:12) e πiτ j ω (cid:12)(cid:12) = 12 − , and for α = k/ j ∗ ∈ − j ∗ Z = Γ ⊥ j ∗ , where k is odd, X j ∈ J : α ∈ Γ ⊥ j ˆ g j ( ω )ˆ g j ( ω + α ) = ∞ X j = j ∗ − j e − πi k j ∗ (2 j − − = e πik − j ∗ ∞ X j = j ∗ − j e − πi j − j ∗ = e πik − j ∗ (cid:18) − − j ∗ + ∞ X j = j ∗ +1 − j (cid:19) = 0 . If one uses N ≥ , N ∈ N in place of N = 2 , then the α -LIC is still not satisfied. However,even though for suitably chosen τ j (the formula is more complicated than for N = 2 , see [5]) ∪ j ∈ N { T γ N − j/ τ j } γ ∈ N j Z is still essentially a reordering of the standard orthonormal basis, every t α -equation is false. The case α = 0 gives t α = N − = 1 , while the cases α = 0 give t α = 0 .We stress that these examples show the existence of generalized translation invariant Parsevalframes for ℓ ( Z ) which do not satisfy the t α -equations. Under special circumstances the local integrability condition will be satisfied automatically. Inthis section we will see that this is indeed the case for TI systems, i.e., Γ j = Γ for all j ∈ J , andfor GTI systems on compact abelian groups G . For brevity, we will only state the correspondingcharacterization results for dual frames, but remark here that the results hold equally for Parsevalframes, in which case, the Bessel family assumption can be omitted.Let us begin with a lemma concerning general GTI systems for LCA groups showing thatthe LIC holds if the annihilators of Γ j possess a sufficient amount of separation. Lemma 3.10. If ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j has a uniformly bounded Calderón integral and if thereexists a constant C > such that for all compact K ⊆ b G X α ∈ S j ∈ J Γ ⊥ j µ b G ( K ∩ α − K ) ≤ C, then ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j satisfies the local integrability condition.
20 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Proof.
By assumption there exists a constant
B > such that P j ∈ J R P j | ˆ g p ( ω ) | dµ P j ( p ) < B for a.e. ω ∈ b G , and we therefore have X j ∈ J Z P j X α ∈ Γ ⊥ j Z K ∩ α − K | ˆ g p ( ω ) | dµ b G ( ω ) dµ P j ( p )= X α ∈ S j ∈ J Γ ⊥ j Z K ∩ α − K X j ∈ J Z P j | ˆ g p ( ω ) | dµ b G ( ω ) dµ P j ( p ) ≤ BC < ∞ . Now, let us consider the case where all subgroups Γ j coincide. In other words, we consider translation invariant systems. Note that this setting includes the continuous wavelet and Gabortransform as well as the shift invariant systems considered in [30, 36]. Theorem 3.11.
Let Γ be a co-compact subgroup in G . Suppose that ∪ j ∈ J { T γ g p } γ ∈ Γ ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ ,p ∈ P j are Bessel families. Then the following statements are equivalent:(i) ∪ j ∈ J { T γ g p } γ ∈ Γ ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ ,p ∈ P j are dual frames for L ( G ) ,(ii) For each α ∈ Γ ⊥ we have t α ( ω ) := X j ∈ J Z P j ˆ g p ( ω )ˆ h p ( ωα ) dµ P j ( p ) = δ α, a.e. ω ∈ b G. (3.11) Proof.
Since Γ ⊥ is a discrete subgroup in b G and since the metric on b G is translation invariant,there exists a δ > so that the distance between two distinct points from Γ ⊥ is larger than δ . Thus, for any compact K ⊂ b G , the set Γ ⊥ ∩ ( K − K ) has finite cardinality because, if not,then Γ ⊥ ∩ ( K − K ) would contain a sequence (take one without repetitions) with no convergentsubsequence which contradicts the compactness of K . Since { α ∈ Γ ⊥ : Kα ∩ K = ∅} is a subsetof Γ ⊥ ∩ ( K − K ) , it is also of finite cardinality. From this together with the Bessel assumptionand Proposition 3.3 we conclude that the assumptions of Lemma 3.10 are satisfied and hence theLIC holds. By Lemma 3.9 the dual α -LIC is satisfied and the result now readily follows fromTheorem 3.4.For TI systems with translation along the entire group Γ = G there is only one t α -equationin (3.11) since G ⊥ = { } . To be precise: Lemma 3.12.
Suppose that
Γ = G . Then assertion (ii) in Theorem 3.11 reduces to X j ∈ J Z P j ˆ g p ( ω )ˆ h p ( ω ) dµ P j ( p ) = 1 a.e. ω ∈ b G. Let us now turn to the familiar setting of [30,36], where Γ is a uniform lattice, i.e., a discrete,co-compact subgroup. Then there is a compact fundamental domain F ⊂ G for Γ , such that G = F Γ , and moreover for any x ∈ G we have x = ϕγ , where ϕ ∈ F, γ ∈ Γ are unique. For auniform lattice we introduce the lattice size s (Γ) := µ G ( F ) , which is, in fact, independent of thechoice of F . Corollary 3.13.
Let Γ be a uniform lattice in G . Suppose that the two generalized translationinvariant systems ∪ j ∈ J { T γ g p } γ ∈ Γ ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ ,p ∈ P j are Bessel families. Then thefollowing statements are equivalent:
21 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups (i) ∪ j ∈ J { T γ g p } γ ∈ Γ ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ ,p ∈ P j are dual frames for L ( G ) , i.e., h f , f i = X j ∈ J Z P j s (Γ) X γ ∈ Γ h f , T γ g p ih T γ h p , f i dµ P j ( p ) , for all f , f ∈ L ( G ) . (3.12) (ii) For each α ∈ Γ ⊥ we have t α ( ω ) = δ α, for a.e. ω ∈ b G , where t α is defined in (3.11) Remark . In the same way, we can state the characterization results for generalized shift-invariant systems. In this case we have countable many uniform lattices Γ j , so we replace s (Γ) inCorollary 3.13 with s (Γ j ) , sum over { j ∈ J : α ∈ Γ ⊥ j } in (3.12), and add the dual α local integra-bility condition as assumption. We obtain a statement equivalent to the main characterizationresult in [36]. In contrast to the result in [36], the lattice size s (Γ) is contained in the reproducingformula rather than in the t α -equations.For compact abelian groups all generalized translation invariant systems satisfy the localintegrability condition. The characterization result is as follows. Theorem 3.14.
Let G be a compact abelian group. Suppose that ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j are Bessel families. Then the following statements are equivalent:(i) ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j and ∪ j ∈ J { T γ h p } γ ∈ Γ j ,p ∈ P j are dual frames for L ( G ) ,(ii) for each α ∈ S j ∈ J Γ ⊥ j we have t α ( ω ) := X j ∈ J : α ∈ Γ ⊥ j Z P j ˆ g p ( ω )ˆ h p ( ωα ) dµ P j ( p ) = δ α, a.e. ω ∈ b G. Proof.
Because G is compact, the dual group b G is discrete. All compact K ⊂ b G are thereforefinite. Let K denote the number of elements in K . From the LIC we then find X j ∈ J Z P j X α ∈ Γ ⊥ j X ω ∈ K ∩ α − K | ˆ g p ( ω ) | dµ P j ( p ) ≤ X j ∈ J Z P j K X ω ∈ K | ˆ g p ( ω ) | dµ P j ( p ) ≤ ( K ) max ω ∈ K X j ∈ J Z P j | ˆ g p ( ω ) | dµ P j . By the Bessel assumption and Proposition 3.3, the Calderón integral is bounded. The far righthand side in the above calculation is therefore finite, and the LIC is satisfied. The result nowfollows from Theorem 3.4 and Lemma 3.9.Finally, let us turn to discrete groups G . In this case, the local integrability condition is not automatically satisfied (as we saw in the examples in the previous section), but it has a simplereformulation: Lemma 3.15.
Suppose G is a discrete abelian group. Then the following statements are equiv-alent:(i) The system ∪ j ∈ J { T γ g p } γ ∈ Γ j ,p ∈ P j satisfies the local integrability condition,(ii) X j ∈ J Z P j µ c (Γ ⊥ j ) k g p k L ( G ) dµ P j ( p ) < ∞ , where µ c denotes the counting measure.
22 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Proof.
Note that if G is discrete, then b G is compact. Hence the discrete groups Γ ⊥ j are alsocompact and therefore finite. By this observation we can easily show the result. If (i) holds, then X j ∈ J Z P j µ c (Γ ⊥ j ) k g p k L ( G ) dµ P j ( p ) ≤ X j ∈ J Z P j Z b G µ c (Γ ⊥ j ) | ˆ g p ( ω ) | dµ b G ( ω ) dµ P j ( p ) . By (3.3) with K = b G the right hand side is finite, and (ii) follows. If (ii) holds, then X j ∈ J Z P j X α ∈ Γ ⊥ j Z K ∩ α − K | ˆ g p ( ω ) | dµ b G ( ω ) dµ P j ( p ) ≤ X j ∈ J Z P j µ c (Γ ⊥ j ) Z b G | ˆ g p ( ω ) | dµ b G ( ω ) dµ P j ( p ) < ∞ . In this section we study applications of Theorem 3.4 leading to new characterization results.Moreover, we will easily recover known results as special cases of our theory. We consider Gaborand wavelet-like systems for general locally compact abelian groups as well as for specific locallycompact abelian groups, e.g., R n , Z n , Z n . We also give an example of characterization resultsfor the locally compact abelian group of p -adic numbers, where the theory of generalized shift invariant systems is not applicable.We will focus on verifying the local integrability conditions and on the deriving the char-acterizing equations, but not on the related question of how to construct generators satisfyingthese equations. The recent work of Christensen and Goh [9] takes this more constructive ap-proach for generalized shift invariant systems on locally compact abelian groups. Under certainassumptions, they explicitly construct dual GSI frames using variants of t α -equations, which areproved to be sufficient. A Gabor system in L ( G ) with generator g ∈ L ( G ) is a family of functions of the form { E γ T λ g } γ ∈ Γ ,λ ∈ Λ , where Γ ⊆ b G and Λ ⊆ G. Note that a Gabor system { E γ T λ g } γ ∈ Γ ,λ ∈ Λ is not a generalized translation invariant systembecause E γ T λ g = T λ (cid:0) γ ( λ ) E γ g (cid:1) cannot be written as T γ g j,p for j ∈ J and p ∈ P j for any { g j,p } .However, by use of Lemma 2.4, we can establish the following two possibilities to relate Gaborand translation invariant systems.Firstly, by Lemma 2.4 with ι = id , U = F and c γ,λ = 1 , we see that the Gabor system { E γ T λ g } γ ∈ Γ ,λ ∈ Λ is a frame if, and only if, the translation invariant system { T γ F − T λ g } γ ∈ Γ ,λ ∈ Λ is a frame. By this observation all results for translation invariant systems naturally carry over toGabor systems. In order to apply the theory established in this paper, we need Γ to be a closed,co-compact subgroup of b G and Λ to be equipped with a measure µ Λ satisfying the standinghypotheses (I)–(III). This approach together with Theorem 3.4 yield t α -equations in the time domain G : for each α ∈ Γ ⊥ we have Z Λ g ( x − λ ) h ( x − λ + α ) dµ Λ ( λ ) = δ α, a.e. x ∈ G. Secondly, by Lemma 2.4 with ι = id , U = id and c γ,λ = γ ( λ ) , we see that the Gabor system { E γ T λ g } γ ∈ Γ ,λ ∈ Λ is a frame if, and only if, the translation invariant system { T λ E γ g } γ ∈ Γ ,λ ∈ Λ is a
23 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups frame. This time we need Λ to be a closed, co-compact subgroup of G and Γ to be equipped witha measure satisfying standing hypotheses (I)–(III). In contrast to the first approach, Theorem3.4 now yields t α -equations in the frequency domain b G : for each β ∈ Λ ⊥ we have Z Γ ˆ g ( ωγ )ˆ h ( ωγβ ) dµ Γ ( γ ) = δ β, a.e. ω ∈ b G. Gabor systems play a major role in time-frequency analysis [27] and it is common to requiresimilar properties on Γ and Λ . In the following theorem we characterize dual Gabor frames,where we combine both of the above approaches and require that Λ and Γ are closed, co-compactsubgroups. If we consider Parseval frames, then the Bessel assumption in Theorem 4.1 can beomitted. Theorem 4.1.
Let Λ and Γ be closed, co-compact subgroups of G and b G respectively and equip Λ ⊥ and Γ ⊥ with the counting measure. Suppose that the two systems { E γ T λ g } γ ∈ Γ ,λ ∈ Λ and { E γ T λ h } γ ∈ Γ ,λ ∈ Λ are Bessel families. Then the following statements are equivalent:(i) { E γ T λ g } γ ∈ Γ ,λ ∈ Λ and { E γ T λ h } γ ∈ Γ ,λ ∈ Λ are dual frames for L ( G ) ,(ii) for each α ∈ Γ ⊥ we have Z Λ g ( x − λ ) h ( x − λ + α ) dµ Λ ( λ ) = δ α, a.e. x ∈ G, (iii) for each β ∈ Λ ⊥ we have Z Γ ˆ g ( ωγ )ˆ h ( ωγβ ) dµ Γ ( γ ) = δ β, a.e. ω ∈ b G. Proof.
By Remark 3 the standing hypotheses are satisfied by the Gabor system. The result nowfollows from Theorem 3.11 together with Lemma 2.4 and the comments preceding Theorem 4.1.From Theorem 4.1 we can derive numerous results about Gabor systems. We begin with anexample concerning the inversion of the short-time Fourier transform.
Example 3.
Let g, h ∈ L ( G ) and consider { E γ T λ g } γ ∈ b G,λ ∈ G and { E γ T λ h } γ ∈ b G,λ ∈ G . We equip G and b G with their respective Haar measures µ G and µ b G . For f ∈ L ( G ) we calculate h f, E γ T λ g i = Z G f ( x ) g ( x − λ ) γ ( x ) dµ G ( x ) = F ( f ( · ) g ( · − λ ))( γ ) . (4.1)With equation (4.1) and since k f k = kF f k , we find Z G Z b G |h f, E γ T λ g i| dµ b G ( γ ) dµ G ( λ ) = Z G Z b G (cid:12)(cid:12) F ( f ( · ) g ( · − λ ))( γ ) (cid:12)(cid:12) dµ b G ( γ ) dµ G ( λ )= Z G Z G (cid:12)(cid:12) f ( x ) g ( x − λ ) (cid:12)(cid:12) dµ G ( x ) dµ G ( λ )= Z G (cid:12)(cid:12) f ( x ) (cid:12)(cid:12) Z G (cid:12)(cid:12) g ( x − λ ) (cid:12)(cid:12) dµ G ( λ ) dµ G ( x ) = k f k k g k .
24 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
The same calculation holds for the Gabor system generated by h . We conclude that bothGabor systems are Bessel families. By Theorem 4.1 the two Gabor systems { E γ T λ g } γ ∈ b G,λ ∈ G and { E γ T λ h } γ ∈ b G,λ ∈ G are dual frames for L ( G ) if, and only if, for a.e. x ∈ G Z G g ( x − λ ) h ( x − λ ) dµ G ( λ ) = Z G g ( λ ) h ( λ ) dµ G ( λ ) = h g, h i = 1 , that is, h g, h i = 1 . This result is the well-known inversion formula for the short-time Fouriertransform [26, 27].
Example 4.
Let G = Γ = R n , Λ = Z n and g ∈ L ( R n ) . We equip G and Γ with the Lebesguemeasure and Λ with the counting measure. Then h f , f i = Z R n X λ ∈ Z n h f , E γ T λ g ih E γ T λ g, f i dγ, for all f , f ∈ L ( R n ) if, and only if, X λ ∈ Z n | g ( x − λ ) | = 1 , a.e. x ∈ R n . Equivalently in the frequency domain, for all β ∈ Z n Z R n ˆ g ( ω + γ )ˆ g ( ω + γ + β ) dγ = δ β, a.e. ω ∈ R n . From the time domain characterization, it is clear that the square-root of any uniform B-splinescan be used to construct such functions g . The Gabor system with Λ = R n and Γ = Z n hassimilar characterizing equations, see [39, Example 2.1(b)]. Example 5.
Let g, h ∈ L ( R ) and a, b > be given. Take Λ = a Z and Γ = b Z . We equip R with the Lebesgue measure and Λ ⊥ ∼ = a Z , Γ ⊥ ∼ = b Z with the counting measure. From thisfollows that the measure on Λ and Γ is the counting measure multiplied with a and b respectively.Theorem 4.1 now yields the following characterizing equation for dual Gabor systems in L ( R ) :If { E γ T λ g } γ ∈ Γ ,λ ∈ Λ and { E γ T λ h } γ ∈ Γ ,λ ∈ Λ are Bessel sequences, then f = ab X λ ∈ a Z X γ ∈ b Z h f, E γ T λ g i E γ T λ h, for all f ∈ L ( R ) if, and only if, for all α ∈ b Z X λ ∈ a Z g ( x − λ ) h ( x − λ + α ) = a δ α, for a.e. x ∈ [0 , a ] . This result is equivalent to the characterization result by Janssen [33]. Higher dimensionalversions can be derived similarly; see Ron and Shen [45] for alternative proofs.One can easily deduce characterization results for Gabor systems in ℓ ( Z d ) following theapproach of the preceding example. We refer to the work of Janssen [16] and Lopez and Han [42]for direct proofs. Finally, we mention the following characterization for finite and discrete Gaborframes.
25 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
Example 6.
Let g, h ∈ C d and a, b, d, N, M ∈ N be such that aN = bM = d . Then f = M − X m =0 N − X n =0 h f, E mb T na g i E mb T na h, for all f ∈ C d if, and only if, N − X k =0 g ( x − nM − ka ) h ( x − ka ) = 1 M δ n, , ∀ x ∈ { , , . . . , a − } , n ∈ { , , . . . , b − } . This result appears first in [49] and has been rediscovered in, e.g., [41].
Following [4], we let
Epick ( G ) denote the semigroup of continuous group homomorphisms a of G onto G with compact kernel. This semigroup can be viewed as an extension of the group oftopological automorphisms on G ; we define the extended modular function ∆ in Epick ( G ) asin [4, Section 6]. The isometric dilation operator D a : L ( G ) → L ( G ) is then defined by D a f ( x ) = ∆( a ) − / f ( a ( x )) . Let A be a subset of Epick ( G ) , let Γ be a co-compact subgroup of G , and let Ψ be a subsetof L ( G ) . The wavelet system generated by Ψ is: W(Ψ , A , Γ) := { D a T γ ψ : a ∈ A , γ ∈ Γ , ψ ∈ Ψ } . Depending on the choice of A and the structure of Epick ( G ) , it might be desirable to extendthe wavelet system with translates of “scaling” functions, that is, { T γ φ : γ ∈ Γ , φ ∈ Φ } for some Φ ⊂ L ( G ) . We denote this extension to a “non-homogeneous” wavelet system by W h (Ψ , Φ , A , Γ) .If Epick ( G ) only contains trivial group homomorphisms, e.g., as in the case of G = Z , it ispossible to define the dilation operator on the dual group b G via the Fourier transform.The two wavelet systems introduced above offer a very general setup that include most ofthe usual wavelet-type systems in L ( R n ) , e.g., discrete and continuous wavelet and shearletsystems [14, 38] as well as composite wavelet systems. Example 7.
Let us consider the general setup as above, where we make the specific choice
Γ = G and Ψ = { ψ j } j ∈ J for some index set J ⊂ Z . For a ∈ A and γ ∈ Γ = G , we have D a T γ ψ j ( x ) = ∆( a ) − / ψ j ( a ( x ) − γ )) = T ¯ γ D a ψ j ( x ) for some ¯ γ ∈ a − Γ so that a (¯ γ ) = γ . It follows that W(Ψ , A , Γ) is a (generalized) translationinvariant system for Γ j = G with j ∈ J and g j,p = g j,a = D a ψ j for ( j, p ) = ( j, a ) ∈ J × A . Forsimplicity we equip each measure space P j = A , j ∈ J , with the same measure; as usual werequire that this measure µ A satisfies our standing hypotheses. Further, we define the adjoint of a by ˆ a ( ω ) = ω ◦ a for ω ∈ b G . Using results from [4], it follows that ˆ a is an isomorphism from b G onto (ker a ) ⊥ and that d D a f ( ω ) = ( ∆( a ) / ˆ f (ˆ a − ( ω )) ω ∈ (ker a ) ⊥ , otherwise .
26 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
As translation invariant systems always satisfy the local integrability condition, we immedi-ately have that
W(Ψ , A , G ) is a Parseval frame, that is, f = X j ∈ J Z A Z G h f, D a T γ ψ j i D a T γ ψ j dµ G ( γ ) dµ A ( a ) for all f ∈ L ( G ) , if, and only if, for a.e. ω ∈ b G , t = X j ∈ J Z A (cid:12)(cid:12) \ D a ψ j ( ω ) (cid:12)(cid:12) dµ A ( a ) = X j ∈ J Z { a ∈A : ω ∈ (ker a ) ⊥ } ∆( a ) (cid:12)(cid:12) ˆ ψ j (ˆ a − ( ω )) (cid:12)(cid:12) dµ A ( a ) = 1 . (4.2)In particular, it follows that W(Ψ , A , G ) cannot be a Parseval frame for L ( G ) regardless of themeasure µ A if b G \ ∪ a ∈A (ker a ) ⊥ has non-zero measure.The Calderón admissibility condition (1.2) is a special case of (4.2). To see this, take G = R and consider the dilation group A = (cid:8) x a − x : a ∈ R \ { } (cid:9) with measure µ A defined on theBorel algebra on R \ { } by dµ A ( a ) = da/a , where da = dλ ( a ) denotes the Lebesgue measure.Higher dimensional versions of Calderón’s admissibility condition are obtained similarly, seealso [20, 40]. Example 8.
We consider wavelet systems in L ( R n ) with discrete dilations and semi-continuoustranslations. Let A ∈ GL ( n, R ) be a matrix whose eigenvalues are strictly larger than one inmodulus, set A = (cid:8) x A j x : j ∈ Z (cid:9) , and let Γ be a co-compact subgroup of R n . The waveletsystem generated by Ψ = { ψ ℓ } Lℓ =1 ⊂ L ( G ) is given by W(Ψ , A,
Γ) := n D A j T γ ψ ℓ = | det A | − j/ ψ ℓ ( A − j · − γ ) : ℓ = 1 , . . . , L, j ∈ Z , γ ∈ Γ o . Any co-compact subgroup of R n is of the form Γ = P ( Z k × R n − k ) for some k ∈ { , , . . . , n } and P ∈ GL( n, R ) . Since W( { ψ } , A, Γ) is unitarily equivalent to W( { D P − ψ } , P − AP, Z k × R n − k ) we can without loss of generality assume that P = I n , i.e., Γ = Z k × R n − k .Clearly, W(Ψ , A, Γ) is a generalized translation invariant system for Γ j = A j Γ with j ∈ J := Z and g j,ℓ = D A j ψ ℓ , where P j = { , . . . , L } . To get rid of a scaling factor in the representationformula, we will use µ P j = | det A | j µ c as measure on P j = { , . . . , L } , where µ c denotes thecounting measure. The standing assumptions are clearly satisfied. Moreover, the local inte-grability condition is known to be equivalent to local integrability on R n \ { } of the Calderónsum [3, Proposition 2.7] and can, therefore, be omitted from the characterization results. Itfollows that two Bessel families W(Ψ , A, Γ) and W(Φ , A, Γ) are dual frames if, and only if, with B = A T , t α ( ω ) = L X l =1 X j ∈ Z : α ∈ B j ( Z k ×{ } n − k ) ˆ ψ l ( B − j ω ) ˆ φ l ( B − j ( ω + α )) = δ α, for a.e. ω ∈ R n , for all α ∈ Z k × { } n − k . For k = n this result was obtained in [11], extending the work ofGripenberg [23] and Wang [48]. Example 9.
Let us finally consider the cone-adapted shearlet systems. For brevity we restrictour findings to the non-homogeneous, continuous shearlet transform in dimension two. Let A a = (cid:18) a a / (cid:19) and S s = (cid:18) s (cid:19)
27 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups for a = 0 and s ∈ R . For ψ ∈ L ( R ) define ψ ast ( x ) := a − / ψ ( A − a S − s ( x − t )) = T t D S s A a ψ. The cone-adapted continuous shearlet system S h ( φ, ψ, ˜ ψ ) is then defined as the collection: S h ( φ, ψ , ψ ) = (cid:8) T t φ : t ∈ R (cid:9) ∪ n T t D S s A a ψ : a ∈ (0 , , | s | ≤ a / , t ∈ R o ∪ n T t D ˜ S s ˜ A a ψ : a ∈ (0 , , | s | ≤ a / , t ∈ R o , where ˜ S s = S Ts and ˜ A a = diag ( a / , a ) . This is a special case of the system W h introduced above.More importantly, this is a GTI system. To see this claim, take J = { , } and Γ = Γ j = R for j ∈ J . Define P = { } and let µ P be the counting measure on P . Define P = n ( a, s ) ∈ R : a ∈ (0 , , | s | ≤ a / o , and let µ P be some measure on P so that our standing hypotheses are satisfied. The generatorsare g ,p = g , = φ for p = 0 ∈ P and g ,p = g , ( a,s ) = D ˜ S s ˜ A a ψ for p = ( a, s ) ∈ P . This provesour claim. By Theorem 3.11 and Lemma 3.12 we immediately have that, if S h ( φ, ψ , ψ ) and S h ( φ, ˜ ψ , ˜ ψ ) are Bessel families, then they are dual frames if, and only if, ˆ φ ( ω ) ˆ φ ( ω ) + Z P a / ˆ ψ ( A a S Ts ω ) ˆ˜ ψ ( A a S Ts ω ) dµ P ( a, s )+ Z P a / ˆ ψ ( ˜ A a ˜ S Ts ω ) ˆ˜ ψ ( ˜ A a ˜ S Ts ω ) dµ P ( a, s ) = 1 for a.e. ω ∈ R . (4.3)A standard choice for the measure µ P in (4.3) is dµ P ( a, s ) = dadsa , which comes from theleft-invariant Haar measure on the shearlet group. The above characterization result generalizesresults from [24, 25, 37]. In this example we consider the additive group of p -adic integers I p . To introducethis group, we first consider the p -adic numbers Q p . Here p is some fixed prime-number. The p -adic numbers are the completion of the rationals Q under the p -adic norm, defined as follows.Every non-zero rational x can be uniquely factored into x = rs p n , where r, s, n ∈ Z and p doesnot divide r nor s . We then define the p -adic norm of x as k x k p = p − n , additionally k k p := 0 .The p -adic numbers Q p are the completion of Q under k · k p . It can be shown that all p -adicnumbers x can be written uniquely as x = ∞ X j = k x j p j , (4.4)where x k ∈ { , , . . . , p − } and k ∈ Z , x k = 0 . The set of all numbers x ∈ Q p for which x j = 0 for j < in (4.4) are the p -adic integer I p . Equivalently, I p = { x ∈ Q p : k x k p ≤ } . In fact, I p isa compact, closed and open subgroup of Q p . Its dual group b I p can be identified with the Prüfer p -group Z ( p ∞ ) , which consists of the union of the p n -roots of unity for all n ∈ N . That is, b I p ∼ = Z ( p ∞ ) := { e πim/p n : n ∈ N , m ∈ { , , . . . , p n − } } ⊂ C .
28 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups
We equip Z ( p ∞ ) with the discrete topology and multiplication as group operation. For moreinformation on p -adic numbers and their dual group we refer to, e.g., [31, §10, §25]. For n ∈ N consider now the subgroups Γ ⊥ n = { e πim/p n : m = 0 , , . . . , p n − } ⊂ Z ( p ∞ ) . Note that all Γ ⊥ n are finite groups of order p n and generated by e πi/p n . Moreover, all Γ ⊥ n are nested so that ⊂ Γ ⊥ ⊂ Γ ⊥ ⊂ · · · ⊂ Z ( p ∞ ) . Let now { g n } n ∈ N ⊂ L ( I p ) . By Theorem 3.14 the generalized translation invariant system { T γ g n } γ ∈ Γ n ,n ∈ N is a Parseval frame for L ( I p ) if, and only if, for each α ∈ S n ∈ N Γ ⊥ n = Z ( p ∞ ) ∞ X k = n ∗ ˆ g n ( ω )ˆ g n ( ωα ) = δ α, for all ω ∈ Z ( p ∞ ) , where n ∗ ∈ N is the smallest natural number such that α ∈ Γ ⊥ n ∗ . Because we consider a GTI sys-tem with countably many generators, the standing hypotheses are trivially satisfied, see Section3. Returning to the p -adic numbers Q p , we note that the only co-compact subgroup of Q p is Q p itself [4]. Therefore any GTI system in L ( Q p ) is, in fact, a translation invariant system ofthe form ∪ j ∈ J { T γ g p } γ ∈ Q p ,p ∈ P j . The equations characterizing the dual frame property of suchsystems are immediate from Theorem 3.11 and Lemma 3.12.Finally, in the product group Q p × I p there are no discrete, co-compact subgroups [4], andthus no generalized shift invariant systems for L ( Q p × I p ) can be constructed. However, anysubgroup of the form Q p × Γ n , where Γ n is a co-compact subgroup of I p as before, is a co-compactsubgroup in Q p × I p , indicating that a large number of generalized translation invariant systemsdo exist in L ( Q p × I p ) .In order to apply Theorem 3.4 to a given GTI system, one needs to verify that the (dual) α -LIC or the stronger LIC holds. By Theorem 3.11 we get this for free for translation invariantsystems. For regular wavelet systems as in Example 8 the LIC has an easy characterization [3,Proposition 2.7]. For certain irregular wavelet systems over the real line a detailed analysis ofthe LIC has been carried out in [35] using Beurling densities. However, for general GTI systemsthere is no simple interpretation of the local integrability conditions. Acknowledgments
We thank O. Christensen for giving access to an early version of [9] and K. Ross for providingan example of an LCA group with no uniform lattices, but with proper co-compact subgroups.
References [1] S. T. Ali, J.-P. Antoine, and J.-P. Gazeau. Continuous frames in Hilbert space.
Ann.Physics , 222(1):1–37, 1993.[2] S. T. Ali, J.-P. Antoine, and J.-P. Gazeau.
Coherent states, wavelets and their generaliza-tions . Graduate Texts in Contemporary Physics. Springer-Verlag, New York, 2000.[3] M. Bownik and J. Lemvig. Affine and quasi-affine frames for rational dilations.
Trans.Amer. Math. Soc. , 363(4):1887–1924, 2011.[4] M. Bownik and K. Ross. The structure of translation-invariant spaces on locally compactabelian groups, preprint.
29 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups [5] M. Bownik and Z. Rzeszotnik. The spectral function of shift-invariant spaces on generallattices. In
Wavelets, frames and operator theory , volume 345 of
Contemp. Math. , pages49–59. Amer. Math. Soc., Providence, RI, 2004.[6] A.-P. Calderón. Intermediate spaces and interpolation, the complex method.
Studia Math. ,24:113–190, 1964.[7] P. G. Casazza, O. Christensen, and A. J. E. M. Janssen. Weyl-Heisenberg frames, translationinvariant systems and the Walnut representation.
J. Funct. Anal. , 180(1):85–147, 2001.[8] O. Christensen.
Frames and bases . Applied and Numerical Harmonic Analysis. BirkhäuserBoston Inc., Boston, MA, 2008. An introductory course.[9] O. Christensen and S. S. Goh. Fourier like frames on locally compact abelian groups,preprint.[10] O. Christensen and A. Rahimi. Frame properties of wave packet systems in L ( R d ) . Adv.Comput. Math. , 29(2):101–111, 2008.[11] C. K. Chui, W. Czaja, M. Maggioni, and G. Weiss. Characterization of general tight waveletframes with matrix dilations and tightness preserving oversampling.
J. Fourier Anal. Appl. ,8(2):173–200, 2002.[12] A. Córdoba and C. Fefferman. Wave packets and Fourier integral operators.
Comm. PartialDifferential Equations , 3(11):979–1005, 1978.[13] C. Corduneanu.
Almost periodic functions . Interscience Publishers [John Wiley & Sons],New York-London-Sydney, 1968. With the collaboration of N. Gheorghiu and V. Barbu,Translated from the Romanian by Gitta Bernstein and Eugene Tomer, Interscience Tractsin Pure and Applied Mathematics, No. 22.[14] I. Daubechies.
Ten Lectures on Wavelets . Society for Industrial and Applied Mathematics,Philadelphia, PA, USA, 1992.[15] F. De Mari and E. De Vito. Admissible vectors for mock metaplectic representations.
Appl.Comput. Harmon. Anal. , 34(2):163–200, 2013.[16] H. G. Feichtinger and T. Strohmer, editors.
Gabor analysis and algorithms . Applied andNumerical Harmonic Analysis. Birkhäuser Boston Inc., Boston, MA, 1998. Theory andapplications.[17] G. B. Folland.
A course in abstract harmonic analysis . Studies in Advanced Mathematics.CRC Press, Boca Raton, FL, 1995.[18] M. Fornasier and H. Rauhut. Continuous frames, function spaces, and the discretizationproblem.
J. Fourier Anal. Appl. , 11(3):245–287, 2005.[19] M. Frazier, G. Garrigós, K. Wang, and G. Weiss. A characterization of functions that gen-erate wavelet and related expansion. In
Proceedings of the conference dedicated to ProfessorMiguel de Guzmán (El Escorial, 1996) , volume 3, pages 883–906, 1997.[20] H. Führ. Generalized Calderón conditions and regular orbit spaces.
Colloq. Math. ,120(1):103–126, 2010.
30 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups [21] H. Führ. Coorbit spaces and wavelet coefficient decay over general dilation groups.
Trans.Amer. Math. Soc. , posted online October 3, 2014, doi: 10.1090/S0002-9947-2014-06376-9(to appear in print).[22] J.-P. Gabardo and D. Han. Frames associated with measurable spaces.
Adv. Comput. Math. ,18(2-4):127–147, 2003. Frames.[23] G. Gripenberg. A necessary and sufficient condition for the existence of a father wavelet.
Studia Math. , 114(3):207–226, 1995.[24] P. Grohs. Continuous shearlet frames and resolution of the wavefront set.
Monatsh. Math. ,164(4):393–426, 2011.[25] P. Grohs. Continuous shearlet tight frames.
J. Fourier Anal. Appl. , 17(3):506–518, 2011.[26] K. Gröchenig. Aspects of Gabor analysis on locally compact abelian groups. In
Gaboranalysis and algorithms , Appl. Numer. Harmon. Anal., pages 211–231. Birkhäuser Boston,Boston, MA, 1998.[27] K. Gröchenig.
Foundations of Time-Frequency Analysis . Appl. Numer. Harmon. Anal.Birkhäuser, 2001.[28] K. Gröchenig, G. Kutyniok, and K. Seip. Landau’s necessary density conditions for LCAgroups.
J. Funct. Anal. , 255(7):1831–1850, 2008.[29] K. Gröchenig and T. Strohmer. Pseudodifferential operators on locally compact abeliangroups and Sjöstrand’s symbol class.
J. Reine Angew. Math. , 613:121–146, 2007.[30] E. Hernández, D. Labate, and G. Weiss. A unified characterization of reproducing systemsgenerated by a finite family. II.
J. Geom. Anal. , 12(4):615–662, 2002.[31] E. Hewitt and K. A. Ross.
Abstract harmonic analysis. Vol. I: Structure of topologicalgroups. Integration theory, group representations . Die Grundlehren der mathematischenWissenschaften, Bd. 115. Academic Press Inc., Publishers, New York, 1963.[32] E. Hewitt and K. A. Ross.
Abstract harmonic analysis. Vol. II: Structure and analysis forcompact groups. Analysis on locally compact Abelian groups . Die Grundlehren der mathe-matischen Wissenschaften, Band 152. Springer-Verlag, New York, 1970.[33] A. J. E. M. Janssen. The duality condition for Weyl-Heisenberg frames. In
Gabor analysisand algorithms , Appl. Numer. Harmon. Anal., pages 33–84. Birkhäuser Boston, Boston,MA, 1998.[34] G. Kaiser.
A friendly guide to wavelets . Birkhäuser Boston Inc., Boston, MA, 1994.[35] G. Kutyniok
The local integrability condition for wavelet frames . J. Geom. Anal. 16 (2006),155–166.[36] G. Kutyniok and D. Labate. The theory of reproducing systems on locally compact abeliangroups.
Colloq. Math. , 106(2):197–220, 2006.[37] G. Kutyniok and D. Labate. Resolution of the wavefront set using continuous shearlets.
Trans. Amer. Math. Soc. , 361(5):2719–2754, 2009.
31 of 32 date/time: 22-Apr-2015/0:33 akobsen, Lemvig Reproducing formulas for GTI systems on LCA groups [38] G. Kutyniok and D. Labate, editors.
Shearlets . Applied and Numerical Harmonic Analysis.Birkhäuser/Springer, New York, 2012. Multiscale analysis for multivariate data.[39] D. Labate, G. Weiss, and E. Wilson. An approach to the study of wave packet systems.In
Wavelets, frames and operator theory , volume 345 of
Contemp. Math. , pages 215–235.Amer. Math. Soc., Providence, RI, 2004.[40] R. S. Laugesen, N. Weaver, G. L. Weiss, and E. N. Wilson. A characterization of the higherdimensional groups associated with continuous wavelets.
J. Geom. Anal. , 12(1):89–102,2002.[41] S. Li, Y. Liu, and T. Mi. Sparse dual frames and dual Gabor functions of minimal time andfrequency supports.
J. Fourier Anal. Appl. , 19(1):48–76, 2013.[42] J. Lopez and D. Han. Discrete Gabor frames in ℓ ( Z d ) . Proc. Amer. Math. Soc. ,141(11):3839–3851, 2013.[43] A. Rahimi, A. Najati, and Y. N. Dehghan. Continuous frames in Hilbert spaces.
MethodsFunct. Anal. Topology , 12(2):170–182, 2006.[44] H. Reiter and J. D. Stegeman.
Classical harmonic analysis and locally compact groups ,volume 22 of
London Mathematical Society Monographs. New Series . The Clarendon PressOxford University Press, New York, second edition, 2000.[45] A. Ron and Z. Shen. Weyl-Heisenberg frames and Riesz bases in L ( R d ) . Duke Math. J. ,89(2):237–282, 1997.[46] A. Ron and Z. Shen. Generalized shift-invariant systems.
Constr. Approx. , 22(1):1–45, 2005.[47] W. Sun. G -frames and g -Riesz bases. J. Math. Anal. Appl. , 322(1):437–452, 2006.[48] X. Wang.
The study of wavelets from the properties of their Fourier transforms . ProQuestLLC, Ann Arbor, MI, 1995. Thesis (Ph.D.)–Washington University in St. Louis.[49] J. Yao, P. Krolak, and C. Steele. The generalized Gabor transform.
IEEE Transactions onimage processing , 4(7):978–988, 1995.
32 of 3232 of 32