Operator Identification on the Circle Gökhan Civan Department of Mathematics University of Maryland College Park, MD 20742 Email: gcivan@math.umd.edu Abstract—We study the identification problem for linear timevariant communication channels on the circle, extending earlier work on the real line. We therefore resolve a suitably formulated periodic version of the conjectures of Kailath and Bello regarding channel identification. I. I NTRODUCTION A linear time-variant communication channel (e.g. mobile communications, radar) can be modeled formally as an integral operator [1] of the form Z f → κ(·, t)f (t) dt, where κ is the kernel function. The problem of identifying such a channel only from its action on a single properly chosen input signal was addressed by Kailath [2]. He considered a family of channels with time delay and Doppler spread confined to a common support set, and conjectured that such a family of channels is identifiable if and only if their common support set has area at most 1. Kailath considered only rectangular support sets; his conjecture was later extended by Bello [3] to include arbitrary support sets. More recently, the channel identification problem was addressed using modern time-frequency analysis techniques. In [4], Kozek and Pfander resolved a suitably formulated version of Kailath’s conjecture. In [5], Pfander and Walnut resolved a suitably formulated version of the more general conjecture of Bello. The theory of ”operator sampling” or ”operator identification” which emerged from these efforts has since been further developed and elaborated on; see [6], [7], and the references therein. We are interested in the extension of the theory to more general groups. In this paper, we study the operator identification problem on the circle. Specifically, we consider operators of the form Z XZ g→ κ(·, t)g(t) dt = η(z, ξ)Mξ Tz g dz, T ξ∈Z T where g is a function on T, κ is a function on T×T, and η is a function on T × Z with compact support; the operators Tz and Mξ are translation and modulation. We shall make precise to which spaces these functions belong presently. Our main results are Theorems 1 and 2, corresponding to the main results in [5], which provide sufficient and necessary conditions for identification, respectively. We follow the formulation and c 2015 IEEE 978-1-4673-7353-1/15/$31.00 proof method in [5], and partly borrow from [4], introducing modifications as needed for our new setting. Therefore, we give credit to the original authors for some of the ideas and constructions that we utilize. II. F OUNDATIONAL T HEORY Our development uses the language and techniques of timefrequency analysis. We suggest [8] as a basic reference. The standard treatment of time-frequency analysis focuses on the b = Z are real line. Here, T = {z ∈ C : |z| = 1} and T the groups that are of interest to us; we take the unit Haar measure on T and the counting measure on Z. The standard theory goes through for these (and more general) groups. We shall therefore be content with mostly stating the basic facts necessary for our discussion. A. The Short-Time Fourier Transform Let G be a group of the form TM × ZN . We have the Schwartz space S(G) of rapidly decreasing functions, and its dual S 0 (G), the space of tempered distributions. Clearly, S(TM ) = C ∞ (TM ). One can show that S 0 (ZM ) is precisely the set of all functions of polynomial growth. Since the Fourier transform is an isomorphism between S 0 (ZM ) and S 0 (TM ), the latter space consists precisely of Fourier series with coefficients of polynomial growth. For f ∈ S 0 (G) and g ∈ S(G) \ {0}, the short-time Fourier b defined by transform (STFT) is a function of (a, â) ∈ G × G Z Vg f (a, â) = hf, Mâ Ta gi = f (t)g(t − a)(−t, â) dt. (1) G Here, the integral is formal unless f ∈ S(G). The operators Ta and Mâ are translation and modulation, respectively. Note that (z, ξ) = z ξ = Eξ (z) for z ∈ T and ξ ∈ Z. The STFT is a continuous function and a tempered distribution in its own right. Let 1 ≤ p ≤ ∞. The modulation space norm of f ∈ S 0 (G) is kf kM p = kVg f kLp . Different window functions g define equivalent norms. The modulation space M p (G) is the set of all f ∈ S 0 (G) with kf kM p < ∞. The space M p (G) is a Banach space invariant under translations, modulations, and (partial) Fourier transforms. (A partial Fourier transform is a Fourier transform with respect to a single variable while all the other variables are fixed.) Of course, here and elsewhere, we shift to the appropriate dual group whenever a (partial) Fourier transform is applied. We have the continuous inclusions S(G) ⊆ M 1 (G) ⊆ L (G) ∩ C0 (G), M 2 (G) = L2 (G), and M p1 (G) ⊆ M p2 (G) for p1 ≤ p2 . The space M 1 (G) is Feichtinger’s algebra (also denoted by S0 (G)), and M ∞ (G) is its dual. The duality is defined via the pairing Z 0 hf, f i = (2) Vg f Vg f 0 , p b G×G where f ∈ M (G) and f 0 ∈ M ∞ (G). This pairing is independent of the chosen window function g as long as kgk2 = 1. Moreover, if f 0 ∈ Lp (G), then f 0 ∈ M ∞ (G), and Z (3) hf, f 0 i = f f 0. 1 G The duality pairing is invariant under (partial) Fourier transforms. Since T is compact, the modulation spaces in this case are much simpler than what the definition might suggest. One can show that the Fourier transform is an isomorphism between M p (TM ) and `p (ZM ) (see [9] for a proof in a different context). In particular, M 1 (TM ) is Wiener’s algebra A(TM ). Note that δ ∈ M ∞ (TM ). On G × G, we have the Fourier transform in the second variable, denoted by F2 , and the coordinate transform (a, t) → (t, t − a), denoted by TG . Finally, we indicate that the theory of the STFT can be extended to include window functions in M 1 (G). B. The Spaces M 1 (T × T) and M 1 (T × Z) Let f ∈ M 1 (T × T). Since M 1 (T × T) ∼ = `1 (Z × Z), X f= cξ1 ,ξ2 Eξ1 ⊗ Eξ2 ξ1 ,ξ2 ∈Z with convergence in M 1 (T × T). Here, {cξ1 ,ξ2 } ∈ `1 (Z × Z). It is almost immediate that, for a ∈ T, f (a, ·) ∈ M 1 (T) ∼ = `1 (Z), and X X f (a, ·) = cξ1 ,ξ2 aξ1 Eξ2 ξ2 ∈Z ξ1 ∈Z 1 with convergence in M (T). We compute X F2 f = cξ1 ,ξ2 F2 (Eξ1 ⊗ Eξ2 ) ξ1 ,ξ2 ∈Z = X bξ cξ1 ,ξ2 Eξ1 ⊗ E 2 = cξ1 ,ξ2 Eξ1 ⊗ δξ2 ξ1 ,ξ2 ∈Z with convergence in M 1 (T × Z). It is almost immediate that F2 f (a, ·) = f (a, ·)b . hg, δi = g(1) (4) The ”almost immediate” conclusions that we have drawn here really follow from the continuity of the inclusion M 1 (G) ⊆ C0 (G) for various choices of G. The last equality may seem like a tautology, but this is the case only when f ∈ S(T × T). (g ∈ M 1 (T)). (5) C. Gabor Analysis in Finite Dimensions Let ωK = e2πi/K for some positive integer K. For c ∈ CK , let [5] A(c) = (A0 |A1 | · · · |AK−1 ), pq K−1 where Ak is the matrix (cp+k ωK )p,q=0 . Here and elsewhere, indices over Z/KZ are modulo K. The columns of A(c) coincide with the vectors Mâ Ta c for a, â ∈ Z/KZ. In other words, A(c) is the Gabor system generated by c. It is shown in [10] (and earlier in [11] for K prime) that there exists c ∈ CK such that every K × K minor of A(c) is invertible; in fact, the set of all such c is a dense open set of full measure. (The result in [11] for K prime is stronger in that the invertibility of every minor is asserted.) III. K ERNELS AND O PERATORS 2 defines an integral operator f → R Every κ ∈ L (T × T) 2 κ(·, t)f (t) dt from L (T) to L2 (T) with operator norm T kκkL2 . These are precisely the so-called Hilbert-Schmidt operators. Alternatively, these operators are characterized as those defined by (infinite) matrices with finite Frobenius norms; the matrices can be defined with respect to the standard orthonormal basis {Eξ }ξ∈Z . As we would like to be able to feed distributional elements into our operators, the L2 framework is not suitable. We next describe a refinement of the above class of operators in order to suit our needs. Our kernel function κ shall be an element of the smaller space M 1 (T × T). In this case, we get a bounded linear operator H : M ∞ (T) → M 1 (T) as follows. We define Hf (a) = hκ(a, ·), f i; we conjugate f so that the resulting operator is linear instead of conjugate-linear. Let us compute this pairing in terms of the series expansions obtained previously. Recall that κ is defined by a sequence {cξ1 ,ξ2 } ∈ `1 (Z × Z). It follows from the characterization M ∞ (T) ∼ = `∞ (Z) that f is defined by a bounded sequence {dξ }ξ∈Z . By (3) and the invariance of (2) under the Fourier transform, X X Hf (a) = cξ1 ,ξ2 aξ1 dξ2 . ξ2 ∈Z ξ1 ,ξ2 ∈Z X Here is another seemingly tautological equality which can be proved using series expansions: ξ1 ∈Z Since the convergence is absolute, we can rearrange this series to obtain X X Hf (a) = cξ1 ,ξ2 dξ2 aξ1 . ξ1 ∈Z ξ2 ∈Z It follows that Hf ∈ M 1 (T), and that H is defined by matrix multiplication via the matrix (cξ1 ,ξ2 )ξ1 ,ξ2 ∈Z . Moreover, the operator norm of H is bounded by CkκkM 1 , where C does not depend on κ. However, we actually have the following stronger form of continuity: If fj → 0 in the weak* topology of M ∞ (T), then Hfj → 0 in the norm topology of M 1 (T). (Since {fj } is convergent in the weak* topology of M ∞ (T), it is bounded in the norm topology of M ∞ (T). This allows one to apply the dominated convergence theorem.) Since M 1 (T) is sequentially weak* dense in M ∞ (T), H is determined by its values on M 1 (T). A simple calculation shows that H satisfies operator identification problem. However, it is reasonable to additionally require that eg be bounded and stable so that errors on one side correspond in a controllable fashion to errors on the other side. (An operator is stable if it is injective with a bounded inverse [5].) Note that the continuity of eg is not automatic, unless g ∈ L2 (T), since we put the L2 norm on HW instead of the M 1 norm. hHg, f i = hκ, f ⊗ gi (f, g ∈ M 1 (T)). We shall use the following notation. For a positive integer K, let IK be the image of [0, 1/K) under the exponential map x → e2πix . The measure of a Borel set will be denoted by the absolute value notation. (6) Here is a summary of our discussion so far. Let H be the set of all linear maps from M ∞ (T) to M 1 (T) that are continuous in the weak* sense described above. We have demonstrated an injective map from M 1 (T×T) to H. It is not too hard to show that this map is a bijection. In other words, we have a version of the Schwartz kernel theorem. Applying F2 and TT , we can recast (6) in the form 1 hHg, f i = hη, Vg f i (f, g ∈ M (T)). ν∈Z Since the convergence is absolute, the function (z, ξ, t) → Mξ Tz g(t) on T × Z × T is continuous and bounded, which guarantees the existence of the double integral in the definition of the following function on T: XZ ϕ:t→ η(z, ξ)Mξ Tz g(t) dz. ξ∈Z IV. S UFFICIENT C ONDITION FOR O PERATOR I DENTIFICATION Let W be a subset of T × Z. Let HW = {H ∈ H : supp ηH ⊆ W } be equipped with the norm kHk = kηH kL2 . For g ∈ M ∞ (T), let eg : HW → L2 (T) be defined by eg (H) = Hg. In other words, eg is evaluation by g. If eg is injective for some g ∈ M ∞ (T), then we have a positive answer to the ck δk , k=0 k where δk is the Dirac distribution on T at the point ωK , and ck ∈ C. Proof. We follow the proof method in [5]. We can assume without loss of generality that W ⊆ T × {0, 1, 2, . . .}. Let K be a positive integer such that W ⊆ T × [0, K − 1]. We also pick K large enough so that W is a subset of a union of at k × {q} (0 ≤ k, q ≤ K − 1); most K sets of the form IK ωK this is possible because of (measure theoretic) outer regularity and the hypothesis that W is compact with |W | < 1. Let g be as in the statement of the theorem with ck chosen so that every K × K minor of A(c) is invertible. Let H ∈ HW . We define T Multiplying the Fourier series expansion of η with (8), we get a series expansion for the integrand. One can integrate this series term by term to discover that ϕ ∈ M 1 (T). Fubini’s theorem now gives hϕ, f i = hη, Vg f i, so ϕ = Hg. This representation of H shows that H can be interpreted as a weighted sum of time-frequency shifts. It can be verified by direct calculation that the measure of supp η quantifies the extent of the timefrequency ”spread” in the output of H. The upshot of this section is that the space M 1 (T × Z) of spreading functions, the space M 1 (T×T) of kernel functions, and the set H are all identified with each other. See [12] for a discussion of the Schwartz kernel theorem over more general groups. K−1 X g= (7) Here, η = F2 TT κ is the so-called spreading function of H. Note that η, Vg f ∈ M 1 (T × Z). Let us express Hg more directly in terms of η. Recall that g is defined by a sequence {cν } ∈ `1 (Z). Then, for t, z ∈ T and ξ ∈ Z, X X Mξ Tz g(t) = cν Mξ Tz Eν (t) = cν tξ+ν z −ν . (8) ν∈Z Theorem 1. Suppose that W is compact with |W | < 1. There exists g ∈ M ∞ (T) such that eg is bounded and stable. In fact, g can be chosen to be of the form −k κH,k (t, z) = κH (t−1 ωK z, z)1IK ×T (t, z) (t, z ∈ T). It is straightforward to check that −k κH,k (t−1 ωK , zt−1 ) = κH (z, zt−1 )1I −1 ω−k ×T (t, z), K K and hence we have the partitioning κH (z, zt−1 ) = K−1 X −k κH,k (t−1 ωK , zt−1 ). k=0 Let ηH,k = F2 κH,k . We can show using (4) that −kξ −k ηH,k (t, ξ) = t−ξ ωK ηH (t−1 ωK , ξ)1IK ×Z (t, ξ) (ξ ∈ Z), or −k ηH,k (t−1 ωK , ξ) = tξ ηH (t, ξ)1I −1 ω−k ×Z (t, ξ), K K and hence we have the partitioning −ξ ηH (t, ξ) = t K−1 X −k ηH,k (t−1 ωK , ξ). (9) k=0 Fix t ∈ IK . For 0 ≤ p ≤ K − 1, we compute p Hg(t−1 ωK )= K−1 X k=0 p k ck κH (t−1 ωK , ωK ) (10) = = = = = K−1 X k=0 K−1 X k=0 K−1 X k=0 K−1 X k=0 K−1 X p+k −k p+k cp+k κH (t−1 ωK ωK , ωK ) p+k cp+k κH,k (t, ωK ) p+k cp+k F2−1 ηH,k (t, ωK ) cp+k K−1 X cp+k (p+k)q ηH,k (t, q)ωK (11) K−1 X ηH,k (t, q)ωK (p+k)q (12) pq kq cp+k ωK ωK ηH,k (t, q), (13) q=0 k=0 = X q∈Z k=0 = Theorem 2. Suppose that W is open with |W | > 1. There exists no g ∈ M ∞ (T) for which eg is stable. p p+k cp+k κH (t−1 ωK , ωK ) K−1 X K−1 X T = k=0 q=0 where (10) follows from (5), (11) follows from (4), and (12) follows from the fact that W ⊆ T × [0, K − 1]. Let kq ηH,k (t, q))0≤k,q≤K−1 . η H (t) = (ωK Then (13) is equivalent to (14) Since W is a subset of a union of at most K sets of the form k IK ωK × {q} (0 ≤ k, q ≤ K − 1), it follows from (9) that at most K entries of η H (t) are nonzero. In fact, we can choose K 2 −K entries that are necessarily zero independent of t (and e H (t) be η H (t) with such K 2 − K entries certainly H). Let η e removed. Let A(c) be A(c) with the corresponding columns removed. Then (14) becomes e η H (t). xH (t) = A(c)e (15) e Since A(c) is invertible, e −1 xH (t). e H (t) = A(c) η a ≤ kxH (t)k22 2 ≤b ξ∈Z Z Z = η(z)δ(ξ)g(tz −1 )tξ−v dt dz T η(z)g(t)t−v z −v dt dz T It follows that |Pcg(ν)| ≤ kĝk`∞ |η̂(ν)|. We now lift the restriction that g ∈ M 1 (T). Let {gj } be a sequence in M 1 (T) such that gj → g in the weak* topology of M ∞ (T). Then Pcg j → Pcg uniformly, and ĝj → ĝ in the weak* topology of `∞ (Z). In particular, ĝj is bounded in the norm topology of `∞ (Z). Let C 0 be a bound. Passing to the limit, we get |Pcg(ν)| ≤ C 0 |η̂(ν)|. Note that the same bound C 0 works for any time-frequency shift of g. We next define a synthesis operator using P as an atom. Observe that Mq+pB P TωK k M−pB ∈ H for 0 ≤ k ≤ K − 1 and p, q ∈ Z. We can show using (7) that ηMq+pB P Tωk ke η H (t)k22 . Integrating these inequalities on IK gives a2 kHk2 ≤ keg (H)k2L2 ≤ b2 kHk2 , which is the desired conclusion. = M(pB,1) T(ωK k ,q) ηP . We define the synthesis operator U : `c (Z × Z × Z/KZ) → H by X K−1 X U (σ) = σ(p, q, k)Mq+pB P TωK k M−pB . p,q∈Z k=0 Here, `c (Z × Z × Z/KZ) is endowed with the `2 norm, and H is endowed with the norm kHk = kηH kL2 . We compute kηU (σ) k2L2 = k X K−1 X 2 σ(p, q, k)M(pB,1) T(ωK k ,q) ηP k 2 L p,q∈Z k=0 X K−1 X X 2 = k σ(p, q, k)M(pB,1) T(ωK k ,q) ηP k 2 L q∈Z k=0 V. N ECESSARY C ONDITION FOR O PERATOR I DENTIFICATION We next prove the following counterpart to Theorem 1. M−pB K (16) e −1 k−1 and b = kA(c)k e Let a = kA(c) 2 . Taking Frobenius 2 norms in (15) and (16) gives ke η H (t)k22 T = η̂(ν)ĝ(ν). and xH (t) = A(c)η H (t). T ξ∈Z XZ Z T p xH (t) = (Hg(t−1 ωK ))0≤p≤K−1 2 Proof. We follow the ideas in [4] and [5]. Let K be a positive integer such that W contains J = K + 2 sets of the form k IK ωK × {q} (0 ≤ k ≤ K − 1, q ∈ Z); this is possible because of the hypothesis that W is open with |W | > 1. Choose J such sets contained in W . Let B = K + 1 and θ = eπi(1/K−1/B) . Note that the center of IB θ coincides with the center of IK . We define an auxiliary operator in H which we shall later use to restrict the identification problem to a more manageable subspace. Let η ∈ C ∞ (T) with 0 ≤ η ≤ 1, η = 1 on IB θ, o and η = 0 outside IK . Let P be that operator in H with ηP = η ⊗ δ. We first study the asymptotic behavior of P . Let g ∈ M ∞ (T). Suppose first that g ∈ M 1 (T). We compute Z XZ Pcg(ν) = ηP (z, ξ)Mξ Tz g(t) dz t−v dt p∈Z (17) = X K−1 X q∈Z k=0 k X p∈Z σ(p, q, k)M(pB,1) ηP k2L2 (18) = X K−1 X X kηP q∈Z k=0 = ≥ = |η(z) T X σ(p, q, k)z pB |2 dz (19) p∈Z X K−1 XZ q∈Z k=0 ≤ C 0 d(λν − (pJ + j)), p∈Z X K−1 XZ q∈Z k=0 = C 0 |η̂(λ−1 (λν − λqj + j − (pJ + j))| σ(p, q, k)M(pB,1) 1k2L2 X | σ(p, q, k)z pB |2 dz IB θ p∈Z X K−1 X XZ |σ(p, q, k)|2 dz (20) IB θ q∈Z k=0 p∈Z VI. C ONCLUSION 1 = kσk2`2 , B where (17) follows from the fact that the indicated translations of ηP have disjoint supports, (18) follows from the translation invariance of the L2 norm, and (20) follows from the Pythagorean theorem. We have shown that U is stable. We now carry out the previous computation in a different direction starting from (19). Note that |η|2 ≤ 1IB + 1IB θ + 1IB θ2 . Then kηU (σ) k2L2 = XZ X K−1 q∈Z k=0 ≤ |η(z) T X σ(p, q, k)z pB |2 dz p∈Z 2 Z X K−1 XX q∈Z k=0 j=0 | IB θ j X σ(p, q, k)z pB |2 dz p∈Z 3 = kσk2`2 . B We have shown that U is bounded. Therefore, U is bounded and stable. Since `c (Z × Z × Z/KZ) is dense in `2 (Z × Z × Z/KZ), and both `2 (Z × Z × Z/KZ) and H are complete, U extends uniquely to a bounded and stable operator U : `2 (Z × Z × Z/KZ) → H. Let {(qj , kj )}J−1 j=0 ⊆ Z × Z/KZ be the collection of those indices corresponding to the J sets chosen at the beginning. Note that the map (p, j) → pJ + j from Z × {(qj , kj )}J−1 j=0 to Z is a bijection. The inverse of this map induces a linear isometry ι : `2 (Z) → `2 (Z × Z × Z/KZ). Let g ∈ M ∞ (T). Let Ag be the composition of the following sequence of maps: ι U eg F `2 (Z) → `2 (Z × Z × Z/KZ) → HW → L2 (T) → `2 (Z). Note that the replacement of H with HW is valid because the image of U ◦ ι is in HW . Since ι, U , and F are all stable, in order to show that eg is not stable, it suffices to show that Ag is not stable. The matrix of Ag is (aν,pJ+j = (P T k ωKj J−1 where d(x) = maxj=0 |η̂(λ−1 (x−λqj +j))|. Since η̂ ∈ S(Z), −s d(x) ≤ Cs (1 + |x|) for all s > 0. We have shown that the entries of Ag decay rapidly away from the slanted diagonal of slope λ > 1, where the slope is defined assuming that the row indices represent the horizontal axis and the column indices represent the vertical axis. As shown in [13], this property implies that Ag is not stable. M−pB g)b (ν − qj − pB))ν,pJ+j∈Z . Let us investigate how the entries of Ag decay. Extend η̂ from Z to R by defining η̂(x) = η̂(dxe). Let λ = J/B > 1. We have |aν,pJ+j | ≤ C 0 |η̂(ν − qj − pB)| We would like to point out certain differences between the operator identification results on the real line and the circle. The identifying signal in Theorem 1 consists of a finite sequence of Dirac impulses whereas an infinite sequence of Dirac impulses is required on the real line. More notably, in Theorem 2, we discretized the right hand side of eg via the Plancherel theorem. On the real line, the discretization of the right hand side is achieved with a Gabor frame generated by a Gaussian window, using the deep result that a Gabor system generated by a Gaussian window below the critical density is necessarily a frame. ACKNOWLEDGMENT I would like to extend many thanks to John Benedetto for helpful discussions, blackboard computations, and his friendship. I would also like to thank Kasso Okoudjou for being available to help me understand certain points. R EFERENCES [1] T. Strohmer, “Pseudodifferential operators and Banach algebras in mobile communications,” Appl. Comput. Harmon. Anal., vol. 20, no. 2, pp. 237–249, 2006. [2] T. Kailath, “Measurements on time-variant communication channels,” IEEE Trans. Inf. Theory, vol. 8, no. 5, pp. 229–236, 1962. [3] P. A. Bello, “Measurement of random time-variant linear channels,” IEEE Trans. Inf. Theory, vol. 15, no. 4, pp. 469–475, 1969. [4] W. Kozek and G. E. Pfander, “Identification of operators with bandlimited symbols,” SIAM J. Math. Anal., vol. 37, no. 3, pp. 867–888, 2005. [5] G. E. Pfander and D. F. Walnut, “Measurement of time-variant linear channels,” IEEE Trans. Inf. Theory, vol. 52, no. 11, pp. 4808–4820, 2006. [6] G. E. Pfander, “Sampling of operators,” J. Fourier Anal. Appl., vol. 19, no. 3, pp. 612–650, 2013. [7] G. E. Pfander and D. Walnut, “Sampling and reconstruction of operators,” 2014, preprint. [8] K. Gröchenig, Foundations of Time-Frequency Analysis. Birkhäuser, 2001. [9] K. A. Okoudjou, “A Beurling-Helson type theorem for modulation spaces,” Journal of Function Spaces and Applications, vol. 7, no. 1, pp. 33–41, 2009. [10] R.-D. Malikiosis, “A note on Gabor frames in finite dimensions,” Appl. Comput. Harmon. Anal., vol. 38, no. 2, pp. 318–330, 2015. [11] J. Lawrence, G. E. Pfander, and D. Walnut, “Linear independence of Gabor systems in finite dimensional vector spaces,” J. Fourier Anal. Appl., vol. 11, no. 6, pp. 715–726, 2005. [12] H. G. Feichtinger and W. Kozek, “Quantization of TF lattice-invariant operators on elementary LCA groups,” in Gabor Analysis and Algorithms, H. G. Feichtinger and T. Strohmer, Eds. Birkhäuser, 1998, ch. 7. [13] G. E. Pfander, “On the invertibility of ”rectangular” bi-infinite matrices and applications in time-frequency analysis,” Linear Algebra and Its Applications, vol. 429, no. 1, pp. 331–345, 2008.