Abstract
We consider Gabor frames generated by a general lattice and a window function that belongs to one of the following spaces: the Sobolev space \(V_1 = H^1(\mathbb {R}^d)\), the weighted \(L^2\)-space \(V_2 = L_{1 + |x|}^2(\mathbb {R}^d)\), and the space \(V_3 = \mathbb {H}^1(\mathbb {R}^d) = V_1 \cap V_2\) consisting of all functions with finite uncertainty product; all these spaces can be described as modulation spaces with respect to suitable weighted \(L^2\) spaces. In all cases, we prove that the space of Bessel vectors in \(V_j\) is mapped bijectively onto itself by the Gabor frame operator. As a consequence, if the window function belongs to one of the three spaces, then the canonical dual window also belongs to the same space. In fact, the result not only applies to frames, but also to frame sequences.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Analyzing the time-frequency localization of functions is an important topic in harmonic analysis. Quantitative results on this localization are usually formulated in terms of function spaces such as Sobolev spaces, modulation spaces, or Wiener amalgam spaces. An especially important space is the Feichtinger algebra \(S_0 = M^1\) [8, 16] which has numerous remarkable properties; see, e.g., [5, Sect. A.6] for a compact overview. Yet, in some cases it is preferable to work with more classical spaces like the Sobolev space \(H^1 (\mathbb {R}^d) = W^{1,2}(\mathbb {R}^d)\), the weighted \(L^2\)-space \(L_{1 + |x|}^2(\mathbb {R}^d) = \{ f : \mathbb {R}^d \rightarrow \mathbb {C}: (1 + |x|) f(x) \in L^2 \}\), or the space \(\mathbb {H}^1(\mathbb {R}^d) = H^1 (\mathbb {R}^d) \cap L_{1 + |x|}^2(\mathbb {R}^d)\) which consists of all functions \(g \in L^2(\mathbb {R}^d)\) with finite uncertainty product
Certainly, one advantage of these classical spaces is that membership of a function in the space can be decided easily. We remark that all of these spaces fall into the scale of modulation spaces (see Sect. 3).
In Gabor analysis, it is known (see e.g., [12, Proposition 5.2.1] and [5, Theorem 12.3.2]) that for a Gabor frame generated by a lattice, the canonical dual frame is again a Gabor system (over the same lattice), generated by the so-called dual window. An important question is what kind of time-frequency localization conditions of the generating window are inherited by the dual window. Precisely, if \(g \in L^2(\mathbb {R}^d)\) belongs to a certain “localization Banach space” V and if \(\Lambda \subset \mathbb {R}^{2d}\) is such that \((g,\Lambda )\) forms a Gabor frame for \(L^2(\mathbb {R}^d)\), then does the canonical dual window belong to V as well? A celebrated result in time-frequency analysis states that this is true for the Feichtinger algebra \(V = S_0(\mathbb {R}^d)\); see [14] for separable lattices \(\Lambda \) and [1, Theorem 7] for irregular sets \(\Lambda \). In the case of separable lattices, the question has been answered affirmatively also for the Schwartz space \(V = \mathcal {S}(\mathbb {R})\) [17, Proposition 5.5] and for the Wiener amalgam space \(V = W(L^{\infty },\ell _v^1)\) with a so-called admissible weight v; see [19]. Similarly, the setting of the spaces \(V = W(C_\alpha ,\ell _v^q)\) (with the Hölder spaces \(C_\alpha \)) is studied in [26]—but except in the case \(q = 1\), some additional assumptions on the window function g are imposed.
To the best of our knowledge, the question has not been answered for modulation spaces other than \(V = M^1_v\), and in particular, not for any of the spaces \(V = H^1(\mathbb {R}^d)\), \(V = L_{1 + |x|}^2(\mathbb {R}^d)\), and \(V = \mathbb {H}^1(\mathbb {R}^d)\). In this note, we show that the answer is affirmative for all of these spaces:
Theorem 1.1
Let \(V \in \{ H^1(\mathbb {R}^d), L_{1 + |x|}^2(\mathbb {R}^d), \mathbb {H}^1(\mathbb {R}^d) \}\). Let \(g \in V\) and let \(\Lambda \subset \mathbb {R}^{2d}\) be a lattice such that the Gabor system \((g, \Lambda )\) is a frame for \(L^2(\mathbb {R}^d)\) with frame operator S. Then the canonical dual window \(S^{-1} g\) belongs to V. Furthermore, \((S^{-1/2} g, \Lambda )\) is a Parseval frame for \(L^2(\mathbb {R}^d)\) with \(S^{-1/2} g \in V\).
We emphasize that we do not show that the inverse frame operator maps V into itself; in fact, it maps the smaller space into itself; see Proposition 4.5 below.
To indicate the practical relevance of Theorem 1.1, recall that a Gabor frame \((g, \Lambda )\) for \(L^2(\mathbb {R}^d)\) allows for the frame expansion \(f = \sum _{\lambda \in \Lambda } \langle f, \pi (\lambda )S^{-1}g \rangle \pi (\lambda )g\) for every \(f \in L^2(\mathbb {R}^d)\), where \(\pi (\lambda )\) denotes the time-frequency shift by \(\lambda \) (see (2.1) below). The sequence \((c_\lambda )_{\lambda \in \Lambda }\) of the frame coefficients \(c_\lambda = \langle f,\pi (\lambda )S^{-1}g\rangle \), \(\lambda \in \Lambda \), in general only belongs to \(\ell ^2(\Lambda )\), which means that if one truncates the sum \(f = \sum _{\lambda \in \Lambda }c_\lambda \pi (\lambda )g\) to N terms (as is necessary in practical applications), the \(L^2\)-approximation error may decay arbitrarily slowly as \(N \rightarrow \infty \).
However, one can impose a faster decay of the coefficients— and therefore improve the decay rate of the \(L^2\)-approximation error— by restricting f and g to certain subspaces of \(L^2(\mathbb {R}^d)\). For example, if \(f,g \in V = S_0(\mathbb {R}^d)\), then it is well-known that \(\bigl ( \langle f, \pi (\lambda ) S^{-1} g \rangle \bigr )_{\lambda \in \Lambda } \in \ell ^1(\Lambda )\). Then, Stechkin’s inequality (see e.g. [11, Propositions 2.3 and 2.11]) implies that if one truncates the sum \(f = \sum _{\lambda \in \Lambda } \langle f, \pi (\lambda )S^{-1}g \rangle \pi (\lambda )g\) to the N terms with the largest frame coefficients, the resulting approximation error will be \({\mathcal {O}}(N^{-1/2})\). If \(f,g \in V = \mathbb {H}^1(\mathbb {R}^d)\) and if \((f, \Lambda )\) is a Bessel sequence, then Theorem 1.1 combined with the proof of Proposition 3.2 shows that \(\bigl ( \lambda _k \langle f,\pi (\lambda )S^{-1}g\rangle \bigr )_{\lambda \in \Lambda } \in \ell ^2(\Lambda )\) for each \(k=1,\ldots ,2d\). Since \(\big ( (1 + |\lambda |)^{-1} \big )_{\lambda \in \Lambda } \in \ell ^{2d, \infty }(\Lambda )\), this implies by Hölder’s inequality for weak Lebesgue spaces (cf. [4, Theorem 5.23]) that \(\bigl ( \langle f,\pi (\lambda )S^{-1}g\rangle \bigr )_{\lambda \in \Lambda } \in \ell ^{\frac{2d}{d+1}, \infty }(\Lambda )\), so that Stechkin’s inequality shows that the error of truncating the frame expansion to the N terms with the largest frame coefficients decays like \({\mathcal {O}}(N^{-1/(2d)})\). At least in dimension \(d = 1\), this is just as good as for \(V = S_0(\mathbb {R}^d)\).
As mentioned above, the corresponding statement of Theorem 1.1 for \(V = S_0(\mathbb {R}^d)\) with separable lattices \(\Lambda \) was proved in [14]. In addition to several deeper insights, the proof given in [14] relies on a simple but essential argument showing that the frame operator \(S = S_{\Lambda ,g}\) maps V boundedly into itself, which is shown in [14] based on Janssen’s representation of \(S_{\Lambda ,g}\). In our setting, this argument is not applicable, because—unlike in the case of \(V = S_0(\mathbb {R}^d)\)— there exist functions \(g \in \mathbb {H}^1\) for which \((g,\Lambda )\) is not an \(L^2\)-Bessel system. In addition, the series in Janssen’s representation is not guaranteed to converge unconditionally in the strong sense for \(\mathbb {H}^1\)-functions, even if \((g,\Lambda )\) is an \(L^2\)-Bessel system; see Proposition A.1. To bypass these obstacles, we introduce for each space \(V \in \{ H^1, L_{1 + |x|}^2, \mathbb {H}^1 \}\) the associated subspace \(V_\Lambda \) consisting of all those functions \(g \in V\) that generate a Bessel system over the given lattice \(\Lambda \).
We remark that most of the existing works concerning the regularity of the (canonical) dual window rely on deep results related to Wiener’s 1/f-lemma on absolutely convergent Fourier series. In contrast, our methods are based on elementary spectral theory (see Sect. 4) and on certain observations regarding the interaction of the Gabor frame operator with partial derivatives; see Proposition 3.2.
The paper is organized as follows: Sect. 2 discusses the concept of Gabor Bessel vectors and introduces some related notions. Then, in Sect. 3, we endow the space \(V_\Lambda \) (for each choice \(V \in \{ H^1, L_{1 + |x|}^2, \mathbb {H}^1 \}\)) with a Banach space norm and show that the frame operator S maps \(V_\Lambda \) boundedly into itself, provided that the Gabor system \((g,\Lambda )\) is an \(L^2\)-Bessel system and that the window function g belongs to V. Finally, we prove in Sect. 4 that for any \(V \in \{ H^1, L_{1 + |x|}^2, \mathbb {H}^1 \}\) the spectrum of S as an operator on V coincides with the spectrum of S as an operator on \(L^2\). This easily implies our main result, Theorem 1.1.
2 Bessel Vectors
For \(a,b\in \mathbb {R}^d\) and \(f\in L^2(\mathbb {R}^d)\) we define the operators of translation by a and modulation by b as
respectively. Both \(T_a\) and \(M_b\) are unitary operators on \(L^2(\mathbb {R}^d)\) and hence so is the time-frequency shift
The Fourier transform \(\mathcal {F}\) is defined on \(L^1(\mathbb {R}^d) \cap L^2(\mathbb {R}^d)\) by \(\mathcal {F}f (\xi ) = {\widehat{f}}(\xi ) = \int _{\mathbb {R}^d} f(x) e^{-2 \pi i x \cdot \xi } \, d x\) and extended to a unitary operator on \(L^2(\mathbb {R}^d)\). For \(z = (z_1, z_2) \in \mathbb {R}^d \times \mathbb {R}^d \cong \mathbb {R}^{2d}\) and \(f \in L^2(\mathbb {R}^d)\), a direct calculation shows that
where
A (full rank) lattice in \(\mathbb {R}^{2d}\) is a set of the form \(\Lambda = A\mathbb {Z}^{2d}\), where \(A\in \mathbb {R}^{2d\times 2d}\) is invertible. The volume of \(\Lambda \) is defined by \({\text {Vol}}(\Lambda ) := |\!\det A|\) and its density by \(d(\Lambda ) := {\text {Vol}}(\Lambda )^{-1}\). The adjoint lattice of \(\Lambda \) is denoted and defined by \(\Lambda ^\circ := JA^{-{\!\top }}\mathbb {Z}^{2d}\).
The Gabor system generated by a window function \(g\in L^2(\mathbb {R}^d)\) and a lattice \(\Lambda \subset \mathbb {R}^{2d}\) is given by
We say that \(g\in L^2(\mathbb {R}^d)\) is a Bessel vector with respect to \(\Lambda \) if the system \((g,\Lambda )\) is a Bessel system in \(L^2(\mathbb {R}^d)\), meaning that the associated analysis operator \(C_{\Lambda ,g}\) defined by
is a bounded operator from \(L^2(\mathbb {R}^d)\) to \(\ell ^2(\Lambda )\). We define
which is a dense linear subspace of \(L^2(\mathbb {R}^d)\) because each Schwartz function is a Bessel vector with respect to any lattice; see [9, Theorem 3.3.1]. It is well-known that \(\mathcal {B}_\Lambda = \mathcal {B}_{\Lambda ^\circ }\) (see, e.g., [9, Proposition 3.5.10]). In fact, we have for \(g \in \mathcal {B}_\Lambda \) that
see [18, proof of Theorem 2.3.1]. The cross frame operator \(S_{\Lambda ,g,h}\) with respect to \(\Lambda \) and two functions \(g,h\in \mathcal {B}_\Lambda \) is defined by
In particular, we write \(S_{\Lambda ,g} := S_{\Lambda ,g,g}\) which is called the frame operator of \((g,\Lambda )\). The system \((g,\Lambda )\) is called a frame if \(S_{\Lambda ,g}\) is bounded and boundedly invertible on \(L^2(\mathbb {R}^d)\), that is, if \(A {\text {Id}}_{L^2(\mathbb {R}^d)} \le S_{\Lambda ,g} \le B {\text {Id}}_{L^2(\mathbb {R}^d)}\) for some constants \(0< A \le B < \infty \) (called the frame bounds). In particular, a frame with frame bounds \(A=B=1\) is called a Parseval frame.
In our proofs, the so-called fundamental identity of Gabor analysis will play an essential role. This identity states that
It holds, for example, if \(f, h \in M^1(\mathbb {R}^d) = S_0(\mathbb {R}^d)\) (the Feichtinger algebra) and \(g,\gamma \in L^2(\mathbb {R}^d)\); see [9, Theorem 3.5.11]. Here, we will use the following version of the fundamental identity:
Lemma 2.1
The fundamental identity (2.5) holds if \(g,h\in \mathcal {B}_\Lambda \) or \(f,\gamma \in \mathcal {B}_\Lambda \).
Proof
The proof can be carried out similarly as [13, Theorem 4.3.2 (ii)] which shows (2.5) under the assumption that \(g,\gamma \in \mathcal {B}_\Lambda \) and \(\sum _{\mu \in \Lambda ^\circ }|\langle \gamma ,\pi (\mu )g\rangle | < \infty \). Note that the latter condition guarantees the absolute convergence of the right-hand side of (2.5). In our case, each of the conditions \(g,h\in \mathcal {B}_\Lambda \) and \(f,\gamma \in \mathcal {B}_\Lambda \) already implies the absolute convergence of both sides of (2.5) (by the Cauchy-Schwarz inequality) so that the proof in [13] is also valid here. \(\square \)
3 Certain Subspaces of Modulation Spaces Invariant Under the Frame Operator
The \(L^2\)-Sobolev-space \(H^1 (\mathbb {R}^d) = W^{1,2}(\mathbb {R}^d)\) is the space of all functions \(f \in L^2(\mathbb {R}^d)\) whose distributional derivatives \(\partial _jf := \frac{\partial f}{\partial x_j}\), \(j \in \{ 1,\ldots ,d \}\), all belong to \(L^2(\mathbb {R}^d)\). We will frequently use the well-known characterization
of \(H^1(\mathbb {R}^d)\) in terms of the Fourier transform. With the weight function
we define the weighted \(L^2\)-space \(L_w^2 (\mathbb {R}^d) := \{ f\in L^2(\mathbb {R}^d) : w (\cdot ) f (\cdot ) \in L^2(\mathbb {R}^d)\}\), equipped with the norm \(\Vert f \Vert _{L_w^2} := \Vert w \, f \Vert _{L^2}\). It is then clear that \(L_w^2 (\mathbb {R}^d) = {\mathcal {F}} [ H^1 (\mathbb {R}^d) ] = {\mathcal {F}}^{-1} [ H^1 (\mathbb {R}^d) ]\). Finally, we define \(\mathbb {H}^1(\mathbb {R}^d) = H^1 (\mathbb {R}^d) \cap L_w^2 (\mathbb {R}^d)\) which is the space of all functions \(f \in H^1(\mathbb {R}^d)\) whose Fourier transform \(\widehat{f}\) also belongs to \(H^1(\mathbb {R}^d)\). Equivalently, \(\mathbb {H}^1(\mathbb {R}^d)\) is the space of all functions \(g \in L^2(\mathbb {R}^d)\) with finite uncertainty product (1.1).
It is worth to note that each of the spaces \(H^1 (\mathbb {R}^d)\), \(L_w^2 (\mathbb {R}^d)\), and \(\mathbb {H}^1(\mathbb {R}^d)\) can be expressed as a modulation space \( M_m^2(\mathbb {R}^d) = \{ f \in L^2(\mathbb {R}^d) : \int _{\mathbb {R}^{2d}} | \langle f,\pi (z)\varphi \rangle |^2 \, |m(z)|^2 \, dz < \infty \} \) with a suitable weight function \(m: \mathbb {R}^{2d} \rightarrow \mathbb {C}\), where \(\varphi \in {\mathcal {S}} (\mathbb {R}^d) \backslash \{ 0 \}\) is any fixed function,Footnote 1 for instance a Gaussian. Indeed, we have
with
see [12, Proposition 11.3.1] and [25, Corollary 2.3].
Our main goal in this paper is to prove for each of these spaces that if the window function g of a Gabor frame \((g,\Lambda )\) belongs to the space, then so does the canonical dual window. In this section, we will mostly concentrate on the space \(H^1(\mathbb {R}^d)\), since this will imply the desired result for the other spaces as well.
The corresponding result for the Feichtinger algebra \(S_0(\mathbb {R}^d)\) was proved in [14] by showing the much stronger statement that the frame operator maps \(S_0(\mathbb {R}^d)\) boundedly into itself and is in fact boundedly invertible on \(S_0(\mathbb {R}^d)\). However, the methods used in [14] cannot be directly transferred to the case of a window function in \(\mathbb {H}^1(\mathbb {R}^d)\) (or \(H^1(\mathbb {R}^d)\)), since the proof in [14] leverages two particular properties of the Feichtinger algebra which are not shared by \(\mathbb {H}^1(\mathbb {R}^d)\):
-
(a)
Every function from \(S_0(\mathbb {R}^d)\) is a Bessel vector with respect to any given lattice;
-
(b)
The series in Janssen’s representation of the frame operator converges strongly (even absolutely in operator norm) to the frame operator when the window function belongs to \(S_0(\mathbb {R}^d)\).
Indeed, it is well-known that \(g\in L^2(\mathbb {R})\) is a Bessel vector with respect to \(\mathbb {Z}\times \mathbb {Z}\) if and only if the Zak transform of g is essentially bounded (cf. [2, Theorem 3.1]), but [2, Example 3.4] provides an example of a function \(g\in \mathbb {H}^1(\mathbb {R})\) whose Zak transform is not essentially bounded; this indicates that (a) does not hold for \(\mathbb {H}^1(\mathbb {R}^d)\) instead of \(S_0(\mathbb {R}^d)\). Concerning the statement (b) for \(\mathbb {H}^1(\mathbb {R}^d)\), it is easy to see that if Janssen’s representation converges strongly (with respect to some enumeration of \(\mathbb {Z}^2\)) to the frame operator of \((g,\Lambda )\), then the frame operator must be bounded on \(L^2(\mathbb {R})\) and thus the associated window function g is necessarily a Bessel vector. Therefore, the example above again serves as a counterexample: namely, the statement (b) fails for such a non-Bessel window function \(g\in \mathbb {H}^1(\mathbb {R})\). Even more, we show in the Appendix that there exist Bessel vectors \(g\in \mathbb {H}^1(\mathbb {R})\) for which Janssen’s representation neither converges unconditionally in the strong sense nor conditionally in the operator norm.
We mention that in the case of the Wiener amalgam space \(W(L^{\infty },\ell _v^1)\) with an admissible weight v, the convergence issue was circumvented by employing Walnut’s representation instead of Janssen’s to prove the result for \(W(L^{\infty },\ell _v^1)\) in [19].
Fortunately, it turns out that establishing the corresponding result for \(V = H^1(\mathbb {R}^d)\), \(L_w^2 (\mathbb {R}^d)\), and \(\mathbb {H}^1(\mathbb {R}^d)\) only requires the invertibility of the frame operator on a particular subspace of V. Precisely, given a lattice \(\Lambda \subset \mathbb {R}^{2d}\), we define
We equip the first two of these spaces with the norms
respectively, where
and \(C_{\Lambda , f}\) is the analysis operator defined in (2.3). Finally, we equip the space \(L_{w,\Lambda }^2(\mathbb {R}^d)\) with the norm
We start by showing that these spaces are Banach spaces.
Lemma 3.1
For a lattice \(\Lambda \subset \mathbb {R}^{2d}\), the spaces \(H^1_\Lambda (\mathbb {R}^d)\), \(L_{w,\Lambda }^2(\mathbb {R}^d)\), and \(\mathbb {H}^1_\Lambda (\mathbb {R}^d)\) are Banach spaces which are continuously embedded in \(L^2(\mathbb {R}^d)\).
Proof
We naturally equip the space \(\mathcal {B}_\Lambda \subset L^2(\mathbb {R}^d)\) with the norm \(\Vert f\Vert _{\mathcal {B}_\Lambda } := \Vert C_{\Lambda ,f}\Vert _{L^2\rightarrow \ell ^2}\). Then \((\mathcal {B}_\Lambda ,\Vert \cdot \Vert _{\mathcal {B}_\Lambda })\) is a Banach space by [15, Proposition 3.1]. Moreover, for \(f\in \mathcal {B}_\Lambda \),
which implies that \(\mathcal {B}_\Lambda \hookrightarrow L^2(\mathbb {R}^d)\). Hence, if \((f_n)_{n\in \mathbb {N}}\) is a Cauchy sequence in \(H^1_\Lambda (\mathbb {R}^d)\), then it is a Cauchy sequence in both \(H^1(\mathbb {R}^d)\) (equipped with the norm \(\Vert f \Vert _{H^1} := \Vert f \Vert _{L^2} + \Vert \nabla f \Vert _{L^2}\)) and in \(\mathcal {B}_\Lambda \). Therefore, there exist \(f\in H^1(\mathbb {R}^d)\) and \(g\in \mathcal {B}_\Lambda \) such that \(\Vert f_n-f\Vert _{H^1}\rightarrow 0\) and \(\Vert f_n-g\Vert _{\mathcal {B}_\Lambda }\rightarrow 0\) as \(n\rightarrow \infty \). But as \(H^1(\mathbb {R}^d)\hookrightarrow L^2(\mathbb {R}^d)\) and \(\mathcal {B}_\Lambda \hookrightarrow L^2(\mathbb {R}^d)\), we have \(f_n\rightarrow f\) and \(f_n\rightarrow g\) also in \(L^2(\mathbb {R}^d)\), which implies \(f=g\). Hence, \(\Vert f_n - f\Vert _{H^1_\Lambda }\rightarrow 0\) as \(n\rightarrow \infty \), which proves that \(H^1_\Lambda (\mathbb {R}^d)\) is complete. The proof for \(L_{w,\Lambda }^2(\mathbb {R}^d)\) and \(\mathbb {H}^1_\Lambda (\mathbb {R}^d)\) is similar. \(\square \)
Proposition 3.2
Let \(\Lambda \subset \mathbb {R}^{2d}\) be a lattice. If \(g,h\in H^1_\Lambda (\mathbb {R}^d)\), then \(S_{\Lambda ,g,h}\) maps \(H^1_\Lambda (\mathbb {R}^d)\) boundedly into itself with operator norm not exceeding \(\Vert g\Vert _{H^1_\Lambda }\Vert h\Vert _{H^1_\Lambda }\). For \(f\in H^1_\Lambda (\mathbb {R}^d)\) and \(j \in \{1,\ldots ,d\}\) we have
where \(d_{j,\Lambda ^{\circ },g,h} \in \ell ^2(\Lambda ^{\circ })\) is defined by
Proof
Let \(f\in H^1_\Lambda (\mathbb {R}^d)\) and set \(u := S_{\Lambda ,g,h}f\). First of all, we have \(u \in \mathcal {B}_\Lambda \). Indeed, a direct computation shows that \(S_{\Lambda ,g,h}\) commutes with \(\pi (\lambda )\) for all \(\lambda \in \Lambda \), and that \(S_{\Lambda , g, h}^*= S_{\Lambda , h, g}\), which shows for \(v \in L^2(\mathbb {R}^d)\) that
and therefore
since \(S_{\Lambda ,h,g} = C_{\Lambda ,g}^*C_{\Lambda ,h}\).
We now show that \(u \in H^1(\mathbb {R}^d)\). To this end, note for \(v \in H^1(\mathbb {R}^d)\), \(a, b \in \mathbb {R}^d\), and \(j \in \{1,\ldots ,d\}\) that
and therefore
Hence, setting \(c_{\lambda ,j} := 2\pi i \cdot \lambda _{d+j}\cdot \langle f,\pi (\lambda )g \rangle \) for \(\lambda = (a,b)\in \Lambda \), we see that
In particular, \((c_{\lambda ,j})_{\lambda \in \Lambda }\in \ell ^2(\Lambda )\) for each \(j\in \{1,\ldots ,d\}\), because \(f,g\in \mathcal {B}_\Lambda \) and \(\partial _j f , \partial _j g \in L^2\).
In order to show that \(\partial _ju\) exists and is in \(L^2(\mathbb {R}^d)\), let \(\phi \in C_c^\infty (\mathbb {R}^d)\) be a test function. Note that \(C_c^\infty (\mathbb {R}^d)\subset \mathcal {B}_\Lambda \). Therefore, we obtain
with \(d_j = d_{j,\Lambda ^\circ ,g,h}\) as in (3.3). Note that \(d_j\in \ell ^2(\Lambda ^\circ )\) because \(g,h\in \mathcal {B}_{\Lambda } = \mathcal {B}_{\Lambda ^\circ }\) and \({\partial _j h , \partial _j g \in L^2}\). Since \(j\in \{1,\ldots ,d\}\) is chosen arbitrarily, this proves that \(u\in H^1(\mathbb {R}^d)\) with
for \(j \in \{1,\ldots ,d\}\), which is (3.2). Next, recalling Eq. (2.4) we get
and \(\Vert C_{\Lambda ^\circ ,f}^*\Vert = {\text {Vol}}(\Lambda )^{1/2}\Vert C_{\Lambda ,f}\Vert \). Therefore,
Hence, with (3.4), we see
and the proposition is proved. \(\square \)
4 Spectrum and Dual Windows
Let X be a Banach space. As usual, we denote the set of bounded linear operators from X into itself by \(\mathcal {B}(X)\). The resolvent set \(\rho (T)\) of an operator \(T\in \mathcal {B}(X)\) is the set of all \(z\in \mathbb {C}\) for which \(T-z := T-z I : X \rightarrow X\) is bijective. Note that \(\rho (T)\) is always open in \(\mathbb {C}\). The spectrum of T is the complement \(\sigma (T) := \mathbb {C}\backslash \rho (T)\). The approximate point spectrum \(\sigma _{{ap}}(T)\) is a subset of \(\sigma (T)\) and is defined as the set of points \(z\in \mathbb {C}\) for which there exists a sequence \((f_n)_{n\in \mathbb {N}}\subset X\) such that \(\Vert f_n\Vert =1\) for all \(n\in \mathbb {N}\) and \(\Vert (T-z)f_n\Vert \rightarrow 0\) as \(n\rightarrow \infty \). By [6, Proposition VII.6.7] we have
Lemma 4.1
Let \((\mathcal {H}, \Vert \cdot \Vert )\) be a Hilbert space, let \(S\in \mathcal {B}(\mathcal {H})\) be self-adjoint, and let \(X\subset \mathcal {H}\) be a dense linear subspace satisfying \(SX\subset X\). If \(\Vert \cdot \Vert _X\) is a norm on X such that \((X,\Vert \cdot \Vert _X)\) is complete and satisfies \(X\hookrightarrow \mathcal {H}\), then \(A := S|_X\in \mathcal {B}(X)\). If, in addition, \(\sigma _{{ap}}(A)\subset \sigma (S)\), then \(\sigma (A) = \sigma (S)\).
Proof
The fact that \(A \in \mathcal {B}(X)\) easily follows from the closed graph theorem. Next, since \(X \hookrightarrow \mathcal {H}\), there exists \(C > 0\) with \(\Vert f \Vert \le C \, \Vert f \Vert _X\) for all \(f \in X\). Assume now that additionally \(\sigma _{{ap}}(A) \subset \sigma (S)\) holds. Note that \(\sigma (S) \subset \mathbb {R}\), since S is self-adjoint. Since \(\sigma (A)\subset \mathbb {C}\) is compact, the value \({r := \max _{w\in \sigma (A)}|\!\Im w|}\) exists. Choose \(z\in \sigma (A)\) such that \(|\!\Im z| = r\). Clearly, z cannot belong to the interior of \(\sigma (A)\), and hence \(z \in \partial \sigma (A)\). In view of Eq. (4.1), this implies \(z \in \sigma _{{ap}}(A) \subset \sigma (S) \subset \mathbb {R}\), hence \(r = 0\) and thus \(\sigma (A) \subset \mathbb {R}\). Therefore, \(\sigma (A)\) has empty interior in \(\mathbb {C}\), meaning \(\sigma (A) = \partial \sigma (A)\). Thanks to Eq. (4.1), this means \(\sigma (A) \subset \sigma _{{ap}}(A)\), and hence \(\sigma (A) \subset \sigma (S)\), since by assumption \(\sigma _{{ap}}(A) \subset \sigma (S)\).
For the converse inclusion it suffices to show that \(\rho (A)\cap \mathbb {R}\subset \rho (S)\). To see that this holds, let \(z\in \rho (A)\cap \mathbb {R}\) and denote by E the spectral measure of the self-adjoint operator S. Since \(\mathbb {R}\cap \rho (A) \subset \mathbb {R}\) is open, there are \(a, b \in \mathbb {R}\) and \(\delta _0 > 0\) such that \(z \in (a,b)\) and \([a-\delta _0, b+\delta _0] \subset \rho (A)\). By Stone’s formula (see, e.g., [21, Thm. VII.13]), the spectral projection of S with respect to (a, b] can be expressed as
where all limits are taken with respect to the norm of \(\mathcal {H}\).
Note for \(w \in \mathbb {C}\setminus \mathbb {R}\) that \(w \in \rho (S) \subset \rho (A)\). Furthermore, \(A - w = (S - w)|_X\), which easily implies \((S - w)^{-1}|_X = (A - w)^{-1}\). Hence, for \(f \in X\),
since the map \(\rho (A)\rightarrow X\), \(z\mapsto (A-z)^{-1}f\) is analytic and thus uniformly continuous on compact sets. This implies \(E((a,b])f = 0\) for all \(f\in X\) and therefore \(E((a,b]) = 0\) as X is dense in \(\mathcal {H}\). But this means that \((a,b) \subset \rho (S)\) (see [21, Prop. on p. 236]) and thus \(z \in \rho (S)\). \(\square \)
For proving the invertibility of \(S_{\Lambda ,g}\) on \(H_{\Lambda }^1, L_{w,\Lambda }^2\), and \(\mathbb {H}^1_{\Lambda }\), we first focus on the space \(H^1_\Lambda (\mathbb {R}^d)\). Note that if \(g \in H^1_\Lambda (\mathbb {R}^d)\), then \(S_{\Lambda ,g}\) maps \(H^1_\Lambda (\mathbb {R}^d)\) boundedly into itself by Proposition 3.2. For \(g \in H^1_\Lambda (\mathbb {R}^d)\), we will denote the restriction of \(S_{\Lambda ,g}\) to \(H^1_\Lambda (\mathbb {R}^d)\) by \(A_{\Lambda ,g}\); that is, \(A_{\Lambda ,g} := S_{\Lambda ,g} |_{H^1_\Lambda (\mathbb {R}^d)} \in \mathcal {B}(H^1_\Lambda (\mathbb {R}^d))\).
Theorem 4.2
Let \(\Lambda \subset \mathbb {Z}^{2d}\) be a lattice and let \(g\in H^1_\Lambda (\mathbb {R}^d)\). Then
Proof
For brevity, we set \(A := A_{\Lambda ,g}\) and \(S := S_{\Lambda ,g}\). Due to Lemma 4.1, we only have to prove that \(\sigma _{{ap}}(A)\subset \sigma (S)\). For this, let \(z\in \sigma _{{ap}}(A)\). Then there exists a sequence \((f_n)_{n\in \mathbb {N}} \subset H^1_\Lambda (\mathbb {R}^d)\) such that \(\Vert f_n\Vert _{H^1_\Lambda } = 1\) for all \(n \in \mathbb {N}\) and \(\Vert (A - z)f_n\Vert _{H^1_\Lambda }\rightarrow 0\) as \(n \rightarrow \infty \). The latter means that, for each \(j \in \{1, \ldots , d\}\),
Suppose towards a contradiction that \(z \notin \sigma (S)\). Since S is self-adjoint, this implies \({\overline{z}} \notin \sigma (S)\). Furthermore, because S is self-adjoint and commutes with \(\pi (\lambda )\) for all \(\lambda \in \Lambda \), we see for \(f \in \mathcal {B}_\Lambda \) that \(C_{\Lambda ,(S - z)f} = C_{\Lambda ,f} \circ (S - \overline{z})\) and hence \(C_{\Lambda ,f_n} = C_{\Lambda ,(S-z)f_n} \circ (S - \overline{z})^{-1}\), which implies that \(\Vert C_{\Lambda ,f_n}\Vert \rightarrow 0\). Hence, also \(\Vert C_{\Lambda ^\circ ,f_n}\Vert \rightarrow 0\) as \(n\rightarrow \infty \) (see Eq. (2.4)). Now, by Eq. (3.2), we have
with some \(d_j\in \ell ^2(\Lambda ^\circ )\) which is independent of n. Hence, the first limit in (4.2) combined with \(\Vert C_{\Lambda ^\circ ,f_n}\Vert \rightarrow 0\) implies that \(\Vert (S - z)(\partial _jf_n)\Vert _{L^2}\rightarrow 0\) and thus \(\Vert \partial _jf_n\Vert _{L^2}\rightarrow 0\) as \(n\rightarrow \infty \) for all \(j \in \{1,\ldots ,d\}\), since \(z \notin \sigma (S)\). Hence, \(\Vert f_n\Vert _{H^1_\Lambda } = \sum _{j=1}^d \Vert \partial _j f_n\Vert _{L^2} + \Vert C_{\Lambda ,f_n}\Vert \rightarrow 0\) as \(n\rightarrow \infty \), in contradiction to \(\Vert f_n\Vert _{H^1_\Lambda } = 1\) for all \(n \in \mathbb {N}\). This proves that, indeed, \(\sigma _{{ap}}(A)\subset \sigma (S)\). \(\square \)
We now show analogous properties to Proposition 3.2 and Theorem 4.2 for \(L_{w,\Lambda }^2(\mathbb {R}^d)\).
Corollary 4.3
Let \(\Lambda \subset \mathbb {Z}^{2d}\) be a lattice. If \(g, h \in L_{w,\Lambda }^2(\mathbb {R}^d)\), then \(S_{\Lambda ,g,h}\) maps \(L_{w,\Lambda }^2(\mathbb {R}^d)\) boundedly into itself.
If \(g = h\) and if \(A^w_{\Lambda ,g} := S_{\Lambda ,g} |_{L_{w,\Lambda }^2(\mathbb {R}^d)} \in \mathcal {B}(L_{w,\Lambda }^2(\mathbb {R}^d))\) denotes the restriction of \(S_{\Lambda ,g}\) to \(L_{w,\Lambda }^2(\mathbb {R}^d)\), then
Proof
We equip the space \(\mathcal {B}_{\Lambda } \subset L^2(\mathbb {R}^d)\) with the norm \({\Vert f \Vert _{\mathcal {B}_\Lambda } := \Vert C_{\Lambda ,f} \Vert _{L^2 \rightarrow \ell ^2}}\), where we recall from Eq. (3.1) that \(\Vert f \Vert _{L^2} \le \Vert f \Vert _{\mathcal {B}_\Lambda }\). Equation (2.2) shows that the Fourier transform is an isometric isomorphism from \(\mathcal {B}_{\Lambda }\) to \(\mathcal {B}_{{\widehat{\Lambda }}}\), where \({\widehat{\Lambda }} := J \Lambda \). Furthermore, it is well-known (see for instance [10, Sect. 9.3]) that the Fourier transform \({\mathcal {F}: L^2 \rightarrow L^2}\) restricts to an isomorphism of Banach spaces \({\mathcal {F}: L_{w}^2(\mathbb {R}^d) \rightarrow H^1(\mathbb {R}^d)}\), where \(H^1\) is equipped with the norm \(\Vert f \Vert _{H^1} := \Vert f \Vert _{L^2} + \Vert \nabla f \Vert _{L^2}\). Taken together, we thus see that the Fourier transform restricts to an isomorphism \(\mathcal {F}: L_{w,\Lambda }^2(\mathbb {R}^d)\rightarrow H_{{\widehat{\Lambda }}}^1(\mathbb {R}^d)\); here, we implicitly used that \({\Vert f \Vert _{H_{{\widehat{\Lambda }}}^1} \asymp \Vert f \Vert _{H^1} + \Vert f \Vert _{\mathcal {B}_{{\widehat{\Lambda }}}}}\), which follows from \(\Vert \cdot \Vert _{L^2} \le \Vert \cdot \Vert _{\mathcal {B}_{{\widehat{\Lambda }}}}\).
Plancherel’s theorem, in combination with Eq. (2.2) shows for \(f \in L^2(\mathbb {R}^d)\) that
Since \( A_{{\widehat{\Lambda }},{\widehat{g}},{\widehat{h}}} = S_{{\widehat{\Lambda }},{\widehat{g}},{\widehat{h}}}|_{H_{{\widehat{\Lambda }}}^1} : H_{{\widehat{\Lambda }}}^1(\mathbb {R}^d) \rightarrow H_{{\widehat{\Lambda }}}^1(\mathbb {R}^d) \) is well-defined and bounded by Proposition 3.2, the preceding calculation combined with the considerations from the previous paragraph shows that \(A_{\Lambda ,g,h}^w = S_{\Lambda ,g,h}|_{L_{w,\Lambda }^2(\mathbb {R}^d)} : L_{w,\Lambda }^2(\mathbb {R}^d)\rightarrow L_{w,\Lambda }^2(\mathbb {R}^d)\) is well-defined and bounded, with
Finally, if \(g = h\), we see \( \sigma (A_{\Lambda ,g,g}^w) = \sigma (A_{{\widehat{\Lambda }},{\widehat{g}},{\widehat{g}}}) = \sigma (S_{{\widehat{\Lambda }},{\widehat{g}},{\widehat{g}}}) = \sigma (S_{\Lambda ,g,g}) , \) where the second step is due to Theorem 4.2, and the final step used the identity \(S_{\Lambda ,g,h} = \mathcal {F}^{-1} \circ S_{{\widehat{\Lambda }},{\widehat{g}},{\widehat{h}}} \circ \mathcal {F}\) from above. \(\square \)
Finally, we establish the corresponding properties for \(\mathbb {H}^1_\Lambda (\mathbb {R}^d) = H^1_\Lambda (\mathbb {R}^d) \cap L_{w,\Lambda }^2(\mathbb {R}^d)\).
Corollary 4.4
Let \(\Lambda \subset \mathbb {Z}^{2d}\) be a lattice. If \(g,h\in \mathbb {H}^1_\Lambda (\mathbb {R}^d)\), then \(S_{\Lambda ,g,h}\) maps \(\mathbb {H}^1_\Lambda (\mathbb {R}^d)\) boundedly into itself. If \(g=h\) and \({\mathbb {A}}_{\Lambda ,g} := S_{\Lambda ,g} |_{\mathbb {H}^1_\Lambda (\mathbb {R}^d)} \in \mathcal {B}(\mathbb {H}^1_\Lambda (\mathbb {R}^d))\) denotes the restriction of \(S_{\Lambda ,g}\) to \(\mathbb {H}^1_\Lambda (\mathbb {R}^d)\), then
Proof
From the definition of \(\mathbb {H}_\Lambda ^1\) and the proof of Corollary 4.3 it is easy to see that \(\mathbb {H}_\Lambda ^1 = H_\Lambda ^1 \cap L_{w,\Lambda }^2(\mathbb {R}^d)\), and \(\Vert \cdot \Vert _{\mathbb {H}_\Lambda ^1} \asymp \Vert \cdot \Vert _{H_\Lambda ^1} + \Vert \cdot \Vert _{L_{w,\Lambda }^2}\). Therefore, Proposition 3.2 and Corollary 4.3 imply that \(S_{\Lambda ,g,h}\) maps \(\mathbb {H}_\Lambda ^1(\mathbb {R}^d)\) boundedly into itself.
Lemma 4.1 shows that to prove (4.3), it suffices to show \(\sigma _{{ap}}({\mathbb {A}}_{\Lambda ,g}) \subset \sigma (S_{\Lambda ,g})\). Thus, let \({z \in \sigma _{{ap}}({\mathbb {A}}_{\Lambda ,g})}\). Then there exists \((f_n)_{n \in \mathbb {N}} \subset \mathbb {H}^1_\Lambda (\mathbb {R}^d)\) with \(\Vert f_n\Vert _{\mathbb {H}_{\Lambda }^1}=1\) for all \(n \in \mathbb {N}\) and \(\Vert (\mathbb A_{\Lambda ,g}-z)f_n\Vert _{\mathbb {H}^1_\Lambda }\rightarrow 0\) as \(n\rightarrow \infty \). Thus, \(\Vert (A_{\Lambda ,g}-z)f_n\Vert _{H^1_\Lambda }\rightarrow 0\) and \({\Vert (A_{\Lambda ,g}^w - z)f_n \Vert _{L_{w,\Lambda }^2} \rightarrow 0}\) as \(n\rightarrow \infty \). Furthermore, there is a subsequence \((n_k)_{k \in \mathbb {N}}\) such that \(\lim _{k\rightarrow \infty } \Vert f_{n_k} \Vert _{H_\Lambda ^1} > 0\) or \(\lim _{k\rightarrow \infty } \Vert f_{n_k} \Vert _{L_{w,\Lambda }^2} > 0\). Hence, \(z \in \sigma (A_{\Lambda ,g})\) or \(z \in \sigma (A_{\Lambda ,g}^w)\). But Theorem 4.2 and Corollary 4.3 show \(\sigma (A_{\Lambda ,g}) = \sigma (A_{\Lambda ,g}^w) = \sigma (S_{\Lambda ,g})\). We have thus shown \(\sigma _{{ap}}({\mathbb {A}}_{\Lambda ,g}) \subset \sigma (S_{\Lambda ,g})\), so that Lemma 4.1 shows \(\sigma ({\mathbb {A}}_{\Lambda ,g}) = \sigma (S_{\Lambda ,g})\). \(\square \)
The next proposition shows that any operator obtained from \(S_{\Lambda ,g}\) through the holomorphic spectral calculus (see [22, Sects. 10.21–10.29] for a definition) maps each of the spaces \(H_{\Lambda }^1(\mathbb {R}^d)\), \(L_{w,\Lambda }^2(\mathbb {R}^d)\), and \(\mathbb {H}_\Lambda ^1(\mathbb {R}^d)\) into itself.
Proposition 4.5
Let \(\Lambda \subset \mathbb {R}^{2d}\) be a lattice, let \(V \in \{ H_\Lambda ^1(\mathbb {R}^d), L_{w,\Lambda }^2(\mathbb {R}^d), \mathbb {H}_\Lambda ^1(\mathbb {R}^d) \}\), and \(g \in V\). Then for any open set \(\Omega \subset \mathbb {C}\) with \(\sigma (S_{\Lambda ,g}) \subset \Omega \), any analytic function \(F : \Omega \rightarrow \mathbb {C}\), and any \(f \in V\), we have \(F(S_{\Lambda ,g})f \in V\).
Proof
We only prove the claim for \(V = H_{\Lambda }^1 (\mathbb {R}^d)\); the proofs for the other cases are similar, using Corollaries 4.3 or 4.4 instead of Theorem 4.2. Thus, let \(g\in H^1_\Lambda (\mathbb {R}^d)\) and set \(S := S_{\Lambda ,g}\) and \(A := A_{\Lambda ,g}\). Let \(f\in H^1_\Lambda (\mathbb {R}^d)\) and define
where \(\Gamma \subset \Omega \setminus \sigma (S)\) is a finite set of closed rectifiable curves surrounding \(\sigma (S) = \sigma (A)\) (existence of such curves is shown in [24, Theorem 13.5]). Note that the integral converges in \(H^1_\Lambda (\mathbb {R}^d)\). Since \(H^1_\Lambda (\mathbb {R}^d)\hookrightarrow L^2(\mathbb {R}^d)\), it also converges (to the same limit) in \(L^2(\mathbb {R}^d)\) and hence, by definition of the holomorphic spectral calculus,
\(\square \)
Our main result (Theorem 1.1) is now an easy consequence of Proposition 4.5.
Proof
(Proof of Theorem 1.1) Using the fact that \(S_{\Lambda ,g}\) commutes with \(\pi (\lambda )\) for all \(\lambda \in \Lambda \), it is easily seen that \((S_{\Lambda ,g}^{-1} \, g,\Lambda )\) is the canonical dual frame of \((g,\Lambda )\) and that \((S_{\Lambda ,g}^{-1/2} g, \Lambda )\) is a Parseval frame for \(L^2(\mathbb {R}^d)\); see for instance, [5, Theorem 12.3.2]. Note that since \((g,\Lambda )\) is a frame for \(L^2(\mathbb {R}^d)\), we have \(\sigma (S_{\Lambda ,g}) \subset [A,B]\) where \(0< A \le B < \infty \) are the optimal frame bounds for \((g,\Lambda )\). Thus, we obtain \(S_{\Lambda ,g}^{-1} \, g \in V_\Lambda \subset V\) and \(S_{\Lambda ,g}^{-1/2} g \,\in \, V_\Lambda \subset V\) from Proposition 4.5 with \(F(z) = z^{-1}\) and \(F(z) = z^{-1/2}\) (with any suitable branch cut; for instance, the half-axis \((-\infty ,0]\)), respectively, on \(\Omega = \bigl \{ x + i y : x \in ( \frac{A}{2}, \infty ) , y \in \mathbb {R}\bigr \}\). \(\square \)
Finally, we state and prove a version of Theorem 1.1 for Gabor frame sequences. For completeness, we briefly recall the necessary concepts. Generally, a (countable) family \((h_i)_{i \in I}\) in a Hilbert space \(\mathcal {H}\) is called a frame sequence, if \((h_i)_{i \in I}\) is a frame for the subspace \(\mathcal {H}' := {\overline{{\text {span}}}} \{ h_i :i \in I \} \subset \mathcal {H}\). In this case, the frame operator \(S : \mathcal {H}\rightarrow \mathcal {H}, f \mapsto \sum _{i \in I} \langle f, h_i \rangle h_i\), is a bounded, self-adjoint operator on \(\mathcal {H}\), and \(S|_{\mathcal {H}'} : \mathcal {H}' \rightarrow \mathcal {H}'\) is boundedly invertible; in particular, \({\text {ran}}S = \mathcal {H}' \subset \mathcal {H}\) is closed, so that S has a well-defined pseudo-inverse \(S^{\dagger }\), given by
where \(P_{\mathcal {H}'}\) denotes the orthogonal projection onto \(\mathcal {H}'\). The canonical dual system of \((h_i)_{i \in I}\) is then given by \((h_i')_{i \in I} = (S^{\dagger } h_i)_{i \in I} \subset \mathcal {H}'\), and it satisfies \(\sum _{i \in I} \langle f, h_i \rangle h_i ' = \sum _{i \in I} \langle f, h_i' \rangle h_i = P_{\mathcal {H}'} f\) for all \(f \in \mathcal {H}\).
Finally, in the case where \((h_i)_{i \in I} = (g,\Lambda )\) is a Gabor family with a lattice \(\Lambda \), it is easy to see that \(S \circ \pi (\lambda ) = \pi (\lambda ) \circ S\) and \(\pi (\lambda ) \mathcal {H}' \subset \mathcal {H}'\) for \(\lambda \in \Lambda \), which implies \(P_{\mathcal {H}'} \circ \pi (\lambda ) = \pi (\lambda ) \circ P_{\mathcal {H}'}\), and therefore \(S^\dagger \circ \pi (\lambda ) = \pi (\lambda ) \circ S^\dagger \) for all \(\lambda \in \Lambda \). Consequently, setting \(\gamma := S^\dagger g\), we have \(S^\dagger (\pi (\lambda ) g) = \pi (\lambda )\gamma \), so that the canonical dual system of a Gabor frame sequence \((g,\Lambda )\) is the Gabor system \((\gamma ,\Lambda )\), where \(\gamma = S^{\dagger } g\) is called the canonical dual window of \((g,\Lambda )\). Our next result shows that \(\gamma \) inherits the regularity of g, if one measures this regularity using one of the three spaces \(H^1, L_w^2\), or \(\mathbb {H}^1\).
Proposition 4.6
Let \(V \in \{ H^1(\mathbb {R}^d), L_w^2(\mathbb {R}^d), \mathbb {H}^1(\mathbb {R}^d) \}\). Let \(\Lambda \subset \mathbb {R}^{2d}\) be a lattice and let \(g \in V\). If \((g,\Lambda )\) is a frame sequence, then the associated canonical dual window \(\gamma \) satisfies \(\gamma \in V\).
Proof
The frame operator \(S : L^2(\mathbb {R}^d) \rightarrow L^2(\mathbb {R}^d)\) associated to \((g,\Lambda )\) is non-negative and has closed range. Consequently, there exist \(\varepsilon > 0\) and \(R > 0\) such that \(\sigma (S) \subset \{ 0 \} \cup [\varepsilon ,R]\); see for instance [3, Lemma A.2]. Now, with the open ball \(B_\delta (0) := \{ z \in \mathbb {C}:|z| < \delta \}\), define
noting that \(\Omega \subset \mathbb {C}\) is open, with \(\sigma (S) \subset \Omega \). Furthermore, it is straightforward to see that
is holomorphic. Since the functional calculus for self-adjoint operators is an extension of the holomorphic functional calculus, [3, Lemma A.6] shows that \(S^{\dagger } = \varphi (S)\). Finally, since \(g \in V_\Lambda \), Proposition 4.5 now shows that \(\gamma = S^{\dagger } g = \varphi (S) g \in V_\Lambda \subset V\) as well. \(\square \)
Notes
The definition of \(M_m^2\) is known to be independent of the choice of \(\varphi \); see e.g., [12, Proposition 11.3.2].
References
Balan, R., Casazza, P., Heil, C., Landau, Z.: Density, overcompleteness, and localization of frames, II. Gabor systems. J. Fourier Anal. Appl. 12, 309–344 (2006)
Benedetto, J.J., Heil, C., Walnut, D.F.: Differentiation and the Balian-Low theorem. J. Fourier Anal. Appl. 1, 355–402 (1995)
Caragea, A., Lee, D.G., Philipp, F., Voigtlaender, F.: A quantitative subspace Balian-Low theorem. Appl. Comput. Harmon. Anal. 55, 368–404 (2021)
Castillo, R.E., Rafeiro, H.: An Introductory Course in Lebesgue Spaces. CMS Books in Mathematics. Springer, Cham (2016)
Christensen, O.: An Introduction to Frames and Riesz Bases. Applied and Numerical Harmonic Analysis, 2nd edn. Birkhäuser/Springer, Cham (2016)
Conway, J.B.: A Course in Functional Analysis, 2nd edn. Springer, New York (1990)
Evans, L.C.: Partial Differential Equations. 2nd ed., Volume 19 of Graduate Studies in Mathematics, American Mathematical Society, Providence, RI (2010)
Feichtinger, H.G.: On a new Segal algebra. Monatsh. Math. 92, 269–289 (1981)
Feichtinger, H.G., Zimmermann, G.: A Banach space of test functions for Gabor analysis. In: Feichtinger, H.G., Strohmer, T. (eds.) Gabor Analysis and Algorithms. Applied and Numerical Harmonic Analysis, pp. 123–170. Birkhäuser, Boston, MA (1998)
Folland, G.B.: Real Analysis, 2nd edn. John Wiley & Sons Inc, New York (1999)
Foucart, S., Rauhut, H.: A Mathematical Introduction to Compressive Sensing. Birkhäuser/Springer, New York (2013)
Gröchenig, K.: Foundations of Time-Frequency Analysis. Applied and Numerical Harmonic Analysis. Birkhäuser, Boston, MA (2001)
Gröchenig, K., Koppensteiner, S.: Gabor frames: characterizations and coarse structure. In: Aldroubi, A., Cabrelli, C., Jaffard, S., Molter, U. (eds.) New Trends in Applied Harmonic Analysis. Applied and Numerical Harmonic Analysis, vol. 2, pp. 93–120. Birkhäuser, Cham (2019)
Gröchenig, K., Leinert, M.: Wiener’s lemma for twisted convolution and Gabor frames. J. Am. Math. Soc. 17, 1–18 (2003)
Han, D., Larson, D.R.: Unitary systems and Bessel generator multipliers. In: Cohen, J., Zayed, A. (eds.) Wavelets and Multiscale Analysis. Applied and Numerical Harmonic Analysis, pp. 131–150. Birkhäuser, Boston (2011)
Jakobsen, M.S.: On a (no longer) new Segal algebra: a review of the Feichtinger algebra. J. Fourier Anal. Appl. 24, 1579–1660 (2018)
Janssen, A.J.E.M.: Duality and biorthogonality for Weyl-Heisenberg frames. J. Fourier Anal. Appl. 1, 403–436 (1995)
Koppensteiner, S.: Characterizations of Gabor frames. Master’s thesis, University of Vienna, (2017). http://othes.univie.ac.at/50275
Krishtal, I.A., Okoudjou, K.A.: Invertibility of the Gabor frame operator on the Wiener amalgam space. J. Approx. Theory 153, 212–224 (2008)
Leoni, G.: A First Course in Sobolev Spaces. 2nd ed., Volume 181 of Graduate Studies in Mathematics, American Mathematical Society, Providence, RI (2017)
Reed, M., Simon, B.: Methods of Modern Mathematical Physics, I: Functional Analysis. Academic Press, New York (1980)
Rudin, W.: Functional Analysis, 2nd edn. McGraw-Hill Inc, New York (1991)
Rudin, W.: Principles of Mathematical Analysis, 3rd edn. McGraw-Hill Inc, New York (1976)
Rudin, W.: Real and Complex Analysis, 3rd edn. McGraw-Hill Inc, New York (1987)
Tinaztepe, R., Heil, C.: Modulation spaces, BMO, and the Balian-Low theorem. Sampl. Theory Signal Image Process. 11, 25–41 (2012)
Weisz, F.: Invertibility of the Gabor frame operator on some function spaces. Acta Math. Hungar. 144, 167–181 (2014)
Acknowledgements
D.G. Lee acknowledges support by the DFG Grants PF 450/6-1 and PF 450/9-1. F. Philipp was funded by the Carl Zeiss Foundation within the project DeepTurb—Deep Learning in und von Turbulenz. F. Voigtlaender acknowledges support by the German Science Foundation (DFG) in the context of the Emmy Noether junior research group VO 2594/1-1.
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Karlheinz Gröchenig.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix A: (Non)-convergence of Janssen’s Representation for \(\mathbb {H}^1\) Windows
Appendix A: (Non)-convergence of Janssen’s Representation for \(\mathbb {H}^1\) Windows
In this appendix we provide a counterexample showing that Janssen’s representation of the frame operator associated to a Bessel vector \(g \in \mathbb {H}^1\) in general does not converge unconditionally with respect to the strong operator topology. We furthermore show that for convergence in operator norm, even conditional convergence fails in general.
For simplicity, we only consider the setting \(d = 1\) and the lattice \(\Lambda = \mathbb {Z}\times \mathbb {Z}\). Thus, given a function \(g \in \mathbb {H}^1 = \mathbb {H}^1(\mathbb {R})\), we say that g is a Bessel vector if the Gabor system \((T_k M_\ell \, g)_{k,\ell \in \mathbb {Z}} \subset L^2(\mathbb {R})\) is a Bessel system. In this case, Janssen’s representation of the frame operator \(S := S_g := S_{\mathbb {Z}\times \mathbb {Z}, g, g}\) is (formally) given by
We are interested in the question whether the series defining Janssen’s representation is unconditionally convergent in the strong operator topology (SOT ), as an operator on \(L^2(\mathbb {R})\). We will construct a function \(g \in \mathbb {H}^1\) for which this fails.
1.1 A.1. Properties of the Zak Transform
The construction of the counterexample is based on several properties of the Zak transform that we briefly recall. Given \(f \in L^2(\mathbb {R})\), its Zak transform \(Z f \in L_{\textrm{loc}}^2(\mathbb {R}^2)\) is defined as
where the series converges in \(L_{\textrm{loc}}^2 (\mathbb {R}^2)\); this is a consequence of the fact that
as shown in [12, Theorem 8.2.3] and of the fact that the Zak transform Zf of a function \(f \in L^2(\mathbb {R})\) is always quasi-periodic, meaning that
for (almost) all \(x,\omega \in \mathbb {R}\) and all \(n \in \mathbb {Z}\); see [12, Equations (8.4) and (8.5)]. Another crucial property is the interplay between the Zak transform and the time-frequency shifts \(T_k M_n\), as expressed by the following formula (found in [12, Eq. (8.7)]):
where we used the functions
Note that \((e_{n,k})_{n,k\in \mathbb {Z}}\) is an orthonormal basis of \(L^2 ([0,1]^2)\).
Finally, we note the following equivalence, taken from [2, Theorem 3.1]:
1.2 A.2. Properties of \(\mathbb {H}^1\)
A further important property that we will use is the following characterization of the space \(\mathbb {H}^1\) via the Zak transform, a proof of which is given in [3, Lemma 2.4].
It is crucial to observe that the Sobolev space \(W^{1,2}(\mathbb {R}^2)\) belongs to the “borderline” case of the Sobolev embedding theorem, meaning \( W^{1,2}_{\textrm{loc}}(\mathbb {R}^2) \not \hookrightarrow L^\infty _{\textrm{loc}}(\mathbb {R}^2)\). In fact, it is easy to verify (see e.g. [7, Page 280]) for \(x_0 := (\frac{1}{2}, \frac{1}{2})^T \in \mathbb {R}^2\) that the function
belongs to \(W^{1,2}( (0,1)^2 )\), but is not essentially bounded. Now, using the chain rule and the product rule for Sobolev functions (see e.g. [20, Exercise 11.51] and [7, Theorem 1 in Sect. 5.2.3]), we see that if \(\varphi \in C_c^\infty ( (0,1)^2 )\) is chosen such that \(0 \le \varphi \le 1\) and such that \(\varphi \equiv 1\) on a neighborhood of \(x_0\), then the function
satisfies \(u \in W^{1,2}(\mathbb {R}^2)\), is continuous and bounded on \(\mathbb {R}^2 \setminus \{ x_0 \}\), but \(\lim _{x \rightarrow x_0} u(x)\) does not exist; this uses that \(\lim _{x \rightarrow \infty } \sin (x)\) does not exist and that on each small ball \(B_\varepsilon (x_0)\), the function \(u_0\) attains all values from \((M,\infty )\), for a suitable \(M = M(\varepsilon ) > 0\).
1.3 A.3. A Connection to Fourier Series
In this subsection, we show that for any fixed window \(g \in L^2(\mathbb {R})\) the unconditional convergence of Janssen’s representation in the strong operator topology implies that the partial sums of a certain Fourier series are uniformly bounded in \(L^\infty \). This connection will be used in the next subsection to disprove the unconditional convergence of Janssen’s representation in the strong operator topology.
Precisely, define \(Q := [0,1]^2\). For \(H \in L^\infty (Q)\), define the associated multiplication operator as
It is well-known that \(\Vert M_H \Vert _{L^2 \rightarrow L^2} = \Vert H \Vert _{L^\infty }\).
Let us fix any window \(g \in L^2(\mathbb {R})\). Given a finite set \(I \subset \mathbb {Z}^2\), we define
Using Eq. A.4 and the isometry of the Zak transform, we then see
where \(I' := \{ (\ell ,-k) :(k,\ell ) \in I \}\) and
In other words, we have
Given \(J \subset \mathbb {Z}^2\), define \(J_*:= \{ (-\ell ,k) :(k,\ell ) \in J \}\) and note \((J_*)' = J\). Now, suppose that \((S_I)_I\) converges strongly to some (bounded) operator, as \(I \rightarrow \mathbb {Z}^2\); this is always the case if Janssen’s representation converges unconditionally (to S or some other operator) in the SOT. Then, given any sequence \((J_n)_{n \in \mathbb {N}}\) of finite subsets \(J_n \subset \mathbb {Z}^2\) with \(J_n \subset J_{n+1}\) and \(\bigcup _{n=1}^\infty J_n = \mathbb {Z}^2\), we see \((J_n)_*\rightarrow \mathbb {Z}^2\) so that the sequence \((S_{(J_n)_*})_{n \in \mathbb {N}}\) converges strongly to some bounded operator. By the uniform boundedness principle, this shows \(\Vert S_{(J_n)_*} \Vert _{L^2 \rightarrow L^2} \le C\) for all \(n \in \mathbb {N}\) and some \(C > 0\). By Eqs. A.2 and A.8, and because of \(((J_n)_*)' = J_n\), this implies
meaning that the partial Fourier sums \(\mathcal {F}_{J_n} [|Z g|^2]\) of the function \(|Zg|^2\) are uniformly bounded in \(L^\infty (Q)\).
1.4 A.4. The Counterexample
In this subsection, we prove the following:
Proposition A.1
There exists a Bessel vector \(g \in \mathbb {H}^1(\mathbb {R})\) such that the series defining Janssen’s representation of the frame operator \(S = S_g = S_{\mathbb {Z}\times \mathbb {Z}, g, g}\) associated to g is not unconditionally convergent in the strong operator topology.
To prove the proposition, we consider the function \(F := u : (0,1)^2 \rightarrow [0,\infty )\) introduced in Eq. A.7. The properties of F that we need are the following:
-
(1)
F has compact support in \((0,1)^2\), say \({\text {supp}}F \subset (\delta , 1-\delta )^2\) for some \(\delta \in (0,\frac{1}{2})\).
-
(2)
F is bounded, but discontinuous at \(x_0 \in (0,1)^2\) (even after adjusting F on a set of measure zero).
-
(3)
\(F \in W^{1,2}\bigl ( (0,1)^2 \bigr )\).
We now extend F by zero to \([0,1)^2\) and then extend 1-periodically in both coordinates to \(\mathbb {R}^2\). Thanks to the compact support of F, it is easy to see that \(F \in W^{1,2}_{\textrm{loc}}(\mathbb {R}^2)\).
Furthermore, we consider the function
where for each \(x\in \mathbb {R}\), \(\lfloor x \rfloor \in \mathbb {Z}\) denotes the unique integer such that \(x \in \lfloor x \rfloor + [0,1)\). It is then straightforward to verify that \(G_0\) is quasi-periodic (see Eq. A.3), i.e., \(G_0(x+m,\omega ) = e^{2 \pi i m \omega } G_0(x,\omega )\) and \(G_0(x,\omega +m) = G_0(x,\omega )\) for \(x,\omega \in \mathbb {R}\) and \(m \in \mathbb {Z}\). Since F is 1-periodic in both coordinates, it is easy to see that \(F \cdot G_0\) is quasi-periodic as well.
Finally, we choose a smooth function \(\psi : \mathbb {R}\rightarrow \mathbb {R}\) satisfying \(\psi (x) = n\) for all \(x \in n + [\delta ,1-\delta ]\) with \(n \in \mathbb {Z}\), and define
Using that \(F(x,\omega ) = 0\) for \(n \in \mathbb {Z}\) and \(x \in [n,n+1] \setminus (n +\delta ,n +1-\delta )\), it is easy to check \(F \cdot G_0 = F \cdot G\), so that \(H := F \cdot G \in W^{1,2}_{\textrm{loc}}(\mathbb {R}^2) \subset L^2_{\textrm{loc}}(\mathbb {R}^2)\) is quasi-periodic.
Since the Zak transform \(Z : L^2(\mathbb {R}) \rightarrow L^2( (0,1)^2 )\) is unitary, there exists a unique function \(g \in L^2(\mathbb {R})\) such that \((Z g)|_{(0,1)^2} = H|_{(0,1)^2}\). Since both Zg and H are quasi-periodic, this implies \(Z g = H\) almost everywhere. Since \(H \in W^{1,2}_{\textrm{loc}}(\mathbb {R}^2)\) is bounded, Eqs. A.5,A.6 show that \(g \in \mathbb {H}^1\) is a Bessel vector. Let us assume towards a contradiction that Janssen’s representation of the frame operator associated to g converges unconditionally in the strong operator topology.
Note that \(|Z g|^2 = |H|^2 = F^2\) is discontinuous at \(x_0 \in (0,1)^2\) (since F is discontinuous there and also non-negative), even after possibly changing \(|Z g|^2\) on a null-set. In particular, this implies that the Fourier coefficients \(c_\alpha := \widehat{|Z g|^2}(\alpha )\) (for \(\alpha \in \mathbb {Z}^2\)) satisfy \(c = (c_\alpha )_{\alpha \in \mathbb {Z}^2} \notin \ell ^1(\mathbb {Z}^2)\), since otherwise the Fourier series of \(|Z g|^2\) would be uniformly convergent. This implies \(\sum _{\alpha \in \mathbb {Z}^2} |\Re c_\alpha | = \infty \) or \(\sum _{\alpha \in \mathbb {Z}^2} |\Im c_\alpha | = \infty \). For simplicity, we assume the first case; the second case can be treated by similar arguments. This implies that there exists an enumeration \((\alpha _n)_{n \in \mathbb {N}}\) of \(\mathbb {Z}^2\) such that \(|\sum _{n=1}^N \Re c_{\alpha _n}| \rightarrow \infty \) as \(N \rightarrow \infty \). Indeed, if \(\sum _{\alpha \in \mathbb {Z}^2} (\Re c_\alpha )_+ < \infty \) or \(\sum _{\alpha \in \mathbb {Z}^2} (\Re c_\alpha )_- < \infty \), this is trivial (for every enumeration); otherwise, existence of the desired enumeration follows from the Riemann rearrangement theorem (see e.g., [23, Theorem 3.54]).
Now, define \(J_n := \{ \alpha _1,\dots ,\alpha _n \}\) for \(n \in \mathbb {N}\). We have seen in Appendix 1 that the partial Fourier sums \(\mathcal {F}_{J_n} [|Z g|^2]\) are uniformly bounded in \(L^\infty \), say \(\Vert \mathcal {F}_{J_n} [|Z g|^2] \Vert _{L^\infty } \le C\) for all \(n \in \mathbb {N}\). Since each \(\mathcal {F}_{J_n} [|Z g|^2]\) is continuous (in fact, a trigonometric polynomial), this implies
which is the desired contradiction.
1.5 A.5. Conditional Divergence of Janssen’s Representation in the Operator Norm
We showed above that unconditional convergence of Janssen’s representation (A.1) in the strong operator topology fails for some Bessel vector \(g \in \mathbb {H}^1(\mathbb {R})\). A similar argument shows that convergence in the operator norm (with respect to any given enumeration) also fails in general: Using Eq. A.8 (or more generally the arguments in Appendix 1), it is relatively easy to see that if for some enumeration \(\mathbb {Z}^2 = \{ \alpha _n :n \in \mathbb {N}\}\) and \(I_n := \{ \alpha _1,\dots ,\alpha _n \}\), the sequence of partial sums \((S_{I_n})_{n \in \mathbb {N}}\) of Janssen’s representation (A.1) converges in operator norm (not even necessarily to S), then the associated sequence \((\mathcal {F}_{I_n'} [|Z g|^2])_{n \in \mathbb {N}}\) of partial Fourier sums of \(|Z g|^2\) is Cauchy in \(L^\infty (Q)\) and thus converges uniformly on Q to a (necessarily continuous) function \({\widetilde{H}} : Q \rightarrow \mathbb {C}\). However, since \(|Z g|^2 = |H|^2 = F^2 \in L^\infty (Q) \subset L^2(Q)\), we know that \(\mathcal {F}_{I_n '} [|Z g|^2] \rightarrow F^2\) with convergence in \(L^2(Q)\). Hence, \(F^2 = {\widetilde{H}}\) almost everywhere on Q, where \({\widetilde{H}}\) is continuous. But we saw above that \(F^2\) is discontinuous on Q, even after (possibly) changing it on a null-set. Thus, we have obtained the desired contradiction:
Proposition A.2
There are Bessel vectors \(g \in \mathbb {H}^1\) for which Janssen’s representation fails to converge conditionally in the operator norm.
However, we leave it as an open question whether Janssen’s representation converges conditionally in the strong sense for Bessel vectors \(g \in \mathbb {H}^1\).
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Lee, D.G., Philipp, F. & Voigtlaender, F. A Note on the Invertibility of the Gabor Frame Operator on Certain Modulation Spaces. J Fourier Anal Appl 29, 3 (2023). https://doi.org/10.1007/s00041-022-09980-0
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s00041-022-09980-0