Abstract
Signal sampling issue has been extensively studied based on a single sampling functional over classical Lebesgue spaces. This paper focuses on discussing the signal reconstruction and error analysis based on multiple sampling functionals over mixed Lebesgue spaces. We firstly explore the stabilities for two kinds of sampling functionals, respectively. Then the corresponding iterative reconstruction algorithms are established. Finally, the error between the reconstruction signal in the presence of noisy and the original signal f is analyzed.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The mixed Lebesgue space is a natural generalization of the classical Lebesgue space, which was firstly in depth introduced by Benedek and Panzone [7]. In fact, it arises from considering a function containing several independent variables of different properties. For instance, a multivariate function depending on both spatial and time variables may belong to a mixed Lebesgue space. Moreover, the flexibility of the separate integrability for each variable is of interests and potentially useful in the study of time-based partial differential equations [10]. The definition of mixed Lebesgue spaces \(L^{p,q}(\mathbb {R}^{d+1})\) [14, 16, 17, 23] is given as follows.
Definition 1.1
Let \(1\le p,q<\infty \), then \(L^{p,q}(\mathbb {R}^{d+1})\) consists of all measurable functions f on \(\mathbb {R}^{d+1}\) such that
The corresponding sequence spaces are
It is easy to check \(L^{p,p}(\mathbb {R}^{d+1})=L^p(\mathbb {R}^{d+1})\) and \(l^{p,p}(\mathbb {Z}^{d+1})=l^p(\mathbb {Z}^{d+1})\), respectively.
Signal sampling and its reconstruction theories are ubiquitous tools in a wide range of applications. The most well-known result is Shannon sampling theorem which gives an explicit reconstruction formula and states that every band-limited function can be reconstructed from its uniform samples. However, due to the slow decay and infinite support of the sinc function, it is often less efficient for numerical implementation. Moreover, there are many data just can be observed on non-uniform sampling set, such as in communication theory, medical imaging, astronomical measurement and among many others [1,2,3,4, 6, 21, 25]. In general, sampling problems have been studied in the following shift-invariant space [3,4,5,6, 11, 20]
where \(1\le p<\infty \). The vector function \(\Phi :=(\phi _1,\cdots ,\phi _r)^T\) is usually called the generator of the space \(V_p(\Phi )\). If \(r=d=1\), \(p=2\) and \(\phi (\cdot )=sinc(\cdot )\), then \(V_2(\phi )\) reduces to the classical space of band-limited functions.
Although the samples are usually supposed to be the exact values of a signal f in classical sampling theory, in fact, only the local average values can be derived. More precisely, the samples of f can be taken near the points which belong to a countable index set. In the last decades, the average sampling theory has drawn considerable attentions including in band-limited signals [8, 13], shift-invariant signals [1, 3,4,5,6, 15, 22], non-decaying signals [18] and multi-channel sampling problems [9, 12].
In addition, the multiple sampling functionals can be traced back to the multi-channel sampling problem [19]. Unser and Zerubia [24] showed that the multi-channel sampling can achieve higher stability and be more suitable for analyzing large bandwidth signals. Recently, Zhang [27] studied the non-uniform average sampling problem in multiply generated shift-invariant subspaces of mixed Lebesgue spaces and provided two fast reconstruction algorithms for two types of average sampled values. Wang and Zhang [26] considered the average sampling problem for signals in shift-invariant subspaces of weighted mixed Lebesgue spaces. More precisely, the sampling stability and iterative reconstruction algorithms are established for two kinds of average sampling functionals. However, the work in Refs. [26, 27] only discussed the average sampling problems by using single sampling functional. Motivated by above literature, this paper investigates the stabilities and reconstruction algorithms of average sampling based on multiple sampling functionals over mixed Lebesgue spaces.
The rest of the paper is organized as follows. In Sect. 2, the definitions and preliminaries are introduced briefly. In order to recover the signal exactly from the average sampling, Sect. 3 provides the sampling stabilities for two kinds of average sampling functionals, respectively. In Sect. 4, the iterative algorithms for the reconstruction are presented. Finally, since the samples are usually contaminated by random noises, the error analysis is discussed in Sect. 5.
2 Definitions and Preliminaries
This section collects some definitions, notations and preliminary results for future convenience. We begin with the mixed Wiener amalgam spaces \(W(L^{p,q})(\mathbb {R}^{d+1})\).
Definition 2.1
[16] Let \(1\le p, q<\infty \), then a measurable function f belongs to \(W(L^{p,q})(\mathbb {R}^{d+1})\) if it satisfies
Moreover, \(W_{0}(L^{p,q})(\mathbb {R}^{d+1})\) denotes the space of all continuous functions in \(W(L^{p,q})(\mathbb {R}^{d+1})\).
For the simpler case, a function f belongs to \(W(L^{p})(\mathbb {R}^{d+1})~(1\le p<\infty )\) if
holds. Furthermore, it is easy to check that \(W(L^{p})\subset W(L^{p,p})\) and \(W(L^p)\subset W(L^q)\subset L^q\) \((1\le p\le q<\infty )\) (the details please see Ref. [3]).
With \(\Phi :=(\phi _{1},\cdots ,\phi _{r})^{T}\in W(L^{1,1})^{(r)}:=\underbrace{W(L^{1,1})\times \cdots \times W(L^{1,1})}_{r~\text {times}}\), the underlying shift-invariant space is given by
where \(C(k_1,k_2):=\left( c_{1}(k_{1},k_{2}),\cdots ,c_{r}(k_{1},k_{2})\right) ^{T}\). On the other hand, the corresponding norm \(\Vert \Phi \Vert \) of a vector function \(\Phi :=(\phi _{1},\cdots ,\phi _{r})^{T}\) stands for \(\Vert \Phi \Vert :=\sum _{i=1}^r\Vert \phi _i\Vert \).
Moreover, we assume that for any \(\xi \in \mathbb {R}^{d+1}\) and all \(k\in \mathbb {Z}^{d+1}\), the sequences
are linearly independent in this paper, where \(\widehat{\phi _i}\) stands for the Fourier transform of \(\phi _i\). For instance, let \(\phi _i~(i=1,\cdots ,r)\) be the \(d+1\)-dimensional tensor product orthonormal wavelets. Then for fixed \(\xi \in \mathbb {R}^{d+1}\) and \(k\in \mathbb {Z}^{d+1}\), \(\{\widehat{\phi }_1(\xi +2\pi k),\cdots ,\widehat{\phi }_r(\xi +2\pi k)\}\) are linearly independent thanks to Theorem 1.4 in Ref. [17]. It is well-known that there exists the dual functions \(\tilde{\phi }_1,\cdots ,\tilde{\phi }_r\in W(L^{1,1})(\mathbb {R}^{d+1})\) such that for any \(f\in V_{p,q}(\Phi )~(1<p,q<\infty )\),
All these claims can be found in Ref. [17].
The following lemma is a natural extension of Theorem 3.1 in Ref. [16].
Lemma 2.1
Let \(\Phi \in W(L^{1,1})^{(r)}(\mathbb {R}^{d+1})\) and \(\{C(k_{1},k_{2})\}\in (l^{p,q})^{(r)}(\mathbb {Z}^{d+1})\) \((1\le p,q<\infty )\). Then for any \(f=\sum _{k_{1}\in \mathbb {Z}}\sum _{k_{2}\in \mathbb {Z}^{d}}C(k_{1},k_{2})^{T}\Phi (\cdot -k_{1},\cdot -k_{2})\),
To prove the sampling stability, we need the definition of relatively separated set.
Definition 2.2
[14] A sampling set \(\Gamma =\{\gamma _{j,k}=(x_{j},y_{k}),~~ (j,k)\in \mathbb {J}:=\mathbb {J}_{1}\times \mathbb {J}_{2}\}\) is called to be \((\delta _{1},\delta _{2})\)-relatively separated, if
and
for some \(\delta _{1}>0\) and \(\delta _{2}>0.\) Here, \(\mathbb {J}=\mathbb {J}_{1}\times \mathbb {J}_{2}\) is a countable index set, \(\chi _{B(x,\delta )}(\cdot )\) denotes the characteristic function on a ball with the center point x and the radius \(\delta \).
Given a relatively separated sampling set \(\Gamma \), two kinds of average sampling schemes based on \(\Gamma \) are considered in the present paper.
\(\bullet \) The first average sampling scheme is given by
where \(\psi ^l\in L^1(\mathbb {R}^{d+1})~(l=1,\cdots ,s)\) satisfy that for \(a>0\),
\(\bullet \) The second average sampling scheme is defined by
where the average sampling functionals \(\{\psi _{j,k}^l,~(j,k)\in \mathbb {J}\}\) \((l=1,\cdots ,s)\) satisfy that
(i). \(\text {supp}~\psi _{j,k}^{l}\subset B(\gamma _{j,k},\widetilde{a})\) for some \(\widetilde{a}>0\);
(ii). There exists a constant \(M>0\) such that \(\int _{\mathbb {R}}\int _{\mathbb {R}^{d}}|\psi _{j,k}^{l}(x,y)|\textrm{d}y\textrm{d}x\le M\) for all \((j,k)\in \mathbb {J}\);
(iii). \(\sum _{l=1}^{s}\int _{\mathbb {R}}\int _{\mathbb {R}^{d}}\psi _{j,k}^{l}(x,y)\textrm{d}y\textrm{d}x=1\) for all \((j,k)\in \mathbb {J}\).
3 Sampling Stability
Before illustrating the main results, the oscillation (or modulus of continuity) of a continuous function \(f\in L^{1}(\mathbb {R}^{d+1})\) is given by
Lemma 3.1
[16] If \(\phi \in W_{0}(L^{1})(\mathbb {R}^{d+1}),\) then the following two statements hold:
and
Denote
Then the next two lemmas are introduced.
Lemma 3.2
Let \(\psi ^{l}\in L^{1}(\mathbb {R}^{d+1})~(l=1,\cdots ,s)\) satisfy \(\sum _{l=1}^{s}\int _{\mathbb {R}^{d+1}}\psi ^{l}(t)\textrm{d}t=1\). Then for any \(\Phi =(\phi _{1},\cdots ,\phi _{r})^{T}\in W_{0}(L^{1})^{(r)}(\mathbb {R}^{d+1}),\)
Proof
According to the definition of \(\Phi ^{a}\) and the fact
one only needs to show \(\Vert \phi _{i}-\phi _{i}*(\sum _{l=1}^{s}\widetilde{\psi _{a}^{l}})\Vert _{W(L^{1})}\rightarrow 0\) as \(a\rightarrow 0^{+}\) for each \(i=1,\cdots ,r\).
It follows from \(\sum _{l=1}^{s}\int _{\mathbb {R}^{d+1}}\psi ^{l}(t)\textrm{d}t=1\) that
For \(I_{1}(x)\), it is clear to see that
Obviously, the above inequality reduces to
Therefore, \(\Vert I_{1}\Vert _{W(L^{1})}\longrightarrow 0\) as \(a\rightarrow 0^+\) follows from \(\psi ^{l}\in L^{1}(\mathbb {R}^{d+1}).\)
For \(I_{2}(x)\), by the definition of oscillation,
This with \(\phi _i\in W_0(L^1)\) and (3.2) shows \(\Vert osc_{\sqrt{a}}(\phi _{i})\Vert _{W(L^{1})}\longrightarrow 0\) as \(a\rightarrow 0^{+}\). Then \(\Vert I_{2}\Vert _{W(L^{1})}\longrightarrow 0.\)
Hence, the above arguments tell that for each \(i=1,\cdots ,r\),
thanks to (3.3). The proof is done.\(\square \)
Lemma 3.3
Under the assumptions of Lemma 3.2, the oscillation of \(\Phi ^a\) satisfies
where \(osc_{\delta }(\Phi ^a)=(osc_{\delta }(\phi ^a_{1}),\cdots ,osc_{\delta }(\phi ^a_{r}))^{T}\) and \(a\in \mathbb {R}\).
Proof
Similar to the discussion of Lemma 3.2, one should prove that for each \(i=1,\cdots ,r,\)
Recall that \(\phi _{i}^{a}:=\phi _{i}-\phi _{i}*\Big (\sum _{l=1}^{s}\widetilde{\psi _{a}^{l}}\Big ).\) Then
For the second term of the right-hand side of (3.4), it is easy to find that
Combining this with (3.2) and \(\psi ^l\in L^1(\mathbb {R}^{d+1})\), one obtains that for each \(a\in \mathbb {R}\),
as \(\delta \rightarrow 0,\) which completes the proof. \(\square \)
Lemma 3.4 is necessary for later discussions.
Lemma 3.4
If \(\Phi \in W(L^{1,1})^{(r)}(\mathbb {R}^{d+1}),\) and \(\{\widehat{\phi }_i(\xi +2k\pi ),~k\in \mathbb {Z}^{d+1}\}\) are linearly independent, then for any \(f\in V_{p,q}(\Phi )(\mathbb {R}^{d+1})~(1<p,q<\infty ),\)
where \(\Vert \widetilde{\Phi }\Vert \) is a positive constant only depending on \(\Phi \).
Proof
For each fixed \(i=1,\cdots ,r\), there exists a dual function \(\tilde{\phi }_i\) such that
thanks to \(f\in V_{p,q}(\Phi )\) (see Ref. [17]). Take \(b=\{b(k_1,k_2),~k_1\in \mathbb {Z},k_2\in \mathbb {Z}^d\}\in l^{p',q'}(\mathbb {Z}^{d+1})\) with \(\frac{1}{p}+\frac{1}{p'}=1\) and \(\frac{1}{q}+\frac{1}{q'}=1\). Then by (3.6), one knows that
This with the Hölder inequality and Lemma 2.1 shows that
Thus, \(\Vert c_i\Vert _{l^{p,q}}\le \Vert f\Vert _{L^{p,q}}\Vert \tilde{\phi }_i\Vert _{W(L^{1,1})}\) which leads to
with \(\Vert \widetilde{\Phi }\Vert :=\sum _{i=1}^r\Vert \tilde{\phi }_i\Vert _{W(L^{1,1})}.\) The proof is finished. \(\square \)
We are in a position to state the first stability theorem.
Theorem 3.1
Suppose that \(\Phi \in W_{0}(L^{1})^{(r)}(\mathbb {R}^{d+1})\) and \(\Gamma \) is a \((\delta _{1}, \delta _{2})\)-relatively separated set. If \(\delta _{1}, \delta _{2}\) and a are chosen such that
then for any signal \(f\in V_{p, q}(\Phi )~(1<p,q<\infty ),\)
where \(V_{d}=\frac{\pi ^{d/2}}{\Gamma (d/2+1)}\) is the volume of d-dimensional unit ball, and \(\Vert \widetilde{\Phi }\Vert \) is the positive constant given in Lemma 3.4.
Proof
For any \(x\in B(x_{j}, \delta _{1})\) and \(y\in B(y_{k}, \delta _{2}),\) it follows from \(f\in V_{p, q}(\Phi )\) that
Moreover, by \(\phi _{i}^{a}:=\phi _{i}-\phi _{i}*\Big (\sum _{l=1}^{s}\widetilde{\psi _{a}^{l}}\Big ),\) the above inequality reduces to
Furthermore, due to Lemma 2.1 and \(W(L^1)\subset W(L^{1,1})\),
thanks to Lemma 3.4 and (3.7).
Define
Then with \(\Gamma \) being \((\delta _1,\delta _2)\)-relatively separated, it is clear to see that
and
On the other hand, it follows from (3.9) that
Taking \(l^{q}\)-norm and \(L^{q}\)-norm about the variable \(k\in \mathbb {J}_{2}\), \(y\in \mathbb {R}^{d}\), respectively, on the both sides of (3.13), one knows that
thanks to \(\Vert \{\beta _k^{\frac{1}{q}}\}_{k\in \mathbb {J}_2}\Vert _{l^q}=1\). Moreover, taking \(l^{p}\)-norm and \(L^{p}\)-norm about the variables \(j\in \mathbb {J}_{1}\), \(x\in \mathbb {R}\), respectively, on the both sides of (3.14), one obtains that
because of \(\Vert \{\alpha _j^{\frac{1}{p}}\}_{j\in \mathbb {J}_1}\Vert _{l^p}=1\) and (3.10).
Combining (3.15) with the left-hand sides of (3.11)–(3.12), the right-hand side of (3.8) is established.
In addition, due to (3.9),
Moreover, similar to the discussions of (3.13)–(3.15), one knows
Clearly, the assumption of \(r_1<1\) is necessary in (3.16). In fact, \(r_1<1\) could follow from Lemmas 3.2–3.3. Therefore, the inequality (3.16) with (3.11)–(3.12) implies that the left-hand side of (3.8) holds, which concludes the desired conclusion. \(\square \)
The next lemma is useful to prove Theorem 3.2.
Lemma 3.5
If \(\Phi \in W_{0}(L^{1})^{(r)}(\mathbb {R}^{d+1}),\) then for any \(f\in V_{p,q}(\Phi )~(1\le p,q<\infty ),\)
Proof
According to the definition of oscillation and \(f\in V_{p,q}(\Phi )\),
Moreover, by Lemma 2.1,
thanks to \(W(L^{1})\subset W(L^{1,1})\), which is the desired conclusion. \(\square \)
The sampling stability for the second kind of average sampling functional is explored by Theorem 3.2.
Theorem 3.2
Suppose that \(\Phi \in W_{0}(L^{1})^{(r)}(\mathbb {R}^{d+1})\) and \(\Gamma \) is a \((\delta _{3},\delta _{4})\)-relatively separated set. If \(\delta _{3}, \delta _{4}\) and \(\widetilde{a}\) are chosen such that
then for any signal \(f\in V_{p, q}(\Phi )~(1< p,q<\infty ),\)
where \(V_{d}=\frac{\pi ^{d/2}}{\Gamma (d/2+1)}\) is the volume of d-dimensional unit ball, and \(\Vert \widetilde{\Phi }\Vert \) is the positive constant given in Lemma 3.4.
Proof
For any \(f\in V_{p,q}(\Phi ),\) it follows from \(\textrm{supp}~\psi _{j,k}^{l}\subset B(\gamma _{j,k},\widetilde{a})\) and \(\sum _{l=1}^{s}\int _{\mathbb {R}}\int _{\mathbb {R}^{d}}\psi _{j,k}^{l}(x,y)\textrm{d}y\textrm{d}x=1\) that
due to \(\int _{\mathbb {R}}\int _{\mathbb {R}^{d}}|\psi _{j,k}^{l}(x,y)|\textrm{d}y\textrm{d}x\le M\). This with Lemma 3.5 and Lemma 3.4 shows
thanks to (3.17).
Define
It follows from (3.19) that
The reminding proof is similar to the arguments in (3.14)–(3.15), one firstly takes the \(l^{q}\)-norm and \(L^{q}\)-norm for the variable k and y, respectively, then takes the \(l^{p}\)-norm and \(L^{p}\)-norm for the variable j and x, respectively. By (3.20),
Therefore, the right-hand side of (3.18) can be concluded from the above inequality, the left-hand side of (3.18) can be obtained by the similar method as (3.16). The proof is done. \(\square \)
4 Iterative Reconstruction Algorithms
In this section, the iterative reconstruction algorithms for recovering original signals are provided. Before introducing the main results, we demonstrate the following important concept.
Definition 4.1
[14] Let \(\mathbb {J}:=\mathbb {J}_{1}\times \mathbb {J}_{2}\) be a countable index set, and \(\Gamma \) be a \((\delta _{1}, \delta _{2})\)-relatively separated set. Then \(\{u_{j,k}(x,y),~(j,k)\in \mathbb {J}\}\) is called a \(\textbf{BUPU}\) (Bounded Uniform Partition of Unity) associated with \(\Gamma \), if
\(\mathrm {(i).}\) \(0\le u_{j,k}(x,y)\le 1\) for all \((x,y)\in \mathbb {R}\times \mathbb {R}^{d}\) and \((j,k)\in \mathbb {J}\);
\(\mathrm {(ii).}\) \(\textrm{supp}~u_{j,k}\subset B(\gamma _{j,k},\sqrt{\delta _{1}^{2}+\delta _{2}^{2}})\) for each \((j,k)\in \mathbb {J}\);
\(\mathrm {(iii).}\) \(\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}u_{j,k}(x,y)=1\) for all \((x,y)\in \mathbb {R}\times \mathbb {R}^{d}\).
Lemma 4.1
Let \(\Phi \in W_{0}(L^{1})^{(r)}(\mathbb {R}^{d+1})\), \(\Gamma \) be a \((\delta '_{5}, \delta '_{6})\)-relatively separated set and \(\{u_{j,k}(x,y),~(j,k)\in \mathbb {J}\}\) be a \(\textbf{BUPU}\) associated with \(\Gamma \). Then for any \(f\in V_{p,q}(\Phi )~(1\le p,q<\infty ),\)
where \(Q_{\Gamma }f:=\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}f(x_{j},y_{k})u_{j,k}.\)
Proof
Clearly,
Then the main work of Lemma 4.1 is to estimate \(\Vert f-Q_{\Gamma }f\Vert _{L^{p,q}}\). According to the definition of \(Q_\Gamma f\), \(\text {supp}~u_{j,k}\subset B(\gamma _{j,k},\sqrt{\delta _{5}^{'2}+\delta _{6}^{'2}})\) and \(\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}u_{j,k}(x,y)=1\),
Then with Lemma 3.5, one concludes
thanks to (3.1).
On the other hand, Lemma 2.1 tells that \(\Vert f\Vert _{L^{p,q}}\le \sum _{i=1}^{r}\Vert c_{i}\Vert _{l^{p,q}}\Vert \phi _{i}\Vert _{W(L^{1})}\). Combining this with (4.1)–(4.2), the proof is completed. \(\square \)
Let P be the projection operator from \(L^{p,q}(\mathbb {R}^{d+1})\) onto \(V_{p,q}(\Phi )\),
where \(\tilde{\phi }_{1},\cdots ,\tilde{\phi }_r\in W(L^{1,1})\) are the dual functions of \(\phi _1,\cdots ,\phi _r\in W(L^{1,1})\), respectively. Then we provide the following lemma.
Lemma 4.2
Define \(A_{\Gamma ,a}f:= \sum _{l=1}^{s}\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}(f*\widetilde{\psi _{a}^{l}})(\gamma _{j,k})u_{j,k}.\) Then there exist a \((\delta _5,\delta _6)\)-relatively separated set \(\Gamma \) and \(a_{0}>0\) such that the operator \(I-PA_{\Gamma ,a}\) is contractive on \(V_{p,q}(\Phi )~(1< p,q<\infty )\) for \(a \le a_0.\)
Proof
For any \(f\in V_{p,q}(\Phi )\), one knows that
On the other hand, it follows from Lemmas 4.1 and 3.4 that
Combining (4.3) with (4.4) and (4.5), one obtains that
By (3.2),
when \(\sqrt{\delta _{5}^{'2}+\delta _{6}^{'2}}\rightarrow 0\). Moreover, Lemma 3.2 shows
Therefore, there exist \(\delta _{5}, \delta _{6}\) and \(a_0>0\) such that
Hence, (4.6) reduces to
which implies that \(I-PA_{\Gamma ,a}\) is a contractive operator.\(\square \)
Theorem 4.1
If \(\Phi \in W_{0}(L^{1})^{(r)}(\mathbb {R}^{d+1})\), then there exist a \((\delta _{5}, \delta _{6})\)-relatively separated set \(\Gamma \) and \(a_{0}>0\) such that each signal \(f\in V_{p,q}(\Phi )~(1< p,q<\infty )\) can be recovered from \(\left\{ \langle f,~\psi _{a}^{l}(\cdot -\gamma _{j,k})\rangle \right\} _{(j,k)\in \mathbb {J}}\) by the following iterative algorithm:
Furthermore,
for some \(\alpha _{1}(\delta _{5},\delta _{6},a_{0},\Phi )<1.\)
Proof
Let \(e_{n}=f-f_{n}\). Then by (4.7),
Using Lemma 4.2 and choosing \(\delta _{5},\delta _{6}\) and \(a_{0}>0\) such that \(\alpha _{1}=\Vert I-PA_{\Gamma ,a}\Vert _{\textrm{op}}<1,\) then
Furthermore, \(\Vert e_{n}\Vert _{L^{p,q}} \longrightarrow 0\) as \(n\rightarrow \infty ,\) which completes the proof.\(\square \)
Similar to Lemma 4.2, we establish the next lemma for the second kind of sampling functional.
Lemma 4.3
Define \(A_{\Gamma }f:=\sum _{l=1}^{s}\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}\langle f,~\psi _{j,k}^{l}\rangle u_{j,k}.\) Then there exist a \((\delta _7,\delta _8)\)-relatively separated set \(\Gamma \) and \(\widetilde{a}_0>0\) such that the operator \(I-PA_{\Gamma }\) is contractive on \(V_{p,q}(\Phi )\) \((1<p,q<\infty )\).
Proof
For any \(f\in V_{p,q}(\Phi ),\) one obtains
According to the definitions of \(Q_\Gamma f,~A_\Gamma f\) and \(\sum _{l=1}^{s}\int _{\mathbb {R}}\int _{\mathbb {R}^{d}}\psi _{j,k}^{l}(x,y)\textrm{d}y\textrm{d}x=1\),
Moreover, by \(\text {supp}~\psi _{j,k}^{l}\subset B(\gamma _{j,k},\widetilde{a})~(l=1,\cdots ,s),\) the above inequality yields
thanks to \(\int _{\mathbb {R}}\int _{\mathbb {R}^{d}}|\psi _{j,k}^{l}(s,t)|\textrm{d}t\textrm{d}s\le M\) and the definition of \(Q_\Gamma \). This with Lemma 4.1 and Lemma 3.4 shows that
Combining (4.8) with (4.4) and (4.9), one concludes
Furthermore, (3.2) implies that
Therefore, there exist \(\delta _{7}, \delta _{8}\) and \(\widetilde{a}_0>0\) such that
Hence, it leads to the conclusion, i.e.,
The proof is finished.\(\square \)
Similar to the proof of Theorem 4.1, we can derive the second iterative algorithm immediately.
Theorem 4.2
If \(\Phi \in W_{0}(L^{1})(\mathbb {R}^{d+1})^{(r)}\), then there exist a \((\delta _{7}, \delta _{8})\)-relatively separated set \(\Gamma \) and \(\widetilde{a}_0>0\) such that each signal \(f\in V_{p,q}(\Phi )~(1< p,q<\infty )\) can be recovered from \(\{\langle f,~\psi _{j,k}^{l}\rangle \}_{(j,k)\in \mathbb {J}}\) by the following iterative algorithm:
Furthermore,
for some \(\alpha _{2}(\delta _{7},\delta _{8},\widetilde{a}_0,\Phi )<1.\)
5 Error Analysis
In many applications, the samples are often contaminated by random noises. Motivated by the work of Aldroubi et al. [5] and Jiang [14], we investigate the error analysis if the samples are destroyed by random noises in this section.
Firstly, we propose the following inequality under mixed norm.
Lemma 5.1
If \(\phi \in W(L^{1})(\mathbb {R}^{d+1})\) and \(f\in L^{p,q}(\mathbb {R}^{d+1})~(1\le p,q<\infty ),\) then
Proof
Obviously, \(\phi \in W(L^{1})(\mathbb {R}^{d+1})\subset W(L^{p})(\mathbb {R}^{d+1})\subset L^{p}(\mathbb {R}^{d+1})\) holds for each \(p\ge 1\) and
For fixed x and \(k_{1}\), denote \(f_{x}(y):=f(x,y)\) and \(\phi _{k_{1}-x}(k_{2}-y):=\phi (k_{1}-x,k_{2}-y)\). Then
This with the generalized Minkowski inequality shows
Let \(f_{x,y}(\cdot ):=f_{x}(\cdot +y)\) and \(\phi _{k_{1}-x,y}(\cdot ):=\phi _{k_{1}-x}(\cdot -y)\). Then it follows from Young’s inequality that
Using the Hölder inequality with \(\frac{1}{q}+\frac{1}{q'}=1\), (5.2) reduces to
Because the function \(\sum _{l\in \mathbb {Z}^{d}}|\phi _{k_{1}-x}(l-\cdot )|\) is \(\textbf{1}\)-periodic, the above inequality yields
Substituting (5.3) into (5.1), one obtains that
Denote \(a_x(n):=\Vert f(x+n,\cdot )\Vert _{L^{q}}\) and \(b_x(n):=\Vert \phi (n-x,\cdot )\Vert _{W(L^{1})}\). Then by the generalized Minkowski inequality again,
Using Young’s inequality and the Hölder inequality with \(\frac{1}{p}+\frac{1}{p'}=1\), one gets
which is the desired conclusion. The proof is completed.\(\square \)
Theorem 5.1
Let \(\Phi \in W_0(L^1)^{(r)}(\mathbb {R}^{d+1})\), \(\Gamma =\{\gamma _{j,k},~(j,k)\in \mathbb {J}\}\) be a relatively separated sampling set and \(\{\varepsilon _{j,k}^{l},~(j,k)\in \mathbb {J},~l=1,\cdots ,s\}\) be random variables satisfying
Then for any initial data \(\{\langle f,~\psi _{a}^{l}(\cdot -\gamma _{j,k}) \rangle +\varepsilon _{j,k}^{l}\}_{(j,k)\in \mathbb {J}}\) or \(\{\langle f, ~\psi _{j,k}^{l}\rangle +\varepsilon _{j,k}^{l}\}_{(j,k)\in \mathbb {J}},\)
where \(C=C(d,p,q,\delta _{5}, \delta _{6}),~\alpha =\alpha _{1}\) for algorithm (4.7); or \(C=C(d,p,q,\delta _{7}, \delta _{8}),~\alpha =\alpha _{2}\) for algorithm (4.10), and the positive constant \(\Vert \widetilde{\Phi }\Vert \) is given in Lemma 3.4.
Proof
Denote \(A:={\left\{ \begin{array}{ll}A_{\Gamma ,a},~~\textit{for} ~\textit{algorithm}~(4.7);\\ A_{\Gamma },~~~~\textit{for}~\textit{algorithm}~(4.10). \end{array}\right. }\) Then Lemmas 4.2 and 4.3 tell \(\alpha =\Vert I-PA\Vert _{\textrm{op}}<1\), which implies
Let \(h_{0}:=\sum _{l=1}^{s}\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}\varepsilon _{j,k}^{l}P(u_{j,k}),\) and the initial data be \(\{\langle f,~\psi _{a}^{l}(\cdot -\gamma _{j,k}) \rangle +\varepsilon _{j,k}^{l}\}_{(j,k)\in \mathbb {J}}\) or \(\{\langle f, ~\psi _{j,k}^{l}\rangle +\varepsilon _{j,k}^{l}\}_{(j,k)\in \mathbb {J}}\), respectively. Then the original signal f can be recovered by the following iterative algorithm
based on Theorem 4.1 and Theorem 4.2, respectively. Moreover, due to (5.4), \(f_n=\widetilde{f}_1+(I-PA)f_{n-1}=\widetilde{f}_1+(I-PA)\Big (\widetilde{f}_1+(I-PA)f_{n-2}\Big )=\cdots =\Big (I+\sum _{k=1}^{n-1}(I-PA)^k\Big )\widetilde{f}_1\) and \(f=\Big (I+\sum _{k=1}^{\infty }(I-PA)^k\Big )PAf=\Big (I+\sum _{k=1}^{\infty }(I-PA)^k\Big )(\widetilde{f}_1-h_0)\), i.e.,
According to \(\alpha =\Vert I-PA\Vert _{\textrm{op}}<1\), one obtains
By the definition of \(h_0\) and the Minkowski inequality, the above inequality reduces to
It follows from \(\text {supp}~u_{j,k}\subset B(\gamma _{j,k},\sqrt{\delta _{5}^{2}+\delta _{6}^{2}})\), \(x\in B(x_{j}, \sqrt{\delta _{5}^{2}+\delta _{6}^{2}})\) and \(y\in B(y_{k}, \sqrt{\delta _{5}^{2}+\delta _{6}^{2}})\) that
thanks to \(|u_{j,k}(x,y)|\le 1\) for all x, y. This implies that \(u_{j,k}\) belongs to \(L^{p,q}(\mathbb {R}^{d+1})\) based on the first kind of sampling functional. Similarly, one can derive \(u_{j,k}\le 2\sqrt{\delta _{7}^{2}+\delta _{8}^{2}}\cdot \Big (V_{d}(\delta _{7}^{2}+\delta _{8}^{2})^{\frac{d}{2}}\Big )^{\frac{p}{q}},\) i.e., \(u_{j,k}\) also belongs to \(L^{p,q}(\mathbb {R}^{d+1})\) based on the second kind of sampling functional.
Furthermore, according to the definition of the operator P and Lemma 2.1, one finds that
This with Lemma 5.1 leads to
where C is given by
Combining (5.6) with (5.7) and \(E\Big (\sum _{j\in \mathbb {J}_{1}}\sum _{k\in \mathbb {J}_{2}}|\varepsilon _{j,k}^{l}|\Big )<N~(l=1,\cdots ,s),\) one concludes that
which completes the proof.\(\square \)
References
Aldroubi, A.: Nonuniform weighted average sampling and reconstruction in shift-invariant and wavelet spaces. Appl. Comput. Harmon. Anal. 13(2), 151–161 (2002)
Aldroubi, A., Gröhenig, K.: Beurling–Landau-type theorems for nonuniform sampling in shift-invariant spline spaces. J. Fourier Anal. Appl. 6(1), 93–103 (2000)
Aldroubi, A., Gröchenig, K.: Nonuniform sampling and reconstruction in shift-invariant spaces. SIAM. Rev. 43(4), 585–620 (2001)
Aldroubi, A., Krishtal, I.: Robustness of sampling and reconstruction and Beurling–Landau-type theorems for shift-invariant spaces. Appl. Comput. Harmon. Anal. 20(2), 250–260 (2006)
Aldroubi, A., Leonetti, C., Sun, Q.Y.: Error analysis of frame reconstruction from noisy samples. IEEE Trans. Signal Process. 56(6), 2311–2325 (2008)
Aldroubi, A., Sun, Q.Y., Tang, W.S.: Nonuniform average sampling and reconstruction in multiply generated shift-invariant spaces. Constr. Approx. 20(2), 173–189 (2004)
Benedek, A., Panzone, R.: The space \(L^{p}\) with mixed norm. Duke Math. J. 28(3), 301–324 (1961)
Butzer, P.L., Lei, J.: Approximation of signals using measured sampled values and error analysis. Commun. Appl. Anal. 4(2), 245–255 (2000)
Djokovic, I., Vaidyanathan, P.P.: Generalized sampling theorems in multiresolution subspaces. IEEE Trans. Signal Process. 45, 583–599 (1997)
Francia, J.L., Ruiz, F.J.: Calderón–Zygmund theory for operator-valued kernels. Adv. Math. 62(1), 7–48 (1986)
Führ, H., Xian, J.: Relevant sampling in finitely generated shift-invariant spaces. J. Approx. Theory 240, 1–15 (2019)
García, A.G., Péerez-Villalón, G.: Dual frames in \(L^{2}(0,1)\) connected with generalized sampling in shift-invariant spaces. Appl. Comput. Harmon. Anal. 20(3), 422–433 (2006)
Gröchenig, K.: Reconstruction algorithms in irregular sampling. Math. Comp. 59(199), 181–194 (1992)
Jiang, Y.C.: Average sampling and reconstruction of reproducing kernel signals in mixed Lebesgue spaces. J. Math. Anal. Appl. 480, 123370, 14 (2019)
Kang, S., Kwon, K.H.: Generalized average sampling in shift invariant spaces. J. Math. Anal. Appl. 377(1), 70–78 (2011)
Li, R., Liu, B., Liu, R., Zhang, Q.Y.: Nonuniform sampling in principle shift-invariant subspaces of mixed Lebesgue spaces \(L^{p, q}(\mathbb{R} ^{d+1})\). J. Math. Anal. Appl. 453(2), 928–941 (2017)
Li, R., Liu, B., Liu, R., Zhang, Q.Y.: The \(L^{p,q}\)-stability of the shifts of finitely many functions in mixed Lebesgue spaces \(L^{p,q}(\mathbb{R}^{d+1})\). Acta Math. Sin. Engl. Ser. 34(6), 1001–1014 (2018)
Nguyen, H.Q., Unser, M.: A sampling theory for non-decaying signals. Appl. Comput. Harmon. Anal. 43(1), 76–93 (2017)
Papoulis, A.: Generalized sampling expansion. IEEE Trans. Circuits Syst. CAS–2(11), 652–654 (1977)
Selvan, A.A., Radh, R.: An optimal result for sampling density in shift-invariant spaces generated by Meyer scaling function. J. Math. Anal. Appl. 451(1), 197–208 (2017)
Sun, Q.Y.: Nonuniform average sampling and reconstruction of signals with finite rate of innovation. SIAM J. Math. Anal. 38(5), 1389–1422 (2006)
Sun, W.C., Zhou, X.W.: Reconstruction of functions in spline subspaces from local averages. Proc. Am. Math. Soc. 131(8), 2561–2571 (2003)
Torres, R.H., Ward, E.L.: Leibniz’s rule, sampling and wavelets on mixed Lebesgue spaces. J. Fourier Anal. Appl. 21(5), 1053–1076 (2015)
Unser, M., Zerubia, J.: A generalized sampling theory without bandlimiting constraints. IEEE Trans. Circuits Syst. 45(8), 959–969 (1998)
Venkataramani, R., Bresler, Y.: Sampling theorems for uniform and periodic nonuniform MIMO sampling of multiband signals. IEEE Trans. Signal Process. 51(12), 3152–3163 (2003)
Wang, S.P., Zhang, J.X.: Average sampling and reconstruction for signals in shift-invariant subspaces of weighted mixed Lebesgue spaces. Math. Methods Appl. Sci. 44(11), 9507–9523 (2021)
Zhang, Q.Y.: Nonuniform average sampling in multiply generated shift-invariant subspaces of mixed Lebesgue spaces. Int. J. Wavelets Multiresolut. Inf. Process 18(3), 2050013, 16 (2020)
Acknowledgements
The authors would like to thank Prof. Youming Liu and Prof. Zhi-Gang Liang for their important comments and suggestions. This paper is supported by the National Natural Science Foundation of China (Nos. 11901019, 12171016), and the Science and Technology Program of Beijing Municipal Commission of Education (No. KM202010005025).
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflicts of interest.
Additional information
Communicated by Rosihan M. Ali.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Wang, J., Yang, S. & Zeng, X. Reconstruction and Error Analysis Based on Multiple Sampling Functionals over Mixed Lebesgue Spaces. Bull. Malays. Math. Sci. Soc. 46, 67 (2023). https://doi.org/10.1007/s40840-023-01462-w
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s40840-023-01462-w