Abstract
Construction of interpolatory wavelets is an important topic in discrete signal processing. In classical wavelet sampling theories, interpolatory wavelets are constructed from Riesz bases in wavelet spaces. Since analytical expressions for such Riesz bases are generally complex or unavailable, it has been difficult to obtain suitable interpolatory wavelets in practice. In this paper, interpolatory scaling functions are used to determine and construct interpolatory wavelets. We first show that there may not exist interpolatory wavelets even when interpolatory scaling functions exist. Then, an inequality in terms of interpolatory scaling functions, denoted as the two-scale condition, is given as a necessary and sufficient condition for existence of interpolatory wavelets. Finally, based on the two-scale condition, a filter bank is constructed for obtaining interpolatory wavelets directly from interpolatory scaling functions. In examples, our theorems are applied to some typical wavelet spaces, demonstrating our construction algorithm for interpolatory wavelets.
Similar content being viewed by others
Abbreviations
- \(W_j :\) :
-
Wavelet space
- \(V_j :\) :
-
Approximation space of MRA
- \(L^2(\mathbb {R}):\) :
-
Square integrable function
- \(l^2(\mathbb {R}):\) :
-
Finite energy discrete signals \(\sum \nolimits _{k=-\infty }^{+\infty } {\left| {f_s (k)} \right| ^2} <+\infty \)
- \(f_s (x):\) :
-
Signal to be recovered
- \(f_{ap} (x):\) :
-
Recovery of signal
- \(\hat{f}(w):\) :
-
Fourier transform of \(f(x)\)
- \(\bar{f}(x):\) :
-
Complex conjugate of \(f(x)\)
- \(\phi (x):\) :
-
Scaling function of \(V_0 \)
- \(\psi (x):\) :
-
Wavelet of \(W_0 \)
- \(\{S^\phi (2^jx-k)\}_{k\in \mathbb {Z}} :\) :
-
Interpolatory basis relative to the samples \(\{f_s (k/2^j)\}_{k\in \mathbb {Z}} \) for \(V_j \)
- \(\{\tilde{S}^\phi (2^jx-k)\}_{k\in \mathbb {Z}} :\) :
-
Dual basis of \(\{S^\phi (2^jx-k)\}_{k\in \mathbb {Z}} \) for \(V_j \)
- \(\{S^\psi (2^jx-k)\}_{k\in \mathbb {Z}}:\) :
-
Interpolatory basis relative to the samples \(\{f_s (k/2^j+1/2^{j+1})\}_{k\in \mathbb {Z}} \) for \(W_j \)
- \(\{\tilde{S}^\psi (2^jx-k)\}_{k\in \mathbb {Z}}:\) :
-
Dual basis of \(\{S^\psi (2^jx-k)\}_{k\in \mathbb {Z}} \) for \(W_j \)
- \(P_\phi (w):\) :
-
\(l^2\)-Sequence defined in (41)
- \(P_s (w)_{:}\) :
-
\(l^2\)-Sequence defined in (3)
- \(Q_\psi (w):\) :
-
\(l^2\)-Sequence defined in (18)
- \(Q_s (w):\) :
-
\(l^2\)-Sequence defined in (14)
- \(E_s (w):\) :
-
Period function defined in (7)
- \(E_\phi (w):\) :
-
Period function defined in (20)
- \(\Delta _{P_s ,Q_s } :\) :
-
Function defined in (16)
- \(T_{W_{j-1} }^{1/2} :\) :
-
Sampling operator defined in (5)
References
A. Aldroubi, M. Unser, Sampling procedures in function spaces and asymptotic equivalence with Shannon sampling theorem. Numer. Funct. Anal. Optim. 15, 1–21 (1994)
R.S. Asamwar, K.M. Bhurchandi, A.S. Gandhi, Interpolation of images using discrete wavelet transform to simulate image resizing as in human vision. Int. J. Autom. Comput. 7(1), 9–16 (2010)
S. Amat, K. Dadourian, J. Liandrat, J. Ruiz, J.C. Trillo, A family of stable nonlinear nonseparable multiresolution schemes in 2D. J. Comput. Appl. Math. 234, 1277–1290 (2010)
M.R. Capobianco, W. Themistoclakis, Interpolating polynomial wavelets on [-1,1]. Adv. Comput. Math. 23(4), 353–374 (2005)
W. Chen, S. Itoh, A sampling theorem for shift-invariant subspace. IEEE Trans. Signal Process. 46(10), 2822–2824 (1998)
W. Chen, S. Itoh, On sampling in shift invariant spaces. IEEE Trans. Inf. Theory 48(10), 2802–2809 (2002)
C.K. Chui, An Introduction to Wavelets (Academic Press, San Diego, 1992)
C.K. Chui, C. Li, Dyadic affince decompositions and functional wavelet transforms. SIAM J. Math. Anal. 27(3), 865–890 (1996)
D.L. Donoho, Interpolating wavelet transforms. David Donoho Personal Website. http://www.stats.stanford.edu/~donoho/Reports/1992/interpol.pdf (1992)
I. Daubechies, Ten Lectures on Wavelets (SIAM, Philadelphia, 1992)
S. Ericsson, Irregular sampling in shift invariant spaces of higher dimensions. Int. J. Wavel. Multiresolut. Inf. Process. 6(1), 121–136 (2008)
A.G. García, G. Pérez-Villalón, Multivariate generalized sampling in shift-invariant spaces and its approximation properties. J. Math. Anal. Appl. 355(1), 397–413 (2009)
A.G. García, M.A. Hernández-Medina, G. Pérez-Villalón, Oversampling in shift-invariant spaces with a rational sampling period. IEEE Trans. Signal Process. 57(9), 3442–3449 (2009)
A.G. García, G. Pérez-Villalón, Approximation from shift-invariant spaces by generalized sampling formulas. Appl. Comput. Harmon. Anal. 24(1), 58–69 (2008)
A.G. García, J.M. Kim, K.H. Kwon, G. Pérez-Villalón, Aliasing error of sampling series in wavelet subspaces. Numer. Funct. Anal. Optim. 29(1–2), 126–144 (2008)
A.G. García, G. Pérez-Villalón, On the aliasing error in wavelet subspaces. J. Comput. Appl. Math. 183(1), 153–167 (2005)
A. Janssen, The Zak transform and sampling theorems for wavelet subspaces. IEEE Trans. Signal Process. 41(2), 3360–3364 (1993)
S. Mallat, A Wavelet Tour of Signal Processing, 2nd edn. (China Machine Press, Beijing, 2003)
Y.M. Robert, An Introduction to Nonharmonic Fourier Series (Academic Press, New York, 1980)
J.F. Reinoso, M. Moncayo, Optimal quality for image fusion with interpolatory parametric filters. Math. Comput. Simul. 81, 2307–2316 (2011)
X. Shen, G.G. Walter, Construction of periodic prolate spheroidal wavelets using interpolation. Numer. Funct. Anal. Optim. 28(3), 445–466 (2007)
K. Schneider, O.V. Vasilyev, Wavelet methods in computational fluid dynamics. Annu. Rev. Fluid Mech. 42, 473–503 (2010)
S. Summers, C.N. Jones, J. Lygeros, M. Morari, A multiresolution approximation method for fast explicit model predictive control. IEEE Trans. Autom. Control 56(11), 2530–2541 (2011)
G.G. Walter, A sampling theorem for wavelet subspaces. IEEE Trans. Inf. Theory. 38(2), 881–884 (1992)
P. Zhao, C. Zhao, P.G. Casazza, Perturbation of regular sampling in shift-invariant spaces for frames. IEEE Trans. Inf. Theory 52(10), 4643–4648 (2006)
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix 1: Proof of Lemma 1
Proof
In (9), the change of variable \(y=2^jx\) gives
Obviously Eq. (9) holds if and only if (37) does. Let \(P_{\tilde{S}^\phi } (w)=\frac{1}{2}\sum \limits _{k\in \mathbb {Z}} {\tilde{p}_k^s \mathrm{e}^{-iwk/2}} \) be a two-scale symbol corresponding to \(\tilde{S}^\phi (x)\) such that
where \(\hat{\tilde{S}}^\phi (w)\) is the Fourier transform of \(\tilde{S}^\phi (x)\).
First, we show that
for the coefficients \(\{\tilde{p}_k^s \}_{k\in \mathbb {Z}} \) in (38).
In [5, 6, 18], it is shown that the Fourier transform \(\hat{S}^\phi (w)\) can be expressed in terms of the Fourier transform \(\hat{\phi }(w)\) as
with \(1/\sum \nolimits _{k\in \mathbb {Z}} {\hat{\phi }(w+2k\pi )} \in L^2[0,2\pi ]\) .
Let
denote a two-scale symbol related to a scaling function \(\phi (x)\). We have
Inserting (42) into (40) yields
Since \(\tilde{S}^\phi (x)\) is also a scaling function, it follows from (38) and (43) that
Since \(\{\tilde{S}^\phi (x-k)\}_{k\in \mathbb {Z}} \) forms the dual basis of \(\{S^\phi (x-k)\}_{k\in \mathbb {Z}} \), we have \(\hat{\tilde{S}}^\phi (w)=\hat{S}^\phi (w)/\sum \nolimits _{k=-\infty }^{+\infty } {\left| {\hat{S}^\phi (w+2k\pi )} \right| ^2} \) (see Eq. 5.6.8 in [7]). Hence, it follows from (44) that
Since \(\sum \nolimits _{k\in \mathbb {Z}} {\hat{S}^\phi (w+2k\pi )} =1\) by (39), Eqs. (45) and (7) imply
Since \(\{S^\phi (x-k)\}_{k\in \mathbb {Z}} \) forms a Riesz basis of \(V_0 \), there are two constants \(A_{Es} \) and \(B_{Es} \) with \(0<A_{Es} \le B_{Es} <+\infty \) such that
It follows from (6), (47) and (48) that
Equation (49) implies \(0<\left| {\sum \nolimits _k {\tilde{p}_{2k}^s \mathrm{e}^{-iwk}} } \right| ^2<+\infty \) for every \(w\in [-\pi ,\pi ]\).
Second, we show there exist unique coefficients \(\{b_k^0 \}_{k\in \mathbb {Z}} \) and \(\{d_k^0 \}_{k\in \mathbb {Z}} \) for which (37) holds.
Assume \(\{c_k^0 \}_{k\in \mathbb {Z}} \) is any element of \(l^2(\mathbb {R})\). Since \(0<\left| {\sum \nolimits _k {\tilde{p}_{2k}^s \mathrm{e}^{-iwk}} } \right| ^2<+\infty \) by (49), the series \(2\sum \nolimits _{k\in \mathbb {Z}} {c_k^0 \mathrm{e}^{-iwk}} /\sum \nolimits _{k\in \mathbb {Z}} {\tilde{p}_{2k}^s \mathrm{e}^{-iwk}} \) is an element of \(L^2[-\pi ,\pi ]\) for any given \(\{c_k^0 \}_{k\in \mathbb {Z}} \in l^2(\mathbb {R})\). Hence there are unique coefficients \(\{b_k^0 \}_k \in L^2(\mathbb {R})\) such that
On the other hand, since \(\{b_k^0 \}_k \in l^2(\mathbb {R})\) in (50), we have that \(\frac{1}{2}\sum \nolimits _{k\in \mathbb {Z}} {b_k^0 \mathrm{e}^{-iwk}} \sum \nolimits _{k\in \mathbb {Z}} {\tilde{p}_{2k+1}^s \mathrm{e}^{-iwk}} \) is also an element of \(L^2[-\pi ,\pi ]\). Hence, there are unique coefficients \(\{d_k^0 \}_k \in l^2(\mathbb {R})\) such that
Equations (50) and (51) imply there always exist two unique sets of coefficients \(\{b_k^0 \}_{k\in \mathbb {Z}} \) and \(\{d_k^0 \}_{k\in \mathbb {Z}} \) in \(l^2(\mathbb {R})\) such that
for any given \(\{c_k^0 \}_{k\in \mathbb {Z}} \in l^2(\mathbb {R})\).
Multiplying (52) by \(\hat{\tilde{S}}^\phi (w/2)\) and inserting (38) yield
for any \(\{c_k^0 \}_k \in l^2(\mathbb {R})\). Taking the inverse Fourier transform on both sides in (53), we obtain
Since the coefficients \(\{b_k^0 \}_k \) and \(\{d_k^0 \}_k \) for which (52) holds are unique, it follows from (54) that, for any given \(\{c_k^0 \}_k \in l^2(\mathbb {R})\), there exist unique \(\{b_k^0 \}_{k\in \mathbb {Z}} \), \(\{d_k^0 \}_{k\in \mathbb {Z}} \in l^2(\mathbb {R})\) for which (37) holds. Hence for a given set of coefficients \(\{c_k^0 \}_k \in l^2(\mathbb {R})\), there exist unique \(\{b_k^0 \}_{k\in \mathbb {Z}} \), \(\{d_k^0 \}_{k\in \mathbb {R}} \in l^2(\mathbb {R})\) for which (9) holds. \(\square \)
Appendix 2: Proof of Theorem 1
Proof
Since \(S^\phi (x)\) is a scaling function, \(f_s (x)\in V_j \) can be written as
where \(S_{j,k}^\phi (x)=S^\phi (2^jx-k)\) and \(\{F_k \}_{k\in \mathbb {Z}} =\{f_s ((2k+1)/2^j)\}_{k\in \mathbb {Z}} \).
It follows from (55) that
where \(\{\tilde{S}_{j-1,k}^\phi (x)=\tilde{S}^\phi (2^{j-1}x-k)\}_k \) is the dual basis of \(\{S_{j-1,k}^\phi (x)=S^\phi (2^{j-1}x-k)\}_k \).
Obviously \(\sum \nolimits _{n\in \mathbb {Z}} {\left\langle {f_s ,S_{j-1,n}^\phi } \right\rangle \tilde{S}_{j-1,n}^\phi } \) in (56) is the projection of \(f_s (x)\) on \(V_{j-1} \). Hence \(f_s (x)\) is an element of \(W_{j-1} \) if and only if
Now we show that for given \(\left\{ {F_k } \right\} _k \), there exists unique \(\{f_s (2k/2^j)\}_{k\in \mathbb {Z}} \) in (56) for (57) to hold, i.e., the sequence \(\left\{ {F_k } \right\} _k \) in (55) determines a unique function \(f_s (x)\) whose projection \(\sum \nolimits _{n\in \mathbb {Z}} {\left\langle {f_s ,S_{j-1,n}^\phi } \right\rangle \tilde{S}_{j-1,n}^\phi } \) on \(V_{j-1} \) vanishes.
Letting
then
Substituting (59) into (55) yields that
It follows from (57) and (60) that
Since \(\{\tilde{S}_{j,k}^\phi (x)\}_k \) is the dual basis of the interpolatory basis \(\{S_{j,k}^\phi (x)\}_k \) in \(V_j \), we have
where \(\bar{S}^\phi (x)\) is the complex conjugate of \(S^\phi (x)\).
Inserting (62) into (61) yields
The Fourier transform of (63) yields
where \(A_f =\sum \limits _{k\in \mathbb {Z}} {F_k \sum \limits _{n\in \mathbb {Z}} {a_n \sum \limits _{l\in \mathbb {Z}} {\bar{S}^\phi (\frac{n+2k-2l+1}{2})\mathrm{e}^{-iwl}} } } \) and \(B_f =\sum \limits _{k\in \mathbb {Z}} f_s (\frac{2k}{2^j})\sum \limits _{n\in \mathbb {Z}} a_n \sum \limits _{l\in \mathbb {Z}} {\bar{S}^\phi (\frac{n+2k-2l}{2})\mathrm{e}^{-iwl}} \).
By the Poisson summation formula
where \(\hat{S}^\phi (w)\) is the Fourier transform of \(S^\phi (x)\).
Inserting (65) into \(A_f \) in (64) yields
Cutting the summation (66) into two parts, one for \(n=2m\) and the other for \(n=2m-1\), yields
Since \(\{S^\phi (x-k)\}_k \) is the interpolatory basis of \(V_0 \), (39) indicates that \(\sum \nolimits _{l\in \mathbb {Z}} {\hat{S}^\phi (w+2l\pi )} =1\). Hence it follows from (67) that
It follows from (58) and Theorem 2.4 in [18] that
Since \(\vert \hat{S}^\phi (u/2)\vert ^2\) is even in \(u\), it follows from (69) that
Similarly, we have
Inserting (70) and (71) into (68) yields
Cutting summation (72) into two parts, one for \(m=2n\) and the other for \(m=2n-1\), yields
Since it follows from (39) that \(\sum \limits _{l\in \mathbb {Z}} {\hat{S}^\phi (w+2l\pi )} =\sum \limits _{l\in \mathbb {Z}} {\hat{S}^\phi (w+4l\pi )} +\sum \limits _{l\in \mathbb {Z}} \hat{S}^\phi (w+ 4l\pi +2\pi ) =1\), Eq. (73) gives
Similarly
Since \(\sum \nolimits _{k\in \mathbb {Z}} {\left| {\hat{S}^\phi (w/2+2\pi k)} \right| ^2} \in L^2[-\pi ,\pi ]\) by the frame theory [19] and \(\sum \nolimits _{n\in \mathbb {Z}} {\hat{S}^\phi (w+4n\pi )} \in L^2[-2\pi ,2\pi ]\), we have
with \(\{\tilde{q}_k \}_{k\in \mathbb {Z}} \in l^2(\mathbb {Z})\), and
where \(P_s (w)\) and \(E_s (w)\) are respectively defined in (3) and (7).
Since \(P_s (w)\) is a two-scale symbol corresponding to the scaling function \(S^\phi (2^jx)\), the sequence \(\{\psi _Q (x-k)\}_{k\in \mathbb {Z}} \) forms a Riesz basis of \(W_0 \), where the Fourier transform of \(\psi _Q (x)\) is \(\hat{\psi }_Q (w)\) (see Eq. 5.6.13 in [7]).
Inserting (76) respectively into (74) and (75) yields
Substituting (78) into (64) gives
Let \(\tilde{\psi }_Q (x)\) be the dual of \(\psi _Q (x)\) in \(W_0 \). Consider the “decomposition relation” of scaling function and wavelet, it follows from (3), (76) and Theorem 5.16 in [7] that
where \(\hat{\tilde{S}}^\phi (w)\) is the Fourier transform of \(\tilde{S}^\phi (x)\) and \(\hat{\tilde{\psi }}_Q (w)\) is the Fourier transform of \(\tilde{\psi }_Q (x)\).
Multiplying the two identities in (80) by \(\sum \limits _{k\in \mathbb {Z}} {(-F_k )\mathrm{e}^{(-iw/2)\times 2k}} \) and \(\sum \limits _{k\in \mathbb {Z}} f_s (\frac{2k}{2^j}) \mathrm{e}^{(-iw/2)\times 2k-1} \) consecutively,
Taking the inverse Fourier transform on both sides in (82),
Obviously, \(f_c (x)\) is an element of \(V_0 \) since it can be represented as a linear combination of \(\{\tilde{S}^\phi (x-m)\}_{m\in \mathbb {Z}} \) in (83). Comparing (83) to (9), we know that \(\{-F_k \}_{k\in \mathbb {Z}} \) and \(\{f_s (2k/2^j)\}_{k\in \mathbb {Z}} \) respectively correspond to \(\{c_k^0 \}_{k\in \mathbb {Z}} \) and \(\{d_k^0 \}_{k\in \mathbb {Z}} \). By (6), it follows from Lemma 1 that the coefficients \(\{f_s (2k/2^j)\}_{k\in \mathbb {Z}} \) for which (83) holds is unique for any given sequence \(\{-F_k \}_{k\in \mathbb {Z}} \). Hence, (83) implies \(\{F_k \}_{k\in \mathbb {Z}} \) uniquely determines \(\{f_s (2k/2^j)\}_{k\in \mathbb {Z}} \) in (56) for (57) to hold.
Since \(\{f_s (2k/2^j)\}_{k\in \mathbb {Z}} \) and \(\{F_k \}_{k\in \mathbb {Z}} \) determine a unique element of \(V_j \) in (55), Eqs. (57) and (83) imply that the sequence \(\{F_k \}_{k\in \mathbb {Z}} \) determines a unique element of \(V_j \) whose projection on \(V_{j-1} \) vanishes, i.e., \(\{F_k \}_{k\in \mathbb {Z}} \) determines a unique element of \(W_{j-1} \). This proves Theorem 1. \(\square \)
Appendix 3: Proof of Lemma 2
Proof
It follows from (17) and (18) that \(\psi (x)\) is an element of \(V_1 \subset L^2(\mathbb {R})\). By Parseval’s identity,
for \(n\in \mathbb {Z}\). Since \(W_0 \) is orthogonal to \(V_0 \) and \(W_0 \oplus V_0 =V_1 \), Eq. (84) implies
if and only if \(\psi (x)\) is an element of \(W_0 \) where \(\oplus \) denotes a direct sum of two vector spaces.
Since \(\{\mathrm{e}^{iwn}\}_n \) is a complete orthonormal basis of \(L^2[-\pi ,\pi ]\), Eq. (85) holds if and only if
almost everywhere.
From (42) and (17), Eqs. (85) and (86) imply
holds if and only if \(\psi (x)\) is an element of \(W_0 \). It follows from (87) and (20) that (19) holds if and only if \(\psi (x)\) is an element of \(W_0 \). \(\square \)
Appendix 4: Proof of Theorem 2
Proof
It follows from (11) and [17] that \(\hat{S}^\psi (w)\) can be expressed as
where \(\hat{\psi }(w)\) is the Fourier transform of a wavelet \(\psi (x)\).
Now, we relate the expressions of \(\hat{S}^\psi (w)\) and \(Q_s (w)\) to that of \(S^\phi (x)\) instead of \(\psi (x)\) in (88).
By (89)
Since \(P_s (w)\) in (3) and \(Q_s (w)\) in (14) form a pair of reconstruction filters in an MRA \(\{V_j \}_{j\in \mathbb {Z}} \), it follows from Lemma 2 and (90) that
On the other hand, applying (89) to (16) yields
Consider the formula
Inserting (92) into (93) yields
It follows from (2) and (7) that
Inserting (96) into (95), we have
Since \(( {P_s (w),Q_s (w)})\) forms a pair of reconstruction filters, (15) holds. Hence, by (97)
Since the sequence \(\{S^\phi (x-k)\}_k \) is a frame, there should exist two constants \(A_{Es} \) and \(B_{Es} \) with \(0<A_{Es} \le B_{Es} <+\infty \) such that
(Theorem 9 in [19]). Applying (15) and (99) to (98) yields
which implies that the two-scale condition holds.
Equation (100) implies
Hence, it follows from (91) and (101) that
which verifies Eq. (21). \(\square \)
Rights and permissions
About this article
Cite this article
Zhang, Z., Kon, M.A. On Relating Interpolatory Wavelets to Interpolatory Scaling Functions in Multiresolution Analyses. Circuits Syst Signal Process 34, 1947–1976 (2015). https://doi.org/10.1007/s00034-014-9937-8
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00034-014-9937-8