Dualizable shearlet frames and sparse approximation (Q530576)

It was proved by Donoho that for cartoon-like functions \(f\), the \(N\)-term approximation \(f_N\) using any frame for \(L^2(\mathbb{R}^2)\) obeys \[ \|f-f_N\|_2 \gtrsim N^{-1}, \quad N\rightarrow \infty. \] The search of frames with optimal approximation rate property is hence an important topic in applied harmonic analysis. Shearlets, as one of the directional multiscale representation systems that are usually constructed to be band-limited, have been shown to provide (nearly) optimal approximation rate for cartoon-like functions: \[ \|f-f_N\|_2\lesssim N^{-1}\log(N),\quad N\rightarrow \infty. \] In both theory and applications, compactly supported shearlets are more desirable. It is known that compactly supported shearlets can be constructed to be a frame and provide (nearly) optimal approximation rate. However, it is not known whether there exist `nice' associated dual frames for compactly supported shearlet frames. In this paper, the authors propose an approach to construct compactly supported shearlet frames having dual frames with closed form as well as nearly optimal approximation rate. Let \(A_j:=\begin{pmatrix} 2^j & 0 \\ 0 & 2^{\lfloor j/2\rfloor}\end{pmatrix}\) and \(S_k=\begin{pmatrix} 1 & k \\ 0 & 1\end{pmatrix}\) be the anisotropic dilation matrix and shear matrix. Starting from a 1D orthonormal wavelet basis \[ \varphi_1(\cdot-m): m\in\mathbb{Z}\cup\{2^{j/2}\psi_1(2^j\cdot-m): j\geq0, m\in\mathbb{Z}\}, \] the paper then forms a 2D shearlet-type wavelet system \[ \begin{aligned} \Psi_s(\varphi_1,\psi_1):=&\{|\det A_{j_0}|^{1/2}\varphi^p(A_{j_0} S_s\cdot-D_pm), m\in\mathbb{Z},p\geq0\}\\ &\cup \{|\det A_{j}|^{1/2}\psi^p(A_{j} S_s\cdot-D_pm), m\in\mathbb{Z},p\geq0, j\geq j_0\} \end{aligned} \] with \(D_p:=\mathrm{diag}(1,2^{-\max\{1-p,0\}})\) and \[ \begin{aligned} \psi^0(x):=&\psi_1(x_1)\varphi(x_2), \\\varphi^0(x):=&\varphi_1(x_1)\varphi_1(x_2),\\ \psi^p(x):=&2^{(p-1)/2}\psi_1(x_1)\psi_1(2^{p-1}x_2),\\ \varphi^p(x):=&2^{(p-1)/2}\varphi_1(x_1)\psi_1(2^{p-1}x_2) \end{aligned} \] for \(p>0\). The shearlet-type wavelet system \(\Psi_s(\varphi_1,\psi_1)\) is then shown to be an orthonormal basis for \(L^2(\mathbb{R}^2)\) for each \(s\) and hence it has frequency tiling spreading over the whole frequency domain. In order to construct a subsystem that is mainly concentrated along the shear direction \(s\in \mathbb{S}\) with \(\mathbb{S}\) being an appropriate collection of shear directions, the paper then proceeds to construct a (compact support) filtering function \(G_s\) mainly concentrating along the shear direction \(s\) with \[ \hat{G}_0(\xi):=|\hat\varphi_0(\xi)|^2+\sum_{j=0}^\infty|\hat{g}(A_j^{-1}\xi)|^2 \text{ and } \hat{G}_s(\xi):=\sum_{j=0}^\infty|\hat{g}(A_j^{-1}S_{-s}^\top \xi)|^2. \] for some compactly supported function \(g\) satisfying certain conditions. With \(\Psi_s(\varphi_1,\psi_1)\) and \(G_s\) at hand, the paper then defines the dualizable shearlet system \(SH(\varphi_1,\psi_1; g)\) to be \[ SH(\varphi_1,\psi_1;g):=\{G_s*h, (G_s*h)\circ R : h\in \Psi_{s}(\varphi_1,\psi_1), s\in\mathbb{S}\} \] with \(R=\begin{pmatrix} 0 & 1 \\ 1 &0\end{pmatrix}\) the permutation matrix for obtaining another cone function. The paper shows that such a system \(SH(\varphi_1,\psi_1;g)\) is indeed a cone-adapted shearlet system \[ \begin{aligned} SH(\phi; \psi,\tilde\psi):=&\{\phi(\cdot-c_1m): m\in\mathbb{Z}^2\}\\ &\cup\{|\det A_j|^{1/2} \psi(S_kA_j\cdot-diag(c_1,c_2)m: j\geq0, |k|\leq 2^{\lceil j/2\rceil}, m\in\mathbb{Z}^2\}\\ &\cup\{|\det \tilde A_j|^{1/2} \tilde \psi(S_k\tilde A_j\cdot-diag(c_2,c_1)m: j\geq0, |k|\leq 2^{\lceil j/2\rceil}, m\in\mathbb{Z}^2\}.\end{aligned} \] Define the normalization filter \(G\) by \[ \hat G := \sum_{s\in\mathbb{S}} (|\hat G_s|^2 + |\hat G_s\circ R|^2). \] The paper then show that one can construct the dual frame of the dualizable shearlet system \(SH(\varphi_1,\psi_1; g)\) to be \[ \widetilde{SH}(\varphi_1,\psi_1; g)=\{\tilde h, \tilde h\circ R: \tilde h\in\widetilde{\Psi}_s(\varphi_1,\psi_1)\}, \] where \[ \widetilde{\Psi}_s(\varphi_1,\psi_1) = \left\{\tilde h: \hat{ \tilde h }:= \frac{\hat h}{\hat G}, h\in \Psi_s(\varphi_1,\psi_1)\right\}. \] The last part of the paper is then devoted to prove the approximation property of the dualizable shearlet frame for cartoon-like function. The paper proves that \[ \|f-f_N\|_2 \lesssim N^{-1+12\rho}\cdot\log(N), \quad N\rightarrow\infty. \] In summary, this paper provides a new approach to construct compactly supported shearlet frames which have nice properties in the sense that its dual can be constructed explicitly as well as nearly optimal approximation rate. However, as pointed out by the paper, the dual is still not a shearlet frame. Moreover, the dual is not compactly supported. It should be noticed that in the rate of \(N\)-term approximation, there is a factor \(12\rho\) appears. This factor will cause a trade-off between the better approximation rate (\(\rho\) needs to be small) and the small support length of the generators (\(\rho\) needs to be large). Construction of compactly supported shearlet frames with compactly supported dual remains a challenge task.