Construction of compactly supported shearlet frames (Q663318)

Fairly general sufficient conditions for a {\textit{cone-adapted shearlet frame}} for \(L^2(\mathbb{R}^2)\) are given. A specific construction of a discrete shearlet frame with compactly supported generating function is then provided, followed by a discussion of \(N\)-term approximation bounds and implementation complexity. The sufficient frame criteria requires the introduction of a \textit{feasible shearlet system} parametrized by sequences \(\{a_j\}_{j\geq 0}\subset \mathbb{R}^{+}\) such that \(a_{j+p}\leq \mu a_{j}\) for some fixed \(p\in \mathbb{N}\) and \(\mu>0\), \(\{s_{j,k}=s_k\sqrt{a_j}\}_{j\geq 0},\, k\in K_j\) with \(K_j\) an appropriate bounded subset of integers containing zero such that \(|s_k|\leq \sqrt{a_j}+1\) when \(k\in K_j\), and \(\{t_{j,k,m}^c\}_{j\geq 0,\;k\in K_j,\;m\in\mathbb{Z}^2}\subset \mathbb{R}^2\). Here, \(t_{j,k,m}^c=S_{s_{j,k}}A_{a_j} M_c m\) where \(S_s\) is the \(2\times 2\) shear matrix with identity along the diagonal and upper right entry \(s\), \(A_a\) is the scaling matrix with diagonal entries \(\sqrt{a}\) and \(1/\sqrt{a}\), and \(M\) is diagonal with entries \(c_1\) and \(c_2\). One assumes a shear bound condition \[ \sup_{x=(x_1,x_2)\in \mathbb{R}^2} \sum_{k\in\mathbb{Z}} \min\left\{1,\left |(S_{s_k}x)_2\right|\right\}\cdot \min\left\{1,\left |(S_{s_k}x)_1\right|^{-\gamma}\right\}\leq C(\gamma)<\infty \] for any \(\gamma>1\). A function \(\psi\in L^2(\mathbb{R}^2)\) is then called a \textit{feasible shearlet} if there exist \(\alpha>\gamma>3\) and \(q>q'>0\) and \(r\in (0,q)\) such that \[ |\hat\psi(\xi_1,\xi_2)|\leq \min\{1,|q\xi_1|^\alpha\}\cdot \min\{1,|q'\xi_1|^{-\gamma}\}\cdot \min\{1,|r\xi_2|^{-\gamma}\}, \] governing localization of the Fourier transform \(\hat{\psi}\) in the frequency domain. For \[ \Phi(\xi,\omega)=\sum_{j\geq 0}\sum_{k\in K_j} |\hat{\psi}(S_{s_k}^T A_{a_j}\xi)||\hat\psi(S_{s_k}^T A_{a_j}\xi +\omega)| \] one sets \[ L_{\text{inf}}=\text{ess}\,\inf_{\xi\in \mathcal{C}} \Phi(\xi,0),\quad L_{\text{sup}}=\text{ess}\, \sup_{\xi\in \mathcal{C}} \Phi(\xi,0) \] where \(\mathcal{C}=\{(\xi_1,\xi_2):\xi_1\geq 1,\,\, |\xi_2/\xi_1|\leq 1\}\), and \[ \Gamma(\omega) =\text{ess}\, \sup_{\xi\in\mathcal{C}} \Phi(\xi,\omega),\quad R(c)=\sum_{m\in\mathbb{Z}^2\setminus \{0\}} [\Gamma(M_c^{-1}m)\Gamma(-M_c^{-1}m)]^{1/2} \] where \(c=(c_1,c_2)\in (\mathbb{R}_+)^2\). It is proved then that the shearlet set \(\mathcal{SH}(\Lambda,\psi)\), \[ \mathcal{SH}_{\phi,\psi,\tilde{\psi}}(f)(t',(a,s,t),(\tilde{a},\tilde{s},\tilde{t})) =(\langle f,\phi_{t'}\rangle, \langle f,\psi_{a,s,t}\rangle, \langle f,\tilde\psi_{\tilde{a},\tilde{s},\tilde{t}}\rangle) \] and \(\psi_{a,s,t}=a^{-3/4}\psi(A_a^{-1}S_s^{-1}(\cdot-t))\), forms a frame for \(L^2(\mathcal{C})\) with frame bounds \(A,B\) if \[ \frac{1}{\det M_c} [L_{\text{inf}}-R(c)]\leq A\leq B\leq \frac{1}{\det M_c} [L_{\text{sup}}-R(c)]. \] To construct an actual example, the authors choose the feasible set of parameters \[ \Lambda =\left\{(2^{-j},k2^{j/2},S_{k2^{j/2}}A_{2^j}M_c m):\, j\geq 0, |k|\leq \lceil 2^{j/2}\rceil,\, m\in \mathbb{Z}^2\right\}, \,\, c\in (\mathbb{R}_+)^2. \] One defines a shearlet function \(\psi\) by \[ \hat\psi=m_1(4\xi_1)\hat\phi(\xi_1)\hat\phi(2\xi_2),\quad m_1(\xi_1)=m_0(\xi_1+1/2) \] in which \(m_0\) is a polynomial low-pass quadrature mirror filter having enough regularity that the scaling function \(\hat\phi(\xi_1)=\prod_{j=0}^\infty m_0(\xi_1/2^j)\) is bounded below in a neighborhood of the origin, has an appropriate decay, and its inverse Fourier transform \(\phi\) has compact support. It is proved that this set of parameters and feasible shearlet indeed give a shearlet frame for \(L^2(\mathcal{C})\). The digital shearlet transform corresponding to these parameters and filter is shown to implement with \(O\left(2^{\log_2 (1/2(L/2-1))}LM^2\right)\) complexity for parameters \(M,L\) that arise in the construction of the filter \(m_0\). One also has an \(N\)-term approximation bound of the form \[ \left\|f-\sum_{\lambda\in\Lambda_N} \langle P_J(f),\psi_\lambda\rangle \tilde\psi_\lambda\right\|_2^2 \leq C (\log N)^3\cdot N^{-2},\quad N\to\infty, \] where \(P_J\) denotes orthogonal projection onto the approximation space of a suitable orthogonal scaling function and \(\tilde\psi_{\lambda}\) is a dual shearlet frame generator for the same feasible parameter set \(\Lambda\).