Tight framelets and fast framelet filter bank transforms on manifolds (Q2278450)

Tight wavelet-type frames on compact, connected smooth Riemannian manifolds \(\mathcal{M}\) are characterized. (Continuous) framelet systems have the form \[ \text{CFS}_J(\Psi)=\text{CFS}_J(\{\alpha;\beta^1,\dots, \beta^r\}):=\{\varphi_{J,y}: y\in\mathcal{M}\}\cup\{\psi_{j,y}^1,\dots, \psi_{j,y}^r,\, y\in \mathcal{M}, j\geq J\}\, . \] The first basic step is to transfer filterbank conditions on \(\mathbb{R}\) to build multiscale systems on \(\mathcal{M}\). Here the authors follow work of \textit{M. Maggioni} and \textit{H. N. Mhaskar} [Appl. Comput. Harmon. Anal. 24, No. 3, 329--353 (2008; Zbl 1242.42027), Theorem 4.1]. One begins with a framelet filter bank on \(\mathbb{R}\), \({\boldsymbol{\eta}}=\{a;b^1,\dots,b^r\}\) defined by a refinable function \(\alpha\) and refinement masks \(a\) and framelet masks \(b^n\) satisfying \(\widehat{a}(2\xi)=\widehat{a}(\xi)\hat\alpha(\xi)\); \(\widehat{\beta^n}(2\xi)=\widehat{b^n}(\xi)\hat\alpha(\xi)\). The continuous framelet elements on \(\mathcal{M}\) are defined by \[\varphi_{j,y}(x)=K_{\widehat\alpha,2^j}(x,y)=\sum_{\ell=0}^\infty \widehat\alpha\Bigl(\frac{\lambda_\ell}{2^j}\Bigr)\overline{u_\ell (y)} u_\ell(x);\] \[ \psi^n_{j,y}(x)=K_{\widehat{\beta^n},2^j}(x,y)=\sum_{\ell=0}^\infty \widehat{\beta^n}\Bigl(\frac{\lambda_\ell}{2^j}\Bigr)\overline{u_\ell (y)} u_\ell(x), n=1,\dots, r, \] where \(u_\ell\) and \(\lambda_\ell\) are eigenpairs for what the authors refer to as ``a certain operator, such as the Laplace-Beltrami operator.'' The characterization is via a type of unitary extension principle (UEP) involving the refinable function \(\alpha\), of the form \[\lim_{j\to\infty} \overline{\widehat\alpha\Bigl(\frac{\lambda_\ell}{2^j}\Bigr)}\widehat\alpha\Bigl(\frac{\lambda_{\ell^\prime}}{2^j}\Bigr)U_{\ell,\ell^\prime}(Q_{N_j})=\delta_{\ell,\ell^\prime},\quad U_{\ell,\ell^\prime}(Q_{N_j})=\sum_{k=0}^{N_j} \omega_{j,k} u_\ell({x}_{j,k})\overline{u_{\ell^\prime}}({x}_{j,k})\, \] and \[ \left[\overline{\widehat{a}\Bigl(\frac{\lambda_\ell}{2^j}\Bigr)} \widehat{\alpha}\Bigl(\frac{\lambda_{\ell^\prime}}{2^j}\Bigr)U_{\ell,\ell^\prime}(Q_{N_{j-1}})+ \sum_{n=1}^r \overline{\widehat{b^n}}\Bigl(\frac{\lambda_\ell}{2^j}\Bigr) \widehat{b^n}\Bigl(\frac{\lambda_{\ell^\prime}}{2^j}\Bigr)U_{\ell,\ell^\prime}(Q_{N_{j}}) \right]=U_{\ell,\ell^\prime}(Q_{N_{j}}) \] when \(\ell/2^j\) and \(\ell^\prime/2^j\) are in the support of \(\hat{a}\), \(j\geq J_0+1\), and \(Q_{N_j}\) define quadratures. These conditions are analogous to UEPs on \(\mathbb{R}\) where a filterbank \(\eta=\{a;b^1,\dots, b^r\}\) associated with \(\Psi=\{\alpha; \beta^1,\dots,\beta^r\}\) satisfies \(|\hat{a}(\xi)|^2+\sum_{n=1}^r |\widehat{b^n}(\xi)|^2=1\) a.e. and \(\overline{\hat{a}(\xi)}\hat{a}\Bigl(\xi+\frac{1}{2}\Bigr)+\sum_{n=1}^r \overline{\widehat{b^n}(\xi)}\widehat{b^n}\Bigl(\xi+\frac{1}{2}\Bigr)=0\) a.e. The second important innovation is a filterbank characterization that relies on assumed quadrature properties. For \(n\geq 0\), a quadrature rule \(Q_{N,n}=\{(\omega_k,y_k)\}_{k=0}^N\) is said to be polynomial exact of degree \(n\) if \(\int_M q(x)\, d\mu(x)=\sum_{k=0}^N \omega_k q(y_k)\), \(q\in \Pi_n\) where \(\Pi_n=\text{span}\,\{u_\ell, \overline{u_\ell}, \lambda_\ell\leq n \}\) (where \(u_\ell\) and \(\lambda_\ell\) are as above) and \(\mu\) is a given probability measure on \(\mathcal{M}\). It is shown (Corollary 2.6) that if \(\mathcal{Q}=\{Q_{N_j}\}_{j\geq J_0}\) is a polynomial exact quadrature rule of order \(2^j\) (\(j\geq J_0\)) then the semi-discrete framelet system \(\text{FS}_J(\Psi,\mathcal{Q})\), where \(y\in\mathcal{M}\) in \( \text{CFS}_J(\Psi)\) is replaced by \(y\in \mathcal{Q}\), is a tight frame for \(L^2(\mathcal{M})\) for all \(J\geq J_0\). One assumes a discrete Fourier transform can be defined by \(({F}_j c)_k=\sum_{\ell=0}^{\Lambda_j} c_\ell\sqrt{\omega_{j,k}} u_\ell({x}_{j,k})\), \(k=0,\dots, N_j\) (here, \(\Lambda_j=\#\{\ell: \lambda_\ell\leq 2^j/c\}\)). The discrete Fourier operator is used to build a fast algorithm to compute discrete multiscale framelet decompositions (Theorem 3.1). A concrete instance of the general theory is provided in the special case \(\mathcal{M}=\mathbb{S}^2\), the two-sphere, including the filterbank decompositions.