Strict positive definiteness of convolutional and axially symmetric kernels on \(d\)-dimensional spheres (Q2135770)

Section 1 introduces the interpolation problem and provides background. It considers kernel functions \(K : \mathbb S^{d-1} \times \mathbb S^{d-1}\to \mathbb C\) that are assumed to be continuous and Hermitian, \(K(\xi,\zeta) = \overline{K(\zeta,\xi)}\), on the \(d\)-sphere (or \((d-1)\)-dimensional sphere) \(\mathbb S^{d-1} := \{ \xi \in \mathbb R^d : \xi_1^2+\cdots +\xi_d^2 = 1\}\), \(d \geq 2\). \(K\) is positive definite (resp. strictly positive definite) if the matrix \[ K_\Xi(\xi,\zeta)_{\xi,\zeta \in \Xi} \tag{1} \] is positive definite (resp. strictly positive definite) for arbitrary finite \(\Xi \subset \mathbb S^{d-1}\). This condition is important because it ensures a unique solution of the interpolation problem: given a function \(f : \Xi \to \mathbb C\), compute \(c_\xi\), \(\xi \in \Xi\), such that the function \(s_f : \mathbb S^{d-1} \to \mathbb C\) \[ s_f(\zeta) := \sum_{\xi \in \Xi} c_\xi \, K(\zeta,\xi) \tag{2} \] satisfies the interpolatory condition \[ s_f(\zeta) = f(\zeta), \ \ \zeta \in \Xi.\tag{3} \] If \(f\) is defined on \(\mathbb S^{d-1}\) then \(s_f\) provides an approximation of \(f\) whose accuracy increases with the smoothness of \(f\) and the density of \(\Xi\). \(\mathbb S^{d-1}\) admits a Riemannian \((d-1)\)-form and under the assumptions above the kernel defines a self-adjoint integral operator \(K : H \to H\) on the Hilbert space \(H := L^2(\mathbb S^{d-1})\). This integral operator is positive definite (resp. strictly positive definite) if \((Kh,h) \geq 0\) (resp. ${>0}$) for every nonzero \(h \in H\). Theorem 2.1 in [\textit{N. Dyn} et al., in: Approximation theory and optimization. Tributes to M. J. D. Powell on the occasion of his 60th birthday. Cambridge: Cambridge University Press. 133--144 (1997; Zbl 1049.41014)] shows that \(K(\xi,\zeta)\) is positive definite iff \(K\) is positive definite. Furthermore, if rank \(K < \infty\) then \(K(\xi,\zeta)\) is not strictly positive definite. However other relations between strict positive definiteness of \(K(\xi,\zeta)\) and the integral operator \(K\) are elusive. The paper derives new conditions that ensure \(K(\xi,\zeta)\) to be strictly positive definite under the assumption that \(K(\xi,\zeta)\) satisfies certain symmetry conditions. \(H = \bigoplus_{j=0}^\infty H_j\) where \(H_j\) is the eigenspace of the Laplace-Beltrami operator \(\Delta\) on \(\mathbb S^{d-1}\) with eigenvalue \[ -\lambda_j = -j(j+d-2) \tag{4} \] and \[ N_{j,d} := \dim H_j = \binom{d+j-1}{d-1}-\binom{d+j-3}{d-1} = \begin{cases} 1, \ \ j = 0, \\ \frac{(2j+d-2)(j+d-3)!}{j!(d-2)!}, \ \ j \geq 1. \end{cases} \tag{5} \] \(H_j\) has an orthonormal basis of spherical harmonic \(Y_j^k, k = 1,\dots,N_{j,d}\); thus \(K(\xi,\zeta)\) admits the expansion \[ K(\xi,\zeta) = \sum_{j,j'=0}^\infty \sum_{k = 1}^{N_{j,d}} \sum_{k' = 1}^{N_{j',d}} a_{j,j',k,k'} \, Y_j^k(\xi,\zeta) \, \overline {Y_{j'}^{k'}(\xi,\zeta)} \tag{6} \] whose coefficients are uniquely determined by \(K Y_{j',k'} = \sum_{j=0}^\infty \sum_{k = 1}^{N_{j,d}} a_{j,j',k,k'} \, Y_j^k\). Section 2 characterizes certain geometric symmetry conditions on \(K(\xi,\zeta)\) in terms of conditions on its expansion coefficients \(a_{j,j',k,k'}\). The paper notes that the case most studied [\textit{S. Hubbert} et al., Spherical radial basis functions, theory and applications. Cham: Springer (2015; Zbl 1321.33017); \textit{A. Lang} and \textit{C. Schwab}, Ann. Appl. Probab. 25, No. 6, 3047--3094 (2015; Zbl 1328.60126); \textit{C. Ma}, Math. Geosci. 47, No. 6, 699--717 (2015; Zbl 1323.62092); \textit{Z. Nie} and \textit{C. Ma}, Proc. Am. Math. Soc. 147, No. 7, 3047--3056 (2019; Zbl 1426.33031)] is when \(K(\xi,\zeta)\) is isotropic, i.e., it depends only on the geodesic distance arccos \(\zeta^T\xi\) between \(\xi\) and \(\zeta\), that this corresponds to radial basis functions, and states that \textit{I. J. Schoenberg} [Duke Math. J. 9, 96--108 (1942; Zbl 0063.06808)] proved that isotropy is equivalent to the condition on coefficients \[ a_{j,j',k,k'} = c_j \delta_{j',j} \delta_{k,k'}.\tag{7} \] The sufficiency of this condition follows from the addition formula \[ \frac{N_{j,d}}{\sigma_{d-1}} P_j^d(\xi^T\zeta) = \sum_{k=1}^{N_{j,d}} Y_j^k(\xi) \overline {Y_j^k(\zeta)}, \tag{8} \] where \(\sigma_{d-1}\) is the surface area of \(\mathbb S^{d-1}\), which is given following Lemma 13 on page 20 of the paper, together with the positive definiteness of the Legendre polynomial \(P_j^d\) of degree \(j\) in \(d\)-dimensions. \(K(\xi,\zeta)\) is convolutional if \(K H_j \subset H_j, \ j \geq 1\), i.e., the matrix of coefficients has the block structure \[ a_{j,j',k,k'} = \delta_{j',j} \, d_j(k,k').\tag{9} \] Axial symmetry depends on the specific choice of polar coordinates \[ \theta := (\theta_1,\dots,\theta_{d-1})^T \in [0,2\pi) \times [0,\pi]^{d-2}\tag{10} \] on \(\mathbb S^{d-1}\). We note that in the paper these are (only implicitly) related to the Cartesian coordinates by \begin{align*} \xi_1 &= \cos \theta_{d-1} \\ \xi_2 &= \sin \theta_{d-1} \cos \theta_{d-2} \\ &\ \vdots \\ \xi_{d-1} &= \sin\theta_{d-1} \dots \sin \theta_{2} \cos \theta_{1} \\ \xi_{d} &= \sin\theta_{d-1} \dots \sin \theta_{2} \sin \theta_{1} \end{align*} For \(d = 3\), (4) follows the convention in physics where \((\theta_1,\theta_2)^T = (\phi,\theta)^T \in [0,2\pi) \times [0,\pi]\), \(\xi_1 = z = \cos \theta\), \(\xi_2 = x = \sin \theta \cos \phi\), \(\xi_3 = y = \sin \theta \sin \phi\). Hence there is a typo error in the assertion \((\theta,\phi) \in [0,2\pi) \times [0,\pi]\) which follows (4). This assertion holds for the convention in mathematics. Using polar coordinates we have that \(K(\xi,\zeta) = \widetilde K(\theta,\theta')\) and the kernel is said to be axially symmetric if \[ \widetilde K((\theta_1,\dots,\theta_{d-1} +\alpha)^T, (\theta'_1,\dots,\theta'_{d-1} +\alpha)^T)\tag{11} \] is idependent of \(\alpha\). The spherical harmonics \(Y_{\ell_1,\dots,\ell_{d-1}}(\xi)\) on \(\mathbb S^{d-1}\) of degree \(\ell_{d-1}\) in \(d\)~variables are expressed using polar coordinates in (12) as \(\frac{1}{2\pi}e^{i\ell_1\theta_1}\) times a product of Legendre functions of the \(\theta_2,\dots,\theta_{d-1}\). Here \(\ell_1,\dots,\ell_{d-1}\) are integers satisfying \(\ell_{d-1} \geq \cdots \geq |\ell_1|\). If \[ \tilde H_{\ell_1} := \operatorname{span} \{Y_{\ell_1,\dots,\ell_{d-1}}: \ell_{d-1} \geq \cdots \geq |\ell_1|\}, \ \ \ell_1 \in \mathbb Z, \tag{12} \] then Theorem 1, which expresses axial symmetry by the coefficient conditions (14), means that \(K(\xi,\zeta)\) is axially symmetric if and only if \[ K \tilde H_{\ell_1} \subset \tilde H_{\ell_1}, \ \ \ell_1 \in \mathbb N. \tag{13} \] This means that the matrix representation for \(K\) has a block structure whose matrix blocks \(c_{\ell_1} : \Lambda_{|\ell_1|} \times \Lambda_{|\ell_1|} \rightarrow \mathbb C\) in (16) are indexed by the sets \[ \Lambda_{|\ell_1|} := \{(\ell_2,\dots,\ell_{d-2}) \in \mathbb N^{d-2} : |\ell_1| \leq \ell_2 \leq \cdots \leq \ell_{d-1}\}\tag{14} \] defined on page 8. Isotropic kernels are both convolutional and axially symmetric. Definition 2 introduces two additional symmetry properties for axially symmetric kernels: longitude reversibility and longitudinal independence. Lemma 5 characterizes these properties in terms of expansion coefficients. We observe that for \(d = 3\) Eqn 11 represents rotation by \(\alpha\) about the \(z\)-axis, but for \(d > 3\) it does not represent rotation about an axis but it does represent a rotation whose fixed points are the \((d-2)\)-sphere \(\{\xi \in \mathbb S^{d-1}: \xi_1^1+ \cdots + \xi_{d-2}^2 = 1\}\). Thus the word `axes' in line 9 on page 7 in the paper should be changed to `polar coordinate'. Sections 3 and 4 characterize positive definite and strictly positive definite kernels for axially symmetric or convolutional kernels under the additional assumption, expressed in (7), which implies that the expansion in Eqn 6 converges pointwise absolutely. Theorem 2 shows that for \(d \geq 3\) an axially symmetric kernel is positive definite iff \(c_{\ell_1}\) is positive definite for every \(\ell_1 \in \mathbb Z\). A similar argument shows a convolutional kernel is positive definite iff the mappings \(K : H_j \to H_j\) are positive definite for \(j \geq 0\). For axially symmetric kernels define \(F := \{\ell_1 \in \mathbb Z : c_{\ell_1} \text{ is strictly positive definite}\}\) and for convolutional kernels define \(F := \{j \geq 0 : K : H_j \to H_j \text{ is strictly positive definite}\}\). Theorem 3 uses the Cholesky-decomposition to derive properties of the expansion coefficients that imply strict positive definiteness for axially symmetric kernels. For axially symmetric \(K(\xi,\zeta)\), \(K\) is positive definite iff \(c_{\ell_1}\) is positive definite for every \(\ell_1 \in \mathbb Z\). Then Lemma 8 shows that \(K(\xi,\zeta)\) is strictly positive definite if it is positive definite and \[ \sum_{\xi \in \Xi} \lambda_\xi Y_{\ell_1,\ell}(\xi) = 0, \, \forall \ell_1 \in F, \ell \in \Lambda_{|\ell_1|} \implies \lambda_\xi = 0, \ \ \xi \in \Xi. \tag{15} \] Proposition 2 says if \(F = \mathbb Z\), then \(K(\xi,\zeta)\) is strictly positive definite. Lemma 9 says if \(K(\xi,\zeta)\) is strictly positive definite, then \(c_{\ell_1} \neq 0\) for infinitely many \(\ell_1\) and Lemma 10 says \(c_0 \neq 0\). Lemma 11 says that if \(K(\xi,\zeta)\) is convolutional and strictly positive definite, then \(K : H_j \to H_j\) are nonzero for infinitely many even \(j\) and infinitely many odd \(j\). Lemma 13 says that if \(K(\xi,\zeta)\) is convolutional and positive definite, then it is strictly positive definite if \[ \sum_{\xi \in \Xi} \lambda_\xi Y_j^k(\xi) = 0, \, \forall j \in F, k = 1,\dots,N_{j,d} \implies \lambda_\xi = 0, \ \ \xi \in \Xi.\tag{16} \] In our opinion the most powerful result in the paper is Theorem 4 which says if \(d > 2\) then condition Eqn 16 holds whenever \(F\) contains infinite many even and infinitely many odd values of \(j\). Its proof uses the addition formula Eqn 8 in Lemma 13 and special properties of Legendre polynomials. This shows that strict positive definiteness of \(K(\xi,\zeta)\) does not imply strict positive definiteness of \(K\) and that strict positive definiteness of \(K\) implies strict positive definiteness of \(K(\xi,\zeta)\) for convolutional kernels. Section 5 discusses convolution kernels on the circle which are the same as axially symmetric kernels. Theorem 5 says they are strictly positive definite whenever \(F\) intersects every arithmetic progression in \(\mathbb N\). Theorem 6 says that if they are strictly positive definite then the set of \(j\) for which \(K : H_j \rightarrow H_j\) is nonzero must intersect every arithmetic progression in \(\mathbb N\). Section 6 briefly comments on numerical issues. The paper is well referenced and the authors describe how their results extend previous results.