Abstract
Axial symmetry for covariance functions defined over spheres has been a very popular assumption for climate, atmospheric, and environmental modeling. For Gaussian random fields defined over spheres embedded in a three-dimensional Euclidean space, maximum likelihood estimation techiques as well kriging interpolation rely on the inverse of the covariance matrix. For any collection of points where data are observed, the covariance matrix is determined through the realizations of the covariance function associated with the underlying Gaussian random field. If the covariance function is not strictly positive definite, then the associated covariance matrix might be singular. We provide conditions for strict positive definiteness of any axially symmetric covariance function. Furthermore, we find conditions for reducibility of an axially symmetric covariance function into a geodesically isotropic covariance. Finally, we provide conditions that legitimate Fourier inversion in the series expansion associated with an axially symmetric covariance function.
Similar content being viewed by others
References
Abramowitz M, Stegun IA (1964) Handbook of mathematical functions: with formulas, graphs, and mathematical tables, vol 55. Courier Corporation, North Chelmsford
Alegria A, Cuevas F, Diggle P, Porcu E (2018) A family of covariance functions for random fields on spheres. CSGB Research Reports, Department of Mathematics, Aarhus University
Barbosa VS, Menegatto VA (2017) Strict positive definiteness on products of compact two-point homogeneous spaces. Integral Transforms Spec Funct 28(1):56–73
Beatson RK, zu Castell W (2017) Dimension hopping and families of strictly positive definite zonal basis functions on spheres. J Approx Theory 221:22–37
Berg C, Christensen JPR, Ressel P (1984) Harmonic analysis on semigroups: Theory of positive definite and related functions, vol 100. Graduate texts in mathematics. Springer, New York
Castruccio S, Stein ML (2013) Global space-time models for climate ensembles. Ann Appl Stat 7(3):1593–1611
Chen D, Menegatto VA, Sun X (2003) A necessary and sufficient condition for strictly positive definite functions on spheres. Proc Am Math Soc 131:2733–2740
Clarke J, Alegria A, Porcu E (2018) Regularity properties and simulations of gaussian random fields on the sphere cross time. Electron J Stat 1:399–426
Daley DJ, Porcu E (2013) Dimension walks and schoenberg spectral measures. Proc Am Math Soc 141:1813–1824
De Iaco S, Posa D (2018) Strict positive definiteness in geostatistics. Stoch Environ Res Risk Assess 32:577–590
Emery X, Porcu E, Bissiri PG (2019) A semiparametric class of axially symmetric random fields on the sphere. Stoch Environ Res Risk Assess 33:1863–1874
Gneiting T (2013) Strictly and non-strictly positive definite functions on spheres. Bernoulli 19(4):1327–1349
Guella JC, Menegatto VA, Peron AP (2016a) An extension of a theorem of schoenberg to a product of spheres. Banach J Math Anal 10(4):671–685
Guella JC, Menegatto VA, Peron AP (2016b) Strictly positive definite kernels on a product of spheres II. In: SIGMA, vol 12(103)
Guella JC, Menegatto VA, Peron AP (2017) Strictly positive definite kernels on a product of circles. Positivity 21(1):329–342
Hitczenko M, Stein ML (2012) Some theory for anisotropic processes on the sphere. Stat Methodol 9:211–227
Huang C, Zhang H, Robeson S (2012) A simplified representation of the covariance structure of axially symmetric processes on the sphere. Stat Probab Lett 82:1346–1351
Jones RH (1963) Stochastic processes on a sphere. Ann Math Stat 34:213–218
Jun M, Stein ML (2007) An approach to producing space-time covariance functions on spheres. Technometrics 49:468–479
Jun M, Stein ML (2008) Nonstationary covariance models for global data. Ann Appl Stat 2(4):1271–1289
Lang A, Schwab C (2013) Isotropic random fields on the sphere: regularity, fast simulation and stochastic partial differential equations. Ann Appl Prob 25:3047–3094
Marinucci D, Peccati G (2011) Random fields on the sphere, representation, limit theorems and cosmological applications. Cambridge, New York
Menegatto VA (1994) Strictly positive definite kernels on the hilbert sphere. Appl Anal 55:91–101
Menegatto VA (1995) Strictly positive definite kernels on the circle. Rocky Mt J Math 25:1149–1163
Menegatto VA, Oliveira CP, Peron AP (2006) Strictly positive definite kernels on subsets of the complex plane. Comput Math Appl 51:1233–1250
Myers D (1992) Kriging, cokriging, radial basis functions and the role of positive definiteness. Comput Math Appl 24:139–148
Porcu E, Alegría A, Furrer R (2018) Modeling temporally evolving and spatially globally dependent data. Int Stat Rev 86:344–377
Porcu E, Castruccio S, Alegría A, Crippa P (2019) Axially symmetric models for global data: a journey between geostatistics and stochastic generators. Environmetrics 30:1327–1349
Schoenberg IJ (1942) Positive definite functions on spheres. Duke Math J 9:96–108
Stein ML (1999) Statistical interpolation of spatial data: some theory for kriging. Springer, New York
Stein ML (2007) Spatial variation of total column ozone on a global scale. Ann Appl Stat 1:191–210
Xu Y, Cheney EW (1992) Strictly positive definite functions on spheres. Proc Am Math Soc 116:977–981
Acknowledgements
Emilio Porcu has been partially supported by CONICYT/FONDECYT/REGULAR/No. 1170290. Ana Paula Peron has been partially supported by FAPESP # 2016/09906-0 and CNPq # 203033/2019-1.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: Mathematical Proofs and Lemmas
Appendix: Mathematical Proofs and Lemmas
1.1 Proof of Proposition 1
Proof
We start by making use of the assumption \(c_m(n,n')=c(n)\delta _{nn'}\) to write
The two sums in (23) can be interchanged provided
To show it, we note that Cauchy-Schwartz inequality in concert with (12) show that, for every \(n\ge 0\),
The fact that \(\textstyle \sum _{n=0}^\infty (n+1/2)c(n)<\infty\) proves that (24) holds true, so that the interchange is legitimate. Therefore, we can swap the two sums in (23) and write:
By (10),
so that:
and therefore (26) yields:
which by (27) and recalling that the cosine is an even function, yields:
which by the spherical harmonics addition theorem (3) yields
where \(\varvec{s},\varvec{s}^{\prime }\) belong to \({\mathbb {S}}^2\), so that \(\varvec{s}=(L,\ell )\) and \(\varvec{s}^{\prime }=(L^{\prime }, \ell ^{\prime })\). \(\square\)
1.2 Auxiliary Lemmas
To provide a constructive proof of Theorem 2, two auxiliary results are required.
Lemma 4
Fix\(m\in {\mathbb {Z}}\)and let\(w_1,\ldots ,w_h\)behdistinct points in\((-1,1)\)such that
for every\(n\in \{\left| {m}\right| ,\left| {m}\right| +1, \ldots \}\)and for some complexh-ple\((z_{1}, \ldots , z_{h})\). Then, \(z_{r}=0\)for every\(r=1,\ldots ,h\).
Proof
The identity (29) implies that
for every \(n\in \{\left| {m}\right| ,\left| {m}\right| +1, \ldots \}\).
At this stage, recall that
where \(k_{m,n}\) is a non zero constant, for each \(n\in \{\left| {m}\right| ,\left| {m}\right| +1,\ldots \}\) and \(x\in (-1,1)\).
Consider now a polynomial p of degree at most \(h-1\) such that
for \(r=1,\ldots ,h\), and let q be a polynomial of degree at most \(h+\left| {m}\right| -1\) such that the \(\left| {m}\right|\)-th partial derivative of q is p. Since the ordinary Legendre polynomials \(P_0,\ldots , P_{\left| {m}\right| +h-2}\) are known to generate the \((\left| {m}\right| +h-1)\)-dimensional polynomial space, there exist \(d_0,\ldots ,d_{\left| {m}\right| +h-2}\in {\mathbb {R}}\) such that
which taking the \(\left| {m}\right|\)–th derivative of both sides and multiplying by \((1-x^2)^{\left| {m}\right| /2}\), yields that:
for every \(x\in (-1,1)\). Hence, we can consider the following function
knowing that by (32) and (33) we have that
as \(r=1,\ldots ,h\). By (30), we have that
which by (34) yields that
We have just proved that (30) implies (35). Similarly, we can prove that (31) implies
and the proof is complete. \(\square\)
Lemma 5
LetKbe an axially symmetric covariance function as in (17) such that for each\(m\ge 0\), \(c_m:\{m,m+1,\ldots \}\times \{m,m+1,\ldots \}\rightarrow {\mathbb {C}}\)is a strictly positive definite function. Then, for\(a=(a_1,\ldots ,a_k)\in {\mathbb {R}}^k\)and\(\varvec{s}_1,\ldots ,\varvec{s}_k\)distinct points in\({\mathbb {S}}^2\)such that the longitude and the latitude of\(\varvec{s}_j\)are\(\ell _j\in (-\pi ,\pi ]\)and\(L_j\in [-\pi /2,\pi /2]\), respectively (\(j=1,\ldots ,k\)), the following assertions are equivalent:
- (i)
\(a^TKa=0\), that is,
$$\begin{aligned} \sum _{j,h=1}^k a_jK(L_j,L_h, \Delta \ell _{jh})a_h=0, \qquad (\Delta \ell _{jh} := \ell _j -\ell _h). \end{aligned}$$(36) - (ii)
The equality
$$\begin{aligned} \sum _{j=1}^k a_j e^{ {\mathsf {i}} m \ell _j}\bar{P}^m_n(\sin L_j )=0. \end{aligned}$$(37)holds for each\(m\in {\mathbb {Z}}\)and each\(n\in \{\left| {m}\right| ,\left| {m}\right| +1,\ldots \}\),
Proof
Note that
Hence, it is clear that (ii) implies (i). Now suppose that (i) holds. Combining (36) and (38), we have that:
where
Since \(c_m\) is positive definite (for every m), then
for every N and every m. Thus, letting N diverge to infinity we have that
for every m. Therefore, by (39), we have that for every \(m\in {\mathbb {Z}}\),
At this stage, for each \(m,N\in {\mathbb {Z}}\) with \(m,N> 0\), let \(C_m(N)\) be the \(N\times N\) matrix whose \((n-m+1,n'-m+1)\) entry is \(c_m(n,n')\) (\(n,n'= m,\ldots , N+m-1\)).
For each \(m,N\in {\mathbb {Z}}\) with \(N, m>0\), the matrix \(C_m(N)\) must be positive definite since by assumption, for each \(m\in {\mathbb {Z}}\), \(c_m(n,n')\) is strictly positive definite. Hence, by Cholesky decomposition, there is a complex \(N\times N\) lower triangular matrix \(A_m(N)\) such that \(C_m(N)=A_m(N) A_m(N)^*\) the diagonal entries of \(A_m(N)\) are real and positive and the matrix \(A_0(N)\) is real. Denoting by \(a^{(m,N)}_{nr}\) the \((n-m+1,r-m+1)\) entry of the matrix \(A_m(N)\) if \(N,m> 0\), the Cholesky decomposition of the matrix \(C_m(N)\) can be written in the following form:
for \(n,n'=m, \ldots , N+m-1\), where \(x\wedge y\) denotes the minimum among x and y. If \(M \ge N\), then by the Cholesky decomposition of the matrix \(C_m(M)\) we have that:
for \(n,n'=m, \ldots , N+m-1\). Note the Cholesky decomposition of \(C_m(N)\) is unique being \(C_m(N)\) positive definite and therefore \(a^{(m,N)}_{nr}=a^{(m,M)}_{nr}\) if \(m\le r \le n \le N+m-1 \le M+m-1\). In particular, we have that \(a^{(m,n-m+1)}_{nr}=a^{(m,M)}_{nr}\) for any quartet (m, r, n, M) where \(m\le r \le n \le M-m+1\). Therefore, we can write \(a^{(m)}_{nr}\) for \(a^{(m,n-m+1)}_{nr}\) and
where for each r, \(a^{(m)}_{rr}>0\). Note that (42) holds also for \(m<0\) if we let \(a^{(-m)}_{nr}=a^{(m)*}_{nr}\). Substituting (42) in (41) we obtain that, for every \(m\in {\mathbb {Z}}\),
that is equivalent to:
which in turn is equivalent to:
which implies that for every \(r\ge \left| {m}\right|\)
Thus, for every fixed \(m\in {\mathbb {Z}}\) and every fixed \(r\ge \left| {m}\right|\) we can write:
for \(s=0,1,2,\ldots\) Recalling that \(a^{(m)}_{r+s,r+s}>0\) for each s, we can define a sequence \((d_s)_{s=0}^\infty\) according to the recursion:
This ensures that:
for \(t=1,2,\ldots\) and therefore by (44),
We have just proved that \(b_{r,m} a^{(m)}_{r,r}=0\) for any \(m\in {\mathbb {Z}}\) and any \(r\ge \left| {m}\right|\). At this stage, recall once again that \(a^{(m)}_{r,r}>0\) for any (m, r). Therefore, for any \(m\in {\mathbb {Z}}\) and any \(r\ge \left| {m}\right|\), \(b_{r,m}=0\), namely by (40)
for each \(m\in {\mathbb {Z}}\) and each \(n\in \{\left| {m}\right| ,\left| {m}\right| +1,\ldots \}\), proving (ii) . \(\square\)
Now we are ready to prove our main result.
1.3 Proof of Theorem 2
Proof
We aim at proving that if \((a_1,\ldots ,a_k)\in {\mathbb {R}}^k\), and \(\varvec{s}_1,\ldots ,\varvec{s}_k\) are k distinct points in \({\mathbb {S}}^2\) such that the longitude and the latitude of \(\varvec{s}_j\) are \(\ell _j\in (-\pi ,\pi ]\) and \(L_j\in [-\pi /2,\pi /2]\), respectively (\(j=1,\ldots ,k\)), and
then \(a_j=0\) for \(j=1,\ldots ,k\).
By Lemma 5, letting \(m=0\), we obtain:
for every \(n\in \{0,1,2,\ldots \}\), where \(\bar{P}_n\) denotes the ordinary normalized Legendre polynomial of degree n.
By denoting \(L_{(1)},\ldots , L_{(h)}\) the distinct elements among \(L_1,\ldots ,L_k\) and letting \(a_{(r)}=\sum _{j\in B_r} a_j\) where \(B_r=\{j=1,\ldots , k: L_j=L_{(r)}\}\), as \(r=1,\ldots , h\), the equation (47) becomes:
for every \(n\in \{0,1,2,\ldots \}\).
At this stage, consider the following function:
where E is a finite or infinite subset of \(\{0,1,2,\ldots \}\), so that by (48),
There exist \(\{d_{n}\in {\mathbb {R}}: n\in E \}\) for some set E such that \(p(L_{(j)})=a_{(j)}\) for \(j=1,\ldots ,h\). Indeed, we can just take \(E=\{0,1,\ldots ,h-1\}\) recalling that \(\bar{P}_0\),..., \(\bar{P}_{h-1}\) generate the h–dimensional polynomial space and that interpolation of arbitrary data at h nodes is possible.
Hence, by (49), \(\textstyle \sum _{j=1}^h a_{(j)}^2=0\) and therefore \(a_{(j)}=0\), namely
Now note that if \(L_j\in \{-\pi /2,\pi /2\}\) then \(\varvec{s}_j\) is a pole and therefore \(L_i\ne L_j\) for any \(i\ne j\) since \(\varvec{s}_1,\ldots , \varvec{s}_n\) are distinct points (i.e., \(B_r=\{j\}\) if \(L_j\in \{-\pi /2,\pi /2\}\)). Hence, by (50), we have just proved that if \(L_j\in \{-\pi /2,\pi /2\}\) then \(a_j=0\).
Hence , the equation (45) becomes:
for each \(m\in {\mathbb {Z}}\) and each \(n\in \{\left| {m}\right| ,\left| {m}\right| +1,\ldots \}\). Since \(\overline{P}_n^m(\pm 1) =0\), (51) becomes:
where \(A=\{r=1,\ldots ,h: L_{r}\notin \{-\pi /2,\pi /2\}\}\), for each \(m\in {\mathbb {Z}}\) and each \(n\in \{\left| {m}\right| ,\left| {m}\right| +1,\ldots \}\). Moreover, since we have just proved that \(a_j=0\) if \(L_j=\pm \pi /2\), then we need to show that \(a_j=0\) for all \(j\in B_r\) and for all \(r\in A\).
By Lemma 4, (52) implies that:
for every \(m\in {\mathbb {Z}}\) and \(r\in A\). At this stage, observe that \(\ell _i \ne \ell _j\) when \(i,j\in B_r\) because in this case \(L_i=L_j\) and \(\varvec{s}_i\ne \varvec{s}_j\).
Moreover, the matrix of the system (53) is Vandermonde–like associated to the distinct points \(e^{{\mathsf {i}} \ell _j}\), \(j\in B_r\) for each \(r\in A\). Therefore, (53) implies that \(a_j=0\) for each \(j=1,\ldots ,k\) such that \(L_j\notin \{-\pi /2,\pi /2\}\) and the proof is complete. \(\square\)
1.4 Proof of Theorem 3
Proof
In order to evaluate the integral in the right hand side of (19), we can plug in (17) and swap the double series and the triple integral if the following is verified:
where:
for every \(m_0\in {\mathbb {Z}}\) and \(n_0,n_0'\ge \left| {m_0}\right|\). Trivially,
At this stage, note that by (12),
for every \(m,n\in {\mathbb {Z}}\) with \(n\ge \left| {m}\right|\), and therefore (55) yields:
which, by the Cauchy–Schwartz inequality and (12), in turn yields:
which by (18) implies that (54) is verified. It is therefore possible to plug (17) into the integral in (19) and swap the double sum with the triple integral. In this way, due to the orthogonality of the associated Legendre polynomials \(\{\bar{P}^m_{n}(\sin L): n\in {\mathbb {N}}\}\) given by (6) and of \(\{e^{im \ell }: m\in {\mathbb {Z}}\}\), and due to (7), (19) is obtained. \(\square\)
Rights and permissions
About this article
Cite this article
Bissiri, P.G., Peron, A.P. & Porcu, E. Strict positive definiteness under axial symmetry on the sphere. Stoch Environ Res Risk Assess 34, 723–732 (2020). https://doi.org/10.1007/s00477-020-01796-y
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00477-020-01796-y