Abstract
We propose a novel basis of vector functions, the mixed vector spherical harmonics that are closely related to the functions of Sheppard and Török and help us reduce the concentration problem of tangential vector fields within a spherical cap to an equivalent scalar problem. Exploiting an analogy with previous results published by Grünbaum, Longhi and Perlstadt, we construct a differential operator that commutes with the concentration operator of this scalar problem and propose a stable and convenient method to obtain its eigenfunctions. Having obtained the scalar eigenfunctions, the calculation of tangential vector Slepian functions is straightforward.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
In our mathematical models, we often assume bandlimitedness for physical or computational reasons, yet also wish that our solutions be localized, with respect to their energy, inside a finite spatial region. Since these are mutually exclusive conditions [31], we need to resort to bandlimited functions with an optimal spatial localization. The goal of the so-called spatial concentration problem [22, p. 75] is to find such functions, and since its first thorough investigation by Slepian, Landau and Pollak in one and multiple Cartesian dimensions [14, 15, 30, 32], it has been revisited many times, including solutions for spherical and planar regions of arbitrary shape [1, 10, 26, 28, 29].
Each individual concentration problem gives rise to an orthogonal set of well-localized functions, which now we refer to under the common name Slepian functions. They have enjoyed increasing popularity in applications involving signal processing, function representation and approximation or the solution of inverse problems. In particular, the scalar spherical Slepian functions have been utilized, for instance, in geodesy and geophysics [1, 11, 26–29], cosmology [5, 6], computer science [16] and mathematics [19].
While they have been widely applied in the last two decades, it was not until recently that the theory of vector Slepian functions began to mature. The first successful construction of bandlimited vector fields, localized to a spherical cap, was reported in the context of biomedical science [18], followed by an application in physical optics [13]. Recently, a more general treatment of the vector spherical concentration problem has been published [23, 24], however, the question on the existence of a commuting differential operator for the spherical cap has been left unresolved. This question is important for the following reason. Slepian functions of a particular problem are eigenfunctions of the so-called concentration operator associated with the spatial region of interest, an integral operator exhibiting a peculiar step-like eigenvalue spectrum. This property makes the direct calculation of its eigenfunctions numerically difficult [4].
A particularly important result of Slepian and his colleagues was the introduction of a Sturm–Liouville differential operator that commutes with the corresponding concentration operator, hence they share a common set of eigenfunctions. Since the differential operator has a simple spectrum with more evenly distributed eigenvalues, it allows a more stable and accurate numerical computation of the eigenfunctions [4]. Two decades after Slepian and his colleagues’ seminal papers, Grünbaum, Longhi and Perlstadt found such a commuting differential operator for the scalar concentration problem within a spherical cap as well [10]. However, we are not aware of a similar proposal for the vectorial problem.
In this paper, we construct a differential operator commuting with the concentration operator of the vector case. This constitutes our main result. We restrict ourselves to tangential vector fields, since the concentration problem of the radial component is equivalent to the scalar concentration problem on the sphere [24], which has already been studied extensively [26, 28]. The key functions in our investigations are the novel mixed vector spherical harmonics \(\varvec{\mathbf {Q}}_{lm}^{\pm }(\theta , \phi )\) which enable us to reduce the vectorial problem to a scalar one involving the special functions \(F_{lm}\) of Sheppard and Török [25]. After that, the problem can be solved in an analogous way to its scalar counterpart [26, 28].
We note that we only consider spatially localized, bandlimited fields here. The symmetric problem of spectrally concentrated, spacelimited functions can be derived by exploiting the analogy to previously published results [24, 28].
2 Preliminaries: Associated Special Functions
In preparation for the concentration problem, we give a detailed survey on the essential properties of important special functions used in our investigations. We start by proving several fundamental relations for the functions \(F_{lm}\) of Sheppard and Török [25], which are needed for subsequent proofs. Next we introduce the mixed vector spherical harmonics and discuss their orthogonality properties. Finally we show how to expand an arbitrary tangential vector field in terms of the mixed vector spherical harmonics and define bandlimitedness in this context, so that we can use these new vector fields as basis functions in the treatment of the concentration problem in Sect. 3.
2.1 Special Functions \(F_{lm}\) of Sheppard and Török
Following Sheppard and Török [25], first we define the functions
where \(U_{lm}\) are the normalized associated Legendre functions of integer degree \(l\) and order \(m\). The classical (unnormalized) associated Legendre functions \(P_l^m\) are thus normalized as [3, p. 757]
where
We can express the conditions for the integer indices \(l\) and \(m\) in definition (1) alternatively as \(-\infty < m < \infty \) and \(l \ge \ell _{m}\), where the minimal degree \(\ell _{m}\) for fixed \(m\) is
The functions \(F_{lm}\) are orthonormal with respect to the usual \(L^2\)-inner product for fixed \(m\) [25], i.e.
This normalization differs by a constant factor from that used by Sheppard and Török [25].
Sometimes it is inconvenient that expression (1) is singular at \(x=\pm 1\) because of the factor \((1-x^2)^{-1/2}\). However, a singularity-free form can also be obtained by using two lesser-known recurrence relations [8, 17, 24]:
where
and
where
We note that everywhere in this paper, where the \(\pm \) or \(\mp \) signs occur, either the upper or the lower one has to be used consistently in the whole expression. The singularity-free form then reads
It is straightforward to show using the symmetry relation [3, p. 743]
that for the special case \(m=0\), the equivalence \(F_{l,\,0}(x) = -U_{l,\,1}(x)\) holds. Moreover, by using symmetry relation (9) and parity relation [3, p. 746]
we can obtain a symmetry relation for \(F_{lm}\), too:
This symmetry is apparent in Fig. 1 where a small subset of \(F_{lm}\) (\(l \le 3\)) is depicted.
We have found the following recurrence relations for \(F_{lm}\), which simplify subsequent proofs:
where
A proof for (12) and (13) is provided in Appendix 1a and 1b, respectively. Relation (14) is straightforward to prove by combining (12) and (13).
Relation (12) can be used to derive a Christoffel–Darboux formula [34, p. 42] specialized for \(F_{lm}\):
See Appendix 1c for the details. Moreover, we can formulate an addition theorem for \(F_{lm}(x)\) as well that will be used later in Sect. 3.2:
The proof can be found in Appendix 1d.
At the interval endpoints \(x = \pm 1\), we can explicitly calculate the values of \(F_{lm}\). Considering that \(U_{lm}\) takes the values [3, p. 746]
and using the singularity-free form (8), we get
We can also formulate a closed-form expression for the \(l = l_m\) case. We combine expression [3, p. 745]
with definition (1) of \(F_{lm}\), and thus get
where
and \((2m - 1)!! = (2m - 1)(2m - 3) \cdots (1)\) is the double factorial. This expression together with recurrence relation (12) provides a stable and efficient method to evaluate \(F_{lm}(x)\) numerically. By setting \(F_{\ell _{m}-1,\,m} = 0\) and starting with the closed-form expression (21), recurrence relation (12) can be used repeatedly in the upward direction until one obtains \(F_{lm}(x)\).
Finally, we turn to the differential equation of the functions \(F_{lm}\). They satisfy the Sturm–Liouville equation
as proven in Appendix 1e. The operator on the left-hand side is closely related to the (surface) vector Laplace–Beltrami operator [33] on the unit sphere \(\Omega :=\left\{ (\theta , \phi ): 0 \le \theta \le \pi ,\ 0 \le \phi < 2\pi \right\} \).
Let \(\varvec{\mathbf {w}}(\theta , \phi ) :=\varvec{\mathbf {v}}(\theta ) \exp (\mathrm {i}m \phi )\) be a tangential separable vector field, where \(\mathrm {i}\) is the imaginary unit, \(\varvec{\mathbf {v}}(\theta ) :=v_{\theta }(\theta ) {\hat{\varvec{\mathbf {\theta }}}} + v_{\phi }(\theta ) {\hat{\varvec{\mathbf {\phi }}}}\), and \(\hat{\varvec{\mathbf {\theta }}}\) and \(\hat{\varvec{\mathbf {\phi }}}\) are unit vectors in the \(\theta \)- and \(\phi \)-directions, respectively, as shown in Fig. 2. When considering the action of the vector Laplace–Beltrami operator on \(\varvec{\mathbf {w}}\), we obtain the fixed-order vector Laplace–Beltrami operator
where
is the fixed-order (surface) scalar Laplace–Beltrami operator. We can diagonalize \(\nabla _{\Omega ,\,m}^2\, \varvec{\mathbf {v}}\) by introducing the tangential basis vectors
which are orthogonal with respect to the complex dot product
where the asterisk denotes the complex conjugate.
Since \(\nabla _{\Omega ,\,m}^2 = \nabla _{\Omega ,\,-m}^2\), in this new basis, we can write (24) as
where \(\varvec{\mathbf {v}} = v_{+}(\theta ,\, \phi ) \hat{\varvec{\mathbf {\tau }}}_{+} + v_{-}(\theta , \phi ) \hat{\varvec{\mathbf {\tau }}}_{-}\) and
Upon substituting \(x = \cos \theta \) in \(\Delta _{\Omega ,\,m}\), we regain the differential operator on the left-hand side of differential equation (23). Thus the functions \(F_{lm}\) are eigenfunctions of \(\Delta _{\Omega ,\,m}\):
2.2 Mixed Vector Spherical Harmonics
It follows from the diagonal form (28) of the fixed-order vector Laplace–Beltrami operator and Eq. (30) that a new kind of vector spherical harmonics can be defined,
We call these vector fields mixed vector spherical harmonics, because their connection to conventional vector spherical harmonics can be expressed as
The tangential vector spherical harmonics \(\varvec{\mathbf {Y}}_{lm}\) and \(\varvec{\mathbf {Z}}_{lm}\) are defined as [20]
where
are the scalar spherical harmonics. We note that a comprehensive discussion of scalar and vector spherical harmonics, although with notations different from ours, can be found in Ref. [9]. Expressions (32) can be verified by substituting definitions (33) and (34) and using definition (1) of \(F_{lm}\) and definition (26) of the unit vectors \(\hat{\varvec{\mathbf {\tau }}}_{\pm }\).
Definition (31) and the orthogonality of \(\hat{\varvec{\mathbf {\tau }}}_{\pm }\) implies that, unlike \(\varvec{\mathbf {Y}}_{lm}\) and \(\varvec{\mathbf {Z}}_{lm}\), the functions \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) exhibit local (vector) orthogonality, regardless of their degree and order, i.e.
This equation together with definition (31) can be used to prove the orthonormality relations
where \(\int _{\Omega } \dots \mathrm{d }\!\Omega :=\int _{0}^{2\pi } \int _{0}^{\pi } \dots \sin \theta \mathrm{d }\!\theta \mathrm{d }\!\phi \).
The values of \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) at the \(\theta \)-coordinate singularities deserve extra attention, since \(\hat{\varvec{\mathbf {\tau }}}_{\pm }\) are not well defined there. We can circumvent this problem by expressing \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) using Cartesian basis vectors (see Fig. 2). Using Eq. (19) for \(F_{lm}(\pm 1)\), we have
Like \(\varvec{\mathbf {Y}}_{lm}\) and \(\varvec{\mathbf {Z}}_{lm}\) [7], the mixed vector spherical harmonics \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) also form a complete basis of the Hilbert space of square-integrable tangential vector fields defined over \(\Omega \). Hence we can expand an arbitrary tangential vector field \(\varvec{\mathbf {v}}\) in terms of \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) as
where the expansion coefficients \(v_{lm}^{\pm }\) can be calculated as
If \(v_{lm}^{\pm } = 0\) for \(L < l < \infty \) and some \(L > 0\), we call \(\varvec{\mathbf {v}}\) bandlimited. The limit \(L\) is the maximal degree of functions \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) that contribute to the expansion (38). Therefore the subspace \(S_{L}\) of bandlimited vector fields is finite dimensional, its dimension is equal to the number of terms in (38):
3 The Concentration Problem of Tangential Vector Fields within an Axisymmetrical Spherical Cap
Having defined all important special functions, we now turn our attention to the main topic of the paper. After a brief discussion of the concentration problem in terms of the mixed vector spherical harmonics, we introduce a reduced scalar problem which can be solved analogously to the theory of scalar spherical Slepian functions [10, 26, 28].
In Sect. 3.2, we analyze the eigenvalue spectrum of the concentration operator and give an illustration on the scalar eigenfunctions. Finally, we propose a fast and numerically stable way to calculate the eigenfunctions by using a differential operator that commutes with the scalar concentration operator obtained previously.
3.1 Formulation of the Vector Problem and its Reduction to a Scalar One
We consider the variational problem of finding a bandlimited, tangential vector field \(\varvec{\mathbf {G}}(\theta , \phi ) \in S_{L}\) that maximizes the fractional energy contained within an axisymmetric spherical cap \(C\):
Without loss of generality, we can center our spherical cap at \(\theta =0\), as seen in Fig. 3:
where \(\Theta > 0\) is assumed.
To solve the vectorial problem (41), we adapt the method of the scalar case [28] and turn (41) into a Rayleigh–Ritz matrix variational problem [12, p. 176]. We can achieve this by expanding \(\varvec{\mathbf {G}}\) in terms of \(\varvec{\mathbf {Q}}_{lm}^{\pm }\):
In the second step, we have interchanged the order of double summation to facilitate the transition to the matrix formulation of (41) later. A visual comparison of the two summation schemes is given in Fig. 4.
Next we insert (43) into (41), interchange the order of summation and integration and use orthonormality relations (35) and (36). Hence the numerator can be written as
while the denominator becomes
The integrals on the right-hand side of (44) can be expressed as
Since \((2\pi )^{-1} \int _{0}^{2\pi } \exp [\mathrm {i}(m' - m) \phi ] \mathrm{d }\!\phi = \delta _{mm'}\), Eq. (46) further simplifies to
where
The integrand of \(K_{m,\,ll'}\) is a polynomial of degree \(l + l'\), hence it can exactly be integrated numerically, for instance, by a Gauss–Legendre formula of \(\lceil (l + l' + 1)/2 \rceil \) nodes.
Taking (47) into account, (41) can be rewritten as
To express (49) in matrix formalism, we construct a column vector \(\mathsf {g}\) of \(2L(L + 2)\) elements as
where the expansion coefficients \(g_{lm}^{\pm }\) are enumerated according to the scheme of Fig. 4b. In addition, we introduce the \(2L(L+2) \times 2L(L+2)\) block-diagonal matrix
where the \(L(L+2) \times L(L+2)\) blocks \(\mathsf {K}^{\pm }\) are themselves block-diagonal:
The elementary blocks
correspond to different orders \(m\) and have an order-dependent size of \((L - \ell _{m} + 1) \times (L - \ell _{m} + 1)\). Note that \(\mathsf {K}^{+}\) contains the same blocks \(\mathsf {K}_{m}\) as \(\mathsf {K}^{-}\), however, the order of the blocks is reversed owing to (47).
We can thus use these constructions to transform problem (49) into the Rayleigh–Ritz matrix variational problem
where the dagger sign denotes the conjugate transpose. Equivalently, we have to find the eigenvector \(\mathsf {g}\) of the eigenvalue problem [12, p. 176]
with the maximal eigenvalue \(\eta \). However, rather than solving the large \(2L(L + 2) \times 2L(L + 2)\) eigenvalue problem (55), the block-diagonal structure of \(\mathsf {K}\) allows us to solve a series of smaller \((L - \ell _{m} + 1) \times (L - \ell _{m} + 1)\) problems instead,
one for each order \(m\).
From (48) follows that \(K_{m,\,ll'} = K_{m,\,l'l}\), implying that the matrices \(\mathsf {K}_{m}\) are symmetric. Hence their eigenvalues are always real. For a given \(m\), we rank-order the \((L - \ell _{m} + 1)\) distinct eigenvalues \(\eta _{mn}\) as \(1 > \eta _{m,\,1} > \eta _{m,\,2} > \cdots > \eta _{m,\,L - \ell _{m} + 1} > 0\). The associated eigenvectors \(\mathsf {g}_{mn}\) can be chosen to be real and orthonormal:
Here we have distinguished between the different eigenvalues and the corresponding eigenvectors by the use of the additional index \(n\) (or \(n'\)). However, we drop this additional index for brevity when we refer to any of the \((L - \ell _{m} + 1)\) eigenvalues or eigenvectors.
Let us denote the elements of an eigenvector \(\mathsf {g}_{m}\) simply by \(g_{lm}\). The definitions (51), (52) and (53) imply that eigenfunctions contain either \(\hat{\varvec{\mathbf {\tau }}}_{+}\) or \(\hat{\varvec{\mathbf {\tau }}}_{-}\), they are composed of mixed vector spherical harmonics of a single order \(m\) only and \(g_{lm}^{\pm } = g_{l,\,\pm m}\). Hence we can considerably simplify our general expansion (43) and write instead
Upon substituting definition (31) of \(\varvec{\mathbf {Q}}_{lm}^{\pm }\) into Eq. (58), we obtain
where
are real functions.
In this way, we managed to reduce the vectorial concentration problem within a spherical cap to equivalent one-dimensional, scalar concentration problems of various orders \(m\). The key idea in this simplification was the choice (31) for our basis functions. The scalar concentration problem for a fixed order \(m\) can be formulated as
where \(G_{m}\) is a bandlimited scalar function belonging to the subspace spanned by \(F_{lm}\). The corresponding Rayleigh–Ritz matrix variational problem is
Instead of the eigenvalue Eq. (56) specifying eigenvectors \(\mathsf {g}_{m}\), we can directly formulate an eigenvalue equation in terms of the functions \(G_{m}\), too. Therefore we first express Eq. (56) component-wise as
Now we multiply both sides by \(F_{lm}(x)\) and sum over \(l\):
The left-hand side can be rewritten as
This way, we obtain a Fredholm integral equation of the second kind for \(G_{m}\),
where the kernel function \(\mathcal {K}_{m}\) is defined as
It follows from the orthogonality relations (57) of the eigenvectors that the scalar eigenfunctions \(G_{m}\) are doubly orthogonal:
The vectorial eigenfunctions \(\varvec{\mathbf {G}}_{m}^{\pm }\) inherit this property as well:
3.2 The Eigenvalue Spectrum and its Peculiarity
The eigenvalue spectrum of Slepian-type concentration problems [29, 30, 32] exhibits a characteristic step-like shape, and the present case is no exception. Figure 5 shows rank-ordered spectra including \(\eta _{mn}\) for all orders \(m\). They correspond to \(\Theta = 30^\circ , 60^\circ , 90^\circ \) and the maximal degree was chosen \(L=18\).
The majority of the eigenvalues for each case is either close to one or zero, corresponding to well-concentrated and poorly concentrated eigenfunctions, respectively. As an illustration, in Fig. 6 we have plotted a small number of scalar eigenfunctions \(G_{mn}\), corresponding to different parts of the eigenvalue spectrum.
Strictly speaking, the solution of the concentration problem (41) is the pair of vectorial eigenfunctions which corresponds to the maximally concentrated \(G_{m}\). However, having solved the equivalent eigenvalue problem (55), we have gained a whole set of well-concentrated, orthogonal pairs of eigenfunctions \(\varvec{\mathbf {G}}_{m}^{\pm }\). How many pairs belong to this set? To answer this question, we first define the partial Shannon number [28]
which gives the approximate number of reasonably well-concentrated (\(\eta \ge 0.5\)) scalar eigenfunctions for a given maximal degree \(L\) and order \(m\). Summing over all possible values of \(m\), we obtain the (total) Shannon number
where \(A_{C} = 2\pi (1 - \cos \Theta )\) is the area of the spherical cap \(C\). In the last equality we substituted definition (66), interchanged the order of double summation and used addition theorem (17).
Hence there are \(N\) pairs of orthogonal vectorial eigenfunctions which are suitable for approximating bandlimited, tangential vector fields localized to \(C\). Equivalently, the use of this basis reduces the number of degrees of freedom from \(\dim S_{L} = 2L(L+2)\) to \(2N\).
3.3 Toward an Efficient Numerical Solution: The Commuting Differential Operator and its Eigenvalue Problem
In Sect. 3.1, we obtained the expansion coefficients \(g_{lm}\) by solving eigenvalue Eq. (56) directly. However, while it is theoretically possible to calculate \(g_{lm}\) this way, the accumulation of the eigenvalues \(\eta \) at one and zero, as seen in Fig. 5, makes the numerical solution of (56) ill-conditioned [4]. In order to circumvent this problem, we set out to construct another matrix with a simple spectrum to supply the expansion coefficients \(g_{lm}\).
Therefore, we first return to the Fredholm eigenvalue Eq. (65). We wish to find a Sturm–Liouville differential operator \(\mathcal {J}_{m}\) that commutes with the concentration (integral) operator on the left-hand side of (65):
for any square-integrable bandlimited function \(u\), so that the two operators share a common set of eigenfunctions [3, pp. 314]. It is known from the Sturm–Liouville theory that \(\mathcal {J}_{m}\) has a simple spectrum of distinct eigenvalues with an accumulation point in infinity [21, p. 724]. If such a differential operator \(\mathcal {J}_{m}\) can be found, its matrix representation can be used to obtain the expansion coefficients \(g_{lm}\) (hence the eigenfunctions) in a numerically stable way.
The same approach was taken by Grünbaum et al. for the concentration problem of scalar functions within \(C\) [10]. They proposed the differential operator
where \(\nabla _{\Omega ,\,m}^2\) is the fixed-order scalar Laplace–Beltrami operator (25). This operator commutes with the concentration operator of the scalar case which contains the kernel function \(\mathcal {D}_{m}(x, x') = \sum _{l={|m |}}^{L} U_{lm}(x) U_{lm}(x')\) [28].
Based on (72), we make the following ansatz on \(\mathcal {J}_{m}\):
where \(\Delta _{\Omega ,\,m}\) is the fixed-order operator (29) related to the fixed-order vector Laplace–Beltrami operator (28). Changing the variable to \(x=\cos \theta \) yields
which is equivalent to
To prove that \(\mathcal {J}_{m}\) satisfies the commutation relation (71), we suggest following the concept of Grünbaum et al. [10]. First, one proves the identity
which holds for any two functions \(u_{1}\) and \(u_{2}\) that are non-singular at the interval endpoints (see Appendix 1f for details). Therefore, the left-hand side of the commutation relation (71) can be rewritten as
Finally, one verifies that
The proof of (78), like the proof of (76), closely resembles its counterpart from the scalar concentration problem [28]. The key steps are the same, with the main difference that the associated Legendre functions are replaced by \(F_{lm}\) together with the corresponding identities. The details can be found in Appendix 1g.
Since \(\mathcal {J}_{m}\) commutes with the integral operator of (65), the functions \(G_{m}\) are eigenfunctions of \(\mathcal {J}_{m}\), too:
Figure 7 shows the \(\chi \)-eigenvalue spectrum of all orders \(m\) for \(\Theta = 30^\circ , 60^\circ , 90^\circ \) and \(L=18\) (cf. Fig. 5). Similarly to the scalar concentration problems [28, 30, 32], the rank-ordering for \(\chi _{mn}\) is the opposite of the rank-ordering for \(\eta _{mn}\). Importantly, the \(\chi \)-spectrum does not exhibit an accumulation of eigenvalues.
To obtain a matrix equation similar to the component-wise eigenvalue Eq. (63) of \(\mathsf {K}_{m}\), we substitute expansion (60) of \(G_{lm}\) in terms of \(F_{lm}\) into eigenvalue Eq. (79), but this time, writing \(l'\) instead of \(l\). After that we multiply by \(F_{lm}(x)\), integrate over \(-1 \le x \le 1\), and invoke orthonormality relation (5) of \(F_{lm}\). In this way, we arrive at the equation
where
Similarly to \(\mathsf {K}_{m}\), we can arrange \(J_{m,\,ll'}\) into a matrix \(\mathsf {J}_{m}\):
However, the only non-zero matrix elements, as proven in Appendix 1h, are
hence \(\mathsf {J}_{m}\) is real, symmetric and tridiagonal. The eigenvalue Eq. (80) can thus be written as
We have already seen in Sect. 2.1 that \(F_{l,\,0} = U_{l,\,1}\), hence in the special case of \(m=0\), matrix \(\mathsf {J}_{0}\) is identical to the matrix of the Grünbaum operator \(\mathcal {G}_{1}\) [28].
In summary, to calculate the scalar eigenfunctions \(G_{m}\) for each order \(m\), we first construct the tridiagonal matrices \(\mathsf {J}_{m}\) using formulae (83) and then solve the corresponding eigenvalue problem (84) numerically. The resulting eigenvectors \(\mathsf {g}_{m}\) contain the expansion coefficients \(g_{lm}\), \(\ell _{m} \le l \le L\), which, substituted into expansion (60) give the eigenfunctions \(G_{m}\). The corresponding energy concentration ratio \(\eta _{m}\) can be calculated using either \(\eta _{m} = \int _{\cos \Theta }^{1} [G_{m}(x)]^2 \mathrm{d }\!x\) or \(\eta _{m} = \mathsf {g}_{m}^{\mathrm {T}}\mathsf {K}_{m} \mathsf {g}_{m}\).
Finally, we demonstrate the numerical stability of the proposed method. We calculated the eigenvectors \(\mathsf {g}_{mn}\) for \(m=1\), \(L=18\) and \(\Theta =30^{\circ }, 60^{\circ }, 90^{\circ }\) in multiple ways. First, as a reference, we used arbitrary precision arithmetic to obtain the eigenvectors of \(\mathsf {K}_{1}\) with the relative error of each coefficient \(g_{l,1}\) being less than \(10^{-23}\). Let \(\mathsf {g}_{1,n}^\text {ref}\) denote these vectors. Then we computed both \(\mathsf {K}_{1}\) and \(\mathsf {J}_{1}\) in double precision and fed them into the divide-and-conquer routines of LAPACK [2] to produce the eigenvectors again. Let \(\mathsf {g}_{1,\,n}^\text {K}\) and \(\mathsf {g}_{1,\,n}^\text {J}\) stand for these results, respectively. In addition, we furthermore assume \({||\mathsf {g}_{1,\,n}^\text {ref} ||} = {||\mathsf {g}_{1,\,n}^\text {K} ||} = {||\mathsf {g}_{1,\,n}^\text {J} ||} = 1\) where \({||\mathsf {v} ||} :=\sqrt{\mathsf {v}^{\mathrm {T}}\mathsf {v}}\).
Figure 8a–c plot the eigenvalue gaps [2, p. 104]
for all three values of \(\Theta \), respectively, where \(1 \le n \le L\). Figure 8d–e show the errors
of the eigenvectors, where we have taken their sign ambiguity into account.
In Fig. 8a–c, we clearly see the accumulation of eigenvalues \(\eta _{1,n}\) of \(\mathsf {K}_{1}\) for both small and large values of \(n\). The decrease in the eigenvalue gap by many orders of magnitude is accompanied by a rapid increase in the error \(\delta \mathsf {g}_{1,\,n}^\text {K}\) [2, p. 104], as seen in Fig. 8d–f. Therefore, with a naïve treatment of \(\mathsf {K}_{1}\), we failed to calculate the well-concentrated eigenfunctions accurately; precisely those that play an important role in the approximation of functions localized to \(C\).
On the contrary, Fig. 8a–c demonstrate again that the eigenvalues \(\chi _{1,\,n}\) of \(\mathsf {J}_{1}\) are well separated, hence we can expect the accuracy of eigenvectors \(\mathsf {g}_{1,\,n}^\text {J}\) to stay reasonably close to machine precision. Indeed, the error is below \(120\epsilon _\text {M}\) for all values of \(n\), as indicated by Fig. 8d–f, where \(\epsilon _\text {M} = 2^{-53} \approx 1.11 \times 10^{-16}\) denotes the machine epsilon in double precision [2, p. 79]. Considering the tridiagonal form of \(\mathsf {J}_{m}\) with the simple expressions (83) for the matrix elements, its superiority over \(\mathsf {K}_{m}\) in the calculation of eigenvectors is justified.
4 Concluding Remarks
We have formulated a scalar problem which is equivalent to the concentration problem of tangential vector fields within a spherical cap, and enables us to treat it analogously to the concentration problem of scalar functions. Hence a construction of a commuting differential operator with a simple spectrum has been made possible. This circumstance, at the same time, opens the way for computing concentrated vector fields in a fast and numerically stable way, as opposed to the direct method based on the ill-conditioned concentration matrix.
The reduction of the vector problem to an equivalent scalar one relies on a new kind of vector spherical harmonics, which we used as basis functions throughout this paper. These exhibit a simple separable form containing the functions \(F_{lm}\) of Sheppard and Török, for which we derived several novel relations. Finally, we note that these relations of \(F_{lm}\) could facilitate the development of a fast vector spherical harmonic transform, too [35].
References
Albertella, A., Sansò, F., Sneeuw, N.: Band-limited functions on a bounded spherical domain: the Slepian problem on the sphere. J. Geodesy 73(9), 436–447 (1999). doi:10.1007/PL00003999
Anderson, E., Bai, Z., Bischof, C., Blackford, S., Demmel, J., Dongarra, J., Croz, J.D., Greenbaum, A., Hammarling, S., McKenney, A.: LAPACK Users’ Guide, 3rd edn. Society for Industrial and Applied Mathematics, Philadelphia, PA (1999)
Arfken, G.B., Weber, H.J., Harris, F.E.: Mathematical Methods for Physicists: A Comprehensive Guide, 7th edn. Academic Press/Elsevier, Waltham, MA (2012)
Bell, B., Percival, D.B., Walden, A.T.: Calculating Thomson’s spectral multitapers by inverse iteration. J. Comput. Graph. Stat. 2(1), 119–130 (1993). doi:10.1080/10618600.1993.10474602
Dahlen, F.A., Simons, F.J.: Spectral estimation on a sphere in geophysics and cosmology. Geophys. J. Int. 174(3), 774–807 (2008). doi:10.1111/j.1365-246X.2008.03854.x
Das, S., Hajian, A., Spergel, D.N.: Efficient power spectrum estimation for high resolution CMB maps. Phys. Rev. D 79(8), 083008 (2009). doi:10.1103/PhysRevD.79.083008
Devaney, A.J., Wolf, E.: Multipole expansions and plane wave representations of the electromagnetic field. J. Math. Phys. 15(2), 234–244 (1974). doi:10.1063/1.1666629
Eshagh, M.: Spatially restricted integrals in gradiometric boundary value problems. Artif. Satell. 44(4), 131–148 (2009). doi:10.2478/v10018-009-0025-4
Freeden, W., Schreiner, M.: Spherical Functions of Mathematical Geosciences: A Scalar, Vectorial, and Tensorial Setup. Springer, Berlin (2009)
Grünbaum, F.A., Longhi, L., Perlstadt, M.: Differential operators commuting with finite convolution integral operators: some non-abelian examples. SIAM J. Appl. Math. 42(5), 941–955 (1982). doi:10.1137/0142067
Han, S.C., Ditmar, P.: Localized spectral analysis of global satellite gravity fields for recovering time-variable mass redistributions. J. Geod. 82(7), 423–430 (2008). doi:10.1007/s00190-007-0194-5
Horn, R.A., Johnson, C.R.: Matrix Analysis. Cambridge University Press, Cambridge, UK (1985). Reprinted with corrections 1990
Jahn, K., Bokor, N.: Vector Slepian basis functions with optimal energy concentration in high numerical aperture focusing. Opt. Commun. 285(8), 2028–2038 (2012). doi:10.1016/j.optcom.2011.11.107
Landau, H.J., Pollak, H.O.: Prolate spheroidal wave functions, Fourier analysis and uncertainty-II. Bell Syst. Tech. J. 40(1), 65–84 (1961)
Landau, H.J., Pollak, H.O.: Prolate spheroidal wave functions, Fourier analysis and uncertainty-III: the dimension of the space of essentially time- and band-limited signals. Bell Syst. Tech. J. 41(4), 1295–1336 (1962)
Lessig, C., Fiume, E.: On the effective dimension of light transport. Comput. Graph. Forum 29(4), 1399–1403 (2010). doi:10.1111/j.1467-8659.2010.01736.x
Liu, Q.H., Xun, D.M., Shan, L.: Raising and lowering operators for orbital angular momentum quantum numbers. Int. J. Theor. Phys. 49(9), 2164–2171 (2010). doi:10.1007/s10773-010-0403-5
Maniar, H., Mitra, P.P.: The concentration problem for vector fields. Int. J. Bioelectromagn. 7(1), 142–145 (2005). URL http://www.ijbem.net/volume7/number1/pdf/037.pdf
Marinucci, D., Peccati, G.: Representations of SO(3) and angular polyspectra. J. Multivar. Anal. 101(1), 77–100 (2010). doi:10.1016/j.jmva.2009.04.017
Moore, N.J., Alonso, M.A.: Closed-form bases for the description of monochromatic, strongly focused, electromagnetic fields. J. Opt. Soc. Am. A 26(10), 2211–2218 (2009). doi:10.1364/JOSAA.26.002211
Morse, P.M., Feshbach, H.: Methods of Theoretical Physics. Part I. International Series in Pure and Applied Physics. McGraw-Hill, New York (1953)
Percival, D.B., Walden, A.T.: Spectral Analysis for Physical Applications: Multitaper and Conventional Univariate Techniques. Cambridge University Press, Cambridge, UK (1993). Reprinted with corrections 1998
Plattner, A., Simons, F.J.: Potential-field estimation using scalar and vector slepian functions at satellite altitude. In: Freeden, W., Nashed, M.Z., Sonar T. (eds.) Handbook of Geomathematics, 2nd edn. Springer, Berlin (2014)
Plattner, A., Simons, F.J.: Spatiospectral concentration of vector fields on a sphere. Appl. Comput. Harmon. Anal. 36(1), 1–22 (2014). doi:10.1016/j.acha.2012.12.001
Sheppard, C.J.R., Török, P.: Efficient calculation of electromagnetic diffraction in optical systems using a multipole expansion. J. Mod. Opt. 44(4), 803–818 (1997). doi:10.1080/09500349708230696
Simons, F.J.: Slepian functions and their use in signal estimation and spectral analysis. In: Freeden, W., Nashed, M.Z., Sonar, T. (eds.) Handbook of Geomathematics, 1st edn, pp. 891–924. Springer, Berlin (2010)
Simons, F.J., Dahlen, F.A.: Spherical Slepian functions and the polar gap in geodesy. Geophys. J. Int. 166(3), 1039–1061 (2006). doi:10.1111/j.1365-246X.2006.03065.x
Simons, F.J., Dahlen, F.A., Wieczorek, M.A.: Spatiospectral concentration on a sphere. SIAM Rev. 48(3), 504–536 (2006). doi:10.1137/S0036144504445765
Simons, F.J., Wang, D.V.: Spatiospectral concentration in the Cartesian plane. Int. J. Geomath. 2(1), 1–36 (2011). doi:10.1007/s13137-011-0016-z
Slepian, D.: Prolate spheroidal wave functions, Fourier analysis and uncertainty-IV: extensions to many dimensions; generalized prolate spheroidal functions. Bell Syst. Tech. J. 43(6), 3009–3057 (1964)
Slepian, D.: Some comments on fourier analysis, uncertainty and modeling. SIAM Rev. 25(3), 379–393 (1983). doi:10.1137/1025078
Slepian, D., Pollak, H.O.: Prolate spheroidal wave functions, Fourier analysis and uncertainty-I. Bell Syst. Tech. J. 40(1), 43–63 (1961)
Swarztrauber, P.N.: The vector harmonic transform method for solving partial differential equations in spherical geometry. Mon. Weather Rev. 121(12), 3415–3437 (1993). doi:10.1175/1520-0493(1993)121<3415:TVHTMF>2.0.CO;2
Szegő, G.: Orthogonal Polynomials, AMS Colloquium Publications, vol. 23, 4th edn. American Mathematical Society, Providence, RI (1975)
Tygert, M.: Recurrence relations and fast algorithms. Appl. Comput. Harmon. Anal. 28(1), 121–128 (2010). doi:10.1016/j.acha.2009.07.005
Winch, D.E., Roberts, P.H.: Derivatives of addition theorems for Legendre functions. J. Aust. Math. Soc. B 37(2), 212–234 (1995). doi:10.1017/S0334270000007670
Acknowledgments
The authors thank Frederik J. Simons and Alain Plattner for helpful discussions. The work reported in the paper has been developed in the framework of the project “Talent care and cultivation in the scientific workshops of BME” project. This project is supported by the Grant TÁMOP-4.2.2.B-10/1–2010-0009.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Hartmut Führ.
Appendix 1: Proofs
Appendix 1: Proofs
1.1 Appendix 1(a): Proof of Recurrence Relation (12)
Proof
First we construct equivalent formulations of \(F_{lm}\) by inserting the following recurrence relations, which can be obtained in a straightforward way from the corresponding relations for \(P_{l}^{m}\) [3, p. 744], into (1):
with
Thus we obtain
Using these expressions of \(F_{lm}\) and recurrence relation [3, p. 744]
we transform the left-hand side (LHS) and right-hand side (RHS) separately so that only terms containing \(U_{lm}\) and \(U_{l-1,\,m}\) remain.
First we rewrite the LHS by inserting (90):
After that we proceed to the RHS. We insert (90) and (91), shifted in index \(l\) by \(+1\) and \(-1\), respectively:
Next we expand \(\zeta _{l+1,\,m}\) and \(\zeta _{lm}\) using their definition (15) and use the relation
which follows from definitions (15) and (89). By straighforward, if lengthy, algebraic calculation, we get
We apply recurrence relation (12) and collect like terms, hence
Taking the difference \(\text {LHS} - \text {RHS}\), it can be shown by further straightforward algebra that the coefficients of \(U_{lm}\) and \(U_{l-1,\,m}\) are zero. Hence \(\text {LHS} = \text {RHS}\).
1.2 Appendix 1(b): Proof of Recurrence Relation (13)
Proof
In this proof, we follow the same strategy as in the previous proof and rewrite the left-hand side (LHS) first. Inserting expression (90) of \(F_{lm}\) from the previous proof yields
Performing the differentiation and using the relation \((1 - x^2) \frac{\mathrm{d }\!^{}}{\mathrm{d }\!x^{}} (1 - x^2)^{-1/2} = x (1 - x^2)^{-1/2}\), we get
Next we insert recurrence relations (87) and (88), shifted in index \(l\) by \(+1\) and \(-1\), respectively. After that we collect like terms and perform some straightforward algebra to obtain
Now we rewrite the right-hand side (RHS). We insert expressions (90) and (91) of \(F_{lm}\), the second one shifted in index \(l\) by \(+1\).
Next we substitute definition (15) of \(\zeta _{lm}\), use (93) and collect like terms. By straightforward algebra we get
Taking the difference \(\text {LHS} - \text {RHS}\), the terms containing \(U_{l-1,\,m}\) cancel. It can be shown that the coefficient of \(U_{lm}\) is zero as well, hence \(\text {LHS} = \text {RHS}\).
1.3 Appendix 1(c): Proof of Christoffel–Darboux Formula (16)
Proof
We start from recurrence relation (12) and multiply both sides by \(F_{lm}(x')\). Then we take the same recurrence relation again, but this time, substitute \(x'\) for \(x\) and multiply both sides by \(F_{lm}(x)\). In this way, we obtain the following two equations:
Taking their difference and summing over \(l\) yields
We can see that consecutive terms cancel in the sum on the right-hand side. Moreover, \(F_{\ell _{m}-1,\,m} = 0\), thus only one term corresponding to \(\zeta _{L+1,\,m}\) remains:
\(\square \)
1.4 Appendix 1(d): Proof of Addition Theorem (17)
Proof
Upon inserting definition (1) of \(F_{lm}\) into the left-hand side of addition theorem (17), we obtain
Next we use the addition theorems
based on the work of Winch and Roberts [36].
Hence the first term on the right-hand side of Eq. (94) yields \((2l + 1)/2\), while the second term vanishes because of symmetry relation (9). Therefore,
\(\square \)
1.5 Appendix 1(e): Proof of \(F_{lm}\) Satisfying Differential Equation (23)
Proof
First let us rearrange (23) and insert \(F_{lm}\):
The left-hand side can be transformed by exploiting recurrence relations (13) and (14) (the second one shifted in index \(l\) by \(-1\)) as follows:
Collecting like terms yields
As expected, the term containing \(F_{l-1,\,m}\) vanishes. Applying definition (15) of \(\zeta _{lm}\) and expanding the fraction by \(l\) yields
By a straightforward, if lengthy, simplification we obtain
\(\square \)
1.6 Appendix 1(f): Proof of Integral Identity (76)
Proof
Inserting expression (75) of \(\mathcal {J}_{m}\) into both sides of integral identity (76) yields
Next we perform integration by parts on the first term of the right-hand side in both equations:
The first term on the right-hand side of both equations vanishes and the rest is identical, hence
Upon inserting (98) into (97a) we find that
\(\square \)
1.7 Appendix 1(g): Proof of Identity (78)
Proof
First we apply expression (74) of \(\mathcal {J}_{m}\) to the kernel function \(\mathcal {K}_{m}(x,\,x')\) and use eigenvalue Eq. (30) of \(\Delta _{\Omega ,m}\):
Likewise, we also apply \(\mathcal {J}_{m}'\) to \(\mathcal {K}(x,\,x')\) and subtract the resulting equation from the previous one, yielding
Using recurrence relation (14) on the terms containing the derivatives of \(F_{lm}\) and performing some straightforward algebra, we get
Applying Christoffel–Darboux formula (16) to the second term on the right-hand side yields
In the last term of the right-hand side, the summation can be interchanged as
Relabeling the sums on the right-hand side of this expression, so that \(l\) becomes \(l'\) and vice versa, and inserting the resulting expression into the right-hand side of (99), we obtain
Since \(\sum _{l'=\ell _{m}}^{L} (2l' + 1) = (L + 1)^2 - l^2\), the right-hand side vanishes. Hence
\(\square \)
1.8 1(h) Proof of Expressions (83) for the Matrix Elements of \(\mathcal {J}_{m}\)
Proof
We start by inserting expression (74) of \(\mathcal {J}_{m}\) into the integral expression (81) for the matrix elements and use the eigenvalue Eq. (30) of \(\Delta _{\Omega ,\,m}\):
The first integral is equal to \(\delta _{ll'}\) because of orthonormality relation (5). The remaining two can be evaluated by using recurrence relations (12) and (13) and orthonormality relation (5):
Thus for (100), we get
Because of the Kronecker deltas, this expression is non-zero for index pairs \((l,\,l)\), \((l+1,\,l)\) and \((l,\,l+1)\) only. The corresponding matrix elements are
hence \(\mathsf {J}_{m}\) is real, symmetric and tridiagonal.
Rights and permissions
About this article
Cite this article
Jahn, K., Bokor, N. Revisiting the Concentration Problem of Vector Fields within a Spherical Cap: A Commuting Differential Operator Solution. J Fourier Anal Appl 20, 421–451 (2014). https://doi.org/10.1007/s00041-014-9324-7
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00041-014-9324-7
Keywords
- Concentration problem
- Spherical cap
- Commuting differential operator
- Vector spherical harmonics
- Bandlimited function
- Eigenvalue problem