Home Search Collections Journals About Contact us My IOPscience

Regularized wavelet-based multiresolution recovery of the harmonic mass density distribution

from data of the Earth's gravitational field at satellite height

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2005 Inverse Problems 21 997

(http://iopscience.iop.org/0266-5611/21/3/013)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 143.167.2.135
This content was downloaded on 07/09/2015 at 14:57

Please note that terms and conditions apply.

INSTITUTE OF PHYSICS PUBLISHING INVERSE PROBLEMS

Inverse Problems 21 (2005) 997–1025 doi:10.1088/0266-5611/21/3/013

Regularized wavelet-based multiresolution recovery of

the harmonic mass density distribution from data of
the Earth’s gravitational field at satellite height
Volker Michel
Geomathematics Group, Department of Mathematics, University of Kaiserslautern,
PO Box 3049, D-67653 Kaiserslautern, Germany

E-mail: [email protected]

Received 22 June 2004, in final form 8 March 2005

Published 29 April 2005
Online at stacks.iop.org/IP/21/997

Abstract
The inverse problem of recovering the Earth’s density distribution from data
of the first or second derivative of the gravitational potential at satellite orbit
height is discussed for a ball-shaped Earth. This problem is exponentially
ill-posed. In this paper, a multiscale regularization technique using scaling
functions and wavelets constructed for the corresponding integro-differential
equations is introduced and its numerical applications are discussed. In the
numerical part, the second radial derivative of the gravitational potential at
200 km orbit height is calculated on a point grid out of the NASA/GSFC/NIMA
Earth Geopotential Model (EGM96). Those simulated derived data out of SGG
(satellite gravity gradiometry) satellite measurements are taken for convolutions
with the introduced scaling functions yielding a multiresolution analysis of
harmonic density variations in the Earth’s crust. Moreover, the noise sensitivity
of the regularization technique is analysed numerically.

1. Introduction

The determination of the structures in the Earth’s interior is usually realized by analysing two
types of data: gravitational data and seismic data. This paper studies the first problem which
is based on the inversion of Newton’s gravitational potential


ρ(x)
dx = V (y),
Earth |x − y|
which is a Fredholm integral equation of first kind. Various publications have studied this
so-called gravimetry problem so far, see, for instance, [1–4, 12, 16, 18–23, 28, 29], and the
references therein. We restrict here our attention to a ball-shaped Earth which is still a common
approximation to the real shape of the planet in this context of a global modelling. Note that
for some geophysical problems even flat Earth models are frequently used.

0266-5611/05/030997+29$30.00 © 2005 IOP Publishing Ltd Printed in the UK 997

998 V Michel

It turns out that each of Hadamard’s criteria for a well-posed problem is violated, where
we first discuss here the theoretical case of continuous data before we treat finitely many
discrete data in numerics:
• The right-hand side V must be harmonic, i.e. it must satisfy the Laplace equation V = 0,
outside the Earth and has to satisfy the corresponding Picard condition. Thus, measuring
errors may modify the right-hand side such that solvability is lost.
• If the problem is solvable, then the space of all solutions ρ corresponding to a fixed
potential V is infinite-dimensional. More precisely, only the harmonic part of the mass
density ρ can be uniquely recovered. The remaining, so-called anharmonic part, cannot
be determined out of gravitational data.
• The (harmonic) density does not continuously depend on the potential V , i.e. the solution
is not stable.
The non-uniqueness can, for instance, be treated by taking other, non-gravitational information
sources, like earthquake travel times, into account. For the modelling of the anharmonic part
of the Earth’s density distribution, which is beyond the scope of this paper, the reader is
referred to [1–4, 12, 19, 21, 29], and the references therein; in [28] a generalized partial
differential equation (fρ) = 0, where f is chosen, is used for a unique spectral density
recovery on a spheroid. Thus, a complete model for the Earth’s interior can be obtained
by combining the method for the harmonic modelling described in this paper with one of
the anharmonic modelling techniques introduced in the above-mentioned publications. The
instability and the possible non-solvability require a regularization technique. It should be
noted that basically three types of gravitational data sources exist: terrestrial, airborne and
spaceborne data. Only the spaceborne data derived from satellite measurements allow a global
covering of the Earth. However, in this case the instability is seriously deteriorated, since the
singular values of the operator mapping the potential to the harmonic density exponentially
diverge. The inverse problem is, therefore, exponentially ill-posed. Moreover, depending
on the measuring technique chosen for a satellite mission the gradient or the Hessian of the
potential becomes available. Hence, we have to deal with scalar, vectorial and matrix-valued
functions, where the latter will also be called tensorial functions here. Note that it is not
the intention of the author to discuss measuring techniques or to treat the derivation of the
corresponding derivatives of the potential out of the true satellite data. Both areas are already
intensively discussed in many other publications by various authors. For theoretical aspects
of the second one of those two areas we refer, for example, to [13].
One standard regularization technique used in physical geodesy for such kinds of inverse
problems is the truncated singular value decomposition. However, this method has several
essential drawbacks that are all connected with the polynomial character of the basis functions
that have to be used. Huge data sets that can be expected from satellite missions have to be
taken completely for the calculation of each orthogonal (Fourier) coefficient of the investigated
functions. Moreover, local changes on the right-hand side V require the new calculation of
every coefficient. Finally, space-dependent measuring accuracies cannot be sufficiently taken
into account, since they contribute to each coefficient. A distinction of ‘good’ and ‘bad’
coefficients is hardly possible.
Multiresolution techniques based on wavelets, which were first introduced for equations
on the sphere in [15], do not have those disadvantages since the used kernel functions are
only in a small part of the domain essentially different from zero. Note that there exist certain
similarities between the localizing scaling functions and wavelets and the so-called covariance
functions in geodesy (see [27]). Following [12] and [18–21], we discuss such a multiresolution
regularization technique for the calculation of the Earth’s harmonic density distribution out of
Regularized multiresolution density recovery from satellite gravity data 999

the gravitational potential, its gradient, its Hessian or certain components of those derivatives,
each given at a spherical satellite orbit, where it should be noted that the spherical shape of
the orbit is not too a serious restriction since there exist tools for mapping the values from the
true orbit to an appropriate sphere, see, for example, [8]. For including the different types of
equations a new generalized concept is used (see also [21]).
Moreover, the localizing character of the used kernels is studied by estimating the omission
error, i.e. the error occurring when only parts of the satellite data are taken into account for
the density recovery. It turns out that the introduced scaling functions and wavelets have a
strong localizing property, such that, for example, for the density recovery at the European
Earth’s crust, satellite measurements obtained above Europe and its immediate neighbourhood
merely have to be taken into account. This localizing behaviour is confirmed by the numerical
calculations shown in section 7. There data of the second radial derivative of the gravitational
potential at a 200 km orbit are given on a point grid. By using the Driscoll–Healy numerical
integration method, the convolutions with the regularizing scaling functions are calculated
yielding approximations of the harmonic density with increasing spatial resolution. For the
first time such a multiscale harmonic density recovery is performed numerically without
requiring a previous Fourier analysis (cf [19]). Note that theoretically it can be shown that
each approximation continuously depends on the given data at orbit height and the sequence
of approximations converges to the exact harmonic solution. This is clearly confirmed by the
new numerical results presented here. Furthermore, for perturbed right-hand sides yielding
non-solvability of the equation a mapping on the image space of the operator is included in
the method. Moreover, the used gravity model is perturbed by white noise. The regularization
proves to be well performing in the numerical experiments since nevertheless good results are
obtained.
We will use here complete orthonormal systems {Yn,j }n∈N0 ;j =1,...,2n+1 ,
 (i)   
yn,j i=1,2,3;n
0i ;j =1,...,2n+1 and y(i,k)
n,j (i,k)∈{1,2,3}2 ,n
0i,k ,j ∈{1,...,2n+1} of scalar, vectorial
and tensorial spherical harmonics. For their definition and important properties as well as for
further notation used in this paper we refer to the appendix.

2. Formulation of the gravimetry problem for SST- and SGG-data

The gravitational potential of the Earth is given by Newton’s law of gravitation

ρ(x)
V (y) = dx,
Bint |x − y|
where ρ represents the Earth’s density distribution and the gravitational constant has been
omitted for reasons of simplicity. It is a well-known fact that the space L2 (Bint ) allows a
decomposition into a direct sum of the closed linear subspace Harm(Bint ) of all harmonic
functions on Bint and the orthogonal space Anharm(Bint ) of all so-called anharmonic functions
(see, for example, [1–4, 12, 18–21, 29]). Well-known complete orthonormal systems are the
system of inner harmonics
    
2n + 3 |x| n x
Hn,j (x) =
B
Yn,j ; x ∈ Bint ,
β 3 β |x|
n ∈ N0 , j ∈ {1, . . . , 2n + 1}; for Harm(Bint ) and the system
  
4m + 2n + 3  m + n + 32
An,j,m (x) =
B
 
β3 m! n + 32
   n  
3 3 |x|2 |x| x
× Gm n + , n + ; 2 Yn,j ; x ∈ Bint ,
2 2 β β |x|
1000 V Michel

n ∈ N0 , j ∈ {1, . . . , 2n + 1}, m ∈ N; for Anharm(Bint ), where  is the gamma function and

Gm (a, b; .) represents the Jacobi polynomial of degree m with respect to the parameters a > 0
and b > a − 1. Those polynomials are uniquely determined by the conditions
(i) Gm (a, b; .) is a polynomial of degree m,
(ii) Gm (a, b; 0) = 1,
1
(iii) 0 x a−1 (1 − x)b−a Gn (a, b; x)Gm (a, b; x) dx = 0 for all n = m.
Modern satellite techniques like SST (satellite-to-satellite tracking) used in the case of the
satellites CHAMP (launch: 2000) and GRACE (launch: 2002) and SGG (satellite gravity
gradiometry) intended for the planned mission GOCE (launch: 2006) yield data from which
derivatives of the gravitational potential can be derived. More precisely, it is possible to obtain
the gradient of V from SST data and the Hessian ∇ ⊗ ∇V from SGG data. For further
details on the mathematical background the reader is referred to, for example, [11–13], and
the references therein. Those gravitational values can, furthermore, be mapped from the true
orbit to a sphere (see, for example, [8]), which will serve here as an approximation to the orbit.
With elementary calculations, it is possible to derive a formula for the gravitational
potential and its first and second derivatives in dependence on a density distribution
F ∈ L2 (Bint ) (see also [13, 19, 21] for the potential and its first and second radial derivatives
and [9] for gradient and Hessian formulae of the spherical harmonics). We obtain for the
gravitational potential at y ∈ Bext caused by a mass density distribution F ∈ L2 (Bint )


∞  n   
β3
1 y  B 
2n+1
F (x) 4π β
dx = Yn,j Hn,j , F L2 (B ) .
Bint |y − x| n=0
2n + 1 |y| 2n + 3 j =1 |y| |y| int

This identity clearly shows the well-known result [29] that only the harmonic part of the density
distribution can uniquely be recovered from the gravitational potential, whereas Anharm(Bint )
represents the null space of the corresponding integral operator. This result remains valid if
more general domains are allowed as Earth model, see [12, 29].
The gradient of this potential at y ∈ Bext is then given by


∞  n 
F (x) 4π β β3 1
∇y dx =
Bint |y − x| n=0
2n + 1 |y| 2n + 3 |y|2

      
2n+1
 (1) y (2) y
× B
Hn,j , F L2 (B ) (−n − 1)yn,j + n(n + 1)yn,j .
j =1
int |y| |y|
It is not surprising (cf [9]) that the gradient part does not contain any spherical harmonic part
of type 3. Finally, using y = |y|η (|y| > β) the Hessian may be represented by


∞  n 
1
 B 
2n+1
F (x) 4π β β3
∇y ⊗ ∇y dx = Hn,j , F L2 (B )
Bint |y − x| 2n + 1 |y| 2n + 3 |y| j =1
2 int
n=0
 √
(n + 1)(n + 2) (1,1) (n + 2) n(n + 1) (1,2)
× yn,j (η) − yn,j (η)
|y| |y|
√
(n + 2) n(n + 1) (2,1) (n + 1)(n + 2) (2,2)
− yn,j (η) − √ yn,j (η)
|y| 2|y|
√ 
n(n + 1)(n(n + 1) − 2) (2,3)
+ √ yn,j (η) .
2|y|
For the following discussions, we will assume that the data are given in a form such that
the inverse problem of determining the harmonic mass density distribution can in general be
Regularized multiresolution density recovery from satellite gravity data 1001

formulated by
T F = G, G ∈ Y given, F ∈ Harm(Bint ) unknown,
where the linear and continuous operator T : Harm(Bint ) → Y has a known singular value
decomposition given by
B
T Hn,j = T ∧ (n)bn,j ; n ∈ N0 , j ∈ {1, . . . , 2n + 1};
with {bn,j }n∈N0 ,j ∈{1,...,2n+1} being a complete orthonormal system in the Hilbert space
(Y , ., .
Y ). Examples of such constellations are listed in table 1. Note that, for instance,
the identities
ξ T y(2,3)
n,j (ξ ) = 0, y(2,3)
n,j (ξ )ξ = 0, ξ ∈ ,
and
(i − ξ ⊗ ξ ) : y(i,k)
n,j (ξ ) = 0, ξ ∈ , (i, k) = (2, 2);
Oξ(2,3) y(i,k)
n,j (ξ ) = 0, ξ ∈ , (i, k) = (2, 3);
oξ(2,3) Oξ(2,3) y(2,3)
n,j (ξ ) = 2n(n + 1)(n(n + 1) − 2)y(2,3) (2,3) (2,3)
n,j (ξ ) = µn yn,j (ξ ), ξ ∈ ,
where
Oξ(2,3) f(ξ ) = ∇ξ∗ · ((i − ξ ⊗ ξ )(∇ξ∗ · ((i − ξ ⊗ ξ )f(ξ ))))
− L∗ξ · ((i − ξ ⊗ ξ )(L∗ξ · ((i − ξ ⊗ ξ )f(ξ )))) − 2∇ξ∗ · ((i − ξ ⊗ ξ )f(ξ )ξ ),
f ∈ c(2) (), hold [9]. Here, we always assumed a simplified spherical orbit S with radius
σ > β. Note that the listed gravitational quantities have already been divided by the
gravitational constant.
Obviously, problem nos 4, 7 and 8 do not allow the reconstruction of the harmonic density
at degree n = 0 and problem no 10 does not even allow the reconstruction of the density at
degrees n ∈ {0; 1}. This is a serious restriction since it is known that the harmonic part of
radially symmetric mass density structures, which are given approximately in the case of the
mantle and the core of the Earth, is only reflected in the degree n = 0 coefficient [19]. Thus,
the problem nos 1–3, 5, 6 and 9, should be preferred. Nevertheless, we will also include
the other problems into the theory presented here. For this purpose we will use the range
{n ∈ N0 |n
n0 } for the degrees, where n0 := 1 in the case of problem nos 4, 7 and 8; n0 := 2
in the case of problem no 10; and n0 := 0 otherwise. Note that here n0 coincides with 0i and
0
i,k , respectively,
∞ for the corresponding basis functions. For reasons of brevity we will write
n instead of n=n0 .
Hadamard’s criteria for the well-posedness of the problem T F = G can now be discussed
as follows:
• Existence: A solution exists if and only if the Picard condition

 G, bn,j
Y 2

2n+1
<∞
n j =1
T ∧ (n)
is satisfied.
• Uniqueness: The null space of the operator T is given by
ker T = Anharm(Bint ) ⊕ span{bn,j |0  n < n0 , 1  j  2n + 1}.
We will now consider T to be restricted to
 B   .
span Hn,j n ∈ N0 , n
n0 , 1  j  2n + 1 L2 (Bint ) ,
such that it is invertible on its image.
 1002
Table 1. Examples of inverse problems resulting from the reconstruction of the Earth’s harmonic density distribution out of satellite data (orbit height σ > β) of the gravitational
field, represented by the corresponding singular value decomposition, here y = |y|η ∈ S = {z ∈ R3 ||z| = σ }

No T ∧ (n) bn,j Y TF Type of data

 n 
4π β β3 S F (x)
1 2n+1 σ 2n+3 Yn,j L2 (S) Bint |y−x| dx Gravitational potential
 n 
β β3 F (x)
2 4π
2n+1 σ
n+1
2n+3 σ
S
Yn,j L2 (S) −η · ∇y Bint |y−x| dx (Negative) first radial derivative (part of SST)
 n   
β β 3 n+1 (1);S F (x)
3 4π
2n+1 σ 2n+3 σ yn,j l21 (S) − η · ∇y Bint |y−x| dx η (Negative) normal part of gradient (part of SST)
 n  3 √
β β n(n+1) (2);S F (x)
4 4π
2n+1 σ 2n+3 σ yn,j l22 (S) (i − η ⊗ η) ∇y Bint |y−x|
dx Tangential part of gradient (part of SST)
 n  3  
β β (n+1)(n+2) F (x)
5 4π
2n+1 σ 2n+3 σ2
S
Yn,j L2 (S) ηT ∇y ⊗ ∇y Bint |y−x| dx η Second radial derivative (part of SGG)
 n  3   
β β (n+1)(n+2) (1,1);S F (x)
6 4π
2n+1 σ 2n+3 σ2
yn,j l21,1 (S) (η ⊗ η) (η ⊗ η) : ∇y ⊗ ∇y Bint |y−x| dx Normal part of the Hessian (part of SGG)
 n  3 √  
β β (n+2) n(n+1) (1,2);S F (x)
7 4π
2n+1 σ 2n+3 σ2
yn,j l21,2 (S) −(η ⊗ η) ∇y ⊗ ∇y Bint |y−x| dx (i − η ⊗ η) (Negative) left normal/right tangential part of
the Hessian (part of SGG)
 n  √  
β β 3 (n+2) n(n+1) (2,1);S F (x)
8 4π
2n+1 σ 2n+3 σ2
yn,j l22,1 (S) −(i − η ⊗ η) ∇y ⊗ ∇y Bint |y−x| dx (η ⊗ η) (Negative) left tangential/right normal part of
the Hessian (part of SGG)
 n    
β β 3 (n+1)(n+2) (2,2);S F (x)
9 4π
2n+1 σ 2n+3
√ yn,j l22,2 (S) 1
2 (η ⊗ η − i) (i − η ⊗ η) : ∇y ⊗ ∇y Bint |y−x| dx Component of (negative) tangential part of the
2σ 2
Hessian (part of SGG)
 n  √  −1   
β3 (2,3);S
10 4π
2n+1
β
σ 2n+3
n(n+1)(n(n+1)−2)
√ yn,j l22,3 (S) µ(i,k)
n oη(2,3) Oη(2,3) ∇y ⊗ ∇y Bint F (x)
|y−x| dx Component of tangential part of the Hessian
2σ 2
(part of SGG)

V Michel
Regularized multiresolution density recovery from satellite gravity data 1003

• Stability: In every case T ∧ (n) contains the factor (β/σ )n such that the inverse operator
of T is not continuous. More precisely, due to the exponential character of T ∧ (n) the
problem is exponentially ill-posed.
In [21], an abstract Hilbert space approach for the regularization of ill-posed problems given by
an operator between separable Hilbert spaces possessing a known singular value decomposition
is introduced. In what follows, we will discuss the realization of this method in the case of the
problem nos 1–10 in more detail and finally show some new numerical results.

3. Convolutions and product kernels

For the construction of the multiscale regularization technique, we need the concept of
convolutions and product kernels introduced below.

Definition 3.1. Let C ⊂ Rn , D ⊂ Rm , E ⊂ Rl and n, m, l, p, q, r ∈ N. The mappings

L2 (C × D, Rp×q ) × L2 (D × E, Rq×r ) → L2 (C × E, Rp×r ),
L2 (C × D, Rp×q ) × L2 (D, Rq ) → L2 (C, Rp ),
defined by


K ∗ L := K(., x)L(x, .) dx, (1)
D


K ∗ F := K(., x)F (x) dx; (2)
D

K ∈ L2 (C × D, Rp×q ), L ∈ L2 (D × E, Rq×r ), F ∈ L2 (D, Rq ); are called convolutions.

In (1) KL stands for the usual matrix multiplication and KF in (2) is the usual
multiplication of a matrix and a vector, which is certainly only a special case of a matrix
multiplication.

Theorem 3.2. The convolutions defined in definition 3.1 are bilinear and continuous with
respect to each argument.

Proof. We restrict our considerations to the convolution defined by (1), since the convolution
in (2) can be regarded as a special case of (1) where r = 1.
Bilinearity. For λ, µ ∈ R we get


(λK1 + µK2 ) ∗ L = (λK1 (., x) + µK2 (., x))L(x, .) dx


D


= λ K1 (., x)L(x, .) dx + µ K2 (., x)L(x, .) dx
D D
= λ(K1 ∗ L) + µ(K2 ∗ L).

Continuity. Let Ki −→ K in L2 (C × D, Rp×q ), i.e.

lim |Ki (x, y) − K(x, y)|2 d(x, y) = 0,
i→∞ C×D
√
where |A| := A : A is the Euclidean norm of a matrix.
1004 V Michel

We have to show that the following expression converges to 0 as i → ∞:

|(K ∗ L)(x, y) − (Ki ∗ L)(x, y)|2 d(x, y) = |((K − Ki ) ∗ L)(x, y)|2 d(x, y)
C×E C×E


2
 
=  (K(x, z) − Ki (x, z))L(z, y) dz d(x, y). (3)
 
C×E D
Representing the components of the involved matrices by
 (i) 
K = (kν,µ ) ν=1,...,p;, Ki = kν,µ ν=1,...,p;, L = (lµ,σ ) µ=1,...,q;,
µ=1,...,q µ=1,...,q σ =1,...,r
one obtains by applying the Cauchy–Schwarz inequality for the (ν, σ )th component (ν ∈
{1, . . . , p}, σ ∈ {1, . . . , r}) of the inner integral in (3) the estimate



q
 
(K(x, z) − Ki (x, z)) L(z, y) dz = kν,µ (x, z) − kν,µ
(i)
(x, z) lµ,σ (z, y) dz
D ν,σ D µ=1
  12   12

q

q
 
 (x, z) dz  (lµ,σ (z, y))2 dz .
2
kν,µ (x, z) − kν,µ
(i)
D µ=1 D µ=1

Hence, the norm of the corresponding matrix satisfies the inequality


2 r 
2
 
p

 (K(x, z) − Ki (x, z)) L(z, y) dz = (K(x, z) − K (x, z)) L(z, y) dz

  i
D ν=1 σ =1 D ν,σ
  

p

q

q
r
 
  kν,µ (x, z) − kν,µ
(i)
(x, z) dz 
2
(lµ,σ (z, y))2 dz
ν=1 σ =1 D µ=1 D µ=1
   

p

r

q
=  (kν,µ (x, z) − kν,µ
(i)
(x, z))2 dz  (lµ,σ (z, y))2 dz
ν=1 D µ=1 σ =1 D µ=1



= |K(x, z) − Ki (x, z)|2 dz |L(z, y)|2 dz.
D D
Insertion of this result into (3) in connection with Fubini’s theorem finally yields


lim |(K ∗ L)(x, y) − (Ki ∗ L)(x, y)|2 d(x, y)
i→∞ C×E



 

 lim |K(x, z) − Ki (x, z)|2 dz |L(z, y)|2 dz dy dx
i→∞ C E D


D
= lim |K(x, z) − Ki (x, z)| d(x, z)
2
|L(z, y)|2 d(z, y)
i→∞ C×D D×E
= 0.
The proof of the properties with respect to the second argument is analogous. 
In classical 1D Fourier analysis, it is a well-known fact that the Fourier transform of the
convolution of two functions is equal to the product of the Fourier transforms of those two
functions. An analogous result can be proved here.
Theorem 3.3. Let {cn,j }n,j be a (not necessarily complete) orthonormal system with respect
to L2 (D, Rq ), D ⊂ Rm . Then the identities
(f ⊗ cν,µ ) ∗ (cρ,θ ⊗ k) = f ⊗ kδνρ δµθ , (f ⊗ cν,µ ) ∗ cρ,θ = f δνρ δµθ
hold for all f ∈ L (C, R ) and all k ∈ L (E, R ) (C ⊂ R , E ⊂ Rl ).
2 p 2 r n
Regularized multiresolution density recovery from satellite gravity data 1005

Proof. Again we restrict our considerations to the convolution of two matrix-valued functions.
We get


(f ⊗ cν,µ ) ∗ (cρ,θ ⊗ k) = (f (.) ⊗ cν,µ (x))(cρ,θ (x) ⊗ k(.)) dx

D

= (f (.)(cν,µ (x))T )(cρ,θ (x)(k(.))T ) dx

D


= f (.) (cν,µ (x))T cρ,θ (x) dx(k(.))T = f (.) ⊗ k(.)(cν,µ , cρ,θ )L2 (D,Rq )
D
= (f ⊗ k)δνρ δµθ ∈ L2 (C × E, Rp×r ).


We will now have a closer look at the spaces Harm(Bint ) and Y needed for the treatment of
the discussed problems. Note that Y consists of scalar, vectorial or tensorial (matrix-valued)
functions on the sphere S, where the corresponding orthonormal system is denoted by bn,j .
With this in mind we are able to define product kernels (see [27] for related kernels).

Definition 3.4. Functions h and l given by

2n+1

h(x, y) = h∧ (n)Hn,j
B
(x)bn,j (y), (4)
n j =1

2n+1

l(y, z) = l∧ (n)bn,j (y) ⊗ bn,j (z), (5)

n j =1

(for almost all) x ∈ Bint , y, z ∈ S, are called product kernels, if their symbols {h∧ (n)}n ,
{l∧ (n)}n (h∧ (n), l∧ (n) ∈ R for all n) satisfy the admissibility condition

2n+1


2n+1

(h∧ (n))2 < ∞, (l∧ (n))2 < ∞. (6)

n j =1 n j =1

Note that the addition theorem for (scalar, vectorial and tensorial) spherical harmonics
essentially simplifies the calculation of such kernels.
Due to the continuity of the convolution (see theorem 3.2) and theorem 3.3 we immediately
get the following lemma.

Lemma 3.5. Let h and l be product kernels of the form (4) and (5), respectively. Then one
gets for all G ∈ Y the identities

2n+1


2n+1

h∗G= h∧ (n) G, bn,j

Y Hn,j
B
, l∗G= l∧ (n) G, bn,j
Y bn,j ,
n j =1 n j =1

where the series are convergent with respect to (., .)L2 (Bint ) and ., .
Y , respectively.

4. Scaling functions and wavelets

˜ J ∈ L2 (Bint × S, Rm ), J ∈ L2 (S × S, Rm×m ), J ∈ Z (m ∈ N
Definition 4.1. Let J , 
appropriate) be product kernels of the forms (4) and (5), respectively. If their symbols satisfy
the conditions
(i) ∧J (n)  ∧J +1 (n) for all n
n0 and all J ∈ Z,
(ii) limJ →∞ ∧J (n) = (T ∧ (n))−1 for all n
n0 ,
1006 V Michel

(iii) ∧J (n)
0 for all n
n0 and all J ∈ Z,

(iv) (refinement equation)
˜ J∧ (n)J∧ (n) = ∧J +1 (n) − ∧J (n)
 (7)
for all n
n0 and all J ∈ Z,
then {J }J ∈Z is called a regularizing scaling function, {J }J ∈Z is called a decomposition
wavelet and {˜ J }J ∈Z is a called reconstruction wavelet.

Note that T ∧ (n) > 0, n
n0 , for all problems in table 1 such that the combination of (ii)
and (iii) is no contradiction. The value m, for instance, equals 1 for problem nos 1, 2 and 5,
equals 3 for problem nos 3 and 4, and equals 9 (more precisely, write 3 × 3 instead of m) for
problem nos 6–10 (see table 1).
The introduced kernels now allow a regularized multiresolution reconstruction of the
harmonic density.

Theorem 4.2. Let {J }J ∈Z , {J }J ∈Z , {

˜ J }J ∈Z be given as introduced above. If G ∈ im T ⊂
Y is given, then the unique harmonic solution F of T F = G is given by
∞

F = lim J ∗ G = J1 ∗ G + ˜ j ∗ j ∗ G,
 J1 ∈ Z,
J →∞
j =J1

where the limits are understood in the L2 (Bint )-sense. Moreover, the mapping
Y → L2 (Bint ), G → J ∗ G
is continuous for all J ∈ Z.

Proof. With lemma 3.5 and the refinement equation (7) we get (J > J1 , J ∈ Z)
J −1


2n+1

J1 ∗ G + ˜ j ∗ j ∗ G =
 ∧J1 (n) G, bn,k
Y Hn,k
B

j =J1 n k=1
J −1
2n+1

+ ˜ j∧ (n)j∧ (n) G, bn,k

Y Hn,k
 B

j =J1 n k=1

2n+1
= ∧J (n) G, bn,k
Y Hn,k
B
.
n k=1

Obviously, the harmonic solution F can be represented with respect to L2 (Bint ) by

2n+1
F = (T ∧ (n))−1 G, bn,k
Y Hn,k
B
.
n k=1

Moreover, the fact that 0  ∧J (n)  (T ∧ (n))−1 for all n
n0 and all J ∈ Z and the Picard
condition allow the interchanging of the following limits:

2n+1

lim F − J ∗ G 2L2 (Bint ) = lim ((T ∧ (n))−1 − ∧J (n))2 G, bn,k

2Y
J →∞ J →∞
n k=1

2n+1
= lim ((T ∧ (n))−1 − ∧J (n))2 G, bn,k
2Y
J →∞
n k=1
= 0.
Regularized multiresolution density recovery from satellite gravity data 1007

Finally, we find the following estimate for the operator norm of G → J ∗ G:

 1/2

2n+1

∧
max J ∗ G L2 (Bint ) = max (J (n)) G, bn,k
Y
2 2
G∈Y G∈Y
G Y =1 G Y =1 n k=1
∧
 max J (n) < +∞.
n
n0 
Hence, we are now able to calculate arbitrarily good approximations


N
J ∗ G = J (., y)G(y) dω(y) ≈ wn J (., yn )G(yn ) (8)
S n=0

to the exact solution F = T −1 G, where those approximations continuously depend on

the right-hand side of the equation T F = G. Moreover, the difference between different
resolutions can be expressed in terms of convolutions with wavelets
2 −1
J

J2 ∗ G − J1 ∗ G = ˜ j ∗ j ∗ G;

j =J1

J1 , J2 ∈ Z, J1 < J2 . As already indicated in (8), the occurring integrals can be discretized with
quadrature methods (see, for instance, [7, 26]; also note the numerical results in section 7) such
that given data sets on (appropriate) point grids can directly be used for the computations. Such
numerical integration techniques are usually exact up to a chosen spherical harmonic degree.
This polynomial exactness suffices in this case for arbitrarily close approximations. Hence,
the calculation of the orthogonal coefficients G, bn,j
Y is not necessary in our approach.
Example 4.3. Examples of applicable scaling functions are the Gauß–Weierstraß scaling
function given by
∧J (n) = (T ∧ (n))−1 exp(−R2−J (n − n0 )(2−J (n − n0 ) + 1)),
n
n0 , J ∈ Z, and the Tykhonov–Philips scaling function given by
T ∧ (n)
∧J (n) = ,
(T ∧ (n))2
+ 2−J n2
n
n0 , J ∈ Z. Note that the regularizing property of the scaling functions is implicitly
contained in the admissibility condition (6), which is equivalent to the requirement that the
used kernels are L2 -functions. The inclusion of a regularization condition can be seen as
follows: a necessary condition for the validity of (6) is the limit relation limn→∞ ∧J (n) = 0
for all J ∈ Z. Thus, we have
∧J (n) −→ (T ∧ (n))−1
J →∞
|| ||
↓n→∞ ↓n→+∞
0 +∞ (exponentially!).
Hence, ∧J (n) approximates (T ∧ (n))−1 for small n but tends to 0 for n → ∞. The larger J is
the better is this approximation. However, the convergence to 0 finally starts in all cases. This
is illustrated in figure 1, where (T ∧ (n))−1 (bold line or bold dots) is plotted in comparison
to ∧J (n) for J = 4 (solid line), J = 5 (dash-dotted line), J = 6 (dashed line) and J = 7
(dotted line) in the case of the Gauß–Weierstraß scaling function. For (T ∧ (n))−1 the symbol
of problem no 1 in table 1 was chosen. Note the logarithmic vertical axis of the first two
illustrations.
1008 V Michel

0 50
10 10

0
10
–5
10
–50
10
–10
10 –100
10

–150
–15 10
10
–200
10
–20
10
–250
10

–25 –300
10 10
0 20 40 60 80 100 0 200 400 600 800 1000
n n

–4 –3
x 10 x 10
3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 5 10 15 20 25 30 0 20 40 60 80 100
n n

Figure 1. Comparison of (T ∧ (n))−1 (bold line or bold dots) to ∧ J (n) for different scales in the
case of the Gauß–Weierstraß scaling function; the horizontal axes always refer to n.

The corresponding wavelets can be constructed in the usual way out of the symbol of the
scaling functions (cf [9]), e.g. as P-wavelets by

˜ J∧ (n) := J∧ (n) := ∧J +1 (n) − ∧J (n)
or as M-wavelets by
   
J∧ (n) := ∧J +1 (n) − ∧J (n), ˜ J∧ (n) :=
 ∧J +1 (n) + ∧J (n)
(or vice versa). ‘P’ stands for ‘packet’ reminding of the origin of those wavelets from the
discretization of scale continuous wavelets to scale discrete wavelets via wavelet packages,
where ‘M’ is an abbreviation for ‘modified’.
The validity of the admissibility condition is easily verified in the case of those two
examples of scaling functions since
 n
σ
(T ∧ (n))−1 = O(nα ) as n → ∞
β
for some α ∈ R. The Gauß–Weierstraß and Tykhonov–Philips scaling functions and their
corresponding M- or P- wavelets are, thus, applicable. Note that the admissibility condition,
indeed, also has to be verified for the wavelets.
Note that the convolutions J ∗ G and ˜ J ∗ J ∗ G can be calculated for all G ∈ Y , no
matter if G ∈ im T or not. This means that the introduced method is able to treat perturbed
Regularized multiresolution density recovery from satellite gravity data 1009

right-hand sides that do not belong to the image, e.g. due to measuring errors. In the described
multiresolution method a mapping on the image space is implicitly performed. Let G ∈ im T
and E ∈ Y \im T . Then
J ∗ (G + E ) = J ∗ G + J ∗ E ,
where

2n+1
J ∗ E = ∧J (n) E , bn,k
Y Hn,k
B
, J ∗ E L2 (Bint )  max ∧J (n) E Y < +∞.
n
n0
n k=1

Finally, it should be mentioned that, in analogy to [19–21], the typical multiresolution

properties can be proved for the corresponding scale spaces and detail spaces.

Theorem 4.4. Let {J }J ∈Z , {J }J ∈Z , {

˜ J }J ∈Z be defined as introduced above. Then the
scale spaces
VJ := {J ∗ G|G ∈ im T }, J ∈ Z,
and the detail spaces
WJ := {
˜ J ∗ J ∗ G|G ∈ im T }, J ∈ Z,
have the properties
(a) VJ ⊂ VJ +1 ⊂ Harm(Bint ) for all J ∈ Z,
 . 2  B  
(b) J ∈Z VJ L (Bint ) = Harm(Bint )  span Hn,j 0  n < n0 , 1  j  2n + 1 ,
−1
(c) VJ2 = VJ1 + Jj 2=J1
Wj for all J1 , J2 ∈ Z with J1 < J2 .

5. Localization property of the multiscale method

A classical truncated singular value decomposition (TSVD) ansatz for the inverse problem
T F = G would consist of calculating the orthogonal coefficients G, bn,j
Y and approximating
the harmonic solution by the truncated Fourier expansion


G, bn,j
Y
N 2n+1
B
Hn,j ,
n=n0 j =1
T ∧ (n)

where the regularization, i.e. the continuous dependence of the approximation on the right-
hand side G, is trivially given due to the finite dimension (bandlimitedness) of the restricted
solution space.
The advantage of the multiresolution technique, presented in this paper, is not only
given by the possibility of theoretically recovering non-bandlimited solutions but also by
the localizing character of the involved kernels. The coefficients G, bn,j
Y , that have to be
calculated for the TSVD namely are L2 -scalar products of the given data and polynomials,
i.e. global functions. This means that a TSVD of that kind is always connected with a global
smoothing of the available data. Moreover, in the case of large data sets the numerical effort
will become immense, since all data have to be used for the calculation of every coefficient.
Finally, a variation of only a few data in a small region, e.g. due to an earthquake, requires the
new computation of all coefficients.
An essential advantage of the method presented here is its localizing property which is
demonstrated by the uncertainty principle [10] and the following considerations:
1010 V Michel

Γ
z

x
Earth

Figure 2. 2D-illustration of the spherical cap. .

Let D ⊂ S be a measurable subset of the spherical orbit domain S. Moreover, consider a

kernel
    

2n + 3 |x| n 2n + 1 x y
∧
K(x, y) = K (n) Pn ·
n
β3 β 4π σ |x| σ
for a fixed point x ∈ Bint \{0} inside the Earth. Such kernels, for instance, occur for the scalar
problem nos 1, 2 and 5.
If we write x = |x|ξ, ξ ∈ , and consider |x| and ξ to be constant, i.e. fixed, we can
interpret the kernel
  

∧ 2n + 3 |x| n 2n + 1  y 
K(x, y) = K (n) Pn ξ ·
n
β3 β 4π σ σ
 y
=L ξ·
σ
as a ξ -zonal function
 (radial
 x y function).
basis 
Now let  := y ∈ S  |x| · σ > δ be a spherical cap around z = |x| σ
x ∈ S with ‘width’
1 − δ ∈ ]0, 2] (see figure 2). Then we obtain for G ∈ C(S) the result



 
 
K(x, y)G(y) dω(y)  G C(S\) |K(x, y)| dω(y)

S\ S\

 
 y 
= G C(S\) L ξ ·  dω(y)
S\ σ


= G C(S\) |L(ξ · η)| dω(η)σ 2
1
(S\)


σ
δ
= G C(S\) 2π |L(t)| dtσ 2 .
−1
The value

δ
E(x, ) := |L(t)| dt
−1
is a measure for the omission error in the case that gravity data outside  are not taken into
account for the recovery of the harmonic density at x.
For the numerical calculation of E(x, ) we use the Gauß–Weierstraß scaling function,
which is given by its symbol
∧J (n) = (T ∧ (n))−1 exp(−R2−J n(2−J n + 1)),
where we consider problem no 5 (second radial derivative, SGG), i.e.
 n 
∧ 4π β β 3 (n + 1)(n + 2)
T (n) = .
2n + 1 σ 2n + 3 σ2
Regularized multiresolution density recovery from satellite gravity data 1011

Figure 3. The estimated omission error E(x, ) in dependence on the relative radius r =
|x|/β ∈ [0, 0.999] of the point x ∈ Bint , where the harmonic density has to be calculated, and on
δ ∈ [−1, 0.99], which is a parameter for the width of the omitted data region (−1: no data are
omitted, 1: all data are omitted). Here, the Gauß–Weierstraß scaling function at scale J = 4 was
chosen.

–11 –10
x 10 x 10
5 7

4.5
6
4

3.5 5

3
4
2.5
3
2

1.5 2

1
1
0.5

0 0
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 –1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1

Figure 4. The estimated omission error E(x, ) in dependence on δ ∈ [−1, 0.99], which is a
parameter for the width of the omitted data region, where the relative radius r = |x|/β was chosen
to be 0.25 (left) and 0.999 (right, situation used in the numerical calculations of section 7); E is
plotted for different scales of the Gauß–Weierstraß wavelet: J = 2 (solid line), J = 3 (dashed
line), J = 4 (dash-dotted line) and J = 5 (dotted line).

Thus, the kernel has the form

∞

 n  2
σ 2n + 3 σ |x| 2n + 1 −J
n(2−J n+1)
L(t) = e−R2 Pn (t),
n=0
(n + 1)(n + 2) β 3 β2 4π

where we choose R = 1.
Figures 3 and 4 show graphs illustrating the dependence of E(x, ) on |x| and δ. In
˜ J ∗ J = J +1 − J
figure 3 the scaling function J and in figure 4 the convolved wavelets 
are discussed. The calculations are realized for an orbit height 200 km. The kernel is
calculated approximately via the Clenshaw algorithm with truncation at degree 360. The
1012 V Michel

Figure 5. Error estimate E(x, ), as in figure 3 but for δ ∈ [−1, 0.9] (left) and δ ∈ [−1, 0.999]
(right).

–10 –9
10 10

–10
10
–11
10
–11
10

–12
10 –12
10

–13
–13 10
10

–14
10
–14
10
–15
10

–15 –16
10 10
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 –1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1
(a) (b)
–2
10

–4
10

–6
10

–8
10

–10
10

–12
10

–14
10

–16
10
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1
(c)

Figure 6. Error estimate E(x, ), as in figure 4 but for scales J = 2 (solid line), J = 4 (dashed
line), J = 6 (dash-dotted line) and J = 8 (dotted line) and for relative radii r = 0.25 (a), r = 0.9
(b), and r = 0.999 (c), note the logarithmic scale of the vertical axes.

numerical integration is performed by means of the trapezoidal rule with 721 sampling points
per integral. The computations within this section were done with Matlab. Note that the
factor 2π σ 2 is neglected here. Moreover, one should take into account that not G but T −1 G is
Regularized multiresolution density recovery from satellite gravity data 1013

approximated here. For this reason the error bound cannot be expected to be 1 in the maximal
case.
Obviously, the kernels used for the method presented here have a strong localizing
property. This localization effect increases with increasing scale of the scaling function and
wavelet, respectively. Hence, if we add details by calculating the convolutions with wavelets
at different scales, it suffices at large scales to take only satellite data into account which were
obtained within a rather small cone around the area of interest by having only a small error due
to the inaccurate calculations. This is precisely the property we expect from a multiresolution
method.
It should be noted that for values of δ that are close to 1 one observes a rather strange
behaviour of the error estimate. On one hand, around a relative radius βr ≈ 0.9 the error
decreases. On the other hand, the estimate increases again as βr approaches 1 and becomes
essentially larger than deep inside the Earth (see figure 5). This effect might be caused by the
inaccuracy of the error estimate due to the integration of the absolute value of the kernel. This
phenomenon should be investigated further in the future.
On the other hand, for larger scales the monotonicity of the error estimate with respect
to δ is still obvious but at a lower level, i.e. this level decreases with increasing scale (with
the exception of relative radii that are very close to 1, which might also be caused by the
inaccuracy of the estimate and be ‘unreal’; note that for r = 0.999 we obtained very good
results in section 7, which strengthens the conjecture that the real error also behaves nicely
for such values—in contrast to the estimate), see figure 6. This stresses the fact that at large
scales the contributed details of the wavelet analysis are quantitatively low. Hence, this is a
good result.

6. An error estimate

Since the real density distribution cannot be measured globally and is, therefore, not available
as a reference for solution appraisal the error in the recovery has to be estimated in a different
way (see also [21]). A common
 B
model for the decay behaviour (with respect to n → ∞) of
the Fourier coefficients P , Yn,j L2 (B)
of the gravitational potential P at the surface B of the
Earth is Kaula’s rule (see, for example, [24])

2n+1
 
B 2
P , Yn,j L2 (B)
= O(ϑ n+1 n−3 ), n → ∞,
j =1

where ϑ = (RBJE /β)2 ∈ ]0, 1[ and RBJE is the radius of the Bjerhammar sphere. For the
general basis {bn,k }n,k in Y we get

∞ 2n+1

T −1 G − J ∗ G 2L2 (Bint ) = ((T ∧ (n))−1 − ∧J (n))2 G, bn,j

2Y
n=n0 j =1
 ∞
 2n+1

= + ((T ∧ (n))−1 − ∧J (n))2 G, bn,j

2Y
n=n0 n=N+1 j =1

N 2n+1 ∞ 2n+1

∧ −1
 ((T (n)) − ∧J (n))2 G, bn,j
2Y + (T ∧ (n))−2 G, bn,j
2Y .
n=n0 j =1 n=N+1 j =1
1014 V Michel

In the case of the Gauß–Weierstraß scaling function, we obtain for the first part of this error
estimate

N 2n+1
1 (N, J ) := ((T ∧ (n))−1 − ∧J (n))2 G, bn,j
2Y
n=n0 j =1


N 2n+1
2
= 1 − exp(−R2−J (n − n0 )(2−J (n − n0 ) + 1)) (T ∧ (n))−2 G, bn,j
2Y
n=n0 j =1

N


2 2n+1 
= 1 − exp(−R2−J (n − n0 )(2−J (n − n0 ) + 1)) T −1 G, Hn,j
B 2
L2 (Bint )
.
n=n0 j =1

Now let R : Harm(Bint ) → L (B) be the operator which maps a harmonic density to the
2

potential at the Earth’s surface B,

F (x)
RF := dx.
Bint |x − .|

β3
Obviously, RHn,j B
= 2n+1
4π
Y B such that
2n+3 n,j

N
 2
1 (N, J ) = 1 − exp(−R2−J (n − n0 )(2−J (n − n0 ) + 1))
n=n0

2n+1
 −1  (2n + 1)2 2n + 3
B 2
× RT G, Yn,j L2 (B)
· ,
j =1
16π 2 β3

where RT −1 G represents the gravitational potential at the surface B corresponding to the

given function G at the sphere S. In combination with Kaula’s rule we get

2n+1
 −1  (2n + 1)2 2n + 3
B 2
RT G, Yn,j L2 (B) 2
· 3
= O(ϑ n+1 n−3 ) · O(n3 ), n → ∞.
j =1
16π β

Hence, there exists a constant γ > 0 such that (for large N)

N
1 (N, J ) ≈ γ ϑ n+1 (1 − exp(−R2−J (n − n0 )(2−J (n − n0 ) + 1)))2
n=n0
∞

γ ϑ n+1 (1 − exp(−R2−J (n − n0 )(2−J (n − n0 ) + 1)))2

n=n0
∞

= γ ϑ n0 ϑ n+1 (1 − exp(−R2−J n(2−J n + 1)))2

n=1

∞ ∞

= γ ϑ n0 ϑ n+1 − 2γ ϑ n0 ϑ n+1 exp(−R2−J n(2−J n + 1))

n=1 n=1

∞
+ γ ϑ n0 ϑ n+1 exp(−R21−J n(2−J n + 1))
n=1

∞
ϑ2
= γ ϑ n0 − 2γ ϑ n0 ϑ exp(n log ϑ − R2−J n(2−J n + 1))
1−ϑ n=1
Regularized multiresolution density recovery from satellite gravity data 1015

0
0 5 10 15

Figure 7. Error bound for the approximation of the harmonic density distribution in the case of a
regularization with the Gauß–Weierstraß scaling function.

+ γ ϑ n0 ϑ exp(n log ϑ − R21−J n(2−J n + 1))

n=1
n0 +2
∞
ϑ
γ − 2γ ϑ n0 +1 exp(n(log ϑ − R(2−2J + 2−J )))
1−ϑ n=1
∞

+ γ ϑ n0 +1 exp(n(log ϑ − R(21−2J + 21−J )))

n=1
ϑ n0 +2
exp(log ϑ − R(2−2J + 2−J ))
=γ − 2γ ϑ n0 +1
1−ϑ 1 − exp(log ϑ − R(2−2J + 2−J ))
exp(log ϑ − R(21−2J + 21−J ))
+ γ ϑ n0 +1 . (9)
1 − exp(log ϑ − R(21−2J + 21−J ))
Obviously, this error estimate tends to 0 as J → ∞ which should be expected. Note that the
real regularization error in the case of the Gauß–Weierstraß scaling function is much smaller,
since we essentially simplified the series for the error bound.
The obtained error bound in (9) is plotted for R = 1 in figure 7 for the scales
J = 1, . . . , 15, where γ = 1, ϑ = 0.9 and n0 = 0 was chosen for simplicity. Note that
the fast convergence of the approximations J ∗ G to the exact solution T −1 G is clearly
demonstrated by the decay of the error bound.
For more sophisticated error estimates in the case of particular regularizations of geodetic
problems see, for example, [14, 25].

7. Numerical experiments

7.1. Noise free gravity model

For numerical purposes, the second radial derivative of the gravity model EGM96 (see [17])
at a 200 km orbit S was calculated on a Driscoll–Healy grid (see [7]) for exact integration of
1016 V Michel

spherical harmonic integrands up to degree m = 719. Note that the degrees 0–2 of EGM96
were omitted as global trend in the calculations. From the values of the second radial derivative
∂2V
∂r 2
at this point grid the convolutions


∂ 2V
J (x, y) 2 (y) dy, |x| = 0.999β (10)
S ∂r

were calculated with the corresponding Driscoll–Healy method for m = 719 (problem no 5).
The evaluation was done on an equiangular grid for x with 360 × 180 points. The chosen
scaling function was the Gauß–Weierstraß scaling function, which was calculated by the
Clenshaw algorithm (see [6]) with truncation at degree 360. The results are shown in
2
figure 8. The regularization at scale J is given by J ∗ ∂∂rV2 , where the details at scale J
are given by
2 2 2
˜ J ∗ J ∗ ∂ V = J +1 ∗ ∂ V − J ∗ ∂ V .
∂r 2 ∂r 2 ∂r 2
Note that the grey scales of the images vary such that more details can be seen. The images
should anyway, due to the missing anharmonic part, be regarded from the qualitative point of
view and not from the quantitative point of view. It is indeed surprising that the pure harmonic
part of the density nevertheless shows so many topographical and tectonic structures.
A movie of a rotating globe illustrating the approximation obtained at scale 10 is available
on the internet at http://www-user.rhrk.uni-kl.de/∼vmichel/gravimetry.html

7.2. Noisy gravity model

To illustrate the regularizing behaviour, we added noise to the used data of the second derivative
of the gravity model EGM96 at a 200 km orbit. This was realized by calculating at each point yi
2
of the discretization of the integral in (10) a random number in the range ∂∂rV2 (yi )·[−10%, 10%]
2
and adding it to the gravity value ∂∂rV2 (yi ). The same calculations that were done in the noiseless
case were again performed where more than one test per scale was made, i.e. for each scale
two different sets of random values were used. From scale to scale, the random values were
also renewed but always within the same range as described above. Figures 9 and 10 show the
results of those calculations. In the last column, the absolute value of the difference between
the recovery from the unperturbed gravitational field and the recovery from the noisy
gravitational field is plotted.
A well-known feature of regularization is the fact that by increasing the resolution one
finally reaches a point where the noise essentially deteriorates the results and the power of the
regularization process reaches its limits. Here, this is obviously the case at scale 8 where the
recovery from the noisy gravitational data does not yield satisfying results. However, up to
and including scale 7 the new method, presented in this work, allows a very stable, i.e. with
low sensitivity with respect to the added noise, solution of the discussed inverse problem.

8. Conclusions

A general concept for recovering the Earth’s harmonic density distribution out of the kth
derivative (k ∈ {0, 1, 2}) of the gravitational potential at orbit height has been presented.
The harmonic density represents this part of the Earth’s mass density that contributes to the
gravitational field. Its representation in different scales as the paper provides allows the
analysis of structures with different spatial resolutions such that at low scales continental
Regularized multiresolution density recovery from satellite gravity data 1017

2 ∂2V
Figure 8. Regularization J ∗ ∂∂rV2 (left) and details 
˜ J ∗ J ∗
∂r 2
(right) at scales J = 3 (top)
to J = 8 (bottom), plotted at a sphere with radius 0.999β.

structures and at high scales some island regions and many tectonic structures can be
seen.
Besides the generality the localizing, multiresolution property of the introduced
regularization method is an essential advantage, since huge data sets can more easily be
handled and local adaptations of the model can quickly be realized. The localization of the
 1018
Scale Density from pure EGM96 Density from noisy EGM96 absolute error

5 -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2 3 4 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045

-3 -2 -1 0 1 2 3 4 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045

6 -10 -5 0 5 10 -10 -5 0 5 10 0.05 0.1 0.15 0.2

V Michel
-10 -5 0 5 10 0.05 0.1 0.15 0.2

Figure 9. Recovered harmonic density in kg m−3 in analogy to the previous calculations from unperturbed EGM96 and EGM96 with 10% artificial noise, plotted at a sphere with radius
0.999β.
 Regularized multiresolution density recovery from satellite gravity data
Scale Density from pure EGM96 Density from noisy EGM96 absolute error

7 -30 -20 -10 0 10 20 30 -30 -20 -10 0 10 20 30 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5

-30 -20 -10 0 10 20 30 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5

8 -100 -80 -60 -40 -20 0 20 40 60 80 -400 -300 -200 -100 0 100 200 300 400 50 100 150 200 250 300 350 400 450

-300 -200 -100 0 100 200 300 400 50 100 150 200 250 300 350 400 450

1019
Figure 10. Recovered harmonic density in kg m−3 in analogy to the previous calculations from unperturbed EGM96 and EGM96 with 10% artificial noise, plotted at a sphere with
radius 0.999β.
1020 V Michel

wavelet-based approach has been proved theoretically and numerically. Moreover, numerical
tests for noisy data showed that the introduced regularization method is indeed able to stabilize
the exponential instability of the discussed inverse problem. For the first time, a multiresolution
recovery of harmonic density variations has been calculated numerically without the need of
knowing the Fourier coefficients of the potential, since data of the (second radial derivative of
the) potential on a discrete point grid have been used as input of the method. Hence, for future
analysis of the gravitational field and the density variations models of Fourier coefficients
are not needed, but discrete gravity values, calculated from the satellite measurements, can
directly be used for the calculation of the wavelet transform of the unknown quantity, which
is here the density distribution.
Challenges for future research are the extension of the method to domains which are
even closer to the true Earth’s shape such as an ellipsoid or even the inner space of a regular
surface (see [12] for some theoretical results with this respect) and the validation of some stop
strategies such as the L-curve method. Moreover, the recovered harmonic part of the density
should be combined with additional information obtainable e.g. from seismology to obtain
also an anharmonic density and, thus, a complete density model.

Acknowledgments

The author wishes to thank Professor Dr Willi Freeden and all members of his Geomathematics
Group, in particular Martin Fengler, for valuable discussions and suggestions. Moreover,
Dipl.-Math. techn. Dominik Michel deserves special thanks for calculating the second radial
2
derivative of the EGM96-potential on a Driscoll–Healy grid. Those data were used for ∂∂rV2 in
the calculation of the convolutions in (10). Last but not least, it should be mentioned that parts
of the considerations of the localization in section 5 were gratefully motivated by discussions
with Madalagama Karawitage Abeyratne.

Appendix. Scalar, vectorial and tensorial function systems on the 2-sphere

The space of positive integers (natural numbers) is denoted by N, such that N0 is the space of
non-negative integers. The symbol for the space of real numbers is, as usual, R. The vector
space Rn , n ∈ N, is equipped with the Euclidean scalar product x · y := √ ni=1 xi yi ; x =
(x1 , . . . , xn )T , y = (y1 , . . . , yn )T ∈ Rn ; and the corresponding norm |x| := x · x, x ∈ Rn .
The standard orthonormal system in R3 with respect to this scalar product is denoted by
     
1 0 0

ε = 0 ,
1  
ε = 1 ,
2  ε = 0 .
3 
0 0 1
The vector product in R3 is defined by
x ∧ y := (x2 y3 − x3 y2 , x3 y1 − x1 y3 , x1 y2 − x2 y1 )T ; x, y ∈ R3 ;
whereas the tensor product x ⊗ y := (xi yj )i,j =1,2,3 ; x, y ∈ R3 ; yields a second-order tensor
(3 × 3-matrix). The identity matrix in R3×3 is denoted by

3
i := εi ⊗ εi = (δij )i,j =1,2,3 ,
i=1

where δij is the Kronecker delta. If A, B ∈ R3×3 are tensors and x ∈ R3 is a vector, then Ax and
AB denote the usual multiplications. Moreover, the double-dot product of A = (aij )i,j =1,2,3

and B = (bij )i,j =1,2,3 is defined by A : B := 3i,j =1 aij bij .
Regularized multiresolution density recovery from satellite gravity data 1021

Such operations can be generalized

  to tensors of larger  orders.
 For instance, the tensor
product of two tensors A := ai1 ...ir i1 ,...,ir =1,2,3 and B := bj1 ...js j1 ,...,js =1,2,3 is defined by
 
A ⊗ B := ai1 ...ir bj1 ...js i1 ,...,ir ,j1 ,...,js =1,2,3 .
Moreover, the multiplication of a fourth-order tensor A and a second-order tensor B is
understood, for example, as the second-order tensor
 3 

AB := aij kl bkl (=: A : B).

k,l=1 i,j =1,2,3

Analogously, one can use the isomorphism R3×3  R9 in this case. For more details on the
calculation with tensors see [5].
We will restrict our attention to a spherical model of the Earth denoted by Bint which is
an open ball with centre 0 and radius β > 0 and is interpreted as the inner space of the sphere
B = {x ∈ R3 : |x| = β}. The corresponding outer space is denoted by Bext . Of particular
importance for our considerations will also be the unit sphere .
The space C(k) (D, Rn ); D ⊂ Rm compact, k ∈ N0 ∪ {∞}, n, m ∈ N; of all k-times
continuously differentiable functions from D into Rn and, in particular, the space C(D, Rn ) :=
C(0) (D, Rn ) of all continuous functions from D into Rn can be equipped with the maximum
norm
F ∞ := F C(D,Rn ) := max |F (x)|, F ∈ C(D, Rn ),
x∈D

where (C(D, Rn ), · ∞ ) is a Banach space. Moreover, the space L2 (D, Rn ); D ⊂ Rm

(Lebesgue) measurable, n, m ∈ N; of all classes of almost everywhere identical square-
integrable functions from D into Rn , equipped with the scalar product


(f, g)2 := (f, g)L2 (D,Rn ) := f (x) · g(x) dx; f, g ∈ L2 (D, Rn );
D
is a Hilbert space. The corresponding norm is given by f 2 := f L2 (D,Rn ) :=

(f, f )L2 (D,Rn ) , f ∈ L2 (D, Rn ). The space L2 (D, R3×3 ); D ⊂ Rm (Lebesgue) measurable,
m ∈ N; of all classes of almost everywhere identical square-integrable functions from D into
R3×3 , equipped with the scalar product


(f, g)2 := (f, g)L2 (D,R3×3 ) := f(x) : g(x) dx; f, g ∈ L2 (D, R3×3 );
D
is a Hilbert space. The corresponding norm is given by f 2 := f L2 (D,R3×3 ) :=

(f, f)L2 (D,R3×3 ) , f ∈ L2 (D, R3×3 ).
To simplify notation we will write C(D) := C(D, R), c(D) := C(D, R3 ), c(D) :=
C(D, R3×3 ), L2 (D) := L2 (D, R), l2 (D) := L2 (D, R3 ), l2 (D) := L2 (D, R3×3 ) etc for
D ⊂ R3 . Analogously, we will usually denote scalar functions by capital letters, vectorial
functions by lower-case letters and (second-order) tensorial functions by bold-face letters.
The Earth model Bint motivates a separation ansatz u(rξ ) = G(r)y(ξ ); r ∈ ]0, β], ξ ∈ ;
such that a more detailed treatment of functions on the unit sphere  is required for our
investigations. An intensive discussion of scalar, vectorial and tensorial functions on the
sphere and their approximations can be found in [9], such that we merely try to give a concise
overview of the most important tools here.
∂2
The gradient grad = ∇ and the Laplace operator , defined by x := 3i=1 ∂x 2 , can be
i
separated into a radial and an angular part in the following way:
∂ 1
∇rξ F (rξ ) = ξ F (rξ ) + ∇ξ∗ F (rξ ),
∂r r
1022 V Michel

∂2 2 ∂ 1
rξ G(rξ ) = G(rξ ) + G(rξ ) + 2 ∗ξ G(rξ );
∂r 2 r ∂r r
r ∈ R+ , ξ ∈ , F ∈ C(1) (D), G ∈ C(2) (D), rξ ∈ D ⊂ R3 . Note that a function G ∈ C(2) (D),
D ⊂ R3 open, satisfying G = 0 in D is called harmonic. The spherical parts ∇ ∗ and ∗ are
called surface gradient and Beltrami operator, respectively. Their representation in spherical
coordinates can easily be derived as well as the representation of the surface curl gradient L∗
defined by the vector product
L∗ξ F (rξ ) := ξ ∧ ∇ξ∗ F (rξ ); r ∈ R+ , ξ ∈ , F ∈ C(1) (D), rξ ∈ D ⊂ R3 .
The divergence and the curl of a vector field f ∈ c(1) (D), D ⊂ R3 , are defined by

3
∂Fj
div f := ∇ · f := ,
j =1
∂xj
 T
∂F3 ∂F2 ∂F1 ∂F3 ∂F2 ∂F1
curl f := ∇ ∧ f := − , − , − ,
∂x2 ∂x3 ∂x3 ∂x1 ∂x1 ∂x2
where f = (F1 , F2 , F3 )T . Moreover, for a function F ∈ L2 ([−1, 1]), ξ ∈ , and
a, b ∈ [−1, 1], a < b, one gets


b
F (ξ · η) dω(η) = 2π F (t) dt.
{η∈|aξ ·ηb} a
An important role in the theory of functions on the sphere is played by the space Harmn ()
of spherical harmonics, i.e. homogeneous harmonic polynomials of degree n ∈ N0 restricted
to the unit sphere . This space has the dimension 2n + 1 such that we are able to assume
that a complete L2 ()-orthonormal system {Yn,j }j =1,...,2n+1 in Harmn () is given. Whenever
we write Y with a double-index we refer to such a complete orthonormal system. Since two
spherical harmonics Yn ∈ Harmn (), Ym ∈ Harmm () of different degrees n = m are always
orthogonal, i.e. (Yn , Ym )L2 () = 0, the whole system {Yn,j }n∈N0 ;j =1,...,2n+1 is automatically an
orthonormal system in L2 (). Moreover, one can prove that such a system is also complete,
such that every function F ∈ L2 () satisfies
 
 


N 2n+1


lim F − (F, Yn,j )L2 () Yn,j  = 0.
N→∞  

 2
n=0 j =1
L ()
∞ 2n+1
We will briefly write ‘F = n=0 j =1 (F, Yn,j )L2 () Yn,j in the sense of L2 ()’.
Obviously, by defining
 
1 x
S
Yn,j (x) := Yn,j ; x ∈ S,
σ |x|
n ∈ N0 , j ∈ {1, . . . , 2n + 1}; we obtain a complete orthonormal system for L2 (S) provided
that S is a sphere with radius σ > 0 and centre at the origin.
Furthermore, the spherical harmonics have the exposed property to be the only on 
infinitely often differentiable eigenfunctions of the Beltrami operator. We have
∗ Yn = −n(n + 1)Yn
for all Yn ∈ Harmn () and all n ∈ N0 .
Using a scalar system {Yn,j }n∈N0 ;j =1,...,2n+1 we obtain a complete orthonormal vectorial
 (i) 
system yn,j i=1,2,3;n
0i ;j =1,...,2n+1
in l2 (), given by

(i)  (i) −1/2 (i) 0, if i = 1
yn,j := µn o Yn,j ; j = 1, . . . , 2n + 1; n
0i :=
1, if i ∈ {2; 3},
Regularized multiresolution density recovery from satellite gravity data 1023

where the operators o(i) ; i = 1, 2, 3; are defined by

oξ(1) F (ξ ) = ξ F (ξ ), ξ ∈ , F ∈ C(),
oξ(2) F (ξ ) = ∇ξ∗ F (ξ ), ξ ∈ , F ∈ C(1) (),
= L∗ξ F (ξ ),
oξ(3) F (ξ ) ξ ∈ , F ∈ C(1) ()
 (i) −1/2
and the normalizing constants µn , i = 1, 2, 3, n
0i , are given by

1, if i = 1
µ(i)
n :=
n(n + 1), if i ∈ {2; 3}.
It is easy to check that vector spherical harmonics o(i) Yn , Yn ∈ Harmn (), n
0i , i ∈
{1, 2, 3}, satisfy
ξ ∧ (o(1) Yn )(ξ ) = 0, ξ · (o(2) Yn )(ξ ) = 0, ξ · (o(3) Yn )(ξ ) = 0,
L∗ξ · (o(2) Yn )(ξ ) = 0, ∇ξ∗ · (o(3) Yn )(ξ ) = 0
(i)
for all ξ ∈ . Not only are the functions yn,j orthonormal in the l2 ()-sense, we also have
the Euclidean orthogonality in the following sense:
 (i)   (j ) 
oξ F (ξ ) · oξ F (ξ ) = 0 if i = j
for F ∈ C(1) (), ξ ∈ , and i, j ∈ {1, 2, 3}.
The application of the Laplace operator  and the Beltrami operator ∗ to vectorial
functions is always meant componentwise.
Finally, a scalar system {Yn,j }n∈N0 ;j =1,...,2n+1 can also be used to construct a tensorial
 
complete orthonormal system y(i,k) 2
n,j (i,k)∈{1,2,3}2 ,n
0i,k ,j ∈{1,...,2n+1} in l () by

  0, if i = k
−1/2
y(i,k) := µ (i,k)
o(i,k)
Y n,j ; j = 1, . . . , 2n + 1; n
0 i,k := 1, if ik ∈ {2, 3}
n,j n

2, if ik = 6,
where the operators o(i,k) ; (i, k) ∈ {1, 2, 3}2 ; are defined by
o(1,1)
ξ F (ξ ) = ξ ⊗ ξ F (ξ ), ξ ∈ , F ∈ C(),
o(1,2)
ξ F (ξ ) =ξ⊗ ∇ξ∗ F (ξ ), ξ ∈ , F ∈ C(1) (),
o(1,3)
ξ F (ξ ) = ξ ⊗ L∗ξ F (ξ ), ξ ∈ , F ∈ C(1) (),
(2,1)
oξ F (ξ ) = (∇ξ∗ F (ξ )) ⊗ ξ, ξ ∈ , F ∈ C(1) (),
o(2,2)
ξ F (ξ ) = (i − ξ ⊗ ξ )F (ξ ), ξ ∈ , F ∈ C(),
(2,3)
oξ F (ξ ) = (∇ξ∗ ⊗ ∇ξ∗ − L∗ξ ⊗ L∗ξ )F (ξ ) + 2(∇ξ∗ F (ξ )) ⊗ ξ, ξ ∈ , F ∈ C(2) (),
o(3,1)
ξ F (ξ ) = (L∗ξ F (ξ )) ⊗ ξ, ξ ∈ ,
F ∈ C(1) (),
 
(3,2)
oξ F (ξ ) = (∇ξ∗ ⊗ L∗ξ + L∗ξ ⊗ ∇ξ∗ )F (ξ ) + 2 L∗ξ F (ξ ) ⊗ ξ, ξ ∈ , F ∈ C(2) (),
 3 

o(3,3)
ξ F (ξ ) = (ξ ∧ εi ) ⊗ εi F (ξ ), ξ ∈ , F ∈ C()
i=1
 −1/2
and the normalizing constants µ(i,k)
n , (i, k) ∈ {1, 2, 3}2 , n
0i,k , are given by

 1, if (i, k) = (1, 1)

2,
(i,k) if (i, k) ∈ {(2, 2), (3, 3)}
µn :=

 n(n + 1), if (i, k) ∈ {(1, 2), (1, 3), (2, 1), (3, 1)}

2n(n + 1)(n(n + 1) − 2) if (i, k) ∈ {(2, 3), (3, 2)}.
1024 V Michel

A fundamental property of spherical harmonics is the addition theorem

2n+1
2n + 1
Yn,j (ξ )Yn,j (η) = Pn (ξ · η), (ξ, η) ∈ 2 ,
j =1
4π

where Pn is the Legendre polynomial of degree n ∈ N0 . The analogous vectorial and tensorial
formulations are

2n+1
 
(i) (k) (k) −1/2 2n + 1 (i) (k)
yn,j (ξ ) ⊗ yn,j (η) = µ(i)
n µn o o Pn (ξ · η)
j =1
4π ξ η
2n + 1 (i,k)
=: p (ξ, η),
4π n

(i,k)
2n+1
 (i,k) (l,m) −1/2 2n + 1 (i,k) (l,m)
yn,j (ξ ) ⊗ y(l,m)
n,j (η) = µn µn o oη Pn (ξ · η)
j =1
4π ξ
2n + 1 (i,k,l,m)
=: Pn (ξ, η),
4π
(ξ, η) ∈ 2 , n ∈ N0 . Explicit representations of p(i,k) n and P(i,k,l,m)
n in terms of Legendre
polynomials and their derivatives can be found in [9].
In analogy to the scalar case complete orthonormal systems in l2 (S) and l2 (S), respectively,
where S is a sphere with radius σ > 0, are obtained by defining
(i);S 1 (i)  x  1 x 
yn,j (x) := yn,j , y(i,k);S
n,j (x) := y(i,k)n,j , x ∈ S.
σ σ σ σ
Moreover, we define
  . l2 (S)
(i);S 
l2i (S) := span yn,j  n
0i , j = 1, . . . , 2n + 1 ,
  . l2 (S)
(i,k);S 
l2i,k (S) := span yn,j  n
0i,k , j = 1, . . . , 2n + 1

for (i, k) ∈ {1, 2, 3}2 .

References

[1] Ballani L, Engels J and Grafarend E W 1993 Global base functions for the mass density in the interior of a
massive body (Earth) Manuscr. Geod. 18 99–114
[2] Ballani L and Stromeyer D 1982 The inverse gravimetric problem: a Hilbert space approach Proc. Int. Symposium
‘Figure of the Earth, the Moon, and other Planets’ (Prague) ed P Holota p 359–73
[3] Ballani L and Stromeyer D 1990 On the structure of uniqueness in linear inverse source problems
Theory and Practice of Geophysical Data Inversion, Proc. 8th Int. Mathematical Geophysics Seminar
on Model Optimization in Exploration Geophysics ed A Vogel, A K M Sarwar, R Gorenflo and
O I Kounchev (Braunschweig: Vieweg)
[4] Ballani L, Stromeyer D and Barthelmes F 1993 Decomposition principles for linear source problems
Inverse Problems: Principles and Applications in Geophysics, Technology, and Medicine (Math. Res. 74)
ed G Anger, R Gorenflo, H Jochmann, H Moritz and W Webers (Berlin: Akademie-Verlag)
[5] Dahlen F A and Tromp J 1998 Theoretical Global Seismology (Princeton, NJ: Princeton University Press)
[6] Deuflhard P 1975 On algorithms for the summation of certain special functions Computing 17 37–48
[7] Driscoll J R and Healy R M 1994 Computing Fourier transforms and convolutions on the 2-sphere Adv. Appl.
Math. 15 202–50
[8] Fengler M J, Freeden W and Michel V 2004 The Kaiserslautern multiscale geopotential model SWITCH03 from
orbit perturbations of the satellite CHAMP and its comparison to EGM96, UCPH2002 02 0.5, EIGEN-1s,
and EIGEN-2 Geophys. J. Int. 157 499–514
Regularized multiresolution density recovery from satellite gravity data 1025

[9] Freeden W, Gervens T and Schreiner M 1998 Constructive Approximation on the Sphere—with Applications to
Geomathematics 1st edn (Oxford: Oxford University Press)
[10] Freeden W and Michel V 1999 Constructive approximation and numerical methods in geodetic research today—
an attempt at a categorization based on an uncertainty principle J. Geod. 73 452–65
[11] Freeden W and Michel V 2001 Basis aspects of geopotential field approximation from satellite-to-satellite
tracking data Math. Methods Appl. Sci. 24 827–46
[12] Freeden W and Michel V 2004 Multiscale Potential Theory (With Applications to Geoscience) (monograph)
(Boston, MA: Birkhäuser)
[13] Freeden W, Michel V and Nutz H 2002 Satellite-to-satellite tracking and satellite gravity gradiometry (advanced
techniques for high-resolution geopotential field determination) J. Eng. Math. 43 19–56
[14] Freeden W and Pereverzev S 2001 Spherical Tikhonov regularization wavelets in satellite gravity gradiometry
with random noise J. Geod. 74 730–6
[15] Freeden W and Windheuser U 1996 Spherical wavelet transform and its discretization Adv. Comput. Math. 5
51–94
[16] Hein G, Sansó F, Strykowsky G and Tscherning C C 1989 On the choice of norm and base functions for
the solution of the inverse gravimetric problem Ricerche di Geodesia Topografia Fotogrammetria CLUP
(Milano) vol 5 pp 121–38
[17] Lemoine F G et al 1996 The development of the NASA GSFC and NIMA joint geopotential model Proc. Int.
Symp. on Gravity, Geoid, and Marine Geodesy (GRAGEOMAR 1996) (The University of Tokyo: Springer)
[18] Michel V 1998 A wavelet based method for the gravimetry problem Progress in Geodetic Science, Proc.
Geodetic Week 1998 ed W Freeden (Aachen: Shaker Verlag) pp 283–98
[19] Michel V 1999 A multiscale method for the gravimetry problem—theoretical and numerical aspects of harmonic
and anharmonic modelling PhD Thesis Geomathematics Group, Department of Mathematics, University of
Kaiserslautern (Aachen: Shaker Verlag)
[20] Michel V 2002 Scale continuous, scale discretized and scale discrete harmonic wavelets for the outer and the
inner space of a sphere and their application to an inverse problem in geomathematics ACHA 12 77–99
[21] Michel V 2002 A multiscale approximation for operator equations in separable Hilbert spaces—case study
reconstruction and description of the Earth’s interior Habilitation Thesis (Aachen: Shaker Verlag)
[22] Nettleton L L 1939 Determination of density for reduction of gravity observations Geophysics 4 176–83
[23] Nettleton L L 1940 Geophysical Prospecting for Oil (New York: McGraw-Hill)
[24] Sansó F and Rummel R (eds) 1997 Geodetic Boundary Value Problems in View of the One Centimeter Geoid
(Lecture Notes in Earth Sciences vol 65) (Berlin: Springer)
[25] Schock E and Pereverzev S 1999 Error estimates for bandlimited spherical regularization wavelets in an inverse
problem of satellite geodesy Inverse Problems 15 881–90
[26] Sloan I H and Womersley R S 2004 Extremal systems of points and numerical integration on the sphere Adv.
Comput. Math. 21 107–25
[27] Tscherning C C 1996 Isotropic reproducing kernels for the inner of a sphere or spherical shell and their use as
density covariance functions Math. Geol. 28 161–8
[28] Tscherning C C and Sünkel H 1981 A method for the construction of spheroidal mass distributions consistent
with the harmonic part of the earth’s gravity potential Manuscr. Geod. 6 131–56
[29] Weck N 1972 Zwei inverse Probleme in der Potentialtheorie Mitteilungen aus dem Institut für Theoretische
Geodäsie (Universität Bonn) vol 4 pp 27–36