Professional Documents
Culture Documents
Geodesy
© Springer-Verlag 1996
Abstract. The currently practiced methods of harmonic today's representations include over 130 thousand terms
analysis on the sphere are studied with respect to aliasing (Rapp and Pavlis 1990).
and filtering. It is assumed that a function is sampled on It is a significant consequence of this fine structure in
a regular grid of latitudes and longitudes. Then, the functions that the harmonic components determined
transformations to and from the Cartesian plane yield from a finite set of function values are biased estimates of
formulations of the aliasing error in terms of spherical the true components. This is a well known phenomenon
harmonic coefficients. The following results are obtained: in signal analysis, where even though one applies
1) The simple quadratures method and related methods are orthogonal operators to the data, they cannot filter out
biased even with band-limited functions. 2) A new harmonics finer in detail than that dictated by the
method that eliminates this bias is superior to Colombo's sampling interval. This is known as "aliasing." Another
method of least squares in terms of reducing aliasing. 3) interpretation, though less precise, is that of attempting to
But, a simple modification of the least-squares model determine all (possibly an infinite number of) harmonics
makes it identical to the new method as one is the dual of from a given number of function values. In the general
the other. 4) The essential elimination of aliasing can case, this is an underdetermined problem and the solution
only be effected with spherical cap averages, not with the invariably lumps together harmonics in a prescribed way.
often used constant angular block averages. Spectral aliasing in functions defined on the line or on the
Cartesian plane is well understood; it is less obvious on a
non-Euclidean surface such as the sphere.
The spherical harmonic analysis is examined here
with the object of clearly understanding the effect of
aliasing in conventional techniques, as well as in the
Introduction
modem techniques developed by Colombo (1981) that are
The harmonic (or spectral) representation of functions used to compute today's models. In so doing, a new
on the sphere has proved useful in modeling the gravity formula for the analysis is rendered that reduces the
field, the magnetic field, the topography, the ocean tides, aliasing effect better than the least-squares approach. It is,
etc, not just for the Earth, but for any roughly spherical however, also shown that the new technique is equivalent
to a modification of the least-squares model, requiring
planet (Rapp 1977; Schmitz and Cain 1983; Kaula 1993).
some extra computations. The results also indicate clearly
Such harmonic representations enable a more precise
which types of averaging (filtering) do and do not
interpretation by wavelength of their magnitudes, their
eliminate aliasing. It is assumed throughout that the data
coherence, and their measurability. Usually, these are provided discretely at every node of a uniform grid.
functions are richly endowed with harmonics to a high (Aliasing in the case of a nonuniform data distribution
degree, decaying relatively slowly with increasing detail. was studied by Sansb 1990.) Though the errors in
For example, the Earth's gravity field, after the first few measuring the function on the sphere are not considered
harmonics (which certainly dominate), decays initially, their effect on the derived procedures is discussed
approximately linearly on the logarithmic scale, and in conclusion.
215
Preparatory Concepts
The infinite sum on the far fight in (4) is an alias of G k
A periodic function integrable over its period may be for any frequency, k, such that Ikl < k N = I(/2, where kr~ is
represented either in terms of its independent (space or the so-called Nyquist frequency. The determination of the
time) variable or, where it is continuous, in terms of its spectnma {Gk} of the function g from the discrete sequence
Fourier spectrum, being the set of coefficients in its {gs}' using (2b), is subject to an aliasing error as
corresponding Fourier series (Priestley 1981): formulated in (4). Note that if the function g does not
contain harmonics with frequency above the Nyquist
g(x) = ~ Gk e i2=kxrr , xE R (la) frequency, then there is no aliasing error. Such functions
k~-~ will be called band-limited. I t is further noted that a
function with infinite bandwidth can be filtered to make it
I"T
Gk = 1J0 g(x) e-i2=kx/rdx , ke Z (lb) band-limited. This procedure is discussed later with
respect to the particular application of spherical harmonic
analysis.
where i2 = -1, T is the period (g(x+T) = g(x)), R i s the set Obviously, similar formulas hold for periodic
of real numbers, Z is the set of integers, and the integer k functions defined on the Cartesian plane. They are
may be termed "frequency," or "wavenumber." In practice, specialized here to the case where g is a real function of
for example from measurements, one knows only discrete two variables, 0 and X, and is periodic in 0 with period
values of g and the corresponding spectrum is the Discrete and in k with period 2re. Then
Fourier Transform (DFT):
K-I o g(O,~.) = ~ ~ Gk,me i(2ke+ r~) , e,~.e R (5a)
gs = ~ Gk e i2=ks/K , s = 0 ..... K-1 (2a) k=-~ m---~
k=0
2~
o K-1 t f
G k = A x Z gs e -i2~ks/K , k = 0 ..... K - 1 (2b) Gk m = _-:-~-| I g(0,k) e -i(2k°+ m~)d0dX , k,me Z
T s=0 • "-~2nJ0 J0
where it is assumed that the function values are known on (5b)
a regular grid, i.e., one with constant spacing, Ax = T/K.
where, with * denoting the complex conjugate,
The unsymmetric transform (2a,b), where k is
nonnegative, is preferred here as being more conventional Gk,-m= G*-lqm (5c)
and avoiding the separate treatment of even and odd K.
Since the DFT is periodic with frequency period K: For a regular grid with constant spacings given by A0 =
o o ~/K and AX = 2~/M, where 0 s = sA0, ~'t = tAX, the
Gk+K = Gk, for any k (2c) Discrete Fourier Transform pair is
the sample spectrum (6b) is, for any k and m, to the coefficients of Heiskanen and Moritz (1967, p.31,
(1-76)) by
o 0o oo
2re (11a)
- - - 1 f0
~/n,m -- f0 g(0,~,) Pn,lmgCOS0) e-im~ sin0 d0 dX ;
4~rn Sn,m(0) = sin0 Pn,m(COS0) and Sn,m(0 + KS) = Sn,m(0)
~- (Pn,m(COSO))2 sin0 dO =Em (9) It is noted that because Pn,mand Sn,m are real functions,
implies that the infinite sequences {p~,m} and {OJ.~} are Substituting (12a) and (8a) into (5b) and taking note of
orthogonal for n ~ j: (18a, b), one obtains
1 2~
i m m' (18a)
More common is an attempt to incorporate the fact that
usually the data are averages over angular cells. In this
-"J0 , if m = m' case, the Legendre spectrum of the m e a n anomaly is
sought and the spherical harmonic function values in (21)
and are replaced by their averages over the cells. Either way,
(20) or its alternative with averaged spherical harmonics is
only an approximation and leads to estimates that are
lf0 ei2(k ' - k)0 dO =
I0,, iifl kk =¢ k'k' (18b) biased due to aliasing not just from frequency components
higher than the Nyquist limit, but from components of all
degrees greater than or equal to the order. Indeed, it is well
and for later use:
known (Pavlis 1988; Albertella et al. 1993) that a band-
limited function analyzed according to (21) is not
M-1
Z eifJ-
= JO, j*m recovered exactly by the estimated spectrum. This feature
m);Zt (18c)
t=o {M, j = m is demonstrated analytically below, but only for the case
218
exemplified by (21). The analogy for the case of cell- The first sum contains exactly K harmonic coefficients for
averages will be evident because the integrals of the any given m, and these are to be solved, given Gk;m. One
Legendre functions have spectra like (17), but multiplied could have separated the first two sums in (26) at some
by (sinkA0)/(kA0). integer larger than Iml + K, but then the problem is
With the aid of (13a) and (17), equation (21) becomes underdetern~oined because there are only K Fourier
coefficienotsGk,m for any given m.
o K-1 o n ImI .2o
yn,m- ~ Z G.~ (-Jk,m (22) Let Gm and Ym be ofinite K-vectors containing,
2~m k--O respectively, coefficients Gt,m and ~,m as follows:
The error in this estimate is derived by substituting (7)
oT [6 o °o]
into (22), adding and subtracting Tn,m' and making use of Gm = O,m, Gl,m ..... GK-1, (27)
(17), (19) and (20). Some manipulations yield
y y = [YI rrl ,m, ~ rrl +l,m..... "~rnl+K-l,m] (28)
~n'm=yn,m+ ~ --Z Z [~kdmlG-nl~lmlyj,m+
28m k=-~j~ ml Also, define the KxK matrix
(23)
0mHml ... 0ml+K'14ml
~-k ]tj ,vM+m
k=0 v=-.o j~ ml •
Rm = (29)
NOW shift the v=0 term to the first sum to get, with (15), [~.~ ml .~lml+K-l,lml
"'" VK-1
O
Tn,m = Yn,m + Then, the collection of equations (26) for a f'Lxedm and for
k = 0 ..... K - l i s
28m =lml~k=0 O
Gk,m = Z Z °LlvM+ml-
Irk yj,vM+m (25)
v=-~ j~ vM+m]
the similarities of the equation pairs (41), (42) and (31), W m = {e "i2rc(sk/K+ tin/M)Ik=0,...,K-1 (46)
Js=0,...,K-1; t=0,...,M-I
(32); however, the solution (41)is aliased more than the
solution (31). Because the Am are (N-Iml)x(N-Iml)
Substituting (45) into (31) and equating with (44) yields
matrices, there are fewer computations associated with
(41) (matrix inverses of diminishing dimension as m --+ A~ (Ym)T g 1 D "1 II/"T
=
~,'~m ,,rag (47)
N-I).
~
which holds for arbitrary g. Therefore, since
m, order *T
Wm (Win) = KM I, where I is the identity matrix,
*T *T
Y h m = ( W m ) (Wm) Rm (48)
Am --- KM Rm Rm (49)
I
K-I
N= K=M/2 n, degree aj,k = M ~ Pj~n~(cOS0s) Pk~ml(COS0s) (50)
Fig,2: Domain of spherical harmonic coefficients estimated by (41), s=0
the conventional least-squares approach. Aliasing is due to
coefficients Ay and ~ in their respective domains, see eq. (42).
One can start with (49), being Parseval's Theorem,
and show, first, that since Am is a full rank matrix, Rm is
A Modification To The Least-Squares Model non-singular; and, second, that solutions (31) and (44) are
equivalent. Thus, there is a perfect duality between these
By now it is evident that the difference between (41)
solutions: (44) operates directly o n the signal, g; (31)
and (31) lies in the model. Augmenting (34) to include
operates on its Fourier transform G; both yield identical
degrees up to Iml + K - 1, the model becomes
results.
M/2
g ~ Z Ym )tin (43)
m;-M/2+l Harmonic Analysis of Averages
In order to make use of all gravity data around the
where Ym is similar to Ym but now each has K columns
world and to keep the computation of spherical harmonic
instead of (N ;..Iml), and Ym is given by (28).
models tractable, usual practice is to transform the data
Correspondingly, Am becomes the KxK matrix Am; and
into mean quantities, i.e., averages over certain defined
the least-squares solution (41) becomes
blocks or areas on the Earth's surface, often delineated by
constant lines of latitude and longitude, e.g., l°xl ° or
30'x30' blocks. This yields a new function, g, in
principle, defined for every point on the sphere, but
Since as many coefficients are computed as data given, sampled, like g, on a regular grid. g has its own unique
this is no longer a "least-squares" solution. If g is band- Legendre spectrum, {yn,m}, and all the previous results
limited, then both (31) and (44) yield the true spectrum, hold with respect to this case. The question, however, is
which means that (31) and (44) are identical for any g. how to relate ~[n,mand Yn,rn, the latter being ultimately
Hence, in general, the aliasing error in this modified sought.
approach is also given by (32). o This relationship depends on the spherical area over
The 2-D Fourier transform Gm (K-vector) is a linear which the averaging is performed. It is derived for the
operator on g: uniformly weighted angular block averages (e.g., l°xl °
mean anomalies) by Gaposchkin (1980). The relationship
o -- for the generally weighted angular block average is deduced
Gm - ~1 Wm
T g (45)
as follows. Let the average be defined in terms of a
where Wm is a KMxK matrix: convolution with the general weighting function, b:
221
-0.2 , , , ~ . . . . . . . , , , , ) , , ,
Now multiply both sides by o3_~, sum over k, and use the
orthogonality (15) to get harmonic degree, n
to reduce the dimensionality of the normal matrices to be estimated coefficients. But the optimality of (61) is
inverted. based on the assumed isotropy and homogeneity of the
If (38) is written as (physical) covariance function of the data. These
assumptions are quite strong and do not hold uniformly
N
g = Y y" (58) for all frequencies. Furthermore, the covariance function
usually is based on a model with questionable fidelity at
*'.-T ~ T the high frequencies. Therefore, in practical applications
where Y =[Y_N+I ..... ~N.1] and 7 =IT-N+ 1..... ~/N-1], then
the least-squares solution with weights based on the noise 1.s.c. (for example, Colombo 1981; Gleason 1987) may
variance/covariance matrix, p-l, is given by not be significantly superior to the estimation based on
the harmonic analysis (31) or (44) with appropriate
filtering, e.g., with (57), because aliasing thus can be well
y ~Y P Pg (59)
controlled and/or eliminated.
In order to take advantage of the orthogonality (40), it is Summary
required that
The methods of harmonic analysis on the sphere were
investigated from the point of view of aliasing. The
YmP =0; ifmel (60)
aliasing error was formulated in terms of spherical
harmonic coefficients in the case when the function is
This happens only if the errors in the d a t a {gs,t} are
sampled on a regular grid of latitudes and longitudes. This
uncorrelated in longitude, and if the correlation in latitude
yielded several results, some of which are known, but
is the same for all longitudes.
perhaps more analytically illustrated, here: 1) The simple
The absence of redundancy of data in the new
quadratures method and related methods are biased even
approach obviates weighting the data, and no restrictions
with band-limited functions. 2) A new method that
need be imposed on the correlations in the data noise for
eliminates this bias is also superior to Colombo's method
either formulation, (44) or (31). The variance/covariance
of least squares in terms of reducing aliasing, because it
matrix of the estimated coefficients is obtained by simple
estimates more coefficients at little added computational
error propagation.
cost. 3) But, a simple modification of the model makes
Colombo's (1981) least-squares collocation (1.s.c)
the least squares method identical to the new method as
method of harmonic analysis, in theory, is supposed to
one is the dual of the other. The new method solves for
yield the best estimator of the Legendre spectrum for a
as many coefficients as data and, in the presence of data
given set of data. In case the data (again, globally and
noise, it need not use fabricated noise covariances to make
regularly distributed on a grid) are perfect measurements
the solution numerically tractable. 4) The essential
(no measurement noise), 1.s.c. minimizes the aliasing
elimination of aliasing can only be effected with spherical
error. The 1.s.c. estimator is also of the quadratures type,
cap averages, not with the often used constant angular
as defined by Colombo (1981, p.43, eq.2.60):
block averages.
K-1 M-1
y_, y_., TIn,m gs,t e (61) Appendix - The C o m p u t a t i o n of t3~'m
s=0 t--0
Analogous to (2b)
where ~i~,mare "weights" to be determined by inverting a K-t
°nm
KxK matrix of discrete spectral components of the Pk' = -1- ~ Pn.m(COS0s)e "i2~ksm , k = 0 ..... K-1
Toeplitz/circulant autocovariance matrix of the data. K s=0
(A.1)
Although originally formulated to estimate only
coefficients up to degree and order N, it is easily extended
to estimate any number of coefficients, such as those in Also define the Discrete Fourier Transforms of cos0 and
(28). Again, the formulation (61) is preserved in the sin0 on the interval (0,re):
presence of data noise only if the noise covariance matrix
o K-I o
has a structure identical to the physical autocovariance ~k = I_._~ COS0s e"i2~ks/I¢, ~k+K = ~k (A.2a)
matrix of the data. K s=0
Considering (37) and (45), it is readily seen that all
K-I o o
the other methods discussed here, namely (22), (31), (41), ~k° = K1s=0
~ sin0s e "i2~ks/K , ~k+K = ~k (A.2b)
and (44) are of the quadratures type, but with different
weights; and those in (61) are the best in terms of
minimizing the aliasing error for a particular set of The recursion formulas for the normalized associated
223