Journal of Geodesy (2002) 76: 279299
DOI 10.1007/s00190-002-0216-2
A unied approach to the Clenshaw summation and the recursive
computation of very high degree and order normalised associated
Legendre functions
S. A. Holmes, W. E. Featherstone
Department of Spatial Sciences, Curtin University of Technology, GPO Box U1987, Perth, WA 6845, Australia
e-mail: holmes@vesta.curtin.edu.au/w.featherstone@curtin.edu.au; Tel.: +61-8-9266-2218; Fax: +61-8-9266-2703
Received: 30 June 2000 / Accepted: 12 June 2001
Abstract. Spherical harmonic expansions form partial
sums of fully normalised associated Legendre functions
(ALFs). However, when evaluated increasingly close to
the poles, the ultra-high degree and order (e.g. 2700)
ALFs range over thousands of orders of magnitude. This
causes existing recursion techniques for computing values
of individual ALFs and their derivatives to fail. A
common solution in geodesy is to evaluate these expan-
sions using Clenshaws method, which does not compute
individual ALFs or their derivatives. Straightforward
numerical principles govern the stability of this technique.
Elementary algebra is employed to illustrate how these
principles are implemented in Clenshaws method. It is
also demonstrated how existing recursion algorithms for
computing ALFs and their rst derivatives are easily
modied to incorporate these same numerical principles.
These modied recursions yield scaled ALFs and rst
derivatives, which can then be combined using Horners
scheme to compute partial sums, complete to degree and
order 2700, for all latitudes (except at the poles for rst
derivatives). This exceeds any previously published result.
Numerical tests suggest that this new approach is at least
as precise and ecient as Clenshaws method. However,
the principal strength of the new techniques lies in their
simplicity of formulation and implementation, since this
quality should simplify the task of extending the approach
to other uses, such as spherical harmonic analysis.
Keywords: Spherical Harmonic Expansions Fully
Normalised Associated Legendre Functions Clenshaw
Summation Recursion Horners Scheme
1 Introduction
Current geodetic practice is witnessing an increase in the
construction and use of ultra-high-degree spherical
Correspondence to: S.A. Holmes
harmonic expansions of the geopotential or topography.
For example, Wenzel (1998) released coecients up to
degree 1800, which were empirically derived to describe
the gravitational potential of the Earth. Wenzel (1998)
states that the maximum degree of 1800 for the spherical
harmonic model was set by the numerical stability of the
recursion algorithm adopted to compute the required
fully normalised associated Legendre functions (ALFs).
The recent interest in synthetic Earth gravity models,
used for comparing and validating gravity eld
determination techniques, has already seen the use of
ultra-high-degree spherical harmonic expansions. These
have taken the form of simple eects models (see e.g.
Featherstone 1999; Novak et al. 2001) for which syn-
thetic geopotential coecients up to degree and order
2700 and 2160, respectively, were produced without re-
ference to a mass distribution. There is also interest in
source models in which synthetic geopotential coe-
cients are generated by analytical or numerical New-
tonian integration over a synthetic global density
distribution and topography (see e.g. Pail 1999).
Hybrids of source and eects models also exist. For
example, Haagmans (2000) combines empirically de-
termined coecients with synthetic ones derived from
numerical integration over isostatically compensated
source masses to degree and order 2160. Lastly, other
scientic disciplines, such as meteorology, quantum
physics and electronic engineering, are also showing
increased interest in high-degree spherical harmonic
modelling and analysis.
The numerical means for including the necessary ALFs
constitutes the principal challenge to evaluating ultra-
high-degree spherical harmonic expansions. Therefore, it
is timely to critically examine the accuracy and numerical
eciency of algorithms that compute individual ALFs
and their partial sums.
1.1 Spherical harmonic expansions
Truncated spherical harmonic expansions of a function,
or its derivatives, reduce to sums S d of ALFs, or their
derivatives, respectively. These are
280

X
M M  n
X a
S d c Xd
m 1 Xma C
nma P
nm h 7
m0 nl
r

where When evaluating gravimetric quantities (e.g. disturbing

2 
potential, geoid heights, gravity anomalies, etc.) in a
X d
Xma cos mk for a 1 sequence of points for which r and h are constant (i.e.
d
Xm d 2
a1 Xma sin mk for a 2 along a geodetic parallel), the form of Eq. (5) is
numerically more ecient than that of Eq. (4)
and (cf. Tscherning et al. 1983). This is because each Xma in
Eq. (3) is independent of k, and thus need only be
X
M
evaluated once for each parallel. If all such computation
d d
Xma E
nma P
nm h 3
nl
points are equally spaced in longitude, further numerical
eciencies can be achieved through application of the
For arguments of spherical polar coordinates (r, k, h) recursion algorithm developed by Rizos (1979). Abd-
and for integer degree n  0 and order 0  m  n: M is Elmotaal (1997) contains a re-derivation of this
the maximum nite degree of the spherical harmonic algorithm which demonstrates that, contrary to the
expansion; l is an integer that may vary with m; c is a approach of Rizos (1979), the algorithm can be applied
real-numbered constant; E
nma is a real number in full without prior rotation of the geopotential
incorporating the fully normalised spherical harmonic coecients.
coecients, C
nm1 and C
nm2 ; P
nm h are the fully
normalised ALFs; the superscript d indicates the
dth derivative with respect to h, or denite integration 1.2 Numerical considerations when evaluating Eq. (1)
(d 1) between two parallels. This paper deals only
with undierentiated functions (d 0) or rst deriva- The simplest approach to evaluating Eq. (1) is to use
tives of these functions (d 1). For d 0, the super- standard recursion relations, such as those found in
script d is omitted. Thus S 0 , X0
0
0 Colombo (1981) and described in Sect. 2.1 of this paper,
m , Xma and Pnm h
to compute the required values of P
nmd
h. These values
are written S, Xm , Xma and P
nm , respectively. First can then be multiplied by the corresponding values of
1
derivatives of these quantities are written S 1 , X1
m , Xma
d
E
nma to yield the intermediate values for Xma in Eq. (3),
d
and P
nm
1
h, respectively. The general notation S d , and hence Xm in Eq. (2), which are then used to
d d
Xm , Xma and P
nm d
h is used whenever a textual or compute the nal sums S d in Eq. (1).
The principal problem with this approach is that,
mathematical reference applies to both the undier- for ultra-high values of M (e.g. 2700), the absolute
entiated quantities and the rst derivatives simulta- values of P
nm h will range over thousands of orders of
neously. magnitude. For example, Fig. 1 shows that, for
The example of a truncated spherical harmonic M 2700, jP
nm d
hj ranges over 5000 orders of
expansion of the gravitational potential V r; h; k is magnitude towards the poles (h ! 0 or h ! 180 ; i.e.
instructive here. Often, it is written as cos h ! 1. This is impractical, because the Institute of
Electrical and Electronic Engineers (IEEE) standard
M  n X
GM GM X a n
754 for binary oating-point arithmetic (cf. Coonen
V r; h; k C
nm1 cos mk 1980) only allocates eight bytes to store each double
r r n2 r m0
precision oating-point number (R). Thus, jRj may
C
nm2 sin mkP
nm h 4 only take values within the range of 10310 < jRj <
10310 . Any computed value where jRj < 10310 will
where GM is the product of the Universal gravitational
constant and the mass of the Earth. Alternatively,
Eq. (4) may be written as
"
M  n
GM GM X M X a


V r; h; k cos mk Cnm1 Pnm h
r r m0 nl
r
#
XM  n
a


sin mk Cnm2 Pnm h 5
nl
r

where l is either 2 or m, whichever is the greater.

Relating Eq. (5) to the form of Eqs. (1)(3) yields
 an

; for a 1
C
E
nma ar n nm1 6

for a 2
r Cnm2 ;
Fig. 1. Logarithm plot of maximum (upper line) and minimum (lower
and line) values of jP
nm hj; 8n; m  2700
 281

underow and be set to zero, whilst any computed 2 Forward recursions for the calculation of ALFs
values for which jRj > 10310 will overow and be
designated not a number (NaN), such that any The most direct approach for evaluating S d [Eq. (1)]
subsequent computation which employs this R will employs a recursive algorithm to compute P
nm h. Values
also be so designated. of P
nm
1
h, if required, are then computed directly from
Underow in the computation of any P
nm d
h excludes two previously computed values of P
nm h. These values
the corresponding (matching degree and order) coe- of P
nm
d
h are multiplied by the corresponding E
nma terms
d
cients from contributing to S d , whereas an overow in to yield the required series values of Xma [Eq. (3)], which
d
the computation of one or more P
nm d
h prevents any subsequently yield Xm [Eq. (2)] and hence S d [Eq. (1)].
d
result for S from being achieved at all. Thus, IEEE The recursion relations for P
nm h can be obtained by
double precision only permits a maximum range of fully normalising standard relations for (un-normalised)
620 orders of magnitude within which to compute and Pnm h, which can be found, for example, in Magnus
store the required P
nmd
h values. Figure 1 shows that, et al. (1966) or Abramowitz and Stegan (1972). The full
for high latitudes, the range of values taken by the normalisation is given by [adapted from Heiskanen and
P
nm h values for M 2700 will eventually exceed Moritz 1967, Eq. (173)
the range of magnitudes capable of being stored within s
the IEEE double-precision format. Scaling all of the k2n 1n  m!
computations upwards by a factor of 10200 allowed P
nm h Pnm h 8
n m!
Wenzel (1998) to compute values of P
nm h to a max-
imum M 1900 for 20  h  160 in IEEE double where k 1 for m 0 and k 2 for m > 0. Similarly,
precision. quasi-normalised values of P~nm h are related to Pnm h
A numerically more stable alternative to these by (cf. Tscherning and Poder 1982)
standard recursion relations is the Clenshaw (1955) s
method (cf. Tscherning and Poder 1982; Gleason n  m!
1985; Deakin 1998). In Sect. 3, it will be shown that P~nm h Pnm h 9
n m!
this can be used to evaluate S in Eq. (1) up to
M 2700 (0  h  180 ), as well as S 1 up to Inspection of Eqs. (8) and (9) shows that
M 2700 (0 < h < 180 ). Standard derivations of p
Clenshaws method (cf. Gleason 1985; Deakin 1998) P
nm h k2n 1 P~nm h 10
utilise matrix algebra, and generally focus on the
means by which this method can be used to evaluate The relationships in Eqs. (8), (9) and (10) also hold for all
partial sums, S d , without computing individual values dth derivatives d
d ~d
nm
h, Pnm h and
Pnm h. However,
of P
nm
d
h. Such derivations are complete, concise and
rigorous, but they also obscure the numerical princi-
ples upon which the stability of the Clenshaw sum-
mation is based. These principles are quite simple,
both in concept and in application.
This paper shows how existing algorithms for
computing ALFs and rst derivatives are easily modied
to incorporate these same numerical principles. The
modied algorithms can be used to compute scaled
ALFs and their rst derivatives, which can then be
combined using Horners scheme (cf. Harris and Stocker
1998) to yield values for the required partial sums, S in
up to M 2700 (0  h  180 ), as well as S 1 up to
M 2700 (0 < h < 180 ). Straightforward examples
and elementary algebra are then used to illustrate the
means by which these numerical principles are
implemented in Clenshaws method.
Results from numerical tests, presented in Sect. 4,
suggest that the modied algorithms are at least as
ecient and precise as the standard Clenshaw techni-
ques for evaluating partial sums of P
nm h or P
nm 1
h.
However, it is the intuitive simplicity of the new ap-
proaches, as well as the fact that they compute in-
dividual scaled values of P
nm d
h, which constitutes their
principal strength over the standard Clenshaw meth-
ods. These two properties should simplify the extension
of these new approaches to other tasks, such as stabi-
lising current techniques for spherical harmonic ana-
lysis (e.g. Lesur and Gubbins 1999).
cosPmk
unlike P
nm h sin mk , P
nm h sin mk is not normalised
d6 0 cos mk
in the correct usage of the word since its average squared
value integrated over the unit sphere is not unity. This
paper focuses solely on the computation of fully normal-
ised ALFs and their rst derivatives. No numerical tests
were conducted for the above quasi-normalisations.
However, it is a trivial task to apply Eq. (10) to the
algorithms presented in this paper.
2.1 Standard forward column methods
The most popular recursive algorithm used for computing
P
nm h in geodesy can be obtained by fully normalising,
for example, Magnus et al. [1966, Eq. 4.3.3(2)]. This full
normalisation yields a recursion that computes non-
sectoral (i.e. n > m) P
nm h from previously computed
P
nm h. This recursion is given as (cf. Colombo 1981)
P
nm h anm t P
n1;m h  bnm P
n2;m h; 8 n > m
11
where t cos h
s
2n  12n 1
anm and
n  mn m
s
2n 1n m  1n  m  1
bnm 12
n  mn m2n  3
282

The sectoral (i.e. n m) P
mm h serve as seed values for 1 1 

P
nm h ntP
nm h  fnm P
n1;m h ; 8 n  m 15
the recursion in Eq. (11). These are computed p using the u
initial values P
0;0 h 1 and P
1;1 h 3u, where
u sin h. The higher degree and order values of where
P
mm h are then computed using the recursion (cf. s
Colombo l981) n2  m2 2n 1
fnm 16
r 2n  1

2m 1

Pmm h u Pm1;m1 h; 8 m > 1 13 For all sectoral P
mm
1
h, fmm 0 and Eq. (15) reduces to
2m
such that 1 t

P
mm h m Pmm h; 8 m  0 17
r u
p Ym
2i 1


Pmm h u m
3 ; 8m1 14 The rst derivative of Eq. (13) with respect to h also
i2
2i gives Eq. (17). For all n  M, each P
nm h of a given m
Q can be computed using Eqs. (11) and (13). These values
where is the product symbol (see e.g. Abramowitz
and Stegan 1972). can be substituted into Eq. (15) to compute, without the
The complete recursion process in Eqs. (11) and (13) need for any further recursion, all P
nm
1
h of the same
may be visualised using the lower triangular matrix in order m and 8 n  m.
Fig. 2, where each circle corresponds to a particular
combination of n and m. Thus, each circle represents a
value of P
nm h, as well as the corresponding pair of re- 2.2 Standard forward row methods
cursive terms anm t and bnm . Note that in Fig. 2, the de-
gree increases in rows down, the order increases in The next approach is termed the standard forward row
columns to the right, and the diagonal elements of the recursion (Fig. 3), and appears to be rarely used in
matrix are the sectoral values. The recursion in Eq. (11) geodesy. As with the standard forward column recursion
computes P
nm h of constant m (a column in Fig. 2) and (Sect. 2.1), the sectoral P
mm h serve as seed values for
sequentially increasing n [or down and away (i.e. for- the standard forward row recursion, and can be
ward) from the diagonal in Fig. 2]. Thus, Eq. (11) will be computed using Eq. (13). However, the standard
referred to as a standard forward column recursion. This forward row recursion computes non-sectoral P
nm h of
nomenclature will be employed throughout the paper. constant n (a row in Fig. 3) and sequentially decreasing
It appears from Eq. (11) that the computation of the m [to the left (i.e. forward) from the diagonal in Fig. 3].
rst value of P
m1;m h forward from the sectoral diag- Full normalisation of Magnus et al. [Eq. 4.3.3(1)] and
onal (Fig. 2) requires a value of P
m1;m h to be multi- substituting P
nm h 1m P
nm h yields
plied by the recursive term bm1;m . This P
m1;m h does 1  t

not exist for ordinary P
nm h. However, the correspond- P
nm h p gnm Pn;m1 h  hnm P
n;m2 h ;
ing bm1;m coecient in Eq. (11) is always zero, thereby j u
allowing the (non-existant) P
m1;m h to be disregarded. 8n>m 18
For a forward column computation of P
nm 1
h, nor-
malisation of Magnus et al. [1966, Eq. 4.3.3(9)] gives (cf. where j 2 for m 0 and j 1 for m > 0, and
Colombo 1981) 2m 1
gnm p
n  mn m 1

Fig. 2. Schematic of the recursion sequences employed in the

standard, rst modied and second modied forward column Fig. 3. Schematic of the recursion sequences employed in the
algorithms to compute P
nm h, P
nm h=um and P
nm h=Pmm h standard and modied forward row algorithms to compute P
nm h
respectively and P
nm h=um , respectively
 283

and magnitudes capable of being stored in IEEE double

s precision, thereby resulting in an underow. The failure
n m 2n  m  1 to compute and store values of P
mm h means that these
hnm 19 cannot serve as seed values for the standard forward
n  mn m 1
column and forward row recursions (Sect. 2.1 and 2.2,
Using the same argument to that introduced for the respectively). This ultimately limits the ranges of h and
forward column recursion, the non-existant value of M over which these recursions can be used, thereby
P
n;n1 h required to compute P
n;n1 h in Eq. (18) may restricting the practical application of spherical harmo-
be disregarded because the corresponding recursion nic expansions of ultra-high M at high latitudes.
coecient, hn;n1 , is always zero.
Note that, to compute P
nm h using the forward row
recursion, Eq. (18) uses the corresponding sectoral values 2.4 Other normalisations and the Edmonds recursion
of the same n, rather than the same m, as seed values. In
this case, these sectoral values are more correctly denoted Belikov (1991) and Belkov and Taybatorov (1991)
by P
nn h, which may be written in the form of Eq. (14) as present a suite of recursive algorithms for computing
r the quantities P^nm
d
, where P^nm
d
are related to un-
p Yn
2i 1 d
normalised Pnm according to the modied normalisation
P
nn h u n
3 ; 8n1 20
i2
2i
n!
P^nm h 2m Pnm h 25
The P
nm
1
h are obtained directly from P
nm h of n m!
matching n by fully normalising Abramowitz and Stegan
[1972, Eq. (8.5.2)] to yield However, this approach is also subject to numerical
limitations. As the computation approaches the poles,
1 t the range of jP^nm
d
hj is comparable to that of jP
nm
d
hj,
P
nm h m P
nm h  enm P
n;m1 ; 8 n  m 21
u thereby resulting in an underow in IEEE double
precision. For example, the sectoral values are given by
where
s P^mm h um sinm h 26
n m 1n  m
enm 22 which for h 1 and M 2700 yields values that range
j
from 1 to 104747 . Therefore, employing the normal-
For all sectoral P
mm
1
h, or equivalently P
nn
1
h, emm 0 isation in Eq. (25) cannot solve the numerical problems
and Eq. (21) reduces to Eq. (17). discussed in Sect. 1.2.
Risbo (1996) claims that the Edmonds (1957) recur-
sion for D-matrices can be used to compute fully nor-
2.3 Numerical problems with the standard malised ALFs up to degree 200 000. However, the
forward methods description of the test results which support this claim
indicates that these computations were only performed
Even when applied in IEEE double precision, both the at the equator, although the point is not clear. Never-
standard forward column [Eq. (11)] and standard theless, Risbo (1996) includes a Fortran 77 subroutine
forward row [Eq. (18)] recursions will underow for for this recursion. It was found that, when implemented
M > 1900 in the co-latitude range 20  h  160 . in IEEE double precision, the Risbo subroutine under-
The numerical instability of both these forward ows for all polar distances h < 50 for M 2700. This
recursions is noted in the geodetic literature (e.g. was not suprising, given that the corresponding ALFs
Gleason 1985) and elsewhere (e.g. Libbrecht 1985). cannot even be stored in IEEE double precision
The cause of this instability is revealed by examining (Sect. 1.2), irrespective of the algorithm used to compute
Eq. (14), which is rst partitioned into the factors um
and Pm , such that
P
mm h um Pm ; 8 m  1 23
where um sinm h, and
r
p Ym
2i 1
Pm 3 8m1 24
i2
2i

Inspection of Fig. 4 shows that the Pm factor introduces

no computational diculties for an arbitrarily ultra-high
value of m 5400. In contrast, the um term in Eq. (23)
becomes increasingly small as u ! 0 (i.e. towards the
poles) and as m increases. Accordingly, the high degree
and order values of P
mm h will exceed the range of Fig. 4. Variation of Pm [Eq. (24)] with order (m)
284
"( "( ) #
them. Therefore, although it may be possible to in- X
M d d
d d XM XM1
corporate the new approaches presented here into the S c Xm c  u M1 u
uM u
Risbo (1996) approach, it is clear that the Edmonds m0
! ) #
recursion alone does not solve the numerical problems d d d
XM2 X2 X d
encountered towards the poles as described in Sect. 1.2. M2 u    2 u 11 u cX0
u u u
31
2.5 The modied forward row method
From Eq. (31), the running total is repeatedly multiplied
d
A simple, yet eective, method by which this problem of by u upon the addition of each Xm =um term. This allows
underowing P
mm h may be avoided is to eliminate the d
the sum S to be computed directly from the values of
um term from the recursion process in Eq. (13). To this d
Xm =um , and so avoids the need to compute under-
end, Libbrecht (1985) adapted the standard forward row owing values of um as u ! 0 and m increases. This will
recursion (Sect. 2.2) to yield a modied forward row be demonstrated numerically in Sect. 4.
recursion that computes the quantities P
nm h=um . A
recursive algorithm which computes the non-sectoral
P
nm h=um is obtained by dividing Eq. (18) by um to give 2.6 Modied Forward Column Method
 
P
nm h 1 P
n;m1 h

2 Pn;m2 h
p gnm t  hnm u ; Values of P
nm
d
h=um may also be computed using what
um j um1 um2 will be termed the rst modied forward column
8n>m 27 recursion. To eect this computation, Eq. (28) is retained

to compute the sectoral values of P
mm h=um . Equations
Equation (27) is seeded by the sectoral Pmmumh, the (11) and (15) are divided by um , to give, respectively
recursive algorithm for which is obtained by dividing
Eq. (13) by um to give P
nm h P
n1;m h P
n2;m h
anm t  bnm ; 8n>m 32
r u m u m um
P
mm h 2m 1 P
m1;m1 h
m
; 8m>1 28  
u 2m um1 P
nm
1
h 1 P
nm h P
n1;m h
p nt  fnm ; 8nm 33
The value P
1;1 h=u1 3 serves as the seed for um u um um
Eq. (28). Equation (23) yields
A variation of the rst modied forward column
P
mm h recursion, herein termed the second modied forward
Pm ; 8 m  1 29 column recursion, is to compute values of P
nm d
h=
um



Pmm h in which the entire sectoral value of Pmm h has
Libbrechts (1985) original formula for computing been eliminated, rather than just the problematic um
P
nm h=um diers
p from Eq. (27) in that it does not component. An immediate result is that all the sectoral
include the 1= j term on the right-hand side. This is values of P
mm h=P
mm h equal 1. Dividing Eqs. (11)
because Libbrecht (1985) employs a dierent normal- and (15) by P
mm h gives, respectively
isation of Pnm h in which k 1; 8 m. Moreover,
Libbrecht (1985) focuses solely on the actual computa- P
nm h P
n1;m h P
n2;m h
a t  b ; 8n>m 34
tion of the values of P
nm h=um . This paper shows how P
mm h
nm
P
mm h
nm
P
mm h
Libbrechts (1985) method is easily extended to compute
the quantities P
nm 1
h=um . It also provides simple  
means by which values of P
nm d
h=um can be applied P
nm
1
h 1 P
nm h P
n1;m h

nt
 fnm
;
in practice without dealing with the unmanageably small Pmm h u Pmm h Pmm h
um terms for u ! 0. 8nm 35
The modied forward row recursion for computing
P
nm
1
h=um is obtained by dividing Eq. (21) by um to Comparison of Eq. (11) with Eqs. (32) and (34) shows
yield these recursions to be of identical form; similarly for the
comparison of Eq. (15) with Eqs. (33) and (35). That is,
P
nm
1
h t P
nm h P
n;m1 h for Eqs. (32) and (34), and for Eqs. (33) and (35), the
m  enm u 8nm 30
u m u u m um1 entire computation has simply been divided by um or
P
mm h, respectively. Thus any computer program that
such that all P
nm 1
h=um may be computed directly
from previously computed P
nm h=um of the same m. already employs Eqs. (11) and (15) to compute P
nm d
h is
In order to evaluate Eq. (1), the values of quickly adapted to compute P
nm h=u or P
nm
d m d
h=
P
nm
d
h=um are multiplied by the corresponding values P
mm h by simply altering the sectoral subroutine to


of Enma [Eq. (3)] and the resulting products summed to return P
mm h=um Pm or P
mm h=P
mm h 1,
d d d
yield values of Xma =um and Xm =um , instead of Xma and respectively.
d
Xm , respectively. In order to compute S d , Eq. (1) is To evaluate S d in Eq. (1), the required P
nm
d
h=
factorised using Horners scheme in terms of u

Pmm h of the same m may be multiplied by their
 285

corresponding values of E
nma [Eq. (3)] and the results for all of the modied forward methods introduced thus
d
added to give values of Xma =P
mm h and subsequently far simply by scaling all of the computations downwards
d d
Xm =P
mm h. These Xm =P
mm h can be multiplied by by a global scale factor of 10280 . This is achieved for all
Pm in Eq. (14) to yield Xd m
for use in Horners of the forward methods by simply multiplying the
m =u
sectoral values of P
mm h=um or P
mm h=P
mm h by
scheme in Eq. (31). Alternatively, inspection of Eq. (14) 10280 , and using these scaled sectoral values as the
shows that P
mm h may be factorised into recursive seeds in place of the original values. As such,
r this scale factor propagates linearly through all sub-
p Ym
2i 1 Y m


Pmm h u 3m
Ui ; 8 m  1 36 sequent computations, thereby generating values of
2i
i2 i1 P
nm
d
h=um  10280 or P
nm
d
h=P
mm h  10280 ,
where respectively. These scaled values are used to form
d
( p Xm =um  10280 or Xd

m =Pmm h  10
280
which,
3
u; i1 d d
when used in place of Xm =u or Xm =P
mm h in
m
Ui q2i1 37
2i u; 8i>1 Horners scheme in Eqs. (31) or (38), respectively, will
yield values of S d  10280 . These are multiplied by
Thus, Eq. (31) may be written using Horners scheme in 10280 to yield S d .
terms of Um Importantly, this global scaling allows, in IEEE
X
M double precision, the computation of spherical harmonic
S d c d
Xm series S for 0  h  180 and S 1 for 0 < h < 180 up
m0 to M 2700. Note that the spherical coordinate system
"( "( ) # renders partial sums S 1 indeterminate at the poles,
d d
XM XM1 since here the meridian tangents no longer uniquely
c  UM

P
M;M h PM1;M1 h dene the direction of the derivative. There are useful
! ) ways around this problem (see e.g. Tscherning 1976),
d d
XM2 X2 but for the sake of continuity they will not be considered
 UM1
UM2   

PM2;M2 h P2;2 h here. Thus, for the remainder of this paper, no partial
# sums S 1 will be computed at the poles.
d
X d
 U2
1 U1 cX0 38
P1;1 h

Due to the numerically stable behaviour of Pm (Fig. 4),

using Pm in this way, rather than multiplying the values
of Xd

m =Pmm h by Pm to yield Xm =u
d m
for use in
d
Eq. (31), gives identical results for S when performed
in IEEE double precision. Moreover, the rst and
second modied forward column recursions are essen-
tially the same, diering only in the treatment of Pm ,
which is irrelevant to the numerical stability of each
algorithm. The second modied forward column
recursion [Eqs. (34) and (35)] and the implementation
of Horners scheme in Eq. (38) are introduced here Fig. 5. Logarithm plot of maximum (upper line) and minimum (lower
line) values of jP
nm h=um j; 8 n; m  2700
primarily because of their relevance to the Clenshaw-
based methods, discussed in Sect. 3. For the purposes of
numerical testing (Sect. 4.2, 4.3 and 4.4), the rst and
second modied forward column methods will be
treated as a single modied forward column recursion.

2.7 IEEE overows and global scale factors

in forward methods

The entire ranges of maximum and minimum values

taken by jP
nm h=um j and jP
nm
1
h=um j, for M 2700,
are shown in Figs. 5 and 6, respectively. Inspection of
Figs. 5 and 6 indicates that further factorisation is
required to prevent the computations from overowing Fig. 6. Logarithm plot of maximum (upper line) and minimum (lower
in IEEE double precision. Overows can be prevented line) values of jP
nm
1
h=um j; 8 n; m  2700
286
3 Clenshaw-based methods
3.1 The forward column factorisation
It is instructive at this stage to consider a simple example
of the summation described in Eq. (3). The rectangle in
Fig. 2 contains circles representing values of P
nm h and
E
nma for m 2 and 2  n  5. A summation of these
elements in the form of Eq. (3) may be expanded as
X
5
X2;a E
n;2;a P
n2 h E
2;2;a P
2;2 h E
3;2;a P
3;2 h
n2
E
4;2;a P
4;2 h E
5;2;a P
5;2 h
The recursion relation in Eq. (11) gives the P
nm h
required in Eq. (39); these are
 
P
3;2 h a3;2 t P
2;2 h 40
P
4;2 h a4;2 tP
3;2 h  b4;2 P
2;2 h
P
5;2 h a5;2 tP
4;2 h  b5;2 P
3;2 h
Substitution of Eq. (40) into Eq. (41) yields
 
P
4;2 h a4;2 a3;2 t2  b4;2 P
2;2 h
Likewise, substitution of Eqs. (40) and (43) into Eq. (42)
gives
 
P
5;2 h a5;2 a4;2 a3;2 t3  a5;2 b4;2 t  a3;2 b5;2 t P
2;2 h
Therefore, each P
n;2 h can be factored into two
components: the seed P
2;2 h value and the aggregation
of al;2 t and bl;2 recursive terms [within the square
brackets in Eqs. (40), (43) and (44)], which constitute
polynomials in t cos h. This factorisation will be used
in Sect. 3.2 to introduce the reverse column algorithms.
Note that l rather than n is used here to denote the
degree of the recursive terms, since each P
n;2 h of degree
n is comprised of an aggregation of all of the al;2 t and
bl;2 for 3  l  n. In general, any P
nm h may be factored
into P
mm h and an aggregation of all of the alm t and
blm terms for m 1  l  n. Inspection of Eqs. (40),
(43) and (44) shows that these aggregations are simply
values of P
nm h=P
mm h, which are generated using
Eq. (34). Substitution of Eqs. (40), (43) and (44) into
Eq. (39) gives
(
   
X2;a E
2;2;a E
3;2;a a3;2 t E
4;2;a a4;2 a3;2 t2  b4;2
)
 
E
5;2;a a5;2 a4;2 a3;2 t3  a5;2 b4;2 t  a3;2 b5;2 t P
2;2 h
The quantity in curly brackets in Eq. (45) equates to
X2;a =P
2;2 h, which is used to form X2 =P
2;2 h or X2 =u2
(Sect. 2.6) for use in Eqs. (31) or (38), respectively.
3.2 Reverse column method
Results from timing tests presented in Sect. 4.3 show the
reverse column methods, described below, to be highly
inecient in comparison with the other approaches
presented in this paper for evaluating the required
partial sums. The primary reason for describing the
reverse column methods here is because these methods
incorporate characteristics of both the modied forward
column methods (Sect. 2.6) and the standard Clenshaw
methods (Sect. 3.3). Thus, the reverse column methods
are used here to highlight the basic similarities and
dierences between these two approaches.
39 To compute any value of P
nm h=P
mm h, the second
modied forward column recursion [Eq. (34)] aggregates
the necessary alm t and blm recursive terms in the se-
quence of increasing degree l (sequentially down each
column in Fig. 2). An alternative is to reverse this process
and apply these same recursive terms in the sequence of
decreasing degree l (sequentially up each column in
41 Fig. 2). That is, a recursion may be employed whereby the
alm t and blm recursive terms for which l n (i.e. the
42 largest value of l) are applied rst and the recursive terms
for which l m 1 (i.e. the smallest value of l) are ap-
plied last. This will be called a reverse column recursion,
and is illustrated schematically in Fig. 7.
43 It is now necessary to introduce the recursive algo-
rithm
slma al1;m t sl1;m;a  bl2;m sl2;m;a ylma 46
where ylma are predetermined, real-numbered constants
44 (described later), and the subscript a (a 1; 2) is
included here as it will be referred to later when
discussing the standard Clenshaw methods. Equation
(46) is used as follows. The recursion begins at the
computation of a chosen snma . For this initial computa-
tion, sn1;m;a and sn2;m;a are set to predetermined values
and then used in the rst recursion to yield snma .
Equation (46) is then used to achieve the recursive
45 Fig. 7. A schematic of the recursion sequences employed in the reverse
column, and the rst and second Clenshaw algorithms to compute the
quantities P
nm h=P
mm h; dP Pmm h=dt, Xma =P
mm h;
Pnm h=P
dXma =PPmm h=dt and Xma =P
mm hqm ; dXma =P Pmm hqm =dt,
respectively

computation of all slma , of constant m (a column in
Fig. 7), and sequentially increasing l [upwards and
towards the diagonal (i.e. reverse) in Fig. 7], until the
recursion is terminated at the computation of the
sectoral smma (on the diagonal in Fig. 7).
Recursion algorithms resembling Eq. (46) form part
of the standard Clenshaw methods (cf. Tscherning and
Poder 1982) for evaluating the partial sums S d without
computing individual scaled values of P
nm d
h (Sect. 3.3).
However, Eq. (46) can be used in a dierent context to
compute individual values of P
nm h=P
mm h. Setting
sn1;m;a 0, snma 1, and all ylma 0 allows the re-
cursive computation of all slma , of constant m and se-
quentially decreasing l, from sn1;m;a to smma . The eect
of using the recursion in Eq. (46) in this way is to se-
quentially aggregate the (alm t) and blm recursive terms, in
the sequence of decreasing l, until the recursion termi-
nates at the computation of smma P
nm h=P
mm h.
One drawback of this approach is that the inter-
mediate values of slma in the recursion do not constitute
actual values of P
lm h=P
mm h. Instead, each
P
nm h=P
mm h value must be computed in isolation
from the others using n  m recursions of Eq. (46). This
necessitates M  m2 =2 recursions to compute all
non-sectoral values of P
nm h=P
mm h of order m and
degree m 1  n  M. This contrasts with the second
modied forward column recursion, where these same
values of P
nm h=P
mm h can be computed using only
M  m recursions of Eq. (34). The relative numerical
eciency of these approaches is tested in Sect. 4.3.
To compute values of P
nm 1
h=P
mm h, the rst
derivative of Eq. (46) with respect to t cos h gives
 
s_lma al1;m s_l1;m;a t sl1;m;a  bl2m s_l2;m;a 47
where s_lma dslma =dt. The seed values for the recursion
in Eq. (47) (i.e. sn;m;a and sn1;m;a ) are dierentiated with
respect to t to give s_nma s_n1;m;a 0. These initial
values allow the recursion in Eq. (47) to compute all
s_lma , of constant m and sequentially decreasing l, from
s_n1;m;a to s_mma . When used in this way, the recursion in
Eq. (47) terminates at the computation of


nm h
d PP
mm h
s_mma
dt
Since sectoral


nm h
d PP
mm h
dt
are zero, their corresponding s_mma are simply set to zero
without the need for any recursion. Application of the
product rule and the chain rule to


nm h
d PP
mm h
dt
yields
P
nm
1
h t
m smma  us_mma ; 8 n  m
P
mm h u that, in this case, the sectoral smma Xma =P
mm h.
287


nm h
where, as above, smma PP
mm h
and


nm h
d PP
mm h
s_mma
dt
To compute all non-sectoral values of P
nm 1
h=P
mm h
of order m and degree m 1  n  M, the reverse
column technique requires M  m2 =2 recursions of
Eqs. (46) and (47), as well as M  m applications of
Eq. (48). In the same manner as the reverse column
computation of P
nm h=P
mm h [Eq. (46)], the numer-
ical eciency of this approach (Sect. 4.3) contrasts poorly
against the second modied forward column recursion
(Sect. 2.6) in which the P
nm 1
h=P
mm h are obtained
from previously computed values of P
nm h=P
mm h
through only M  m applications of Eq. (35).
As with the modied forward column and forward row
recursion techniques (Sect. 2.6 and 2.5, respectively), a
global scale factor must be applied to the reverse column
computations to prevent an overow in the computation
of P
nmd
h=P
mm h. This is achieved by simply setting
the initial value, snma , to 10280 (rather than 1) for use in
Eq. (46). This scale factor of 10280 will propagate linearly
through the subsequent computations to yield values of
P
nm
d
h=P
mm h  10280 . Importantly, this global
scaling allows, in IEEE double precision, the computation
of spherical harmonic series S for 0  h  180 and S 1
for 0 < h < 180 up to M 2700.
3.3 Standard Clenshaw methods
The standard Clenshaw methods, summarised below,
closely resemble the reverse column recursions
(Sect. 3.2). The Clenshaw (1955) approach, which was
formulated originally to evaluate partial sums of
Chebyshev polynomials, was adapted for use in geodesy
by Gulick (1970) to compute partial sums of P
nm d
h.
Section 3.3.1 introduces a simple implementation of the
Clenshaw (1955) approach, whilst Sect. 3.3.2 presents
the implementation that is used more commonly in
geodesy (cf. Gleason 1985).
3.3.1 The rst Clenshaw method
The simplest implementation of the Clenshaw (1955)
technique, herein termed the rst Clenshaw method,
uses the recursions in Eqs. (46) and (47) to compute,
d
directly, the intermediate sums Xma =P
mm h, without
evaluating individual values of P
nm d
h=P
mm h.


Undierentiated values of Xma =Pmm h may be com-
puted using Eq. (46). Setting sM1;m;a sM2;m;a 0 and
all ylma E
lma [ylma 0; 8l < l in Eq. (3)], allows the
recursive computation of all slma , of constant m (a
column in Fig. 7), and sequentially increasing l
(upwards and towards the diagonal in Fig. 7), from
sMma to smma . As in the reverse column recursion
(Sect. 3.2), the recursive process terminates at the
48 computation of smma (on the diagonal in Fig. 7), except
288

The rst Clenshaw method also may be extended to For undierentiated values of S, the required
compute partial sums S 1 , S 2 and S 1 , where d 1 Xm =P
mm hqm quantities are computed using a mod-
denotes denite integration (cf. Tscherning and Poder ied version of the reverse recursion algorithm in Eq.
1982). However, this study is conned to the computa- (46); this is (Gleason 1985)
tion of S 1 . For this task, the rst Clenshaw method
uses the recursion in Eq. (47). The rst derivative with slma al1;m t qsl1;m;a  bl2;m q2 sl2;m;a ylma 51
respect to t cos h of the seed values used above
(sM1;m;a sM2;m;a 0) gives s_M1;m;a s_M2;m;a 0. To produce the required values of Xma =P
mm hqm , the
These seed values allow the recursive computation of all recursion in Eq. (46) is initiated exactly as for the rst
s_lma , of constant m and sequentially decreasing l, from Clenshaw method (sM1;m;a sM2;m;a 0), except that
s_Mma to s_mma . As in the reverse column recursion all ylma are set to E
lma =ql [ ylma 0; 8l < l in Eq. (3)],
(Sect. 3.2), this recursive process terminates at the rather than E
lma . Thus, to evaluate the truncated
computation of s_mma , except that in this case expansion of geopotential in Eq. (5), these values of
  ylma are set to C
lma , rather than a=rl C
lma , as for the rst
Xma
d P
mm Clenshaw method. This allows the recursive computa-
h tion of all slma , of constant m (a column in Fig. 7), and
s_ mma
dt sequentially increasing l (upwards and towards the
diagonal in Fig. 7), from sMma to smma , where
Application of the product rule and the chain rule to the smma Xma =P
mm hqm .
1
quantity Similarly, values of Xma =P
mm hqm may be com-
  puted by dierentiating Eq. (51) with respect to t to give
Xma (Gleason 1985)
d P
mm h
dt s_lma al1;m qs_l1;m;a t sl1;m;a  bl2m q2 s_l2;m;a 52
yields (adapted from Gleason 1985) Dierentiation (with respect to t) of the seed values used
1 above (sM1;m;a sM2;m;a 0) gives s_M1;m;a s_M2;m;a
Xma t 0. These values allow the recursive computation of all
m smma  us_mma 49
P
mm h u s_lma , of the same m and sequentially decreasing l, from
s_Mma to s_mma . The recursive process terminates at the
where, as above, smma Xma =P
mm h and computation of s_mma , where
   
Xma
d P
mm h d P
mmXhq
ma
s_mma s_mma
m

dt dt
Application of the product rule and the chain rule to the
3.3.2 The second Clenshaw method quantity
In geodesy, what is termed the second Clenshaw method
 
in this paper is more commonly used to evaluate
d P
mmXhq
ma
spherical harmonic expansions such as Eq. (5) (e.g. m

Gleason 1985; Deakin 1998). For this task, Eq. (38) is dt

reformulated using Horners scheme in terms of Um q,
yields [adapted from Gleason (1985)]
where q a=r
1
X
M Xma t
S d
c Xd m smma  us_mma 53
m P
mm hqm u
m0
"( "( ) # where, as above, smma Xma =P
mm hqm and
d d
XM XM1
c  UM q
 


PM;M hqM PM1;M1 hqM1 d P
mmXhq
ma
m
! s_mma
d
XM2 dt
 UM1 q
d
PM2;M2 hqM2 Values of Xma =P
mm hqm , obtained from the second
) d
Clenshaw method, are used in place of Xma in Eq. (2) to
d
X2 d
form Xm =P
mm hqm . The standard approach (cf.
 UM2 q   

P2;2 hq2 Gleason 1985; Deakin 1998) is to combine these values
# using the implementation of Horners scheme given in
d
X1 d Eq. (50). Alternatively, values of Xm d
=P
mm hqm can
 U2 q
U1 q cX0 50 d
P1;1 hq either be multiplied by q to yield Xm =P
mm h, which
m

are combined using the implementation of Horners

Note that in the geodetic literature [see e.g. Gleason 1985, scheme in Eq. (38), or multiplied by qm Pm to yield
d
Eqs. (2.38), (2.39) and (2.40)], the use of Horners scheme values of Xm =um , which are combined using the
in Eq. (50) is presented as an implementation of Eq. (46). implementation of Horners scheme given in Eq. (31).
 289

The resulting numerical values of S d are identical in all cision. The tests of numerical eciency will compare the
cases (Sect. 4.2). execution times of the algorithms. The tests of numerical
accuracy will use analytic solutions for the sum of the
square of P
nm
d
h, to compare the modied forward row
3.4 IEEE overows and global scale factors and modied forward column algorithms only.
in the standard Clenshaw methods
4.2 Relative numerical precision
Overows can be prevented in the both rst and second
Clenshaw methods by multiplying all ylma by the global The rst step in comparing dierent methods for
scale factor of 10280 . This scale factor propagates computing the partial sums S d in Eq. (1) is to choose
linearly through subsequent recursions to nally pro- some appropriate values for E
nma and k. For E
nma , we
duce values of
d
Xma =P
mm h  10280 and might use empirically generated coecients such as
d EGM96 (Lemoine et al. 1998) and/or GPM98B (Wenzel
Xma =P
mm hqm  10280 . These are used in place of 1998) to compute the lower degree E
nma . Higher-degree
d
Xma in Eq. (2) to form Xd

m =Pmm h  10
280
and coecients could be generated synthetically to conform
d
m 280
Xm =Pmm hq  10 , respectively. Again, this with the predicted spectral characteristics of the Earths
global scaling allows, in IEEE double precision, the gravity eld (e.g. Tscherning and Rapp 1974). However,
computation of spherical harmonic series S for the ubiquity of high-degree spherical harmonic expan-
sions across multiple scientic disciplines favours a more
0  h  180 and S 1 for 0 < h < 180 up to general approach. For this reason, the following
M 2700. comparisons have employed the testing regime utilised
by Gleason (1985), which is to set all E
nma equal to 1 and
4 Numerical tests k equal to 0. This testing regime has the advantage of
being straightforward to present, as well as being
4.1 Viable methods suciently general for a rst assessment of the new
algorithms. In this approach, the computed estimates
The previous derivations have presented six apparently sd of the partial sums S d [Eq. (1)] reduce to
viable methods for computing S0  h  180 and 2700 X
X n
S 1 0 < h < 180 for M  2700. These algorithms are sd d
P
nm h 54
summarised in Table 1. n0 m0
As mentioned in Sect. 2.6, the rst (MFC-1) and
second (MFC-2) modied forward column recursions Each algorithm in Table 1 was evaluated in IEEE double
are, essentially, a single method. Therefore, for the re- precision on a Sun Ultra-10 workstation. The algorithms
mainder of this section they will be treated as one al- were encoded in Fortran 77 and compiled using the
gorithm, termed the modied forward column recursion. SparkworksTM v3:0:1 Fortran compiler. The values of s
This leaves, to this point, ve separate methods for were computed for integer values of colatitude
computing spherical harmonic expansions. The purpose 0  h  180 and values of s1 were computed for
of this section is to provide an initial, general assessment integer values of colatitude 1  h  179 . These were
of the relative merits of these algorithms using tests of compared with the corresponding control values,
precision, numerical eciency and accuracy. The tests of obtained from the second Clenshaw summation (Sect.
precision will compare partial sums, S d , computed in 3.3.2) which was implemented in IEEE extended double
IEEE double precision and IEEE extended double pre- precision (i.e. 16 bytes to store each oating-point
number) (cf. Coonen 1980). The relative precision (RP)
for each h was calculated using
 d 
Table 1. Summary of algorithms for computing partial sums S d s double  sd extended
RP    55
sd extended 
Methoda Section Formulae Computed
quantities
for S for S(d) where sd (double) is the value of Eq. (54) for the
d summation computed in double precision by each
P
nm h
MFC-1 2.6 (28) (32) (33) um respective method and sd (extended) is the result for
d
P
nm h the same sum computed using the second Clenshaw
MFC-2 2.6 (34) (35) P
mm h
d
P
nm h method in extended double precision. The values of RP
MFR 2.5 (28) (27) (30) um
d
are computed under the assumption that the results
P
nm h
RC 3.2 (46) (47) (48) P
mm h obtained from IEEE extended double precision are
CLEN-1 3.3.1 (46) (47) (49)
d
Xnma correct to at least one signicant gure more than those
P
mm h
d
Xnma
obtained from IEEE double precision.
CLEN-2 3.3.2 (51) (52) (53) P
mm hqm The term precision is used here since any systematic
a
MFC-1: rst modied forward column; MFC-2: second modied errors common to computations in both the double
forward column; MFR: modied forward row; RC: reverse precision and extended double precision formats will not
column; CLEN-1: rst Clenshaw; CLEN-2: second Clenshaw be revealed during such a comparison. Moreover, to
290

insure that such systematic errors did not bias the con-
trol sd (extended) values in favour of the sd (double)
values computed using the second Clenshaw method, a
second set of control values was computed in extended
double precision using the modied forward column
method. The two sets of sd (extended) control values
agreed to a minimum of 25 signicant gures.
For each of the algorithms, the computed quanti-
ties (Table 1) were globally scaled to yield values
of Xm =um  10280 ; Xm =P
mm h  10280 or

Fig. 8. Logarithm of the

P relative
Pn precision
[Eq. (54)] to evaluate 2700
n0


m0 Pnm h
using the rst Clenshaw (solid line) and the
second Clenshaw (dashed line) methods

Fig. 9. Logarithm of the

P relative
Pn precision
[Eq. (54)] to evaluate 2700
n0


m0 Pnm h
using the modied forward column (solid
line) and the second Clenshaw (dashed line)
methods

Fig. 10. Logarithm ofP the relative

Pn precision
[Eq. (54)] to evaluate 2700
n0


m0 Pnm h
using the reverse column (solid line) and
second Clenshaw (dashed line) methods

Fig. 11. Logarithm ofP the relative

Pn precision
[Eq. (54)] to evaluate 2700
n0


m0 Pnm h
using the modied forward row (solid line)
and second Clenshaw (dashed line) methods
 291

Xm =P
mm h qm  10280 , which were then substituted performance of the other ve methods can be assessed.
into the implementations of Horners scheme in To test the second Clenshaw method, an extreme, but
Eqs. (31), (38) and (50), respectively, to yield values of realistic, value for q was chosen to provide a contrast to
sd  10280 and then sd . Recall that E
nma 1 and the rst Clenshaw method. Thus, q was set to
k 0, so Eqs. (31),(38) and (50) reduce to Eq. (54). It bGRS80 =aGRS80 6 356 752:3141=6 378 137, where
was found that, once the values of Xm =um  10280 ; aGRS80 and bGRS80 are the semi-major and semi-
Xm =P
mm h  10280 or Xm =P
mm hqm  10280 minor axes of the GRS80 ellipsoid (Mortiz 1980).
were computed, they could be appropriately factorised The relative precision signatures of the ve algo-
for use in any of the three implementations of Horners rithms for computing s are shown in Figs. 8 to 11, whilst
scheme with no change to the nal computed value of the relative precision signatures of the ve algorithms
sd . That is, all combinations of the ve algorithms with for computing s1 are shown in Figs. 12 to 15. To fa-
the three implementations of Horners scheme showed cilitate easier visual comparisons, the error signatures
that the choice of implementation of Horners scheme obtained from the second Clenshaw method in the
was irrelevant to the observed precision of the algorithm. computation of s and s1 have been superimposed (da-
Of the ve algorithms to be tested in this way, the shed line) over the corresponding error signatures ob-
second Clenshaw method is the most widely used in tained from each of the other four methods. P2700
geodesy. Thus, it is useful to employ the RP error sig-
PnFor the computation of the quantities s n0
nature of this method as a benchmark against which the

m0 Pnm h, inspection of Figs. 8 and 9 shows no sys-

Fig. 12. Logarithm ofP the relative

Pn precision
[Eq. (54)] to evaluate 2700
n0

1
m0 Pnm h
using the rst Clenshaw (solid line) and the
second Clenshaw (dashed line) methods

Fig. 13. Logarithm ofP the relative

Pn precision
[Eq. (54)] to evaluate 2700
n0

1
m0 Pnm h
using the modied forward column (solid
line) and second Clenshaw (dashed line)
methods

Fig. 14. Logarithm ofP the relative

Pn precision
[Eq. (54)] to evaluate 2700
n0

1
m0 Pnm h
using the reverse column (solid line) and
second Clenshaw (dashed line) methods
292
tematic dierences between the performance of the
Clenshaw methods and the modied forward column
method. One interesting feature is that the relative pre-
cision signatures obtained from all of the column
methods for the northern latitudes are almost identical,
whilst the signatures for the southern latitudes are not
(cf. Figs. 8 to 10). This is particularly evident for the
relative precision signature for the reverse column re-
cursion (Fig. 10), which contrasts poorly, in the south-
ern latitudes, against the signature from the second
Clenshaw method. Nevertheless, the relative precision is
still < 109 .
Inspection of the relative precision signature for the
modied forward row recursion (Fig. 11) reveals a
slight, but clear, improvement in precision over the other
algorithms. This is particularly evident as computation
approaches the poles. P The ve methods that successfully computed S d , for
PnFor 1 the computation of the quantities s1 2700 n0 M 2700 and for integer values of h to the poles, were


m0 Pnm h, Fig. 12 reveals no clear dierences be-
tween the relative precision signatures of the rst and
second Clenshaw methods. Figure 13 shows, in the
northern latitudes only, a moderately improved relative
precision signature for the modied forward column
method over that of the second Clenshaw method. Si-
milarly, the relative precision signature for the modied
forward row recursion (Fig. 15) shows an increasing
improvement over the second Clenshaw method towards
the north pole. There is also a slight improvement to-
wards the south pole. As with the relative precision sig-
natures for s, the reverse column method gives a relative
poor relative precision signature for s1 in the southern
latitudes (Fig. 14). This result is to be expected since the
values of P
nm1
h=um are computed from P
nm h=um .
None of the ve recursion tested delivered a relative
error of >109 in the computation of either s or s1 . No
analysis was conducted to explain the hemisphere-de-
pendent performance observed (Figs. 8 to 15) during
these tests. However, this phenomenon is not observed
in the accuracy tests for the two modied forward al-
gorithms (Sect. 4.4). These both produced relative ac-
curacy signatures (Figs. 16 and 17) that are noticeably
more symmetric about the equator than the precision
signatures (Figs. 9, 13 and 11, 15) for the corresponding
algorithms.
For each algorithm, the accuracy tests used the same
values of P
nm d
h=um as those used in the precision
tests. However, the accuracy tests squared these values
Fig. 15. Logarithm ofP the relative
Pn precision
[Eq. (54)] to evaluate 2700
n0

1
m0 Pnm h
using the modied forward row (solid line)
and second Clenshaw (dashed line) methods
before combining them using Horners scheme. This
suggests that, at least for the two modied forward al-
gorithms, the actual computation of P
nm h=um and
P
nm
1
h=um is performed equally well in both hemi-
spheres. In this case, the lack of symmetry in the relative
precision signatures returned by these algorithms might
result from the change in sign, across the equator, of half
the P
nm h=um and P
nm 1
h=um . This would create
dierent cancelling eects in each hemisphere when
these values are combined using Horners scheme. Such
an eect would not be present when squared terms are
combined. Further work may validate this explanation.
4.3 Numerical eciency
tested for their relative numerical eciency. Consider-
able attention was given to eliminating all redundant
computations from each algorithm. For example, the
square roots and inverted square roots required to
construct the recursion coecients were computed once
by each algorithm and then stored for multiple use in the
synthesis subroutines. However, the latitude-indepen-
dent components of the recursion coecients were not
computed and stored in this way, but were generated
from the square roots and inverted square roots as they
were required. This was done so that the eciency
results would be applicable to the many PCs, still widely
used, which possess insucient RAM to store
~7.3 million recursion coecients.
Table 2 shows the CPU time required by each of the
ve algorithms to computer the following:
P Pn
(1) s 2700
n0


m0 Pnm h, for integer values of
 
0  h  180 P Pn
(2) s as above and s1 2700n0

1
m0 Pnm h, for integer
 
values of 1  h  179 simultaneously.
The CPU times for the reverse column algorithm are
excessively large and so were extrapolated from the
computation times for a single parallel. All computations
were performed, once again, on a Sun Ultra-10(333 MHz)
workstation that uses a virtual or swapped RAM
conguration, which is slower than actual RAM.
It should be noted that these CPU times, in addition
to showing the relative eciency of each approach, are
 293
P Pn
Table 2. CPU time required to compute S 2700
n0


m0 Pnm h; 8 from which results
 
integer values of 0  h  180 , and the CPU time required to
 
computeP bothPn S, 18 integer values of 1  h  179 , and  n 
M X
X 2
S 1 2700
n0


m0 Pnm h; 8 integer values of 1  h  179 ,

R P
1 h
together M
n0 m0
!
MM 12 M 2
(1)
Task S (sec) S and S (sec)
; 0 < h < 180 59
CLEN-1 192 282 4
CLEN-2 192 282
MFC 186 408 For M 2700, Eq. (57) gives R2700 7 295 401 and Eq.
MFR 174 258 (59) gives R2700 13 305 717 113 850. However, these
RC 41 862 67 778
analytic values of R2700 and R2700 cannot be used to
a
CLEN-1: rst Clenshaw; CLEN-2: second Clenshaw; MFC: verify the accuracy of the rst and second Clenshaw
modied forward column; MFR: modied forward row; RC: methods, because these methods do not compute
reverse column individual, scaled values of P
nm d
h. An alternative test
for the standad Clenshaw approaches is to compute
partial sums of second derivatives, S 2 , and then use
also functions of the computer architecture, compiler
these to evaluate Laplaces equation Df 0. However,
and programming language employed, as well as the
this study does not extend to the computation of second
programmers implementation of these algorithms.
derivatives, and so will be conned to testing the
Variations in any of these, particularly algorithmic
accuracy of the modied forward column approach
implementation, can slightly improve or worsen the
and the modied forward row approach only.
relative performance of each algorithm. However, the
To implement this accuracy test, both methods were
results presented in Table 2 are sucient for the current
applied, in IEEE double precision, to compute, for
purpose, which is to demonstrate that all algorithms,
M 2700, integer values of P
nm h=um  10280 0 
except the reverse column algorithm, appear to be of
h  180 and P
nm 1
h=um  10280 1  h  179 .
comparable numerical eciency in evaluating the
However, these values cannot be squared in IEEE
required parital sums. Lastly, for the reasons outlined in
double precesion without an underow. This neces-
Sect. 3.2, the reverse column algorithm is extremely in-
sitates that the squaring of P
nm d
h=um  10280 ,
ecient when compard with the other four approaches,
as well as their combination using Horners scheme,
and thus will be excluded from further examination.
be performed in IEEE extended double precision.
This means that the results from these tests will
only reect the accuracy with which P
nm d
h=um 
4.4 Accuracy 280
10 are computed, since this is the only operation
which can be performed in IEEE double precision for
As mentioned, the numerical evaluations presented in M 2700.
Sect. 4.2 are tests of precision only, since both the tested The values of P
nm
d
h=um  10280 were converted
algorithms and the control algorithm computed in to IEEE extended double precision, squared, and then
IEEE extended double precision may contain shared rescaled by 10560 to yield values of P
nm d
h2 =u2m . All

d 2 2m
systematic errors. These could be due to any one of the values of Pnm h =u of the same m were sum-
compiler, computer architecture or programming errors, med to give the quantities
for example. Therefore, it is prudent to supplement tests
of precision with accuracy assessments that utilise exact X Pnm h2
2700

identities (i.e. analytic results) incorporating the com- Cm
nm
u2m
puted quantities. For this purpose, starting with the
well-known identity
and
X
n
P
nm h2 2n 1; 8 h 56 X P h2
2700
1
m0 Cm nm

nm
u2m
this can be shown to yield
These Cm were then combined using Horners scheme
M X
X n
R 2 2
P
nm h M 1 ; 8 h 57     
M C2700 2 C2699 2 C2698 2
n0 n0 R comp  u 5398 u 5396 u
2700 u5400 u u
 
Eq. (56) gives C2 C1
   4 u2 2 u2 C0 60
n 
X 2 nn 12n 1 u u
1


Pnm h ; 0 < h < 180 where R2700 (comp) are the computed estimates of R2700 .
m0
2
Exchanging Cm for Cm in Eq. (60) yields R2700 (comp),
58 which are the computed estimates of R2700 .
294
The numerical accuracy (NA) of P
nm h was calculated
using the relation
 
R2700 comp  7 295 401
NA 61
7 295 401
whereas the numerical accuracy NA of P
nm h was
calculated using the relation
   the fundamental recursive algorithms, and then employ-
R2700 comp  13 305 717 113 850
NA 62
13 305 717 113 850
The resulting R2700 (comp) error signatures of both
algorithms are plotted in Fig. 16 for integer values of
0  h  180 . The R2700 (comp) error signatures of both
algorithms are plotted in Fig. 17 for integer values of
1  h  179 .
Figures 16 and 17 show the accuracy of both meth-
ods to be almost identical. In the computation of both
R2700 (comp) (Fig. 16) and R2700 (comp) (Fig. 17), the
modied forward column algorithm performs slightly,
but consistently, better than the modied forward row
algorithm in the lower latitudes. However, the modied
forward column algorithm becomes increasingly less
accurate than the modied forward row algorithm as the
computation approaches the poles. This observation is
consistent with the results presented in Sect. 4.2, in
which, near the poles, the modied forward row algo-
rithms remained relatively stable in comparison with the
other methods tested. Lastly, note that neither the
modied forward row method, nor the modied forward
column method, delivered values of NA or NA greater
than 1011 . These results support those obtained for
these two algorithms in the precision tests (Sect. 4.2).
Fig. 16. Logarithm of the relative
accuracy
P2700 Pn[Eq. (61)] to evaluate

2
n0 m0 Pnm h using the modied
forward row (solid line) and modied
forward column (dashed line) algorithms
Fig. 17. Logarithm of the relative
accuracy
P2700 Pn[Eq. (62)] to evaluate

1 2
n0 m0 Pnm h using the modied
forward row (solid line) and modied
forward column (dashed line) algorithms
5 Summary, conclusions and recommendations
This paper has shown that standard Clenshaw methods
for evaluating high-degree spherical harmonic expan-
sions derive their stability from simple numerical
principles. IEEE underows are avoided by rst
eliminating the numerically problematic um term from
ing Horners scheme to gradually reintroduce this term
into the nal computed value for the paritial sums S d .
Moreover, existing algorithms for computing individual
1
values of Pnm h and Pnm h are easily modied to
incorporate these two fundamental characteristics of the
standard Clenshaw methods.
This last statement is strongly supported by the re-
sults of numerical tests. These show that the two new
algorithms (the modied forward row and the modied
forward column methods) can be applied in IEEE
double precision to compute the partial sums S, up to
M 27000  h  180 , as well as S 1 , up to
M 27000 < h < 180 , without IEEE underow or
overow. Moreover, the results also suggest that the new
algorithms are equivalent to the standard Clenshaw
methods in both precision and eciency. No doubt a
more rigorous testing regime, specic to geodesy, will
incorporate realistic geopotential coecients into the
computed partial sums S d . The relatively stable per-
formance of the modied forward row method, near the
poles, might also warrant further examination.
Perhaps the most interesting characteristic of the new
methods is their relative simplicity. Unlike the standard
Clenshaw methods, both the modied forward row and
modied forward column algorithms are easily for-

mulated using elementary algebra. More importantly,
the mechanisms within the computation process are
highly intuitive and transparent. These qualities should
simplify the process of adapting these approaches to
other tasks, such as evaluating partial sums of even
higher degree and order (e.g. 5400), for all latitudes tes-
ted in this study. Two other useful adaptions include, for
M 2700, evaluating partial sums of second derivatives,
and evaluating quantities that have been integrated over
geographic squares bordered by meridians and parallels.
Another potentially useful property of the new
methods is the fact that they compute individual, scaled
values of P
nm
d
h. This property, in conjunction with the
inherent simplicity of the principles presented here,
renders the new approach an attractive starting point for
extending the maximum M over which existing algo-
rithms for spherical harmonic analysis can be applied.
Acknowledgements. The authors would like to oer their sincerest
thanks to Christian Tscherning and the other two reviewers. Their
suggestions were of considerable use to the authors in allowing
them to produce a much more rigorous and integrated document.
Appendix A
This Appendix deals with some miscellaneous points on
implementing the new methods on a computer.
A.1 RAM conservation for the modied
forward row method
The modied forward row method (Sect. 2.5) computes
P
d h=um of matching degree. However, ecient
nm
evaluation of the partial sums S d requires that
P
nm
d
h=um of matching order be combined to yield
d
Xm =um for use using Horners scheme in Eq. (31). A
straightforward approach is to precompute and store,
for all degrees and orders, all the P
nm d
h=um and then
combine the P
nm h=u of matching order. However,
d m
many PCs do not have sucient RAM to store, say,
3.7 million values of P
nm h=um and/or P
nm 1
h=um
for M 2700.
The alternative is to use the modied forward row
method to compute P
nm d
h=um for one order at a time.

d m standard Clenshaw methods, since they do not auto-
The sectoral Pnm h=u are computed and stored as
before. Referring again to Fig. 3, computation of all
non-sectoral P
nm d
h=um , of order m and degree
m 1  n  M (a complete column in Fig. 3), re-
quires only that previously computed P
n;m1 h=um1
and P
n;m2 h=um2 have been stored.
Three arrays are used in rotation. Arrays 1 and 2
contain all the previously computed values of
P
n;m1 h=um1 and P
n;m2 h=um2 , respectively.
These are used in Eq. (27) to compute all the required non-
sectoral P
nm h=um , which are stored in array X . Values
295
of P
nm1 h=um1 are then used in Eq. (30) to compute
all the required P
nm
1
h=um , which are stored in array 2.
m 1 m
Xm =u and Xm =u are obtained from the values of
P
n;m h=um and P
n;m
1
h=um , repectively. For the next
d
round of computation P
n;m1 h=um1 , (corresponding
to the next column to the left in Fig. 3), the old array X
becomes the new array 1, the old array 1 becomes the new
array 2, and the old array 2 becomes the new array X .
This method employs the same number of mathe-
matical operations as that which precomputes and stores
all the required P
n;m h=um and P
n;m1
h=um . However,
the method of three rotated arrays requires a RAM
allocation of only 3M array elements to compute and
use the required non-sectoral P
n;m h=um and
P
n;m
1
h=um . This contrasts with the pre-compute
approach, which requires a RAM allocation of
MM 1=2 array elements for the same task.
A.2 Underows from problematic coecients
For spherical harmonic synthesis of very high degree
(e.g. M 2700), all of the methods presented here will
underow for suciently small values of Enma in Eq. (3).
For example, all methods will report underows (8h)
when EGM96 coecients are employed for the lower
degrees. The sole cause of this is C
360;360  4:5  1025 ,
which underows when combined with P
360;360 h=
u360  10280 using Horners scheme. In this case, setting
C
360;360 to zero prevents the underow message and
yields an error which is undetectable in IEEE double
precision. Of course, an entire coecient set that is
relatively homogenous in magnitude can be scaled
upwards or downwards as needed. Otherwise, a set of
coecients which dier by 20 orders of magnitude or
more can be partitioned, according to magnitude, into
subsets. Each subset is then scaled as a whole and then
used to compute a corresponding partial sum. The
resulting partial sums are then rescaled and combined to
yield the nal result.
A.3 Combining components by degree
The modied forward column and modied forward
row recursions are immediately more versatile than the
matically combine quantities of the same order m. This
feature is necessary, for example, to form intermediate
sums of quantities which share the same n, rather than
the same m. For example, consider the spherical
harmonic expansion of gravitational potential V in
Eq. (4). Set
M  n
GM GM X a
V r; h; k Xd
n A1
r r n2 r
where
296

X
n p
where q accounts for the k term [Eq. (8)] in the full
Xnd Zmd A2 normalisation of the Pnm h, and is given by q 2 for
m0
m 1 and q 1 8 m > 1. Note that, similar to the
and modied forward row recursion [Eq. (27), Sect. 2.5],
Eq. (A5) is ppresented
in Libbrecht [1985, Eq. (4)]
Zmd C
nm1 cos mk C
nm2 sin mkP
nm
d
h A3 without the q term, due to a dierent normalisation
which uses k 1; 8 m > 0. A schematic of this staggered
For ultra-high values of M, the modied forward recursion algorithm is given in Fig. B1.
column or modied forward row algorithms are used To use the recursion in Eq. (A5), both the sectoral
to generate values of P
nm
d
h=um . These replace P
nm
d
h P
mm h=um ; 8 m  M (upper diagonal in Fig. (31), and
d m
in Eq. (A3) to yield Zm =u , which are combined using zonal P
n;0 h=u0 P
n;0 h; 8 n  M (leftmost column
Horners scheme in Fig. B1), must be computed independently before-
X
n hand. The sectoral values are computed as before using
Xnd Zmd Eq. (28), whilst the zonal values may be computed using
m0 the modied forward column algorithm [Eq. (32)].
"( " d
# d
! Eq. (A5) can then be used to compute each of the re-
Zmd Zm1 Zm2 maining P
nm h=um from the adjacent P
n1;m h=um
 u m1 u m2 u
um u u (immediately above in Fig. B1), and P
n1;m1 h=um1
) # (diagonally above and to the left in Fig. B1). That is,
d d
Z Z d once all of the sectoral and zonal values are known, this
   22 u 11 u Z0 A4
u u is sucient to compute all P
n;1 h=u1 (second column
to the right in Fig. B1), which may then be used to
to give each separate Xd
n ; 8 2  n  M. Inspection of compute all P
n;2 h=u2 (third column to the right) and
Eq. (A1) shows that this algorithm provides an ecient so on up to m M  1.
means for computing multiple values of V (or any other Libbrecht (1985, p. 372) claims that, provided that
gravimetric quantity) at multiple points along the the zonal P
n;0 h values are computed with sucient
geocentric radial through h and k. accuracy using a ...rapidly converging trigonometric
expansion..., ...one would have to go up to a very high
l [degree] and m [order] indeed before roundo errors
Appendix B
became a problem. That is, the claim seems to be that
the overall accuracy of this computing values of
In addition to the forward row recursion in Eq. (27),
P
nm h=um is superior to that of implied modied for-
Libbrecht (1985) provides a second algorithm for
ward row recursion. However, no numerical results are
computing values of P
nm h=um . Fully normalising
provided by Libbrecht (1985) for this staggered algo-
Magnus et al. [1966, Eq. 4.3.3(6)] and then dividing by
rithm. Moreover, results from numerical tests conducted
um gives
in the current study contradict this claim. For these
p  tests, sectoral P
mm h=um were computed using
P
nm h 2n 1 p P
n1;m h
m
p
nmt Eq. (28), whilst the zonal P
nm h were computed using
u 2n  1n m um
 the standard forward column recursion [Eq. (32)] ap-
p P
n1;m1 h plied in IEEE extended double precision. A procedure
qn m  1 ; 8 n > m; m > 0 identical to that used for the precision trials of the other
um1
recursion methods (Sect. 4.2) was employed to yield a
A5 relative precision signature for the computation of

Fig. B1. Schematic of the recursion sequences employed in staggered Logarithm

Fig. B2. P Pn of the relative precision [Eq. (54)] to
algorithm to compute P
nm h=um evaluate 2700
n0


m0 Pnm h using the staggered algorithm
 297
P2700 Pn
n0 m0 P
nm h. The plot of the relative precision C.1.1 Standard forward row factorisation:
statistic is shown in Fig. B2. P
nm h=P
nn h
Inspection of Fig. B2 shows that the staggered algo- The rectangle in Fig. 3 for the forward row recursions
rithm for computing P
nm h=um is highly unstable, ex- contains all P
nm h, for which n 3 and 0  m  3. For
cept for points proximal to the poles and the equator. these values of P
nm h, the standard forward row
However, the relative precision of this algorithm close to recursion [Eq. (18)] gives
the poles does not exceed that of any of the approaches h  t i
presented in Sect. 4.2. Therefore, the staggered recursion P
3;2 h g3;2 P
3;3 h A6
algorithm should not be used to compute ultra-high u
degree and order spherical harmonic expansions.  
 t 2


P3;1 h g3;1 g3;2 h3;1 P
3;3 h A7
u
Appendix C
   t 3 t
This Appendix explores the possibility of formulating
1
P3;0 h p g3;0 g3;1 g3;2 g3;0 h3;1
Clenshaw methods, based on row recursions rather than 2 u u
column recursions, for evaluating the partial sums S d .  t 
These formulations have proven, at present, less g3;2 h3;0 P
3;3 h A8
u
successful than the other approaches presented in the
main body of the paper. The general approach in
where the aggregations of gnp t=u and hnp terms within
outlined below in the event that it may yet prove useful
the square brackets are equal to P
nm h=P
nn h. Here,
for future developments in this area.
p has been used instead of m to denote the order of each
recursive term gnp t=u and hnp , since each P
3;m h or
order m is comprised of an aggregation of all gnp t=u
C.1 Two forward row factorisations
and hnp terms for m  p  2. In general, any P
nm h may
be factored into P
nn h and an aggregation of all of the
In Sect. 2.6, each of the standard [Eq. (11)], rst
gnp t=u and hnp terms for m  p  n  1. Note that the
modied [Eq. (32)] and second modied [Eq. (34)]
denominator of P
nm h=P
nn h is not P
mm h, but
forward column recursion algorithms are of identical
P
nn h, which is the sectoral value of the same n, rather
form such that only the initial sectoral seed values dier.
than that of the same m.
That is, Eqs. (11), (32) and (34) all employ the same
The quantity of P
nm h=P
nn h appears to be of no
recursive terms (alm t) and blm but use dierent sectoral
practical use. Inspection of Fig. C1, which gives the
seed values. In contrast, the modied forward row
range of magnitudes taken by P
nm h=P
nn h, shows
recursion [Eq. (27)] for computing P
nm h=um diers in
that there is no global scale factor, capable of storage in
form from the standard row recursion [Eq. (18)] for
IEEE double precision, that will allow these quantities
computing P
nm h because the denominator um varies
to be computed for 0 < h < 180 up to M 2700. As
with the order of each P
nm h=um in any given row.
such, the quantity P
nm h=P
nn h will not be used to
That is, the recursive terms, gnm t=u and hnm , used in
compute very high degree and order spherical harmonic
Eq. (18), are dierent from the equivalent terms, gnm t
expansions.
and hnm u2 , used in Eq. (27). Thus, the standard forward
row recursion and the modied forward row recursion
will each yield a separate factorisation of P
nm h. These
factorisations serve to introduce Q the quantities
P
nm h=P
nn h and P
nm h=um n , which will be
examined in the following two subsections.

Fig. C2. A schematic of the recursion sequences employed in the

p P
nm h
reverse row algorithm to compute the quantities j Pn um and
Fig. C1. Logarithm plot of maximum (upper line) and minimum p dP
nm h=Pn um
(lower line) values of jP
nm h=P
nn hj; 8 n; m  2700 j
dt
298

C.1.2 Modied forward row factorisation: j values (Sect. 2.2) is determined by the value of m in
P
nm h=um Pn P
nm h.
Expanding on the example in Sect. C.1.1, recall that the To compute P
nm 1
h=um Pn , dierentiating Eq.
rectangle in Fig. 3 for the forward row recursions (A12) with respect to t cos h gives
contains P
nm h for which n 3 and 0  m  3. For
these values of P
nm h, the modied forward row s_npa gn;p1 ts_n;p1;a sn;p1;a
recursion [Eq. (27)] gives  hn;p2 u2 s_n;p2;a  2tsn;p1;a A13
P
3;2 h The seed values for the recursion in Eq. (A13), snma and
g3;2 tPn3 A9 sn;m1;a , are dierentiated with respect to t to give
u2
s_nma s_n;m1;a 0. These seed values allow the recursion
P
3;1 h in Eq. (A13) to be used to compute all s_npa , of the same
g3;1 g3;2 t2  h3;1 u2 Pn3 A10 n, and sequentially increasing p, from s_n;m1;a to s_nna .
u1 When used in this way, the recursion in Eq. (A13) will
  terminate at the computation of
P
3;0 h 1
p g3;0 g3;1 g3;2 t3  g3;0 h3;1 tu2 

u0 2 Pnm h
p d um Pn

2
s_nna j
 g3;2 h3;0 tu Pn3 A11 dt
The sectoral
where quantities in square brackets are P
nm h=um Pn . 

Note that the Pn term in the denominator renders any d Punnn Phn
P
nm h=um Pn dierent from the corresponding
P
nm h=um (Fig. 5) by no more than one order of dt
magnitude. are zero, and so their corresponding s_nna are simply set
to zero without the need for any recursion. Application
of the product rule and the chain rule to
C.2 Reverse row methods 

Pnm h
d
p um Pn
To compute any value of P
nm h=um , the modied j
dt
forward row recursion begins with the seed value Pm
and then aggregates the necessary gnp t and hnp u2 yields
recursive terms in order of decreasing p (sequentially left
across each row in Fig. 3). In a manner similar to the P
nm
1
h 1  t 
m
p m snna  us_nna ; 8 n  m A14
reverse column recursion [Eq. (46)], an alternative to the u Pn j u
modied forward row recursion [Eq. (27)] is to reverse
where, as above
this process and apply these same recursive terms in the
sequence of increasing p (sequentially right across each p P
nm h
row in Fig. C2). That is, a recursion may be employed snna j
Pn um
whereby the gnp t and hnp u2 recursive terms, for which
p m, and are applied rst and the recursive terms, for and
which p n  1, are applied last. This will be called a 

reverse row recursion and is illustrated schematically in Pnm h
p d um Pn
Fig. C2. s_nna j
The reverse row recursion is applied using the algo- dt
rithm In terms of the number of recursions required, the
eciency of the reverse row techniques for computing
snpa gn;p1 tsn;p1;a  hn;p2 u2 sn;p2;a ; ynpa A12 values of P
nm h=um Pn and P
nm
1
h=um Pn is the
same as for the reverse column methods (Sect. 3.2) for
To compute P
nm h=Pn um , the recursion in computing values of P
nm h=P
mm h and P
nm 1
h=
Eq. (A12) is used as follows. The seed values are set P
mm h, respectively.
to snma 1; sn;m1;a 0, and all ynpa 0. This allows Similarly, overows in the nal computed values of
the recursive computation of all snpa , of constant n (a P
nm
d
h=um Pn are prevented in the reverse row
row in Fig. C2), and sequentially increasing p (across
method by setting the seed value snma to 10280 (rather
to the right and towards the diagonal in Fig. C2),
from sn;m1;a to snna . The eect of using the recursion than 1) for use in Eq. (A12). This scaling propagates
p
d
in Eq. (A12) in this way is to sequentially aggregate linearly though subsequent computationspof  j P
the gnp t and hnp u2 terms, in the sequence of
m
nm h=Pn u to generate values of j P
nm
d
h=
increasing p, until the p recursion terminates at the Pn um  10280 . However, while this will prevent an
computation of snna jP
nm h= Pn um . Here, the overow, the reverse row method cannot be applied over

the same ranges of M and h as the reverse column al-
gorithm due to underow problems during the compu-
tation. For example, for M 2700, the reverse row
method will underow for h < 76 and h > 104 . No
investigation of this underow was conducted.
C.3 Row Clenshaw methods
The standard Clenshaw methods do not translate well to
row-type recursions. The rst Clenshaw method uses the
reverse column recursion in Eqs. (46) and (47) to
d Haagmans RRN (2000) A synthetic Earth for use in geodesy.
compute Xma =P
mm h. An equivalent utilisation of
the reverse row recursion in Eq. (A12) proves to be of
little use. In this case, setting sn;m1;a sn;m2;a 0 and
all ynpa E
npa allows the recursive computation of all
snpa , of constant n (a row in Fig. C2), and sequentially
increasing p (across to the right and towards the
diagonal in Fig. C2), from snma to snna . In this case
recursion terminates at
X n p  National Imagery and Mapping Agency (NIMA) geopotential

P
nm h
snna j Enma m
m0
u Pn
p for geophysical applications. Geophys J Int 139: 547555
The j term notwithstanding, the principal problem
with the quantity
X n p  Magnus W, Oberhettinger F, Soni RP (1966) Formulas and theo-
P
nm h
snna j E
nma m
u Pn
m0
is that the denominator of P
nm h=um Pn varies with
both n and m. This means that this quantity has eectively
summed components for which the scale factor is not
constant, thereby preventing such sums from being
combined and rescaled to achieve the nal sums S.
References
Abd-Elmotaal HA (1997) An ecient technique for the computa-
tion of the gravimetric quantities from geopotential earth
models. Paper presented to the IAG Scientic Assembly, 39
September, Rio de Janeiro
Abramowitz W, Stegan IA (1972) Handbook of mathematical
functions, 9th edn. Dover Publications, New York
Belikov MV (1991) Spherical harmonic analysis and synthesis with
the use of columnwise recurrence relations. Manuscr Geod 16:
384410
Belikov MV, Taybatorov KA (1991) An ecient algorithm for
computing the Earths gravitational potential and its derivatives
at satellite altitudes. Manuscr Geod 17: 104116
Clenshaw CW (1955) A note on the summation of the Chebyshev
series. Math Tab Automat Comput 9: 118120
Colombo C (1981) Numerical methods for harmonic analysis on
the sphere. Rep 310, Department of Geodetic Science and
Surveying, The Ohio State University, Columbus
Coonen JT (1980) An implementation guide to a proposed stan-
dard for oating point arithmetic. Computer 13(1): 6879
299
Deakin RE (1998) Derivatives of the Earths potentials. Geom Res
Aust 68: 3160
Edmonds AR (1957) Angular momentum in quantum mechanics,
Princeton University Press, Princeton, NJ
Featherstone WE (1999) Tests of two forms of Stokess integral
using a synthetic gravity eld based on spherical harmonics. In:
Krumm F, Schwarze VS (eds) Quo vadis geodesia. Institute of
Geodesy, University of Stuttgart, pp 101112
Gleason DM (1985) Partial sums of Legendre series via Clenshaw
summation. Manuscr Geod 10: 115130
Gulick LJ (1970) A comparison of methods for computing grav-
itational potential derivatives. Tech rep 40, Environmental
Sciences Services Administration, US Department of Com-
merce, Washington, DC
J Geod 74: 531551
Harris JW, Stocker H (1998) Handbook of mathematics and
computational science. Springer, Berlin Heidelberg New York
Heiskanen WA, Moritz H (1967) Physical geodesy. Freeman, San
Francisco
Lemoine FG, Kenyon SC, Factor JK, Trimmer RG, Pavlis NK,
Chinn DS, Cox CM, Klosko SM, Luthcke SB, Torrence MH,
Wang YM, Williamson RG, Pavlis EC, Rapp RH, Olson TR
(1998) The development of the joint NASA GSFC and the
model EGM96. NASA/TP-1998-206861, National Aeronautics
and Space Administration, Washington, DC
Lesur V, Gubbins D (1999) Evaluation of fast spherical transforms
Libbrecht KG (1985) Practical considerations for the generation of
large order spherical harmonics. Solar Phys 99: 371373
rems for special functions of mathematical physics, 3rd edn.
Springer, Berlin Heidelberg New York
Moritz H (1980) Geodetic reference system 1980. Bull Geod 54(4):
395405
Novak P, Van cek P, Veronneau M, Holmes SA, Featherstone WE
(2001) On the accuracy of Stokess integration in the precise
high-frequency gravimetric geoid determination. J Geod 74:
644654
Pail R (1999) Synthetic global gravity model for planetary bodies
and applications in satellite gravity gradiometry. PhD thesis,
Technical University of Graz
Risbo T (1996) Fourier transform summation of Legendere series
and D-functions. J Geod 70: 383396
Rizos C (1979) An ecient computer technique for the evaluation
of geopotential from spherical harmonic models. Aust J Geod
Photogram Surv 31: 161169
Tscherning CC (1976) Computation of second-order derivatives of
the normal potential based on the representation by a Legendre
series. Manuscr Geod 1: 7192
Tschering CC, Poder K (1982) Some geodetic applications of
Clenshaw summation. Boll Geos Sci A 4: 351364
Tscherning CC, Rapp RH (1974) Closed covariance expressions for
gravity anomalies, geoid undulations, and deections of the
vertical implied by anomaly degree variance models. Rep 208,
Department of Geodetic Science and Surveying, The Ohio State
University, Columbus
Tscherning CC, Rapp RH, Goad C (1983) A comparison of
methods for computing gravimetric quantities from high degree
spherical harmonic expansions. Manuscr Geod 8: 249272
Wenzel G (1998) Ultra-high degree geopotential models GPM98A,
B, and C to degree 1800. Paper presented to the joint meeting of
the International Gravity Commission and International Geoid
Commission, 712 September, Trieste