I. Quont Sprcirosc Radial Transfer Vol 30. No. ?, pp 143.161.

1983
Prmted in Great FIrmin

NUMBER OF TERMS REQUIRED IN THE FOURIER

EXPANSION OF THE REFLECTION FUNCTION FOR
OPTICALLY THICK ATMOSPHERES

MICHAEL D. KING

Laboratory for Atmospheric Sciences, Goddard Space Flight Center. NASA. Greenbelt. MD 20771. U.S.A.

(Received 8 October 1982)

Abstract-Computational results have been obtained for the separate terms in the Fourier expansion of the
reflection function of an optically thick, conservatively scattering, atmosphere composed of cloud particles.
The computations were obtained by successive applications of the invariant imhedding. doubling and
.
asymptotic fitting methods to cover the range from very thin to very thick atmospheres. Result\ are
presented which illustrate the magnitude of the separate terms in the Fourier expansion of the phase
function and the Fourier expansion of the reflection function of a semi-infinite atmosphere as a function of
the zenith angles of incidence and reflection. The azimuthally independent reflection fun&n i\ enhanced by
as much as a factor of I15 over the first-order reflection function, whereas the azimuth-dependent reflection
functions generally result from less multiple scattering. These results are compared with those for an
atmosphere having a Henyey-Greenstein phase function with the same asymmetry factor (4 = 0.84123) as
in the cloud model. The relative difference in the escape function and azimuthally independent reflection
function is generally less than a few per cent, though differences up to 70% occur in the reflection function
at angles where single scattering is important, Results are also presented which chow the number of terms
required in the Fourier expansion of the reflection function to be assured an accuracy of tI.lci. The number
of terms required depends strongly on the zenith angles of incidence and reflection a\ well as on details of
the phase function.

I. INTRODUCTION

The scattering of sunlight in planetary atmospheres is often characterized by anisotropic phase

functions expressible as finite expansions of Legendre polynomials. For atmospheres composed
of particles large compared to the wavelength of the incident light, as in terrestrial clouds at
visible wavelengths, this series may consist of up to a few hundred terms. For multiple
scattering calculations it is convenient to transform the reference system from the plane of
scattering to two vertical planes containing the directions of incidence and scattering. Using the
addition theorem for spherical harmonics the phase function is thereby expressible as a Fourier
series in the cosine of the azimuthal angle. Though the number of terms required in the Fourier
expansion of the phase function may be as many as in the Legendre series at some angles,
Dave,’ Dave and Gazdag,’ and Herman and Browning3 have shown that the required number of
terms varies substantially with the angles of incidence and scattering.
Regardless of the number of terms required to express the phase function as a Fourier series
in the cosine of the azimuthal angle, the maximum number of terms required to describe the
reflected and transmitted intensities will not exceed the number of terms required to describe
the phase function3 This led many authors’-’ to suggest that multiple scattering computations
could be made less time consuming by restricting the computations to a few azimuthal terms.
Since the number of terms required to describe the intensity field to a given level of accuracy
depends not only on the angles of incidence and scattering but also on the optical thickness of
the atmosphere, it is not in general possible to estimate a priori the number of azimuthal terms
to be carried in the computations.
van de Hulst’ has presented empirical results for the separate terms in the Fourier
expansion of the reflection function of semi-infinite atmospheres. These results, which were
based on the analytic phase function first introduced by Henyey and Greenstein,’ suggest that if
either the zenith angle of incidence (0,) or reflection (0) is well removed from the grazing
directions, the required number of Fourier terms is small. Furthermore, van de Hulst’
concludes that the required number of terms varies in accordance with the expression
M(B, 0,) = 25 sin O,, where 0, is the smallest of 8 and 0”. Hansen and Travis’ present selected
results for an optically thick atmosphere (T,- = 128) composed of a polydisperse collection of
aerosol particles. In contrast to the results obtained using the Henyey-Greenstein phase
function6 the Mie theory results of Hansen and Travis’ show that the required number of terms
143
144 M. D. KING

may be larger when 0 = 8” than when 0 is either smaller or larger than B,,. Their illustration of
the separate terms in the Fourier expansion of the reflection function for selected values of 0
and B,, is noticeably more variable than the comparable Henyey-Greenstein results presented
by van de Hulst6 Since the separate terms in the Fourier expansion of the reflection function
for a Mie theory phase function undergo frequent reversals of sign and since their magnitude
decreases less rapidly with increasing order of the Fourier series than does the comparable
Henyey-Greenstein results, it appears clear that conclusions drawn for a Henyey-Greenstein
phase function may differ from those drawn for a Mie theory phase function.
The intent of this paper is to present computational results for the separate terms in the
Fourier expansion of the phase function and the Fourier expansion of the reflection function of
a semi-infinite, conservatively scattering, atmosphere composed of cloud particles. From these
results we determine the ratio of the total reflection function to the first-order (single scattering)
reflection function and the number of terms required to describe the reflection function to an
accuracy of 0.1%. These results are compared with similar results obtained for a Henyey-
Greenstein phase function having the same asymmetry factor as in the cloud model.

KLEGENDRE EXPANSION OFTHEPHASE FUNCTION

In many radiative transfer applications it is convenient to express the product of the single
scattering albedo wgand the phase function @(cos 0) as a finite expansion in Legendre polynomials
of the form

O”@(COS
0) = 9 01P,(cos O),
I=0

where 0 is the scattering angle and P,(cos 0) a Legendre polynomial of order 1. AS a

consequence of the orthogonality of the Legendre polynomials, the L + 1 coefficients o, may
formally be obtained from

(21+1) ’
01 = ~ wn O(cos 0) P,(cos 0) d(cos 0).
2 I -I

With these definitions, the phase function satisfies the normalization condition

C’,(cos0) d(cos 0) = 1, (3)

with the asymmetry factor g related to the Legendre coefficient o, by

i I_’ @(COS 0) cos 0 d(cos 0) = g = 0,/(3wn).

I

We employ two phase functions in calculations of the reflection function of optically thick
atmospheres. The first is a Mie phase function for a wavelength 0.754 Frn, refractive index
1.333, and a size distribution of particles of a given radius proportional to r6 exp (- 1.6187r),
where r is the particle radius in pm. This distribution of particles is a gamma distribution with
effective radius 5.56 pm, effective variance 0.111, and an asymmetry factor 0.84123. This phase
function, illustrated in Fig. 1, is typical of fair weather cumulus (FWC) clouds in the visible
wavelength region (see Hansen’). The second phase function is the widely used analytic phase
function first introduced by Henyey and Greenstein and given by

@(cos 0) = (1 - g’)/( I + g’ - 2g cos a)=. (5)

This function is illustrated in Fig. 1 for the same asymmetry factor as in the FWC model.
Though it lacks the structure of the Mie phase function, especially the rainbow and glory
features apparent in Fig. 1, its extensive use in radiative transfer modeling studies makes it a
valuable phase function with which to compare results obtained with the FWC model.
 The reflection function for optically thick atmospheres 145

104 r 1 I I I I I I I r 1 I 1 I I I I
m=l.333 - O.Oi
g=O.S4123

1oJ - FWC, X=0.754 pm

-- HENYEY- GREENSTEIN 5

SCAlTERING ANGLE (DEGREES)

Fig. 1. Phase function as a function of scattering angle for a fair weather cumulus (FWC) size distribution
given by n(r) (I r6 exp( - 1.6187r), where A = 0.754 pm and m = 1.333 -O.Oi. Also shown is the Henyey-
Greenstein phase function for the same asymmetry factor as in the FWC model.

One of the features of the Henyey-Greenstein phase function which makes it especially
attractive for radiative transfer applications is the simple expression which results for the
Legendre coefficients,

01 = (21+ l)g’wo. (6)

For a general Mie phase function, no such relationship exists.

Dave’ and Kattawar et al9 present formulas for calculating the coefficients of the Legendre
series expansion of a Mie phase function for a given size parameter and refractive index. These
modified Mie expressions are far more tedious to compute than the well-known Mie expres-
sions for the phase function, due primarily to the fact that the expressions for the Legendre
coefficients involve infinite series imbedded within infinite series. As an alternative to using
these expressions for hundreds of particle sizes and subsequently integrating over a size
distribution, it is sufficient to evaluate the L + 1 coefficients by numerically integrating over the
phase function computed at very small angular intervals. Kattawar”’ evaluated the Legendre
coefficients by both methods and found “excellent agreement”, though no details were provided
on either the quadrature formula used to evaluate the integrals in Eq. (2) or on the difference in
computational time between the methods.
In the present investigation, we adopt the quadrature formula

(7)

where pj are the abscissas and cj the weights for Gaussian quadrature on the interval [-1, 11.
With the phase function a polynomial of degree L, the integrand in Eq. (2) is at most a
polynomial of degree 2L. Since Gaussian quadrature is exact for polynomials of degree less
than 25, it is essential that the order of the quadrature formula (.I) exceed the number of
significant Legendre coefficients (L) in the phase function, a number which is not known a
priori.
Figure 2 illustrates the variation of the coefficients of the Legendre polynomial expansion of
the FWC and Henyey-Greenstein phase functions illustrated in Fig. 1. The abscissas and

QSRT Vol. 30,No. 2-D

146 M.D. KING

m = 1.333 - O.Oi
g =0.64123

10-3 0 ' I ' 0 ' I ' I ' 1 ' 1 ' 1"' 1 ' 1
0 20 40 60 SO 100 120 140 160 160 200 220

ORDER II)

Fig. 2. Coefficients of the Legendre polynomial expansion of the FWC and Henyey-keenstein phase
functions illustrated in Fig. 1.

weights used in evaluating Eq. (7) for the FWC model were taken from Stroud and Secrest”
for a Gaussian quadrature of order .I = 512. This results in the first few scattering angles being
0.27”, 0.62”, 0.97” and 1.32”. Based on the somewhat arbitrary criterion that or < 1O-9tin for I > L,
we find that the Legendre expansions are of length L = 152 for the Henyey-Greenstein model and
L = 229 for the FWC model.
The use of Eq. (7) has several built in accuracy checks. These include the necessity that w.
equal the single scattering albedo (unity in the present investigation), that w, = 3g00, where g is
computed directly from the Mie theory, and that the values of the phase function reconstructed
from the L + 1 term Legendre series agree with the values computed from Mie theory. All of
these tests were met to an accuracy of at least six significant figures. Though some authors’2-‘3
recommend using Lobatto quadrature, rather than Gaussian quadrature, in order to include 0”
and 180” as explicit quadrature points, we found this to be unnecessary. The time required to
evaluate the phase function at 512 scattering angles and at 500 particle sizes, to integrate the
phase function over a particle size distribution, and to evaluate the Legendre coefficients was
5.42 min on an IBM 360/91 computer.

3. FOURIER EXPANSION OF THE PHASE FUNCTION

For multiple scattering calculations it is necessary to transform the reference system from
the plane of scattering to two vertical planes containing the directions of incidence and
scattering. Using the addition theorem for spherical harmonics we obtain the well-known
expression6. I4

ti&‘(COS O) = h”(n, U) + 2 5 hm(n, U) COS m+, 03)

m=l
 The reflection function for optically thick atmospheres 147

where the azimuth-dependent redistribution functions h’“(u, u) are given by

hrn(u, u) = 5 01 Y,“(u) Y,“(u). (9)

I=m

In these expressions u and u are cosines of the zenith angles of scattering and incidence with
respect to the downward normal (- 15 U, v 5 l), Y’“‘(u) the renormalized associated Legendre
polynomials expressible in terms of the associated Legendre polynomials P’“(u) byI

1 P,“(u),
l/2
Y,“(u) =
[ ff$-$ (10)

and

l
. cos 0 = uu + (1 - u2)l’* (1 - i?)“2 cos f#‘. (11)

Figure 3 illustrates the azimuthally independent redistribution function for reflection,

Ir”( - CL,po), where p. is the cosine of the solar zenith angle and p the cosine of the zenith angle
of reflection with respect to the outward normal (0 5 CL,~~ I 1). The redistribution function for
the FWC model is shown in the left portion of the figure while the corresponding function for the
Henyey-Grenstein phase function is shown in the right portion. Since /I’( - IL, po) represents
the product of the single scattering albedo and phase function averaged over azimuth, the
values of h”( - CL,cam)for CL= 1 or p. = 1 equal those of the phase function itself. The prominent
maxima and minima in the backward hemisphere of the FWC phase function (see Fig. 1) lead
directly to corresponding features in the redistribution function. For example, the minimum
value of h’(-p, 1) at p = 0.264 corresponds to the minimum of the phase function at 0 = 105.3”,
whereas the maxima at p = 0.812, 0.998 and 1.0 correspond to the rainbow and glory features
apparent in Fig. 1. Values of h”( - CL,po) along the diagonal k = k. represent averages of the
phase function over the range of scattering angles 180” - 2eo 5 0 I 180”, where B. = cos-’ cam.
Thus the broad minimum at p = k. ~0.7 for the FWC model is associated with the broad
minimum in the phase function at 0 = 105.3”. For the Henyey-Greenstein phase function, the
redistribution function lacks any maxima or minima since the phase function itself is a
monotonically decreasing function as the scattering angle increases. In fact the azimuthally
independent redistribution function for the Henyey-Greenstein phase function can be expres-

0.8

0.8

0.0
0.0 0.2 0.4 0.6 0.8 1.0
w

Fig. 3. Azimuthally independent redistribution function for reflection, hO( - I*, /.~a), where the figure on the
left applies to the FWC model and the figure on the right to the Henyey-Greenstein model.
148 hf. D. KING

sed in terms of the complete elliptic integral of the second kind, as noted by Wiscomber6 and
van de Hulst.6
Unlike the azimuthally independent term, the azimuth-dependent terms in the Fourier
expansion of the phase function may contain negative as well as positive values. Figure 4
illustrates the first azimuth-dependent redistribution function for reflection, h’( - CL,po), where
again the left portion of the figure applies to the FWC model and the right portion to the
Henyey-Greenstein model. Due to the monotonically decreasing nature of the Henyey-
Greenstein phase function, h’( - /.L,/.L~)is positive for all values of /J and /.L~.In contrast, the
broad minimum in the FWC phase function at 0 = 105.3” gives rise to large regions where
h’( - CL,po) is negative. These regions are denoted by shading in Fig. 4. When either p = 1 or
p. = 1 the azimuth-dependent redistribution functions h”( - p, po) = 0 for all m 2 1, For these
special cases it is adequate to terminate the series in Eq. (8) after the first, azimuthally
independent, term. This boundary condition is clearly evident in Fig. 4.
Dave and Gazdag’ noted that Eq. (8) can be rewritten by replacing the fixed upper limit L of
the Fourier series representation of the phase function by a variable upper limit N(u, u). By
examining the absolute magnitude of h”‘(u, u) as a function of m for selected values of u and u,
Dave and Gazdag2 concluded that N(u, v) is a strong function of u and u, that all I. + 1 terms
are required only if u = u = 0, and that the maximum number of terms required for a fixed angle
of incidence is required when either u = 2; (transmission) or u = - u (reflection). This latter
conclusion was found to be a function of size distribution parameters, for the number of terms
required in the Fourier expansion of the phase function for a single particle showed no such
maxima at u = u and u = - u. For a single particle the required number of terms was found to
be a maximum when u = 0, regardless of the angle of incidence.

4. NORMALIZATION OF THE REDISTRIBUTION FUNCTION

In performing multiple scattering calculations it is convenient to subdivide the angular

interval [0, l] into K Gaussian quadrature points 0 < /.L’< . . . < pk < 1 with mirror-symmetric
points on the interval [ - l,O] for a total of 2K streams. Then, if the Gaussian weights are
Cl,. **, ck, it follows that

I-I
h”‘(u, /+)Ymm(U) du =
I
o’ [h”‘- K /+I + h”‘(P* /+)lYm”‘(PCL)dF

=: [h”‘(- pir/+) + h”‘(pi, /+)IYm’“(Pi)Ci (12)

i=l

Fig. 4. The same as Fig. 3, except for h’( - p, ~0).

 The reflection function for optically thick atmospheres 149

for all j=l,..., K. Replacing h”(u, CL’)by its explicit form [Eq. (9)] and noting the ortho:
gonality of the renormalized associated Legendre polynomials, it follows that

I_‘,
h”‘(u,
pj)Ym”‘(u)
du
=,im
wlYr”‘bj)
1’
-IYF(u)Ym”‘(u)du
=& WM Yrnm(Pj). (13)

Combining Eqs. (12) and (13) we find that the phase function must satisfy the following
normalization condition in its quadratured form:
K
2
C [h”( - /Jai,&) + hm(/Li, /Lj)]Y,“(/.Li)Ci = 2m Wn Ym”(Pj) (i = l, *. 9K).
i=l

The need to satisfy this equation for the azimuthally independent redistribution function
(m = 0) has long been recognized in order to conserve flux (wO= 1) or to obtain accurate flux
divergence values in multiple scattering calculations.” This normalization condition for the
azimuth-dependent terms of the redistribution function has not been previously noted in the
literature to the author’s knowledge. If the phase function is highly asymmetric (large L) and
the order of the Gaussian quadrature (K) is too small, Eq. (14) will be poorly satisfied. As a
consequence the results of multiple scattering calculations will be inaccurate, especially for
optically thick layers with nearly conservative scattering. On the other hand, if K is chosen too
large, Eq. (14) and consequently multiple scattering calculations will be highly accurate but
enormous amounts of computer time may have been expended. In the doubling method, for
example, computer time escalates roughly as K3 (Hansen’) while storage requirements escalate
as K2. This led several investigator’s’“‘8,8.3 to limit the order of their angular discretization and
to compensate for the potential loss in accuracy by renormalizing the azimuthally independent
redistribution function until Eq. (14) is satisfied. To assure accurate intensity computations as
well as accurate flux computations, however, it is equally important that Eq. (14) be satisfied for
all m 20.
For a fixed angle of incidence the redistribution function is a polynomial of degree I. while
Y,“‘(p) is a polynomial of degree m. Since Gaussian quadrature is exact for polynomials of
degree less than 2K, a sufficient condition for Eq. (14) to be satisfied is for 2K to exceed L + m.
Due in part to the small magnitude of the high order Legendre coefficients (see Fig. 2) and in
part to the use of Gaussian quadrature on the angular interval [0, 11, we find that Eq. (14) may
adequately be satisfied provided
K h 0.35L. (13

A Gaussian quadrature of order K = 80 was used for both the Henyey-Greenstein and FWC
models. For the FWC model (I. = 229) the difference between the left and right sides of Eq. (14)
varied from 4.9 x 10m9to 1.2 x lo-l2 for the azimuthally independent term, depending on the
direction cosine of the angle of incidence (CL’).This error decreased with increasing Fourier
frequency so that at m = 30 the maximum normalization error was 2.1 x 10-15. Since the same
order of angular discretization was used for the Henyey-Greenstein model (L = 152) the
normalization errors were generally less than for, the FWC model. At m = 0 the maximum
normalization error was 1.9 X 10mL4,reducing to 3.7 x 10-l’ at m = 30. In no instance was any
renormalization performed.

5. FOURIER EXPANSION OF THE REFLECTION FUNCTION

Having determined the azimuth-dependent redistribution functions, multiple scattering

calculations were performed to determine the separate terms in the Fourier expansion of the
reflection function R(T,; CL,p,,, 4) and transmission function T(r,; CL,po, 4), where

R(T,; CL,po, 4) = R’(T~; p, po) + 2 5 R”(T,; p, ~~~cos m4, (16)

m=l
150 M. D. KING

and 7Cis the total optical thickness of the atmosphere. In terms of these functions, the reflected
I(0, - CL,4) and transmitted I(T,, CL,4) intensities from a horizontally homogeneous atmosphere
illuminated from above by a parallel beam of radiation of incident flux density F. may be
expressed in the forms

Ito, -CL, 4) = (/#oh) R(T,; P, PO, 41, (18)

r(Tc, /-b 4) = boFo/r) T(T, ; CL,CLov

4). (19)

By virtue of the Helmholtz principle of reciprocity each of the azimuth-dependent terms of the
reflection and transmission functions is symmetric in p and pa.
The reflection and transmission functions were calculated using the doubling method.5” In
this method the reflection and transmission functions of a single layer of optical thickness TV are
combined with those of a similar layer to obtain the reflection and transmission functions of a
combined layer of optical thickness 27,. In applying the doubling method it is necessary to
obtain the reflection and transmission functions of an initial layer of infinitesimal optical ’,
thickness. Many different methods have been used as initializations in the doubling method.
Wiscombe’6 described and compared the majority of these methods and concluded that the
error in the computational results for most values of the optical thickness can be reduced by .
many orders of magnitude when the best starting technique is used.
In the present investigation we have solved the integrodifferential equations satisfied by the
reflection and transmission functions, known as the principles of invariance, by a second order
Runge-Kutta method. Replacing the integrals on the angular interval [O, 11 by a Gaussian
quadrature formula with abscissas 0 < 11’< . . . < FK < 1 and corresponding weights ck, the
principles of invariance” may be written as

aR m(~c; CL>
PO) _ ’
Rm(7c; P,I*o)+- h”(-F,po)
a7, -- 4PcLo

hm(p., !dRm(Tc; pk, pO)ck +L 2po gl R"'(c; CL, Pk)hm( - Pk, - r*O)ck

+ i: 5 Rm(7c; PL, Pdhm(- pk, CLJR~(T,; /4, ~O)ck% (20)

k=l I='

aT"h; CL, PO) _ 1 1

--~T'"bc:wo)+ - e-‘c’p hm(p, po)
a? 4PPo

h”(- p, /4)Rm(T,; FL, /b,)ck +& gl T”‘(7,; Lb /dhm(- /Jkv - pO)ck *

K K

+ Tm(7c; P, Pk)hm( - kk, dR”(~c; PI, ~o)ckG. (21)

Integrating these differential equations from the origin to an optical depth rC using a second
order Runge-Kutta method2’ we find that the discretized reflection and transmission functions
for a thin initial layer may be expressed for all i, j = 1,. . ., K as

+&$, [h”(-CLi,~k)hrn(~k~~i)+hrn(-~i~
-Irk)hm(-ILk,~j)]~+O(7,)),

(2-a

where O(T,~) denotes terms of order 7c3 or higher.

 The reflection function for optically thick atmospheres 151

Equations equivalent to Eqs. (22) and (23) were derived by Wiscombe’6 by expanding the
matrix inversion equations of his diamond initialization in powers of rC3and retaining terms out
to Tc2. Wiscombe refers to this method as the expanded diamond initialization, though our
derivation from the principles of invariance makes it clear that these equations are equivalent to
an invariant imbedding initialization. The first term in each equation represents the contribution
from single scattering while the second term represents the contribution from multiple scatter-
ing. Equations (22) and (23) are therefore more complex than a single scattering initialization
but far less complex than an initialization based on the sum of single plus second order
scattering.
Since the computer time required in the invariant imbedding initialization escalates roughly
as K3, as in the doubling method itself, the time required to initialize according to Eqs. (22) and
(23) is fixed relative to the time required to perform a single doubling. The diamond in-
itialization recommended by Wiscombe’6 and the fourth order Runge-Kutta method recom-
mended by Kattawar and Plass2’ are both more accurate than the second order Runge-Kutta
method presented here, but the computer time required to initialize with these methods
escalates roughly as K4 and KS, respectively. This makes them less attractive in the absence of
phase function renormalization (large K), since the increase in computer time required to
initialize the reflection and transmission functions rapidly exceeds the time saved by initializing
at a larger initial T,.
We have followed Wiscombe’s2* recommendation that the initial layer have an optical
thickness rC - p,/lOO, where pl is the smallest Gaussian quadrature point. This results in an
initial layer of optical thickness 2-19, rather than 2-” as in our earlier work with the single
scattering initialization.23 Since the time required to perform the invariant imbedding in-
itialization is nearly equal to the time required to perform 2 doublings, we reduce the number of
doublings required by 11 with a modest increase in the time required to perform the in-
itialization.
In addition to the invariant imbedding and single scattering initializations we performed a
limited set of calculations using the diamond initialization with an initial layer of optical
thickness 2-13 (i.e., T, - CL’). Using the more time consuming diamond initialized doubling
results as representative of an exact solution, we compared the reflection and transmission
functions, plane albedo and conservation of energy at 7, = 32 and found in all instances that the
invariant imbedding initialized results exceeded in accuracy the single scattering initialized
results. Though the single scattering results conserved energy very well, the reflection and
transmission functions at most angles of incidence and reflection were less accurate than the
invariant imbedding results. This is in agreement with the findings of Kattawar and Plass2’ who
showed that the use of energy conservation as a test of accuracy is misleading, for the single
scattering initialization conserves energy exactly, except for round-off in the calculations.
When r, has been made large enough by the doubling method, the numerical results must
agree with known asymptotic expressions for the reflection and transmission functions of very
thick layers. In the case of conservative scattering these expressions are given by24.h

R(T,; FL,PO,4) = R&, PO,4) - UT,; CL,PO,41, (24)

T(T,; CL,
PO,4) = ~K(PL)K(Po)/M~
- d(Tc + ho)l, (25)

where R,(p, all, 4) is the reflection function for a semi-infinite atmosphere, K(p) the escape
function, and q. the extrapolation length for conservative scattering, where q’ = (1 - g)q, is
known to range between 0.709 and 0.715 for all possible phase functions.6 Since the trans-
mission function for optically thick layers is azimuthally independent [see Eq. (25)], the entire
azimuthal dependence of the reflection function is contained in the Fourier expansion of the
reflection function for a semi-infinite atmosphere, given by

R&, po, 4) = R,‘(F, po) + 7. i R=“‘h c~o)COsm+ (26)

*=I

For the azimuth-independent term of the reflection and transmission functions doubling
computations were performed from an initial layer of optical thickness 2219 to a final layer of
152 M. D. KING

optical thickness 32 (24 doublings), from which R_‘(p, F~), K(p) and q. were determined by the
asymptotic fitting method of van de Hulst. 24V6 Figure 5 illustrates the escape function K(p) as a
function of p for both the FWC and Henyey-Greenstein models. It is evident from this figure
that the transmitted radiation at the base of an optically thick atmosphere is 4.86 times greater
at the zenith (CL= 1) than at the horizon (cc =0) for the FWC model (4.33 for the Henyey-
Greenstein model), Though the single scattering phase functions differ substantially between
these models (see Fig. l), the relative difference in the escape function is generally less than a
few per cent, with the maximum difference of 11% occurring at the horizon.
Once the escape function has been determined, the reduced extrapolation length q’ is
obtained from the moment integral6

4’=(1-g)q,=2 r W~CL* dcL, (27)

I0

where the integral is evaluated by Gaussian quadrature. The reduced extrapolation length for
the FWC model is 0.71478 and for the Henyey-Greenstein model is 0.71394. Since q’ shows
little sensitivity to details of the phase function, and since K(p) must obey the normalization
condition

1= 2 ’ K(p.)p dp,
I0

it is not surprising that K(p) shows little sensitivity to the high order Legendre coefficients of
the phase function.
Figure 6 illustrates the azimuthally independent reflection function for a semi-infinite
atmosphere, R,‘(p, po), where the left portion of the figure applies to the FWC model and the
right portion to the Henyey-Greenstein model. R,‘(p, po) was determined by applying the
asymptotic fitting method24X6 to doubling computations at an optical thickness of 32, viz.,
R,‘(p, po) = R’(32; p, po) + ‘i”‘(32; CL,po), The difference between the reflected flux density (or
plane albedo) of the FWC and Henyey-Greenstein models is generally less than 0.2% at an
optical thickness of 32, rising to 0.66% only at grazing incidence. The reflection function

,,I0.0 0.2 0.4 0.6 0.6 1.0

Fig. 5. Comparison of the escape function K(p) for the FWC and Henyey-Greenstein models.
 The reflection function for heres

1.0 "'b.0 0.2 0.4 0.6 0.0

P cc

Fig. 6. Azimuthally independent reflection function for a semi-infinite atmosphere, R,‘(F, M), where the
figure on the left applies to the FWC model and the figure on the right to the Henyey-Greenstein model.

R=‘(p) po) itself shows differences between the FWC and Henyey-Greenstein models of up to
25% for perpendicular incidence ~~~ = 1) and 72% for grazing incidence (/.L~= 0). At some
values of p0 (viz., 0.30 9 po5 0.35, 0.63 5 ko-< 0.74) the differences in R_‘(p, po) are less than
10% for all values of CL.In general the largest differences in the reflection function occur at
angles for which the contribution from multiple scattering is small compared to the contribution
from single scattering. Since the reflection function at p = p0 = 0 is entirely determined by
single scattering,6 the largest difference (72%) occurs at grazing incidence and reflection. For all
other values of p. the contribution form multiple scattering is the smallest for grazing
reflection, so that the largest difference for a given p. occurs for /.L- 0.
. Combining the results of Figs. 3 and 6, the ratio of the total reflection function to the
first-order (single scattering) reflection function may readily be determined. This ratio, given by

L
Lrn(~, PO)= 4(~ + PO)Ra”‘h ~o)lh”(- PCL,
PO), (29)
.

is illustrated in Fig. 7 for the azimuthally independent term (m = 0), where again the left portion
of the figure applies to the FWC model and the right portion to the Henyey-Greenstein model.
It is evident from Fig. 7 that the multiple scattering enhancement factor Xx0&, po) approaches
unity in the limit p = p. = 0, as required,6 and that the smallest multiple scattering enhancement
for a given value of p. occurs for CL- 0. Furthermore, the enhancement of the first-order
reflection function by multiple scattering generally increases as p and pa increase. Comparing
the Henyey-Greenstien results in Figs. 3 and 7 along the diagonal p = /.L~,one sees that
/I’( - CL,po) monotonically decreases as p increases from 0 to 1, whereas the enhancement of
the single scattering reflection function by multiple scattering, as measured by Xx0@, pO),
monotonically increases. This trade off gives rise to the well defined saddle point in the
reflection function R,‘(p, po) at k = p. - 0.6 (see Fig. 6).
For the azimuth-dependent terms of the reflection and transmission functions doubling
computations were performed from an initial layer of optical thickness 2-19 to a final layer of
sufficient thickness that the internal and transmitted intensities are negligibly small, as required
by Eq. (25). For 1 5 m I 3 calculations were performed to a final layer of optical thickness 32
(as in the azimuth-independent term). As m increased further it was sufficient to terminate the
calculations at an optical thickness of 16 for 4 5 M 5 27 and 8 for 28 5 m I 30. This necessarily
reduced the number of doublings required from 24 (m I 3) to 22 (m 2 28).
Figure 8 illustrates the azimuth-dependent reflection functions Rxm(~., po) for m = l-4 and
for the FWC model. These functions, which are valid for TV2 32 when m = 1-3 and
0.0 0.2 0.4 0.6 0.8 1.0 --&- 0.2 0.4 nc na
U.” v.0 1.0 _
P
Ir

Fig. 7. The same as Fig. 6, except for the ratio of the total reflection function to the first-order (single
scattering) reflection function: X=“‘(w, ~0) = 4(p + ~0) R,“‘(F, pg)/h”‘( - k, PO) for m = 0.

I.” “.L 0.4 0.6 m 1.0

0.8

.4 0.6 0.8 1.0

P

Fig. 8. Azimuth-dependent reflection functions R,‘“(p. ~0) for m = l-4 and for the FWCmodel
 The reflection function for optically thick atmospheres 155

7,. 2 16 when m = 4, necessarily obey symmetry in /.L and po, as required by the principle of
reciprocity. Comparing the results for R_‘(F, po) with the redistribution function h’( - p, po),
illustrated in Fig. 4, one sees that the distribution of positive and negative values is similar in
overall shape but the demarcation between positive and negative values has been shifted in
angular position. With the single exception of Rx2(p, po), the reflection functions for increasing
Fourier frequency develop an increasing frequency of positive and negative values. At
R,3”(p, I*,,),for example, nearly half of the angles have negative values for the reflection function,
though the absolute magnitude of the reflection function is quite small for the majority of angles.
The azimuth-dependent reflection functions for the Henyey-Greenstein phase function are
presented in Fig. 9 for m = 1 and 2. The reflection functions for m = 3 and 4 (not illustrated)
appear similar to those for m = 1 and 2, except that the magnitude of R,‘“(p, po) as a function of
m monotonically decreases for all values of /.L and ko, especially for /.L and k. near 1 (see
Section 6). Unlike the FWC results presented in Fig. 8, the azimuth-dependent reflection
functions for the Henyey-Greenstein model are positive for all values of p and p. and for all
values of m. Due to azimuthal symmetry when either p = 1 or ~~ = 1, it is necessary that
Rxrn(p, I) = R,‘“(l, po) = 0 for all m 2 1. This condition, which is particularly obvious in Fig. 9,
is also present in Fig. 8.
As in the case of the azimuth-independent term, it is instructive to examine the
ratio of the tota! reflection function to the first-order reflection function for the m = 1 term. This
ratio, given by X,‘(p, CL”),is presented in Fig. 10 for both the FWC (left) and Henyey-
Greenstein (right) models. For the Henyey-Greenstein model X,‘(b, @o)5 X_‘(/L, po) for all
values of IJ-and k,, (see Fig. 7). As m increases further the enhancement of the single scattering
reflection function by multiple scattering continues to decrease so that at m = 3, X,‘(p, po) is
everywhere less than IO.
van de Hulst” was the first to observe that the contribution from successive orders of
scattering decreases rapidly with increasing order of the Fourier expansion of the reflection
function. This led him to suggest that multiple scattering computations might be made less time
consuming for large values of m if this observation could be utilized. In addition to the angles
of incidence and reflection, however, the azimuth-dependent enhancement factors X_“‘(p, po)
depend strongly on the details of the phase function, on the single scattering albedo, and on
optical thickness. For the Henyey-Greenstein model, for example, the enchancement factor
X,3(O.S, 0.5) = 5;9Ol when g = 0.84123. This indicates that higher-order scattering contributes
nearly 5 times as much as single scattering to the value of the reflection function at these
angles. In contrast, van de Hulst” presents results for a conservatively scattering Henyey-

0.8

0.6 0.6

nn
_.- 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1 .o
P P

Fig. 9. Azimuth-dependent reflection functions ~0)

R,"'(p. for m = I-? and for the Henyey-Greenstein
model.
 156 M. D. KING

0.8 0.8

0.6 0.6

PO PO

0.4 0.4

0.2 0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 -
v Ir

Fig, 10. Ratio of the total reflection function to the first-order reflection function Xxm(~, 11”) for m = 1.
where the figure on the left applies to the FWC model and the figure on the right to the Henyey-Greenstein
model.

Greenstein model having g = 0.5, a case for which Xx3(0.5,0.5) = 1.272. From these results,
together with those presented by van de Hulst6 as a function of T, and oo, we conclude that the
azimuth-dependent enhancement factors for reflection increase with increasing g, increasing T,,
and increasing wo. For a conservatively scattering, optically thick, atmosphere having a large
asymmetry factor, the approach of R,“‘(p, po) to a first-order reflection function as m increases
is so slow as to be of little help in reducing the labor involved in making multiple scattering
computations.
For the FWC model, illustrated in the left portion of Fig. 10, the ratio of the total reflection
function to the first-order reflection function for the m = 1 term shows large regions where
X,‘(,u, po)< 0, as well as angular combinations where Xa’(p, ko) = to. Comparing the dis-
tribution of positive and negative values of R,‘(p, po) with h’( - CL,po), presented in Figs. 8 and
4, respectively, one sees that the shift in the angular position where these functions are zero
accounts for the distribution of positive and negative values of X,‘(p, po). At angular positions
where R,‘(p, pO) = 0 it is necessary that X_‘(F, ko) equals zero, whereas X,‘(F, po) equals
infinity when h’( - CL,po) equals zero. The large negative region in X,‘(p, po) arises from
positive values of R,‘(p, po) and negative values of h’( - CL,po), whereas the small negative
regions arise from negative values of R,‘(p, po) and positive values of h’( - CL,po). As m
increases the demarcation between positive and negative values of R,“‘(p, ko) increasingly
coincides with angular positions where h”( - p, po) equals zero. As a consequence the negative
regions in X,“‘(p, po) become narrower as m increases, but the negative regions and sin-
gularities persist even at m = 30. It is therefore difficult to uniformly replace multiple scattering
computations with single scattering computations as m increases for a general asymmetric
phase function.

6. NUMBER OF FOURIER TERMS REQUIRED

The number of terms required in the Fourier expansion of the reflection function to achieve
a given level of accuracy depends strongly on CL,~~ and T,. Restricting our attention to optically
thick atmospheres, cases for which Eqs. (24)-(26) apply, it follows that Eq. (26) may be
rewritten by replacing the fixed upper limit L by a variable upper limit M(p, cam)such that

Lb, PO, 4) = Rzo(wL,PO) + 2 2 Rx’%, ~~1 cos mi. (30)

m=l

When p0 and p are both near grazing incidence and reflection and single scattering dominates,
 The reflection function for optically thick atmospheres 157

M(p, pO) necessarily approaches L. On the other hand the azimuthally independent term is the
only term required when either p = 1 or p. = 1.
The observation that M(F, po) is a strong function of p and p. led Dave and Gazdag* and
Hansen and Pollack4 to suggest that multiple scattering computations could be made less time
consuming by terminating the calculations at some value of m which depends on F and /.L~.In
addition to the angles of incidence and reflection, however, the values of M(/L, ko) depend on
the single scattering albedo and phase function as well as on the criterion used to terminate the
series. As a consequence it is not in general possible to estimate a priori the number of
azimuthal terms to be carried in the computations, except for the trival case when either /.L= 1
or I*~= 1.
Figure 11 illustrates the separate terms in the Fourier expansion of the reflection function as
a function of m for p. = 0.5 and for selected values of CL.Filled symbols indicate positive
values while open symbols indicate negative values. For the Henyey-Greenstein model,
presented in the right portion of Fig. 11, the azimuth-dependent reflection functions are all
k* positive. Furthermore, the decrease of R_‘“(k, po) with increasing m is nearly linear on a
semi-logarithmic scale for m 2 2, as noted by van de Hulst,6 with the steepness of the decrease
being directly related to the steepness of the angles of incidence and reflection. For a fixed angle of
,
, incidence the fewest number of azimuthal terms are required for reflection to zenith (CL= 1) while
the largest number of terms are required for grazing reflection (/.L= 0).
For the FWC model, presented in the left portion of Fig. 11, the azimuth-dependent
reflection functions contain both positive and negative values (see Fig. 8). In addition, the
absolute values of R,“(p, clo) undergo oscillations as a function of m which are not present in
atmospheres obeying the Henyey-Greenstein phase function. As a consequence, the number of
Fourier terms required to describe the reflection function to a given level of accuracy depends
on the criterion used to terminate the series.
In the case of the fair weather cumulus model, the number of terms required to attain a
relative accuracy of 0.1% can best be assessed by evaluating R&L, po, 4) as a function of the
largest term retained in the Fourier series expansion. Figure 12 illustrates this Fourier series

3
lo.6C
0 5 10 15

l-n
20 25 30 0 5' 10

Fig. 11. Separate terms in the Fourier expansion of the reflection function of a semi-infinite
15
m
20

atmosphere for
25

,UO= 0.5 and for selected values of p. Filled symbols indicate positive values while open symbols indicate
negative values. The figure on the left applies to the FWC model and the figure on the right to the
Henyey-Greenstein model.
158 M. D. KING

3.0 , , , ( , / , , , , / , ( ( , ( , , , , , ( , , , , , , ,

1
2’5
t I
I f
p=0’4’3

I/~i . . ,. . ..*.... -* ~ l ...lr=o.5400

:
gJv* .
. l . l
-a-r---a--- --,---*-*-*-*-*-*-*-*-*-*r=o.-
~-*-*-*-*-*-*-*r=~ ,413 .
p=o.&so
;:** *-4 & .
---+-+*~~,,,**.~-0.5000
. .-.-a.--*
. -* .
...
.l .p=o,,4,3

0.3’ ” ” “1 ” ’ “1 ’ “1 ” ‘1’ ‘1 ““I

0 5 10 15 20 25 30
M

Fig. 12. Reflection function R&L, PO, 4) as a function of the largest term retained in the Fourier series
expansion of the reflection function for ~0 = 0.5 and for selected values of p and 4, Results apply to the

FWC model.

reconstruction for p. = 0.5 and for selected values of p and 4. For each combination of p and
p. the values of R&J,, po, 4) which would be obtained if M(p, CL,,)= L were estimated for each
of four azimuthal angles. Having determined estimates of R&L, po, $I) at selected values of 4,
we determined values of the largest term required in the series to be assured an accuracy of
0.1%. These results were intercompared for each azimuth angle to determine a representative
value for M(p, ko). As CLand p. depart further from unity the inter-azimuth variability in the
required number of Fourier terms tends to increase. Thus the selection of a representative
value for M(F, po) contains some subjectivity. In the results presented below the values for the
required number of terms should be read give or take a few terms, with generally less terms
required for azimuth angles 0” and 60” and more terms required for azimuth angles 120” and
180”. Furthermore, as the required number of terms increases it becomes increasingly more
difficult to estimate asymptotic values for R&L, po, 4) since multiple scattering calculations were
performed only out to m = 30. As a consequence, values of M(p, po) are reported out to a
maximum value of 23.
This description of the criterion used to terminate the Fourier series representation of the
reflection function may be made more concrete by examining in detail the results presented in
Fig. 12. The curves marked with the circles approach within 0.1% of their asymptotic values for
all four azimuth angles at M(p., po) = 19. In this example, corresponding to the case when
).L= p. = 0.5, the 4 = 0” and 60” azimuth planes are characterized by oscillations about asymp-
 The reflection function for optically thick atmospheres 159

totic values of R,(p, ko, 0) for values of M less than 19, though the magnitudes of the
oscillations exceed 0.1% in both cases. In the case of the 4 = 120” and 180” azimuth planes
more terms are required to approach asymptotic values of R&, ~~~4) but once approached the
amplitudes of the oscillations are within the desired threshold level of 0.1%. Similarly, we find
on examination of Fig. 12 that the Fourier series representation of the reflection function may
be terminated at M(p, @,J = 18 for p = 0.1413 and M(F, CL,,)= 11 for p = 0.8660, Comparing
these results with the magnitude of the azimuth-dependent reflection functions presented in the
left portion of Fig. 11, we find that somewhat different results are obtained than might have
been had we based our selection criterion solely on the amplitude of the Fourier coefficients.
Using the specified selection criterion, the number of terms required in the Fourier
expansion of the reflection function was determined for the FWC model and for each value of
p and ~~~ These results, presented in the left portion of Fig. 13, yield the surprising result
that for most values of pa the largest number of Fourier terms are required when p = @o,with
generally fewer terms required when CLis either smaller or larger than po. The local maximum
-* in the vicinity p = p. - 0.95 is associated with angles for which the azimuthal variation of the
reflection function contains multiple scattering signatures associated with the rainbow (4 = 0’)
a and glory (4 = 180”). The results presented in Fig. 13 confirm the expectations of Dave and
3 Gazdag* and Hansen and Pollack4 that multiple scattering calculations need be performed for
fewer azimuthal terms than required to expand the phase function in Legendre polynomials. In
the FWC model, for example, L = 229 and yet it is sufficient to perform multiple scattering
calculations out to m = 20, provided one is interested in applications for which either p or p. is
greater than about 0.42. For less anisotropic phase functions arising from distributions contain-
ing either smaller particles or absorbing particles, it is likely that 21 terms would be sufficient
for even a larger range of p and p. values.
For the Henyey-Greenstein model the criterion used to terminate the Fourier series was
that M(p, ~~~ be that value of m such that all R,“(p, po) with m > M(p, po) be less than 10u3
times Rzo(p, po). This criterion was adequate for the Henyey-Greenstein model due to the
simple monotonically decreasing behavior of the azimuth-dependent reflection functions (see
Fig. 11). These results, presented in the right portion of Fig. 13, conform with the commonly
held belief that for a given value of p. the number of terms required to attain a given level of
accuracy monotonically increases as p decreases. Furthermore, fewer terms are required for
c the Henyey-Greenstein model than for the FWC model at all values of p and cam.For many
.
aircraft and satellite applications involving scanner instruments restricted to scan angles from 0”

PC

0.0 0.2 0.4 0.6 0.8 1.0

Fig. 13. Index of largest term required in the Fourier expansion of the reflection function of a semi-infinite
atmosphere, where the figure on the left applies to the FWC model and the figure on the right to the
Henyey-Greenstein model.
160 M. D. KING

to 45”, the number of azimuthal terms required to describe the reflected intensity field for
optically thick atmospheres will generally not exceed 16 (m = 15).

I.CONCLUSIONS

In the foregoing sections, results have been presented for the separate terms in the Fourier
expansion of the reflection function of a semi-infinite, conservatively scattering, atmosphere
composed of cloud particles. These results have been compared with those for an atmosphere
having a Henyey-Greenstein phase function with the same asymmetry factor as in the cloud
model. The relative difference in the azimuthally independent reflection function, though
generally less than a few per cent, can be as large as 70% at angles where single scattering is
important (viz., p = po- 0). The moments (plane albedo, escape function) and bimoments
(spherical albedo, extrapolation length) of the azimuthally independent reflection function are
generally similar for both models, whereas the azimuth-dependent terms of the reflection
function are generally dissimilar. These results emphasize the fact that the similarity relations
discussed by van de Hulst6 and Kingz3 are the most applicable for the integrated quantities such
as the spherical albedo and the least applicable for details of the reflected intensity as a function
of azimuth angle.
As a prelude to performing multiple scattering calculations it is necessary that each term in
the Fourier expansion of the phase function satisfy a normalization condition in quadraturized
form. Though this requirement has long been recognized for the azimuth-independent term,”
quadraturized normalization conditions for the azimuth-dependent terms are presented here for
the first time. A criterion has been introduced whereby the order of the angular discretization
can be estimated as a function of the number of terms required to expand the phase function in
a Legendre series. Using this criterion the azimuth-dependent normalization conditions are
sufficiently well satisfied that accurate intensities as well as accurate flux densities result from
multiple scattering computations.
The intego-differential equations satisfied by the reflection and transmission functions,
known as the principles of invariance, were solved for an optically thin initial layer by a second
order Runge-Kutta method. This initialization method, which we have referred to as invariant
imbedding, permits doubling computations to be initiated at a larger optical thickness than
required for the single scattering initialization, with little increase in the time required to
perform the initialization. Though our derivation from the principles of invariance is new, the
resulting expressions for the reflection and transmission functions are entirely equivalent to
Wiscombe’sr6 expanded diamond initialization, once allowance is made for differences in the
definition of his reflection and transmission operator’s. In applications involving highly aniso-
tropic phase functions, cases for which a large order of angular discretization is required, the
invariant imbedding initialization presented here is especially attractive. Under these conditions
the more accurate fourth order Runge-Kutta” and diamond16 initializations require more
computer time to initialize than is saved by reducing the number of doublings required.
The azimuth-dependent reflection functions of a semi-infinite atmosphere were obtained for
both the fair weather cumulus and Henyey-Greenstein models by successive applications of the
invariant imbedding, doubling and asymptotic fitting methods. One important finding of the
present investigation is that the reflection function of a semi-infinite atmosphere can be
represented by a Fourier series whose upper limit depends strongly on the angles of incidence
and scattering. These results, presented in Fig. 13, show that the number of terms required to
describe the reflection function is larger for a Mie theory phase function than for a Henyey-
Greenstein phase function. Furthermore, the Henyey-Greenstein results show that for a fixed
solar zenith angle the required number of terms increases monotonically as the zenith angle
increases from 0” to 90”. On the other hand, the FWC model generally requires more terms
when CL= ~~ than when p is either smaller or larger than po. For aircraft or satellite
applications involving scanning radiometers for measuring the reflected intensity field at nadir
angles from 0” to 45”, the number of terms required in the Fourier expansion of the reflection
function for semi-infinite atmospheres will generally not exceed 16 (m = 15).
For atmospheres of sufficient optical thickness that asymptotic expressions for the reflection
and transmission functions apply [( 1 - g)r, 2 1.21, it is only the azimuthally independent
reflection function which varies with optical thickness. As a consequence, the magnitude of the
 The reflection function for optically thick atmospheres 161

m = 0 terms in Fig. 11 will decrease with decreasing optical thickness whereas the m > 0 terms
will remain unchanged. Thus in order to maintain a relative accuracy of 0.1% in the reflection
function of optically thick atmospheres, more terms may be required in the Fourier series
expansion of the reflection function than required for a semi-infinite atmosphere.

Acknowledgements-The author is grateful to H. G. Meyer and K. Govindaraju for aid in performing the computations.

REFERENCES
1. .I: V. Dave, Appl. Opt. 9, 1888 (1970).
2. J. V. Dave and .I. Gazdag, Appl. Opt. 9, 1457 (1970).
3. B. M. Herman and S. R. Browning, J. Afmos. Sci. 32, 1430 (1975).
4. J. E. Hansen and J. B. Pollack, J. Atmos. Sci. 27, 265 (1970).
5. J. E. Hansen and L. D. Travis, Space Sci. Rev. 16, 527 (1974).
6. H. C. van de Hulst, Multiple Light Scattering. Tables, Formulas, and Applications, Vols. 1 and 2. Academic Press,
New York (1980).
7. L. C. Henyey and J. L. Greenstein, Astrophys. J. 93,70 (1941).
r 8. J. E. Hansen, J. Atmos. Sci. 28, 1400 (1971).
<- 9. G. W. Kattawar, S. J. Hitzfelder, and J. Binstock, J. Atmos. Sci. 30, 289 (1973).
10. G. W. Kattawar, JQSRT 15, 839 (1975).
11. A. H. Stroud and D. Secrest, Gaussian Quadrature Formulas. Prentice-Hall, Englewood Cliffs, New Jersey (1966).
I. 12. G. E. Hunt, JQSRT 10, 857 (1970).
13. W. J. Wiscombe, J. Atmos. Sci. 34, 1408 (1977).
14. S. Chandrasekhar, Radiative Transfer. Oxford University Press, Oxford (1950).
15. J. V. Dave and B. H. Armstrong, JQSRT 10, 557 (1970).
16. W. J. Wiscombe, JQSRT 16,637 (1976).
17. S. Twomey, H. Jacobowitz, and H. B. Howell, J. Atmos. Sci. 23,289 (1%6).
18. G. N. Plass, G. W. Kattawar, and F. E. Catchings, Appl. Opt. 12, 314 (1973).
19. K. N. Liou, An Introduction to Atmospheric Radiation. Academic Press, New York (1980).
20. P. J. Davis and I. Polonsky, In: Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical
Tables (Edited by M. Abramowitz and I. A. Stegun). National Bureau of Standards Applied Mathematics Series No.
55, Washington, D.C. (1964).
21. G. W. Kattawar and G. N. Plass, JQSRT 13, 1065 (1973).
22. W. J. Wiscombe, JQSRT 18,245 (1977).
23. M. D. King, J. Almos. Sci. 38, 2031 (1981).
24. H. C. van de Hulst, J. Comput. Phys. 3, 291 (1%8).
25. H. C. van de Hulst, JQSRT 11,785 (1971).

QSRT Vol. 30, No. Z--E