Why is it so difficult to solve the radiative

transfer equation?
Bernard Rutily and Loc Chevallier
(July 2007)

Introduction

The purpose of this Note, which is a slightly modified version of [1], is to show
that the Radiation Transfer Equation (RTE) is a two-faced equation, easy to
solve at times, and difficult to solve at other times, depending on the values
assumed by its coefficients. It is rather tractable when applied to the deep layers of atmospheres and for frequencies from their continuum spectrum, but it is
difficult to solve close to the surface of atmospheres, especially within spectral
lines. It follows that the photons most rich in information that reach our telescopes are also the most difficult to interpret.
It is the escape of photons by the free surface of atmospheres which is at the
origin of the difficulties that will be discussed below. As a result, these difficulties are present whatever the model of atmosphere we are dealing with. This
fact incites us to choose the simplest model to illustrate our matter, that of a
static and stationary plane-parallel atmosphere, with isotropic and monochromatic light scattering.
Under these assumptions, using the optical depth as the space variable and
the angular variable u = cos as the direction variable, the RTE reads [2]
I
(, u) = I(, u) S( ),

the source function S of the radiative field being given by

u

S( ) = [1 a( )]B ( ) + a( )J( ).

(1)

(2)

In the right-hand side of this expression, a( ) is the albedo for single scattering
of the atmosphere, i.e. the ratio of the scattering coefficient to the extinction
coefficient at level . It is a number ranging between 0 and 1, close to 0 on
weakly scattering layers and close to 1 in the contrary case. B ( ) describes the
thermal emission of the atmosphere and coincides with the Planck function in
the deep layers of the atmosphere. J( ) stands for the mean intensity of the field,
which is derived from the specific intensity I(, u) by an angular integration:
Z
1 +1
I(, u)du.
(3)
J( ) =
2 1
1

In this very simple model atmosphere, the number of variables is decreased from
7 (in the general case) to 2, without mutilating the RTE of its essential terms.
Let us notice that the frequency variable does not appear in the above equations,
because it is not transformed by the scattering process we have chosen. The
preceding equations are written for a given frequency, which is not specified.
The absence of change of the photon frequency during a scattering event is not
realistic in spectral lines, since it amounts to adopting a rectangular profile (see
Ivanov [3], p. 57). But our main purpose here is to understand the multiple
scattering of photons in the surface layers of an atmosphere, independently of
the effet of redistribution in direction or frequency of photons. Therefore, the
simplest indicatrix is suitable enough.
The problem (1)-(3) contains all the difficulties we wish to bring to the fore, in
spite of its apparent simplicity. These difficulties come from the fact that the
RTE is both an integrodifferential and a singular equation. The integrodifferential character has been underlined many times in the literature. It appears by
substituting the definition (3) of the mean intensity into the expression (2) of
the source function, then the latter in the right-hand side of the RTE (1). The
specific intensity I(, u) is seen to obey an equation involving its derivative with
respect to the variable and its integral with respect to the u variable. The
solution of this equation is simple when the scattering integral term a( )J( )
can be neglected in the right-hand side of (2), because it is easy to solve a differential equation of the form (1) for a given source function S. This simplification
is possible in the weakly scattering layers of the atmosphere, those where the
albedo a( ) tends to 0. Conversely, in strongly scattering media, a( ) tends to
1 and the integral term of the RTE dominates the other ones. The solution is
then modelled by the integral structure of the problem and becomes far more
complicated. In some cases, the integral term does not dominate separately the
extinction term I(, u) and the emission terms [1 a( )]B ( ), but their difference, which generates many numerical difficulties in strongly scattering and
optically thick media.
Another difficuty appears in the left-hand side of Eq. (1): the advection operator u/ , which describes the streaming of photons along rectilinear rays,
involves the product of the spatial derivative of the unknown function by the
variable u, which vanishes in the domain [1, +1]. It is this term which is responsible for the singular character of the RTE. We shall see that because of this
term, the derivative of the source function necessarily diverges on the boundary
planes. This doesnt look very bright when discretizating the space variable,
especially if one needs to calculate the solution of the RTE close to the surface
of the atmosphere, where the gradients are very high. But it is precisely these
boundary layers which are most interesting for astrophysicists, because most
photons received by their telescopes are coming from there, especially those
emitted in spectral lines.
The RTE thus contains all the ingredients which make its resolution difficult,
even if it is written in the very simple form (1)-(3). We have seen that we must
be wary of the superficial layers of strongly scattering and optically thick media, where the difficulties should be maximal. This will be confirmed in what
2

follows from more quantitative arguments taken from the solar atmosphere. As
expected, it is the transfer in spectral lines which will pose the greatest problems, for the three conditions previously mentioned are simultaneously satisfied
at the central frequencies of the spectral lines.

All is simple in the deep and weakly scattering

layers of a moderately thick atmosphere

Let us come back to what we called an atmosphere in our preceding article [4].
An atmosphere consists of the surface layers of an object open on the vacuum
side, thus loosing photons. It is when the losses are high that the coupled-step
solution of the RTE and the structural equations of the atmosphere becomes
difficult. In other words, in the deep layers of an irradiated object, those where
photons are trapped by the medium itself, this type of problem has a simple
solution. The densities of the particles take values close to those they would
have in a medium in local thermodynamic equilibrium, because matter and radiation are uncoupled. The radiation field can be described in various more or
less equivalent approximations, all of which are appropriate enough: diffusion
approximation, two-stream approximation, Eddington approximation... The
problem becomes considerably complicated when one approaches the surface
layers of the objects which lose photons, and none of the preceding approximations is any longer valid. The mean free path of photons lengthens as they feel
the possibility to escape freely from the boundary surface, and the transport
of photons loses the local character it had in the optically thick layers, which
complicates it considerably. It is under these conditions that the integral term
of the RTE becomes predominant. Before specifying how the solution of the
RTE behaves in this context, we need to go back to the circumstances which
make this solution easier, in order to stress the fact that they only can be local.
Let us place in the deep layers of an atmosphere, which is supposed to be made
up of a mixture of atoms, ions and electrons to simplify. In these optically thick
layers, the loss of photons is weak and the medium is in a state close to local
thermodynamic equilibrium, which means that:
(a) all particles have maxwellian distribution functions at the same temperature
T,
(b) the distribution of atoms over their various states of excitation and ionization is described by the Boltzmann and Saha relations, respectively.
These conditions prevail in media dense enough for the collisional processes to
dominate the radiative processes and be microreversible. If this is the case, the
absorption and emission coefficients of the RTE are related by the KirchhoffPlanck relation, which allows to write on the right-hand side of (2)
B ( ) = B [T ( )]

(LTEM).

(4)

B [T ( )] designates the Planck function at temperature T ( ) [2]. Media in

which this condition is satisfied will be said to be in Local Thermodynamic
3

Equilibrium for Matter (LTEM).

The preceding definition is that of Local Thermodynamic Equilibrium (LTE)
used in most works of the years 1960-70, for example in Mihalas [2] or Ivanov
[3]. The language has evolved since then, and today LTE often applies to the
configurations in which the source function S coincides with Planck function at
the local temperature, i.e.
S( ) = B [T ( )]

(LTE).

(5)

This definition of LTE is obviously more restrictive than the one of the past, and
a certain confusion reigns in the current literature about LTE. One often finds
statements that amount to saying that the conditions (a) and (b) are fulfilled,
and in which it is concluded that the medium is in LTE in the sense of (5),
whereas it is only in LTEM in the sense of (4).
To understand to what extent LTE is more restrictive than LTEM, let us suppose
that the LTEM condition holds and let us write relation (2) in the form
S( ) = B [T ( )] + a( ){J( ) B [T ( )]}.

(6)

It may be seen that under the condition of LTEM, the medium is in LTE if, and
only if, the condition a|J/B 1| 1 is met everywhere. That is the case in
a weakly scattering medium (a 0) but also in a medium with a high albedo
and large optical thickness, so that the mean intensity is close to the Planck
function. The first situation corresponds to that of a continuum in the deep
layers of an atmosphere, the second applies to spectral lines in the same layers.
The important point is that LTEM and especially LTE are possible only locally,
in the deepest layers of the atmospheres, and not globally. To solve a problem
of transfer by supposing LTE everywhere can give only bad results when one
seeks to describe the radiation leaving an atmosphere: this radiation comes from
layers which by definition (i.e. because of the loss of photons) are not in LTE.
We will give further two examples to illustrate this last point quantitatively.
We have seen that LTE does not inevitably mean the absence of scattering. On
the other hand, a non-scattering medium in LTEM is necessarily in LTE according to (6), and so the radiation field it generates is the same as that of a medium
in LTE. It follows that the assumptions LTE and a = 0 are different, but
they lead to the same radiation field.
There are other approximate treatments of the radiative transfer in the optically thick layers of an atmosphere, for example the diffusion approximation
or the Eddington approximation. The first one reads I(, u) = B [T ( )] +
u(B / )[T ( )], it is stronger than LTE because it is equivalent to LTE plus
the condition 2 B / 2 [T ( )] negligible everywhere. The Eddington approximation provides a very convenient relation to close the moment equations derived from the RTE, and because it is weaker than the diffusion approximation,
it has a wider domain of validity [5].
Let us turn to the principal issue: all of these approximations of the source function - the heart of the solution of the transfer equation - are only appropriate in
the deepest layers of an atmosphere, which are indeed nearly in LTE. As soon
4

as photons leave an atmosphere, the source function falls short of the Planck
function, and the emissive layers cannot be in LTE. In other words, there is no
atmosphere which is in LTE as understood today: these two terms - atmosphere
and LTE - are incompatible. We illustrate these remarks with the help of two
examples.
Let us first consider the case of a scattering medium of constant albedo a, limited by a sphere of optical radius b, containing at the center an isotropic point
source of luminosity L : a star at the center of a spherical nebula. Let L(a, b)
be the luminosity of the nebula, i.e., the product of its surface by the radiative flux through each unit area. This flux can be calculated analytically in
the general case, but it is useless to refer to this calculation, because we will
be interested in two extreme situations concerning the albedo: either close to
unity or close to 0. In the first case, it may be seen that the luminosity of the
nebula is L(a, b) = (1/a)L , which shows that this luminosity is close to that
of the central star if a is close to unity. It is obvious that there is practically
no absorption, and most of the radiation from the central star eventually leaves
the nebula. On the other hand, if it is supposed that the albedo is close to 0,
the radiation of the central star suffers an extinction of about exp(b) before
leaving the nebula: this is nothing else than the extinction law for totally absorbing media, known since Bouguers work during the 18th century. Since the
flux suffers the same extinction as the intensity, the luminosity of the nebula
is close to its value for a = 0, i.e., L(a, b) = L exp(b). Let us suppose now
that we observe the nebula at a frequency where its albedo is close to unity,
for example at the center of a spectral line. To calculate the luminosity of the
nebula assuming it to be in LTE amounts to supposing its albedo to be nil, in
keeping with our finding that both assumptions lead to the same radiation field.
Accordingly, the luminosity which one obtains is equal to the real luminosity
divided by exp(b), which is equal to 2.7 if b = 1, 22000 if b = 10 and 2.7 1043
for b = 100. High values of b are quite realistic at the central frequency of
strong spectral lines, which goes hand in hand with high albedos. One sees
through this example that to describe the radiation leaving a nebula under the
assumption of LTE amounts to taking into account one photon out of 2.7, 22000
or 2.7 1043 depending on whether b = 1, 10 or 100: such a treatment is not
acceptable even for b = 1.
Let us examine a second example, that of a star located on the median layer of a
plane-parallel homogeneous nebula with albedo a and optical thickness 2b. One
can calculate analytically the flux leaving through one of the boundary planes,
as a function of a and b. Calculating the ratio of the flux for an albedo close to
1 and the flux corresponding to an albedo close to 0, one obtains an estimate of
the error involved by supposing that the slab is both strongly scattering and in
LTE. The ratio of the two preceding fluxes is about 1/E2 (b), which is equal to
6.7 for b = 1 and behaves like (b + 2) exp(b) when b 1. This second example
confirms the conclusion of the first one, rendering it even worse.

All becomes complicated in the strongly scattering surface layers of a very thick atmosphere

Let us place now near the surface of an atmosphere, where the radiative transfer
of energy is generally dominated by the multiple scattering of photons: a( ) 1
when 0. In the next section, this behavior of the albedo will be explained
on the example of the solar atmosphere. When a( ) 1, the scattering integral
term dominates the source term of the RTE (1)-(3), and it is appropriate to
transform this equation into an integral equation satisfied by the source function
S.
This integral equation is derived by solving the differential equation (1) for a
given source function S. The solution is unique for given incoming boundary
conditions, viz.
I(0, u) =
I(b, u) =

I (u) if

1 u < 0,

(7)

0 < u 1.

(8)

I (u) if

The solution is
1

I(, u) = I + (u)e(b )/u +

1
u

I(, u) = I (u)e

/u

S( )e(

)/u

d ,

(9)

for u < 0, and

b

S( )e(

)/u

d .

(10)

for u > 0. Inserting these expressions into the definition (3) of the mean intensity, the latter becomes
Z
1 b
E1 (| |)S( )d ,
(11)
J( ) = Jext ( ) +
2 0
where Jext denotes the corresponding contribution from the external sources
Z
Z
1 1 +
1 0
I (u)e /u du +
I (u)e(b )/u du,
(12)
Jext ( ) =
2 1
2 0
and E1 stands for the first exponential integral function defined by
Z 1
du
( > 0).
E1 ( ) =
e /u
u
0

(13)

The integral equation satisfied by the source function S may be deduced by

substituting the expression (11) for the mean intensity in the definition (2) for
the source function, thus giving
Z b
S( ) = S0 ( ) + a( )
K( )S( )d .
(14)
0

The free term of this equation

S0 ( ) = [1 a( )]B ( ) + a( )Jext ( )

(15)

consists of radiation emitted by the internal sources and scattered by every

volume element irradiated from outside. The convolution kernel defined by the
function
1
K( ) = E1 (| |) ( 6= 0),
(16)
2
describes the photon scattering process. This kernel is weakly singular on the
diagonal of the domain [0, b] [0, b], as the function K has a weak singularity
at = 0:
K( ) (1/2) ln(| |) when | | 0+ ,
(17)
which does not prevent it from being integrable on R
Z +
K( )d = 1.

(18)

We thus transformed the initial integrodifferential equation into an integral one,

in which the free term incorporates the boundary conditions of the initial problem. Since the RTE (1) is singular, because the factor u multiplying the derivative I/ may vanish, the integral equation (14) inherits that property. Once
solved, we insert the solution into Eqs. (9)-(10) in order to get the specific
intensity at every point of the atmosphere. This transformation of the RTE is
standard, it is done for example in the Section 4 of [6].
There is a long-standing controversy as to whether it is more suitable to solve
the RTE directly in its integrodifferential form (1)-(3), or indirectly by resorting
to the integral form (14). We resolutely belong to the camp of those who advocate the integral formulation, since at a numerical level it is easier to integrate
accurately than to derivate accurately, while from an analytical standpoint the
theory of integral equations has reached far more advanced stage than the theory of partial differential equations. Note that both approaches must face a
major difficulty: the source function S( ) is difficult to compute when 0,
and the specific intensity I(, u) is difficult to compute when 0 and u 0.
The latter difficulty is induced by the former, as I(, u) S( ) when u 0
from Eq. (1). The calculation of the source function gets more complicated
close to the surface of an atmosphere, because this function is not differentiable
on the right at = 0: S ( ) when 0+ . Indeed, by deriving the two
members of equation (14) with respect to , one obtains
Z b
S ( ) = S0 ( ) + a ( )
K( t)S(t)dt
0
"Z
#
b

K( t)S (t)dt + S(0)K( ) S(b)K(b ) .

+a( )

(19)

It appears that even if the functions S0 and a are differentiable at = 0, the

source function S is not, since it behaves like a(0)S(0)K( ) when 0. This
7

result is independent of the nature of the functions S0 and a, and hence of the
model to which the problem (14) corresponds.1
The fact that the source function necessarily has an infinite derivative at = 0 is
one of the points allowing us to understand why the RTE poses such a difficult
problem to solve in the surface layers of atmospheres. Indeed, if the source
function varies rapidly when 0, its numerical evaluation by discretization
of the variable becomes arduous in the neighborhood of = 0. Not only
does the literature skip this major difficulty, but it is also delusive insofar as it
displays curves of the source function in which the optical depth is shown on a
logarithmic scale [see e.g. the Chapt. 11 of [2]]. By choosing such a scale, the
derivative of the source function in the neigborhood of 0 becomes
dS
dS d
dS
a(0)
=
=
a(0)S(0) K( )
S(0) ln( ),
d ln( )
d d ln( )
d
2
i.e., it tends to 0 as 0+ : the change of scale has converted the infinite
derivative into a vanishing derivative! This is illustrated on Fig. 1, taken from
[7], which shows the source function S of a homogeneous and isothermal slab of
temperature T , not illuminated from the outside: S0 ( ) = (1 a)B (T ), with
constant a and T . We have chosen a = 0.9999 and b = 2000, values typical of an
average spectral line. The source function has been normalized with Plancks
function, it is thus the ratio S( )/B (T ). Note the violent decline of this ratio
when is smaller than the thermalization depth of the atmosphere, equal to 58
for the type of scattering envisaged and for the adopted albedo value [7]. This
illustrates the developments in Sec. 2 regarding the example of an isothermal
slab, thereby favouring a priori the LTE conditions.
We now wish to focus our attention on the solution of the problem (14), which
constitutes the main step in solving the RTE (1)-(3). Specialists in integral
equations are few, perhaps because the theory of integral equations is more advanced than other subjects like partial differential equations or statistics, which
concentrate the efforts of mathematicians. Though not numerous, a community
does exist and has a hard core made up by about a hundred mathematicians
which regularly attend the IMSE conferences (IMSE = Integral Methods in
Science and Engineering). The group comes together every two years in meetings organized by turns by its members. We had the chance to participate at
the meeting of 2002 at Saint-Etienne (France), organized by M. Ahues and A.
Largillier, members of the team of Numerical Analysis of the University Jean
Monnet of Saint-Etienne (www.univ-st-etienne.fr/anum). We thereby were able
to draw the attention of the specialists in integral equations on the vast field of
application offered by theories of transfer and transport [8]. During the IMSE
2002, many discussions took place around the problem (14), from which the
following circumstances making its solution difficult to reach have emerged:
(i) the function K is not regular on the interval where it is defined,
1 We

note that the source function has also an infinite derivative at = b, since it behaves
like a(b)S(b)K(b ) when b . We suppose here that the source function does not
vanish on the boundary planes: S(0) 6= 0 and S(b) 6= 0.

500

1000
1

10

-1

0,5

10

linear scale
log scale
-2

10
-4
10

0
10

-3

10

-2

10

-1

10

10

10

10

Figure 1: Source function S( ) in a slab with a = 0.9999 and b = 2000. Right

and top axes display the linear plot with infinite derivative at = 0. Left and
bottom scales are logarithmic.
(ii) the free term S0 varies rapidly in the neighborhood of one or more points,
(iii) the albedo a is close to unity,
(iv) the domain of the integral is extended: b 1.
The two first points are easy to understand. The mathematical origin of the
latter two may be sensed by referring to the existence and uniqueness condition
of the solution of (14) [9]:
sup {a( )} [1 E2 (b/2)] < 1.

(20)

[0,b]

One sees that the condition is the more difficult to fulfill the more the maximum
value of the albedo approaches unity and the greater the optical thickness b of
the atmosphere.2
We shall see that the preceding four conditions are simultaneously fulfilled in
the surface layers of an optically thick atmosphere that scatters strongly, both
conditions of which are particularly well fulfilled at the centre of spectral lines.
By way of example, we shall invoque the case of the solar atmosphere, since
that was the earliest field of application of the pioneers of transfer theory. The
example of the Sun will enable us to understand the physical reasons behind
the simultaneous emergence of the conditions (i)-(iv) in the surface layers of a
stellar atmosphere.
2 We

remind that E2 ( ) =

R1
0

e /u du, which shows that E2 ( ) 0 as +.

The example of solar atmosphere

Let us examine what become the four points (i)-(iv) when we solve the RTE
in the Suns atmosphere. To calculate the coefficient of the RTE for the solar
atmosphere, we have used the model of Vernazza et al. [10], still currently used,
and which extends as high as the base of the transition region with the corona.
Condition (i):
The function K is not regular at = 0, where it diverges like ln(| |). This
divergence stems from the presence of the advection operator u/ on the
left-hand side of the RTE, which describes the rectilinear propagation of light
between two interactions. The induced singularity lies at the heart of the transfer equation and must be addressed.
Conditions (ii) and (iii):
We have drawn in Fig. 2 the curves representing the functions B and a needed
to calculate the free term S0 of equation (14) thanks to its expression (15).
These functions were reckoned for three wavelengths : in the continuum at
= 500 nm, and at the center of the lines H and Ly at = 656.3 nm and
= 121.6 nm respectively. In the continuum, the function B was calculated
for an atmosphere in LTEM, i.e., by writing B = B (T ), the temperature being available in tabulated form in [10]. The opacity was deduced from the
tabulated optical depth, and the scattering coefficient was calculated by taking into account Thomson scattering of photons by free electrons and Rayleigh
scattering by the atoms of hydrogen. Whence a = /.
At the center of the two spectral lines ij (i < j), the opacity has been calculated
by means of the following classical relation:
ij [1

gi nj hij
]
Bij ni ij (ij ),
gj ni 4

(21)

which holds when complete frequency redistribution is assumed in the line, and
when the opacity of the underlying continuum is neglected [2]. The coefficients
gi and gj denote the statistical weights of the atomic states i and j, and the
ni and nj are the number densities of the lower and upper states, respectively.
Bij is Einsteins absorption probability for transition i j and ij is the line
absorption profile. In order to simply calculate the functions B and a at the
center of the line, we assumed that the levels i and j are mainly populated by
the transition i j, which allows to write B Bij (T ) and
aij

Aji
Aji + [1 exp(

hij
kT )]Cji

(22)

In this formula, Aji denotes Einsteins spontaneous-emission probability for

transition j i and Cji the collisional rate of this transition [2].
Let us recall that scattering in a spectral line is not truly a scattering event
but a photo-excitation followed by a radiative de-excitation. True scattering
is generally neglected in a spectral line. But another scattering integral appears when calculating the emission term of the RTE by taking into account
10

the equations of statistical equilibrium, an essential step for the convergence

of the system of coupled equations of transfer and statistical equilibrium [2].
Physically, the occurrence of a scattering integral in the emission term of the
RTE comes from the fact that an excited atom bears in mind the way in
which it has been excited. If it was excited by absorption of a photon, the probability that it shall de-excite spontaneously depends on the specific intensity
of the exciting radiation field, thereby leading to the awaited scattering term,
of albedo given by (22). The latter is then interpreted as the probability that
an absorbed photon is re-emitted rather than being destroyed by a collisional
de-excitation process. In other words aij = 1 ij , where ij is the photon
destruction probability in the line [2].
The curves of B in Fig. 2 show that this function can vary very rapidly within
the chromosphere, suggesting the use of adaptive grids in for its calculation.
The albedo curves show that in the three cases the albedo tends to 1 at the
surface, the convergence being much slower in the continuum than in the lines.
In the continuum, the albedo covers the whole interval [0, 1], going from 0.0006
on the internal boundary plane to 0.99 at the surface. In the H line, the albedo
is equal to 0.2 on the internal boundary plane and to 1 106 at the surface,
while in the Ly line it is always close to unity, going from 0.99 in the deep
layers to 14109 at the surface! The albedo curve in the Ly line is of a very
bad foreboding for the numerical solution of Eq. (14), which will be confirmed
in what follows.
The physical reasons of the albedos convergence to 1 at the upper layers of the
atmosphere are as follows: in the continuum, the contribution from Thomson
scattering to the opacity is predominant in the hotter layers of the chromosphere, which explains why the ratio of these two quantities approaches 1. In
the two spectral lines, the collisional processes are widely dominated by the radiative processes within the less dense layers of the atmosphere, and in (22) one
may assume Cji Aji in these layers, thus letting the albedo approach 1. The
great closeness of the albedo to 1 in a spectral line means that a photon emitted
in that line has all the chances of having been emitted by a photoexcited atom,
implying that in the line a scattered photon is in fact involved. It is thus solution
of a transfer equation in which the source term is strongly dominated by the
scattering integral.
Condition (iv):
Finally, the domain [0, b] of the optical depth variable is relatively limited in
the continuum (b = 7.445), but it is very extended in both of the spectral lines
considered: b = 4 105 at the center of the H line and b = 2 1011 at the
center of Ly! The opacity is high in a transition in which the lower level is
highly populated, for ij ni according to (21). That is the case in the Ly
transition, as the lower level is the fundamental one.
We have outlined the circumstances making the solution of the problem (14)
difficult and have shown that those circumstances are present simultaneously
in the simplest problems of transfer that an idealized solar atmosphere poses.
11

-4

10

10

B*(tau)
10

albedo(tau)

-5

10

-1

-6

10

10
10

10

10

-2

-7

10

-8

-9

10

-3

-4

-8

10

-7

10

-6

10

-5

10

-4

10
tau

-3

10

-2

10

-1

10

10

10 -10
-9
-8
-7
-4
-3
-2
-1
0
1
-6
-5
10 10 10 10 10 10 10 10 10 10 10 10
tau

1.1
10

-5

B*(tau)

1.0
0.9
albedo(tau)

-6

10

0.8
0.7

10

0.6

-7

0.5
0.4
10

-8

0.3
0.2

10

-9
-8

10
10

10

-4

-5

-3

-2

-1

0.1 -8 -7
-4
-3
-2
-1
0
1
2
3
4
-6
-5
5
6
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
tau

-3

1.00

-5

-6

10
10
10

-7
-8
-9

-10

10
10

-6

-4

10

10

-7

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
tau

-11

-12

10
10

10
10

-13

-14

albedo(tau)

B*(tau)

-15

-16

10

-17
-3

-2

-1

10

11

10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
tau

12

0.99
-3
-2
-1
0
1
2
3
4
7
8
9
10 11 12
5
6
10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
tau

12
Figure 2: Variations with optical depth of the functions B (left panels) and a
(right panels) in the Suns atmosphere for the 500 nm continuum (upper panels),
the center of the H line (middle panels) and the center of the Ly line (bottom
panels).

10

10
10
10
10

-2

-4

-6

continuum
average line
strong line
1/k()

-8

-10
-4

-3

-2

10 10 10 10

-1

10 10 10 10 10 10 10 10 10

Figure 3: Relative error d of a ALI code described in [7], as a function of .

The three curves are relative to three different couples (, b), where = 1 a:
(0.01, 20) in the continuum, (104 , 2000) in an average spectral line and (108 ,
2 108 ) in a strong line. Black dots show the thermalization depth 1/k() in
the three cases, where k() is explicited in [7].
This point encourages us to solve carefully the RTE in a stellar atmosphere,
justifying the approach developed by the Transfer group: solve accurately a
few idealized problems of radiative transfer theory, in order to establish reference solutions for the validation of codes takling more realistic problems [11].
In [12] many tests performed on the ALI method - by far the most utilized in
astrophysics - are presented. The tests consist in comparing the analytical and
numerical solutions of the RTE in a non-isothermal, plane-parallel atmosphere.
They confirm that the accuracy of ALI is degraded in the above-mentioned circumstances (i)-(iv). More precisely:
- Condition (i): singularity of the integral equation satisfied by the source function. The singularity of the kernel induces that of the source function, which has
an infinite derivative at = 0. One has thus to expect difficulties when calculating this function close to the surface of an atmosphere. This is confirmed by
Figure 3, taken from [7]: in a strong line (solid line), ALI calculates the source
function of an isothermal atmosphere with a precision going from 108 in the
deep layers to 5 103 in the surface layers of the atmosphere. The precision
found on the surface remains satisfactory, but this good result was obtained
only after optimizing the entry parameters of the code ALI and after a very
high number of iterations. By removing these two precautions, i.e. by using the
default numerical parameters of the code and a standard stopping criterion of
the iterations, the surface accuracy of ALI becomes poor, only 15 %.
- Condition (ii): influence of high gradients close to the surface of atmospheres.
13

10

10

106

20
20

50
50

100

200

500

1000

190
0
15
00

108

300
104
200

102

10 0
100
10-12

20
10-10

20
10-6

10-8

50

10-4

10-2

100

Figure 4: Number of iterations Nc ensuring an accuracy of 1% everywhere, as

a function of = 1 a and = b.
This influence can be highlighted by solving the integral equation (14) for free
terms of the form S0 ( ) = n (n = 0, 1, 5) and S0 ( ) = e . We have established
the exact solution of this equation for polynomial or exponential distributions
of sources. The numerical input parameters of our ALI code are fixed at their
default values as assigned by the program TLUSTY developed by I. Hubeny
and Th. Lanz (http://tlusty.gsfc.nasa.gov/). For the values a = 0.9999 and
b = 1000 typical of average lines, the maximum relative error for S0 = n is
a few percent for n = 0 or 1, 15-20 % for n = 5 and it can reach 30% for
the exponential free term. This maximum error generally occurs close to the
surface, except in the case S0 ( ) = e where it is reached in the deep layers of
the atmosphere. This free term simulates the sudden decrease in temperature
when one penetrates into the Suns transition region.
- Conditions (iii) and (iv): degradation of the precision when a 1 or b +.
This degradation is clearly visible on the Fig. 4, also taken from [7]. This figure shows the number of iterations Nc which must be exceeded to warrant an
accuracy better than 1% everywhere, as a function of the parameters = 1 a
and = b. We see that the ALI code converges after about ten iterations for
= 0.1 whatever the value of b, but it requires 200 iterations for convergence
when = 106 and b = 104 , and 1000 iterations when = 109 and b = 106 .
We refer to the tables and curves of [12] for details on the tests performed
on the ALI method. All indicates that the accuracy of ALI is degraded in the
14

difficult cases, as expected. Moreover, these tests show in which circumstances

the relative errors are important, i.e. higher than a value fixed in advance that
one does not wish to exceed, for example 1%. That makes it possible to adjust
the numerical parameters of the code to get results with the required precision
even in the difficult cases, of course at the cost of considerable slowing down
of the calculations. The interest of a code like ALI is to solve numerically the
exact, i.e. nonapproximate equation of transfer. It is thus possible to increase,
within a certain limit, the precision of ALI by refining the grids of discretization
of the variables and u, or by modifying the stopping criterion of the iterations.
This is no longer possible when using methods that solve an approximate form
of the RTE, most of time exactly. We did not perform any test with these
methods, convinced as we are that no approximate method may account in a
satisfactory way for the complex behavior of the exact solution when it is difficult to calculate. On the other hand, all these methods give good results in the
deep layers of the atmospheres...
We come back on the methods which solve numerically the exact form of the
RTE. We have seen that their accuracy can be improved at the expense of their
computing time. There then arises the problem of the duration of calculations
carried out with these codes. The computation time required for solving only
one equation of transfer cannot be too long, because a very high number of such
equations must be solved in order to build a model atmosphere, as seen in [4].
Let us recall that we have to solve as much RTE as there are frequencies in the
grid of the adopted model, the number of frequencies going from 1000 to 100000
depending on the chemical complexity of the atmosphere. Note also that in the
general case, i.e. in a 3D medium without particular symmetry, the number of
space and angular variables is 5, which requires great stocking capacities. In
any problem with frequency redistribution of photons during a scattering event,
the frequency variable has to be added, and the time variable is also present in a
time evolution problem. Moreover, solving the RTE represents only one step in
the iterative resolution of a system including a very high number of equations,
those leading to the volumic coefficients of the RTE together with its solution.
Precision and speed are thus the two almost contradictory requirements to which
any method for solving the RTE must satisfy. We know only one category of
methods that are at the same time precise and fast, namely the analytical methods. But the latter can solve only relatively simple problems in comparison to
those posed to modellers, which prevents their use in concrete cases. In spite of
that, these methods must be developed because they will finally be able to deal
with the realistic problems of our discipline, and they are essential to the validation and the improvment of the numerical methods which solve these problems
today.

Conclusion

The example of the solar atmosphere teached us at least two things:

- The coefficients of the RTE cover the whole of the domain containing them,

15

i.e., there is no empty region. Thus, the variable can take on values very
close to the ends of the interval [0, b], and particularly very close to 0, for example = 109 b. Of course, the extreme values = 0 and = b are not excluded.
The albedo a can take on values in [0, 1], approaching 0 in the continuum and
in the deep layers of an atmosphere, while approaching unity in a spectral line
and within the surface layers. The optical thickness b may be relatively modest
in the continuum, of the order of a few units, and very high in a spectral line:
b = 2 1011 for the solar atmosphere at the center of the Ly line.
- According to the values assumed by the coefficients within their domains of
variation, the RTE may be easy or difficult to solve. In the deep layers of an
atmosphere, the coefficients of the RTE take on values that render the solution
easy, in contrast to what happens at the surface layers, from where the photons
we observe originate.
It appears important to us to distinguish between the situations for which the
RTE is easy to solve from those where it is not. It is indeed only in the easy cases
that our transfer codes converge fast and accurately, and that our approximate
treatments are valid. But there is no numerical or approximate method which
solves the RTE satisfactorily everywhere. Some are appropriate for certain frequencies or for certain regions of the atmospheres, but are ill-suited for others.
There are always problems in the most external part of the atmospheres, in
which the various continua and spectral lines originate. It is this fact that calls
for caution: our telescopes receive photons that are largely difficult to interpret.
The tendency today is to ascertain the satisfactory behavior of our transfer
codes in the simpler cases, and then to apply them to the difficult cases, which
is not justified. It is not rare to see codes tested for media with a small albedo
(even 0!), or bearing on the deep layers of an atmosphere where asymptotic
behavior is good-natured, or even in degenerated cases where the RTE loses
its substance and admits as an exact solution the Eddington approximation, for
example. This by no means signifies that these codes are able to solve the RTE
in the difficult cases. For these, the only valid test is the confrontation with
exact solutions, or, at the very least, to seek the reassurance of identical results
obtained with two completely different codes.
The current situation with respect to the approximate treatment of radiative
transfer problems is most alarming. Because of the absolute necessity to solve
the RTE in 2D or 3D geometry, approximations are made which are incompatible with the astrophysical applications under consideration. All approximations
bearing on the source function - the heart of the RTE - are not innocent, working less well the better services they render (Murphys law?). A most striking
example is the current tendency to forget the scattering term of the RTE,
which amounts to solving the equation in a medium globally in LTE. It is by
no means certain that it may be called judicious to make 2D or 3D transfer at
such a price, since the fact remains that the scattering term dominates largely
in the layers which lose photons. Moreover, the experience gained when scattering is neglected is of no use in tackling the problems with scattering included,
because the RTE poses two completely different problems, according to whether
scattering takes place or not. There is no hope to solve the quadratic equation
16

ax2 + bx + c = 0 with a code capable of solving it only for a = 0...

A majority of astrophysicists thinks that a good fit between a synthetic spectrum of atmosphere and an observed spectrum is a good validation test for the
model atmosphere they built. We think that this is true only if the model is
both physically and numerically self-consistent, which is never the case in
practice. The fact that the majority of our model atmospheres are physically
incomplete or inconsistent was already noted in a recent review of Gustafson
[13], to which the reader is referred. To Gustafsons arguments we should like to
add the statement that our model atmospheres are also numerically (and even
mathematically) inconsistent. We dont know how errors are propagated when
implementing an atmosphere code, especially errors due to the solution of the
RTE. The rare attempts to estimate such errors have the merit to be extant in
the literature [14, 15], but in our opinion they are much too optimistic. Let us
recall that the RTE must be solved typically a few thousand or ten thousand of
times at each iteration of an atmosphere code. The fact of misrepresenting the
solution in certain spatial or spectral domains, or of solving it in an insufficient
number of frequencies, must necessarily shift the populations, the electron
density and the temperature of the atmosphere from their correct values, without preventing the code from converging.
To illustrate this last point, let us examine the way one calculates the temperature by solving the energy equation of a stellar atmosphere. Let us take
the simplest possible example, that of a static and stationary atmosphere, in
radiative and local thermodynamic equilibrium. The energy equation then reads
Z
(~r, ){B [T (~r)] J(~r, )}d = 0.
(23)
0

It is a nonlinear integral equation for the temperature T , on which the opacity

and the mean intensity J depend. Mathematicians do not know how to solve this
type of equation, and astrophysicists have solved it only for gray atmospheres,
i.e., when is independent of . In a non-gray medium, the radiative energy
balance is determined through an integral over frequency which is very difficult
to compute. One solves the equation (23) by determining empirically, at each
point ~r, the value of the temperature which satisfies the equation, without being
sure that its coefficients have been well calculated. But the reasons to calculate
them wrongly are numerous: the mean intensity J can be wrongly calculated
because the RTE is wrongly solved, the opacity may well be inappropriate
because of a mistaken chemical composition of the atmosphere, or a misleading
calculation of the populations or the cross sections, and finally the integration
over in (23) is extraordinarily tricky, because of very fast variations in opacity
with the frequency of radiation. In spite of all these sources of uncertainty, a
temperature satisfying (23) will be found without difficulty, but without any
warning as to the correctness of both the coefficients of the energy equation
and its solution. Our atmosphere codes turn in closed-circuits, which is why
we wrote that they seemed to us numerically and, as far as problem (23) is
concerned, mathematically inconsistent. It is high time to face this difficulty by
proposing benchmark atmosphere models, which, though highly idealized, will
17

make it possible to validate our codes internally, and not by artificial fits with
observations. Works in this direction are still rare, but they do exist: see [16] or
[17]. Unfortunately, neither the energy equation nor the hydrostatic equilibrium
equation are solved in these references, and we are still far from being able to
establish complete benchmark models, even if we simplified chemistry to the
extreme and also had good cross sections available.

References
[1] B. Rutily, L. Chevallier, Why is it so difficult to solve the radiative transfer equation? In Transfert radiatif et exploitation des TGE, third GRETA
meeting, Frejus, France, 11-13 Mai 2005, Ed. Ph. Stee, 1-23 (2006).
[2] D. Mihalas, Stellar Atmospheres, Freeman and Co, San Francisco, 2nd edition (1978).
[3] V.V. Ivanov, Transfer of radiation in spectral lines, National Bureau of Standards Special Publication no 385, U. S. Government Printing Office, Washington, DC (1973).
[4] B. Rutily, La theorie du transfert et ses applications astrophysiques, in this
site (2007).
[5] G.C. Pomraning, The Equations of Radiation Hydrodynamics, Pergamon,
Oxford (1973).
[6] I.W. Busbridge, The mathematics of radiative transfer, Cambridge University Press, Bristol, UK (1960).
[7] L. Chevallier, F. Paletou, B. Rutily, On the accuracy of the ALI method for
solving the radiative transfer equation, Astronomy and Astrophysics 411,
221-227 (2003).
[8] B. Rutily, Multiple scattering theory and integral equations. In Integral Methods in Science and Engineering, Eds. C. Constanda, M. Ahues, A. Largillier,
Birkh
auser, Boston, 211-232 (2004).
[9] A. Osses, O. Titaud, Finite rank approximation based method for solving
the radiative transfer equation in stellar atmospheres and application to an
inverse problem. In Transfert radiatif et exploitation des TGE, third GRETA
meeting, Frejus, France, 11-13 Mai 2005, Ed. Ph. Stee, 77-97 (2006).
[10] J.E. Vernazza, E.H. Avrett, R. Loeser, Structure of the solar chromosphere.
III - Models of the EUV brightness components of the quiet-sun, Astrophysical Journal Supplement Series 45, 635-725 (1981).
[11] B. Rutily, Solution analytique de lequation de transfert, in this site (2007).

18

[12] L. Chevallier, Resolution analytique de lequation de transfert: description

du code ARTY et tests numeriques, in this site (2007).
[13] B. Gustafsson, The current status in the modelling of stellar atmospheres.
In Modelling of Stellar Atmospheres, IAU Symp. Vol. 210, Eds. N.E.
Piskunov, W.W. Weiss, D.F. Gray, 3-18 (2003).
[14] L.H. Auer, P. Fabiani Bendicho, J. Trujillo Bueno, Multidimensional radiative transfer with multilevel atoms. I: ALI method with preconditioning
of the rate equations, Astronomy and Astrophysics 292, 599-615 (1994).
[15] P. Fabiani Bendicho, J. Trujillo Bueno, L. Auer, Multidimensional radiative transfer with multilevel atoms. II. The non-linear multigrid method.,
Astronomy and Astrophysics 324, 161-176 (1997).
[16] R.G. Athay, E.H. Avrett, H.A. Beebe, H.R. Johnson, A.I. Poland, Y. Cuny,
Calculations of solar hydrogen lines: comparative solutions for a standard
line transfer problem. In Resonance lines in Astrophysics, Eds. R.G Athay, J.
Mathis & A. Skumanich, National Center for Atmospheric Research, Boulder, 169-212 (1968).
[17] L. Chevallier, La thermalisation des electrons dans une atmosph`ere stellaire,
Th`ese de lUniversite Claude Bernard-Lyon 1 (2000).

19