THE ROSSELAND LIMIT FOR RADIATIVE

TRANSFER IN GRAY MATTER

François Golse, Francesco Salvarani

To cite this version:

François Golse, Francesco Salvarani. THE ROSSELAND LIMIT FOR RADIATIVE TRANSFER IN
GRAY MATTER. 2008. �hal-00268799�

HAL Id: hal-00268799

https://hal.science/hal-00268799v1
Preprint submitted on 1 Apr 2008

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
THE ROSSELAND LIMIT FOR RADIATIVE TRANSFER IN
GRAY MATTER

FRANÇOIS GOLSE AND FRANCESCO SALVARANI

Abstract. This paper establishes the Rosseland approximation of the

radiative transfer equations in a gray atmosphere — i.e. assuming that
the opacity is independent of the radiation frequency. The problem is
set in a smooth bounded domain Ω of the Euclidian space R3 , assuming
that the incoming radiation intensity at the boundary of Ω is Lipschitz
continuous and isotropic — including for instance the case of a black-
body radiation with nontrivial temperature gradient.

MSC: 85A25 35K55 (82A70, 35B25, 45K05)

Key-words: Radiative transfer, diffusion approximation, Rosseland equa-

tion, Relative entropy, Compensated compactness

1. Presentation of the model

Radiative transfer is an important phenomenon in many branches of
physics and, in many situations, the most significant mode of transmission
of energy.
As established in all textbooks on radiation hydrodynamics (see, for ex-
ample, [15]), radiative transfer in a host medium can be described by the
following set of partial differential equations, with unknowns the radiative
intensity I = I(t, x, ω, ν) and the temperature T = T (t, x). The indepen-
dent variables in these equations are the time t ∈ [0, τ ], τ > 0, the position
x ∈ Ω ⊆ R3 , the spherical angle with respect to a suitable reference frame
ω ∈ S 2 and the frequency ν ∈ R+ . With these notations, the radiative
transfer equations are

1 ∂I


 + ω · ∇x I + Σ(T, ν)(I − Bν (T )) = 0
 c ∂t

(1) Z +∞
 ∂E
Σ(T, ν)(I¯ − Bν (T )) dν,


 (T ) =
∂t 0

where
αν 3
Z
1
Bν = I¯ = I(t, x, ω, ν) dω.
ehν/kT − 1 4π S2
1
2 F. GOLSE AND F. SALVARANI

The parameter c > 0 is the speed of light and Bν (T ) is Planck’s formula for
the black-body radiation at temperature T , the parameters h and k being
the Planck and the Boltzmann constants respectively.
We have introduced the normalization constant
 4
h
α = 15 ,
kπ
so that Stefan’s law assumes the simple form:
Z +∞
Bν (T ) dν = T 4 .
0
The function Σ = Σ(T, ν), is the opacity — or absorption cross-section
— of the host medium. In general, the opacity takes into account how
energy is transferred from the incoming radiation to the electrons in the
background medium. Its expression is complicated in general, as it must
take into account the contributions of bound electrons, of free electrons as
well as transitions from bound to free states. The respective importance of
each such transition in the opacity strongly depends upon the nature of the
background and the temperature. (For instance, bound-bound transitions
are more important for heavy elements; however, their importance decreases
at high temperatures as most electrons are in free states). We refer to
chapter VII of [18] and to [1] for more information on this subject.
Thus we seek to establish the validity of the Rosseland approximation
under the mildest possible assumptions on the opacity, since so little is
known about it in general.
There are however, two particular features of the opacity that complicate
the Rosseland approximation, and that we wish to take into account in the
present paper.
Specifically, our proof of the Rosseland approximation allows considering
opacities that may tend to infinity as T → 0+ . As we shall see, such opacities
lead to a limit equation that is parabolic degenerate, and hence has singular
solutions.
This kind of behavior of the opacity can be seen on the example of the
Kramers opacity (an explicit formula for the opacity in the case of free-free
transitions and for a hydrogen-like atom). The Kramers opacity is of the
form
hν
1 − e− kT
  
kT
ΣK (T, ν) = C 1 + O
(hν)3 (kT )1/2 hν
— see fla (7.72) on p. 174 in [18]. In this example Σ(T, ν) ∼ T −1/2 .
Otherwise, the opacity can oscillate strongly because of bound-free and
bound-bound transitions.
In view of the above considerations, we only assume in this paper that
the opacity satisfies Σ(T, ν) ∼ T −λ as T → 0, with λ ∈ (0, 1].
The other important nonlinear term in system (1) is the internal energy
E = E(T ). This quantity depends on the temperature, but the experimental
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 3

curves suggest that its dependence upon T is simpler than that of the cross
section.
Henceforth, we assume that E is increasing and such that E 0 (T ) ∼ T γ
as T → 0, where γ ∈ [0, 2]. This assumption includes the perfect gas law,
E(T ) = cV T . Without loss of generality we further assume that E(0) = 0.
In some relevant physical situations, for example in the external layers of a
stellar atmosphere [15], the opacity is independent of the photons frequency.
It is therefore convenient to integrate system (1) over frequencies and obtain
the following system of equations, called the gray model:

1 ∂u

4
 c ∂t + ω · ∇x u + Σ(T )(u − T ) = 0



(2) Z
 ∂E 1
(T ) = Σ(T )(ū − T 4 ),


 ū = u(t, x, ω) dω.
∂t 4π S2

Here, the unknowns are the photon density

Z +∞
u = u(t, x, ω) = I(t, x, ω, ν) dν
0

and, again, the temperature T = T (t, x).

In this paper, we restrict our attention to the gray system (2). The
mathematical study of system (1) will be left for future research.
We consider situations where the interaction between radiation and mat-
ter is the dominant phenomenon.
Through multiple interactions, the photon experiences a sort of random
walk in the host medium. Since the equilibrium (Planck’s function) has
mean zero velocity, it is therefore natural to investigate a diffusive regime
for the photons gas.
From a mathematical point of view, this would mean that it is possible
to obtain diffusion-type equations through the usual parabolic scaling:

ε ∂uε 1

4
 c ∂t + ω · ∇x uε + ε Σ(Tε )(uε − Tε ) = 0



(3) Z
 ∂E 1
 ε2 (Tε ) = Σ(Tε )(ūε − Tε4 ),

 ūε = uε (t, x, ω)dω.
∂t 4π S2

Here ε > 0 is a non-dimensional quantity which represents the ratio between

the mean free path of the photons and some characteristic length of the
host medium (for example, in the case of the Sun, the mean free path is
typically of the order of magnitude of a few centimeters, which is negligeable
if compared to the size of the Sun).
It is therefore obvious to look at the behavior of the angular average of
the photon density, governed by system (3), when ε → 0.
4 F. GOLSE AND F. SALVARANI

The first step in obtaining the formal limit is the definition of the macro-
scopic density
Z
1
ρε (t, x) = ūε (t, x, ω) = uε (t, x, ω) dω
4π S 2
(1) (2) (3)
and the flux vector jε (t, x) = (jε (t, x), jε (t, x), jε (t, x)):
Z
1
jε(i) (t, x) = ω (i) uε (t, x, ω) dω,
4πε S 2
where ω (i) denotes the i-th component of the vector ω.
Then, by integrating the first equation of System (3), it is easy to deduce
that
∂ h ρε i
+ E(Tε ) + ∇x · jε = 0.
∂t c
Moreover, integrating the first equation of System (3) with respect to ω in
S 2 after multiplying each side of this equation by ω (i) gives
(i)
ε2 ∂jε
Z
1
+ ω (i) ∇x · (ωuε ) dω + Σ(Tε ) jε(i) = 0.
c ∂t 4π S 2
Assume the existence of the two limits:
ρ = lim ρε , j = lim jε .
ε→0 ε→0

If jε is bounded in some sense and that

lim ε2 Et (Tε ) = 0 lim ε2 (jε )t = 0,
ε→0 ε→0

the second equation of (3) implies that

lim Tε4 = ρ.
ε→0

At leading order in ε, System (3) forces u to be independent of ω, so that

ρ = u. Moreover, ρ solves the following diffusion-type equation, which is
known in the literature under the name of Rosseland equation:
 
hρ i ∇x ρ
+ E(ρ1/4 ) − ∇x · = 0.
c t 3Σ(ρ1/4 )
One major difficulty with the Rosseland approximation is about bound-
ary conditions. As is well known, natural boundary conditions for transport
equations and for diffusion equations are of a very different nature — for in-
stance, Dirichlet boundary conditions are admissible for diffusion equations,
not for transport equations. A typical instance of natural boundary condi-
tion for a transport equation consists of prescribing the density of particles
entering the spatial domain where the equation is to be solved.
So far, the Rosseland approximation has been established in the case of
matter radiating in vacuum — see [3].
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 5

In the more complicated case of a prescribed incident radiation intensity,

the Rosseland approximation is based on asymptotic analysis involving non-
trivial boundary layer analysis: see [12], and [2] for a mathematical justifica-
tion of these asymptotic ansatz, based on somewhat unrealistic monotonicity
assumptions on the opacity.
Here, we discuss the case of an incident radiation field that is a Planck
distribution (thereby avoiding boundary layers), however with a possible
temperature gradient at the boundary. This type of boundary condition
cannot be handled with the ideas of either [2, 3] for realistic opacities.

2. Mathematical assumptions and results

We start this section by declaring some notation which we will use in the
sequel.
We will work in a bounded domain Ω ⊂ R3 with boundary Γ = ∂Ω of
class C 1 .
If we denote with nx the outer normal to the boundary of Γ in x, we can
define the sets Γ− , Γ+ and Γ0 , subsets of Γ, as:
Γ− = {x ∈ Γ : ω · nx < 0}, Γ+ = {x ∈ Γ : ω · nx > 0}
and
Γ0 = {x ∈ Γ : ω · nx = 0}.
Moreover, since in the paper we will heavily use indices, in order to avoid
confusion, any vector v ∈ Rn will be written in the following way, when it
will be convenient to give the components in explicit form:
v = (v (1) , . . . , v (n) ).
We are now ready to give a brief presentation of our results. Goal of
the paper is the mathematical study of system (3): we prove existence of a
weak solution and we deduce rigorously the Rosseland approximation as a
diffusive limit of such system.
We assume that the opacity and the internal energy satisfy the following
properties:
Definition 2.1. Let Σ = Σ(T ) and E = E(T ) be the opacity and the
internal energy of system (3). We say that Σ and E are admissible if and
only if
(1) Σ(y) > 0 for any y > 0 and is of class C 1 ((0, +∞));
(2) lim Σ(y) = 0;
y→+∞

(3) lim Σ(y) = +∞ and Σ(y) ∼ y −β when y → 0, with β ∈ (0, 1].

y→0

(4) E is continuous in [0, +∞) and belongs to the class C 1 ((0, +∞)).
Moreover E(0) = 0 and E 0 (y) ≥ 0 for all y > 0.
(5) E 0 (y) ∼ y α for y → 0, where α ∈ [0, 2].
6 F. GOLSE AND F. SALVARANI

In what follows, we will consider the initial-boundary value problem for

System (3), with (t, x, ω) ∈ (0, τ ) × Ω × S 2 , supplemented with appropriate
initial and boundary conditions.
Our first theorem concerns an existence result:
Theorem 2.2. Let us consider System (3) for (t, x, ω) ∈ (0, τ ) × Ω × S 2 ,
with initial conditions
uε (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )

Tε (0, x) = T in (x) ∈ L∞ (Ω)

and boundary data
uε (t, x, ω)|Γ− = ub (t, x, ω) ∈ L∞ ((0, τ ) × ∂Ω × S 2 ),
where Ω ⊂ R3 is a bounded domain with boundary of class C 1 and ε > 0.
Then System (3) admits a weak solution
(uε (t, x, ω), Tε (t, x)) ∈ L∞ ((0, τ ) × Ω × S 2 ) × L∞ ((0, τ ) × Ω).
This existence result improves on the earlier references [14, 8] which used
unrealistic monotonicity assumptions on the opacity σ, even in the gray
case. It also improves on the existence result in [3], which further reduces
the system of two equations (2) to a scalar equation on the single unknown
u by neglecting the time-derivative of the internal energy in the second
equation of (2). An earlier attempt at treating the system (2) can be found
in [6], with however a lesser degree of generality in the opacity and internal
energy admissible.

Moreover, we give a rigorous justification of the Rosseland limit equation

of the radiative transfer system. Here, in order to avoid additional difficulties
dues to the presence of a boundary layer, we do not allow that the boundary
data depend on ω.
The result is summarized in the following theorem:
Theorem 2.3. Let us consider System (3), posed for (t, x, ω) ∈ (0, τ ) × Ω ×
S 2 , with initial conditions
uε (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )

Tε (0, x) = T in (x) ∈ L∞ (Ω)

and boundary data
uε (t, x, ω)|Γ− = ub (t, x) ∈ W 1,∞ ((0, τ ) × ∂Ω),
where Ω ⊂ R3 is a bounded domain with boundary of class C 1 and ε > 0.
Let {uε } be a family of solutions of System (3) which satisfy the same initial
and boundary conditions.
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 7

Then, there exists a function ρ ∈ L∞ ((0, τ ) × Ω) such that ρ = limε→0 ūε

strongly in L2 ((0, τ )×Ω). The limit function ρ moreover solves the following
initial-boundary value problem in (0, τ ) × Ω, for τ > 0:
 
hρ i ∇x ρ
+ E(ρ1/4 ) − ∇x · = 0.
c t 3Σ(ρ1/4 )
with initial conditions ρ(0, x) = ūin (x, ω) and boundary data ρ(t, x)|∂Ω =
ub (t, x).
For technical reasons, the regularity condition on the boundary data for
the family {uε } in Theorem 2.3 is slightly more stringent than the require-
ments on the boundary conditions in Theorem 2.2.
It is interesting to note that, even if only the incoming density on the
boundary is prescribed for the family {uε }, we obtain that the boundary
condition for the limit is fully defined.
The Rosseland approximation has been established in [2] by a method
based on a multiscale expansion in powers of ε, assuming that the opacity
Σ is decreasing, while T 7→ Σ(T )T 4 is increasing.
In [3], the same Rosseland approximation for a reduced version of the
gray model (2) neglecting the time derivative of E(T ) in the energy bal-
ance equation. The method in [3] used compactness arguments based on
velocity averaging, and hence could handle the most general class of opaci-
ties imaginable. However, it assumed an absorbing boundary condition —
corresponding to radiation expanding in the vacuum. The same method
could also handle the case of a incoming black-body radiation at a constant
temperature. The case of a temperature gradient at the boundary, treated
in the present paper, requires a new idea for controling the radiation flux,
which is the key to the Rosseland approximation.
Finally, we would like to mention the reference [11], where a nonlinear
diffusion approximation is established by compensated compactness — at
variance with the argument in [3] based on velocity averaging instead.
Because of the conditions on the opacity given in Definition 2.1, the Rosse-
land equation is degenerate parabolic, so that the temperature may display
some singularity in the form of a propagating front when radiation pene-
trates into cold matter. Hence result above proves that the Rosseland ap-
proximation is robust in the sense that it holds even in situations in which
such singularities may occur — something that is not at all clear with clas-
sical asymptotic expansions.
The method for proving existence of a weak solution (Theorem 2.2) uses
a compactness argument based on a maximum principle and velocity aver-
aging. The main difficulty is that the absorption-emission nonlinear term is
not a Lipschitz continuous perturbation of the transport operator, since the
opacity tends to infinity as the temperature vanishes.
The method for proving the Rosseland approximation is based on using
the relative entropy to control the radiation flux. Usually, the relative en-
tropy method is used when solutions of the target equation are known to
8 F. GOLSE AND F. SALVARANI

be smooth. This is not the case here, as temperature may fail to be differ-
entiable at interfaces with cold matter. Here, as in the case of Carleman’s
kinetic model [10], we use both relative entropy and a compactness argument
that allows for possibly singular solutions of the target equations. Relative
entropy is not used all the way through the proof to control the distance
between the radiative intensity and the (fourth power of) the temperature
that solves the Rosseland limiting equation, but only to control the radiation
flux. This control is in turn one essential step in the following compactness
argument.
That it is possible to use relative entropy within a compactness argument
to establish the validity of such a macroscopic (or hydrodynamic) limit seems
to be special to the case of limits leading to a pure diffusion equation (de-
generate or not), but without streaming term. In particular, the strategy
presented here does not seem to apply to the derivation of the incompressible
Navier-Stokes equations from the Boltzmann equation.
From a mathematical point of view, it will be more convenient to work
with two new unknowns: the photon density u = u(t, x, ω) and the normal-
ized temperature θ = θ(t, x), the last one being defined as θ = T 4 . Moreover,
we will set c = 1.
We write now the radiative transfer equations (3) with respect to the two
new unknowns.
We first define the normalized cross section σ and the internal energy E
as
σ(θ) = Σ(θ1/4 ), E(θ) = E(θ1/4 ).
The pair (σ, E), must satisfy the same constraints we have imposed on
the pair (Σ, E). The precise meaning of admissible cross section and energy
is given in the following definition:
Definition 2.4. The normalized opacity
σ : R+ → R+
and the internal energy
E : [0, +∞) → [0, +∞)
are admissible if and only if
(1) σ(y) is strictly positive for any y > 0 and is of class C 1 ((0, +∞));
(2) lim σ(y) = 0;
y→+∞

(3) lim σ(y) = +∞ and σ(y) ∼ y −β when y → 0. We will always

y→0
suppose that β ∈ (0, 1/4].
(4) E is continuous in [0, +∞) and is of class C 1 ((0, +∞)). Moreover
E(0) = 0 and E 0 (y) ≥ 0 for all y > 0.
(5) E 0 (y) ∼ y α for y → 0, where α ∈ [−3/4, −β], where β is the same
exponent that gives the behavior of σ when y → 0.
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 9

Then, System (3) becomes, in the new unknowns,

∂u

 ε2 ε + εω · ∇x uε + σ(θε )(uε − θε ) = 0

∂t


(4) Z

2 1
 ε Et (θε ) = σ(θε )(ūε − θε ), ūε = uε (t, x, ω)dω.


4π S 2
Our results, already given in Theorems 2.2 and 2.3, will hence be proved
under the following form:
Theorem 2.5. Let us consider System (4), posed for (t, x, ω) ∈ (0, τ ) × Ω ×
S 2 , with initial conditions
uε (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )
θε (0, x) = θin (x) ∈ L∞ (Ω)
and boundary data
uε (t, x, ω)|Γ− = ub (t, x, ω) ∈ L∞ ((0, τ ) × ∂Ω × S 2 ),
where Ω ⊂ R3 is a bounded domain with boundary of class C 1 and ε > 0.
Then System (4) admits a weak solution
(uε (t, x, ω), θε (t, x)) ∈ L∞ ((0, τ ) × Ω × S 2 ) × L∞ ((0, τ ) × Ω).
Theorem 2.6. Let us consider System (4), posed for (t, x, ω) ∈ (0, τ ) × Ω ×
S 2 , with initial conditions
uε (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )
θε (0, x) = θin (x) ∈ L∞ (Ω)
and boundary data
uε (t, x, ω)|Γ− = ub (t, x) ∈ W 1,∞ ((0, τ ) × ∂Ω),
where Ω ⊂ R3 is a bounded domain with boundary of class C 1 and ε > 0.
Let {uε } be a family of solutions of System (4) which satisfy the same initial
and boundary conditions.
Then, there exists a function ρ ∈ L∞ ((0, τ ) × Ω) such that ρ = limε→0 ūε
strongly in L2 ((0, τ )×Ω). The limit function ρ moreover solves the following
initial-boundary value problem in (0, τ ) × Ω, for τ > 0:
 
∇x ρ
(5) [ρ + E(ρ)]t − ∇x · = 0.
3σ(ρ)
with initial conditions ρ(0, x) = ūin (x, ω) and boundary data ρ(t, x)|∂Ω =
ub (t, x).

3. Existence proof
This section is devoted to the proof of Theorem 2.5. In order to simplify
the notation, in this section we will specialize system (4) by considering
ε = 1, and by eliminating all the subscripts in the unknowns. Obviously,
the results are valid for all parameters ε > 0.
10 F. GOLSE AND F. SALVARANI

3.1. The maximum principle. As a first step, we prove that (4) preserves
the non-negativity of the solutions, which are uniformly bounded in time.
This property is proved in the following proposition:
Proposition 3.1. Let us suppose that there exists a solution (u, θ) for the
radiative transfer system (4) in (t, x, ω) ∈ (0, τ ) × Ω × S 2 , with bounded
initial conditions
uε (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )
θε (0, x) = θin (x) ∈ L∞ (Ω)
and bounded boundary data
uε (t, x, ω)|Γ− = ub (t, x, ω) ∈ L∞ ((0, τ ) × ∂Ω × S 2 ).
Then there exists a constant M , which depends only on the initial and
boundary data, such that 0 ≤ u, θ ≤ M a.e. (uniformly in x, ω and t).
Proof. We develop first the part of the proof which establishes the bound
from above. We define
n o
M = max kuin kL∞ (Ω×S 2 ) , kθin kL∞ (Ω) , kub kL∞ (Γ− ×S 2 ×(0,τ ]) .

We consider now the pair of constant functions (v = M, S = M ), solutions

of the system

∂v
 ∂t + ω · ∇x v + σ(θ)(v − S) = 0



(6)

 ∂

 E(S) = σ(θ)(v̄ − S),
∂t
where θ satisfies the radiative transfer system (4), with initial and boundary
conditions
v(x, ω, t = 0) = M v(x, ω, t)|Γ− = M.
S(x, t = 0) = M.
We subtract the first equation of system (6) from the first equation of system
(4) and subtract the second equation of system (6) from the second equation
of system (4).
We then multiply the two obtained equations by sign+ (u−v) and sign+ (θ−
S) respectively, where sign+ (y) = 1 when y ≥ 0 and sign+ (y) = 0 when
y < 0.
Finally we add the two equations and integrate with respect to x and ω
in the domain Ω × S 2 . We note that the term
Z Z
ω · ∇x (u − v) dωdx = (u − v)+ ω · nx dsdω,
Ω×S 2 Γ×S 2
+
where (u − v)+ = (u − v) sign (u − v), is always non-negative: indeed, in the
region Γ0 ∪ Γ+ , the factor (ω · nx ) is non-negative, as well as (u − v)+ ; on the
other hand, in the region Γ− , the term (ω · nx ) is negative, but there u < v
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 11

because of the boundary conditions imposed to v, and therefore (u−v)+ = 0

and the integral vanishes.
Since E 0 is non-negative, we can moreover use the identity
sign+ (θ − S) = sign+ (E(θ) − E(S)).
We hence obtain
Z Z 
d + +
(u − v) dωdx + (E(θ) − E(S)) dωdx ≤ 0,
dt Ω×S 2 Ω×S 2
since Z
σ(θ)(θ − S)( sign+ (u − v) − sign+ (θ − S)) dωdx−
Ω×S 2
Z
σ(θ)(u − v)( sign+ (u − v) − sign+ (θ − S)) dωdx ≤ 0.
Ω×S 2
But in t = 0 we have that
Z Z
+
(u(0) − v(0)) dωdx + (E(θ(0)) − E(S(0)))+ dωdx = 0,
Ω×S 2 Ω×S 2
and therefore the bound from above is proved.
In order to prove that System (4) preserves the sign of the solution,
we apply a similar strategy: we consider the function sign− (y), defined
as sign− (y) = 0 when y ≥ 0 and sign− (y) = −1 when y < 0; then
we multiply the first equation of (4) by sign− (u) and the second one by
sign− (θ). We integrate the equations in Ω × S 2 with respect to the space
variable x and ω and deduce (notice that, by hypothesis on E, the identity
sign− (E(θ)) = sign− (θ) is always satisfied):
Z Z Z
d
u− dxdω+ u− (ω·nx ) dsdω+ σ(θ)(u−θ) sign− (u) dxdω = 0
dt Ω×S 2 Γ×S 2 Ω×S 2

and
Z Z
d −
(E(θ)) dxdω = σ(θ)(u − θ) sign− (θ) dxdω = 0,
dt Ω×S 2 Ω×S 2

where we have defined the function y − as y − = 0 when y ≥ 0 and y − = −y

when y < 0.
By summing the two equation, with the same strategy used to prove the
maximum principle, we deduce that, separately,
Z
u− (ω · nx ) dsdω ≥ 0
Γ×S 2
and Z
σ(θ)(u − θ)( sign− (u) − sign− (θ)) dxdω ≥ 0.
Ω×S 2
We can hence conclude that
Z
d
[u− + (E(θ))− ] dxdω ≤ 0.
dt Ω×S 2
12 F. GOLSE AND F. SALVARANI

Since the initial and boundary conditions are non-negative, E is monotone

and E(0) = 0, this implies that u ≥ 0 and E(θ) ≥ 0 a.e. in Ω for all
t ∈ [0, τ ]. 
3.2. The truncated problem. In the proof on the existence of a weak
solution of System (4), we use a trick introduced by Mercier in [14], by
working with the solutions of a truncated approximation of (4).
Then, thanks to the maximum principle of the truncated problem, it
will be possible to construct sequences of solutions of such approximated
problems which will be relatively compact in L∞ -weak∗ . Once passed to the
limit, the existence proof will be achieved if one proves that the limit of the
solutions of the approximated problems solves System (4).
We first consider the following truncated approximation of System (4):

∂un

 ∂t + ω · ∇x un + σn (θn )(Un − θn ) = 0



(7) Z
 1
 Et (θn ) = σn (θn )(Ūn − θn ), Ūn = Un (t, x, ω)dω,


4π S2
where
σn (y) = min{σ(y), n}, Un (t, x, ω) = max{0 , min{un , M }}, n ∈ N.
System (7) has the form of a Lipschitzian perturbation of the free trans-
port equations. Therefore, thanks to Theorems 6.1.2 and 6.1.6 in [17],
System (7) admits a strong solution (un , θn ) ∈ C 0 ([0, τ ]; L2 (Ω × S 2 )) ×
C 0 ([0, τ ]; L2 (Ω)).
By applying the same strategy of Lemma 3.1, we can prove that, if we
suppose that there exists a solution (un , θn ) for the truncated approximation
(7) in (t, x, ω) ∈ (0, τ )×Ω×S 2 , with the same initial and boundary conditions
of Theorem 2.5, then there exists a constant M , which depends only on the
initial and boundary data, such that 0 ≤ un , θn ≤ M a.e. in Ω, uniformly
in x, ω and t.
Thanks to the maximum principle for System (7), the solution of such set
of equations coincides therefore with the solution of the following system:
∂un

 ∂t + ω · ∇x un + σn (θn )(un − θn ) = 0



(8) Z
 1
 Et (θn ) = σn (θn )(ūn − θn ), ūn = un (t, x, ω)dω,


4π S 2
where
σn (y) = min{σ(y), n}, n ∈ N.
We hence deduce the following Lemma:
Lemma 3.2. Consider System (8) for (t, x, ω) ∈ (0, τ ) × Ω × S 2 , with initial
conditions
un (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 13

θn (0, x) = θin (x) ∈ L∞ (Ω)

and boundary data
un (t, x, ω)|Γ− = ub (t, x, ω) ∈ L∞ ((0, τ ) × ∂Ω × S 2 ).
Then, there exists a strong solution (un (t, x, ω), θn (t, x)) of System (8) which
belongs to the space C 0 ([0, τ ]; L2 (Ω×S 2 ))×C 0 ([0, τ ]; L2 (Ω)). Moreover, there
exists a constant M > 0 such that
kun kL∞ ((0,τ )×Ω×S 2 ) , kθn kL∞ ((0,τ )×Ω×S 2 ) ≤ M
uniformly with respect to n.
Thanks to Lemma 3.2, the sequences {un } and {θn } are relatively com-
pacts in L∞ -weak∗ . Nevertheless, this property is not enough to pass to
the limit because of the nonlinearities of the problem. We need strong com-
pactness for the sequences {ūn } (obtained through an averaging lemma) and
{θn }.
We prove therefore the following result:
Lemma 3.3. Let {un } and {θn } be sequences of solutions of the truncated
approximation (8). Then ūn is relatively compact in Lploc (t, x) for all p ∈
[1, +∞).
Proof. Since θn is uniformly bounded in L∞ t,x and the internal energy E is of
class C 0 ([0, +∞)), the second equation of System (8) shows that we deduce
that Z Z
E(θn )(τ, x) dx − E(θn )(0, x) dx+
Ω Ω
Z τZ Z τZ
σn (θn ) θn dtdx = σn (θn ) ūn dtdx.
0 Ω 0 Ω
Therefore, there exists a constant C > 0 such that
Z
σn (θn )un dtdxdω ≤ C.
(0,τ )×Ω×S 2

By the first equation of System (8), we hence deduce that (∂t + ω · ∇x )un is
bounded in L1loc (t, x, ω).
Since un bounded in L∞ (t, x, ω), we then use simply the averaging lemma
proved in [9] to conclude. 
As a next step we prove that the unknown θn of System (8) depends only
on ūn in a very special way:
Lemma 3.4. Consider System (8). Then there exists a . . . functional F :
Lploc ((0, τ ) × Ω) → Lploc ((0, τ ) × Ω), p ∈ [1, +∞) such that θn = F(ūn ).
Proof. Let us consider the second equation of (8). Since E is of class C 1 , we
can introduce the function Hn = Hn (θn ) defined in such a way that
E 0 (θn )
(9) Hn0 (θn ) = .
σn (θn )
14 F. GOLSE AND F. SALVARANI

This allows us to rewrite the second equation of (8) in the form

∂Hn
(10) (θn ) + θn = ūn .
∂t
Thanks to the hypotheses on E and σn , we can deduce that Hn is one-to-
one and of class C 1 . Indeed, by definition of Hn0 , we have that Hn0 is always
non-negative. The only point which can give troubles is θn = 0. But
E 0 (θn )
lim Hn0 (θn ) = lim ∼ θnα+β ,
θ→0 θn →0 σn (θn )

because E and σ are admissible internal energy and opacity respectively.

Since β ∈ (0, 1/4] and α ∈ [−3/4, −β], we can deduce that
lim (Hn−1 )0 (y)
y→0

exists and is always finite.

The quantity Hn being defined by integration of Hn0 , we can choose the
arbitrary constant; the best choice is to put Hn (0) = 0:
Z θn 0
E (ζ)
Hn (θn ) = dζ.
0 σn (ζ)
Thanks to the maximum principle, the behavior of Hn when θn → +∞
plays no role. We have then proved that Hn ∈ C 1 ([0, Hn (M )]), Hn−1 ∈
C 1 ([0, M ]) and Hn is invertible. By setting
φn = Hn (θn ),
Equation (10) becomes
∂φn
(11) + Hn−1 (φn ) = ūn .
∂t
We build the difference between the previous equation and a shifted ver-
sion of the same equation. We can deduce that
1d
[φn (t, x + h) − φn (t, x)]2 +
2 dt
[Hn−1 (φn (t, x + h)) − Hn−1 (φn (t, x))][φn (t, x + h) − φn (t, x)] =
[ūn (t, x + h) − ūn (t, x)][φn (t, x + h) − φn (t, x)].
We integrate respect to x in Ω and obtain, thanks to the monotonicity of
Hn (which is a consequence of the monotonicity of E):
1 1
kφn (t, x + h) − φn (t, x)k2L2x ≤ kφn (0, x + h) − φn (0, x)k2L2x +
2 2
Z tZ
[ūn (s, x + h) − ūn (s, x)][φn (s, x + h) − φn (s, x)] dsdx
0 Ω
1
≤ kφn (0, x + h) − φn (0, x)k2L2x +
2
Z t
kūn (s, · + h) − ūn (s, ·)kL2x kφn (s, · + h) − φn (·, x)kL2x ds.
0
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 15

Thanks to Gronwall’s Lemma, in the limit |h| → 0 we obtain that

kφn (t, · + h) − φn (t, ·)k2L2x → 0,
Moreover, since ūn is relatively compact in Lploc , p ∈ [1, +∞),
Z t
kūn (s, · + h) − ūn (s, ·)kL2x ds → 0
0
when |h| → 0. Finally, since ūn is bounded in L∞ −1
t,x and Hn (φn ) is also
bounded in L∞ ∞
t,x , we are guaranteed that φ̇n is bounded in Lt,x . Hence,
we can deduce that the family of functions {φn } is relatively compact in
Lploc ((0, τ ) × Ω), and therefore there exists a subsequence, which converges
to a limit point φ in Lploc ((0, τ ) × Ω) and a.e.. 
3.3. Proof of theorem 2.5. Thanks to Lemma 3.2 and the Banach-Alaouglu
theorem, un and θn are weakly* compact. This means that, by subsequences,
there exists two limits u and θ such that
un * u ∈ L∞ ([0, τ ] × Ω × S 2 ) θn * θ ∈ L∞ ([0, τ ] × Ω)
weakly-*.
The next step consists in proving that, if {un } and {θn } are two sequences
of solutions of System (8) with initial and boundary data as in Theorem
2.5, then the two weak* limits u and θ solve the non-truncated version of
Equations (4) with initial and boundary conditions
u(x, ω, t = 0) = uin (x, ω) u(x, ω, t)|Γ− = ub (t, x, ω)
and
θ(x, t = 0) = θin (x).
The previous equations are highly nonlinear; hence the passage to the
limit is not straightforward. We analyze therefore each term of the two
equations.
i) We have easily that
σn (θn ) θn = O(1)
inL∞ ([0, τ ]
× Ω): indeed, in any interval of type [a, +∞) with a > 0, σ is
uniformly bounded and, when y → 0, we have that
σn (y) y ∼ y 1−β
with β ∈ [0, 1/4).
ii) We look now at the term
σn (θn ) un .
Thanks to the second equation of system (8), we deduce, by integrating with
respect to (t, x, ω) in [0, τ ] × Ω × S 2 ) that
Z τZ Z Z τZ
1
σn (θn ) un dtdxdω = σn (θn ) ūn dtdx =
4π 0 Ω S 2 0 Ω
16 F. GOLSE AND F. SALVARANI
Z τ Z
[Et (θn ) + σn (θn ) θn ] dtdx.
0 Ω
We already know that σn (θn ) θn belongs to L∞ ([0, τ ] × Ω). We now consider
the other term of the last integral: we have that
Z τZ Z Z
Et (θn ) dtdx = E(θn )(x, τ ) dx − E(θin )(x) dx.
0 Ω Ω Ω
Thanks to the hypotheses on E, we have proved that also σn (θn ) un belongs
to L∞ ([0, τ ] × Ω) since Ω is bounded.
In order to pass to the limit, we use the compactness properties of ūn and
θn . By Lemma 3.3, we have that ūn is relatively compact in Lploc ((0, τ ) × Ω)
for all p ∈ [1, +∞) and, by Lemma 3.4, that θn is relatively compact in
Lploc ((0, τ ) × Ω) for all p ∈ [1, +∞).
We know that
(1) un * u in L∞
t,x,ω -weak
∗

p
(2) θn * θ in L∞ ∗
t,x,ω -weak and, by Lemma 3.4, that θn → θ in Lloc and
a.e..
Consequently, σn (θn ) θn * σ(θ) θ in in Lploc and a.e..
We can now apply the dominated convergence theorem to control the
nonlinear terms. We recall moreover that the only term which depends
on the angular variable ω is un . Therefore we can deduce that, modulo
extraction of a subsequence, there exists a measure µ such that
σn (θn ) un *∗ µ.
We only need to identify µ with the limit product σ(θ) u.
For any m < n, we have that
σm (θn ) un ≤ σn (θn ) un ,
and therefore, in the limit m, n → +∞
σ(θ) u ≤ µ.
On the other hand, the second equation of (8), says us that
Z Z Z τZ
in
E(θn )(τ, x) dx − E(θ )(x) dx + σn (θn ) θn dtdx =
Ω Ω 0 Ω
Z τ Z Z
1
σn (θn ) un dtdxdω.
4π 0 Ω S2
The l.h.s converges to
Z Z Z τ Z
in
E(θ)(τ, x) dx − E(θ )(x) dx + σ(θ) θ dtdx,
Ω Ω 0 Ω
whereas the r.h.s converges to
Z τ Z
µ dtdx.
0 Ω
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 17

But in the limit, by strong convergence

σ(θ) ū = Et (θ) + σ(θ) θ,
and therefore Z τ Z Z τ Z
σ(θ)θ u dtdx = µ dtdx.
0 Ω 0 Ω
We have hence proved that the term σn (θn ) un converges weakly∗ to
σ(θ) u.
Hence, the limit (u, θ) of the sequence (un , θn ) solves System (4).
iii) The proof will be complete once we show that the initial and boundary
conditions for the limit problem are satisfied (in the sense of traces).
We know that the sequence un is bounded in Lp ((0, τ ) × Ω × S 2 ) for any
p ∈ [1, +∞). If we can find an exponent q > 1 such that (∂t + ω · ∇x )un is
bounded in Lq ((0, τ ) × Ω × S 2 ), we can deduce that, thanks to Cessenat’s
results [4, 5], un has a trace in Lq -sense and that un * u weakly* in Lq ({t =
0} × Ω × S 2 ∪ ∂Ω × S 2 × (0, τ ]).
Let us therefore consider the first equation of System (8). Thanks to the
behavior in zero of σn , we deduce that σn (θn ) θn is bounded in L∞ ((0, τ )×Ω).
Therefore we deduce that (∂t + ω · ∇x )un is bounded in Lq ((0, τ ) × Ω × S 2 )
if we can prove that σn (θn ) un ∈ Lq ((0, τ ) × Ω × S 2 ); we have
Z Z
q q q−1
σn (θn ) un dtdxdω ≤ kun kL∞ σnq (θn ) ūn dtdx.
(0,τ )×Ω×S 2 (0,τ )×Ω

By the second equation of System (8), we have

Z Z
∂θn
σnq−1 (θn )E 0 (θn ) dtdx + σnq (θn ) θ dtdx =
(0,τ )×Ω ∂t (0,τ )×Ω
Z
σnq (θn ) ūn dtdx.
(0,τ )×Ω
But the first term in the l.h.s. of the previous equation is bounded thanks to
fundamental theorem of integral calculus and the second term in the l.h.s.
is bounded thanks to the behavior in zero of σn , provided that q ∈ (1, 4].
Hence, we deduce that
Z
σnq (θn ) uqn dtdxdω
(0,τ )×Ω×S 2

is bounded and therefore un * u weakly* in Lq ({t = 0} × Ω × S 2 ∪ ∂Ω ×

S 2 × (0, τ ]) for any q ∈ (1, 4].
In order to prove, finally, that the limit θ satisfies the initial condition, we
consider the second equation of System (4). Since E is of class C 1 and strictly
increasing and the second member of the equation is bounded in Lploc and
a.e., we deduce that θ(x, t) → θin (x) strongly in Lp (Ω) for all 1 ≤ p < ∞.
Hence the proof of Theorem 2.5 is complete.
18 F. GOLSE AND F. SALVARANI

4. Relative entropy estimates

This section is devoted to give the tools which will permit to handle the
limiting behavior of system (4).
From now on, we set
(12) ν = kub kW 1,∞ (∂Ω×[0,τ ]) and m = min σ(y).
0≤y≤M

The main mathematical tool we will need is a functional, which is nothing

but the relative entropy with respect to a gauge profile.
There are many ways to define this gauge profile. Its most important
property concerns the behavior on the boundary. For example, the gauge
function f = f (t, x) could be the solution of the Laplace equation with
boundary data given by ub (t, x):

 ∆x f (t, x) = 0 x∈Ω
(13)
f (t, x)|∂Ω = ub (t, x) ∈ L∞ ((0, τ ] × ∂Ω).


Classical theory on the Laplace equation ensures that f exists, is unique and
satisfies the strong maximum principle. In particular, 0 ≤ f ≤ M = kub k∞
in Ω̄ and f ∈ W 2,p (Ω) for any p ∈ [1, +∞). Obviously, here the time t acts
only as a parameter.

4.1. An elementary inequality. The natural entropy generating function

to control the L2 -norm of radiation flux should be y 2 /2.
However, one needs to control boundary terms, and the natural choice
stated before does not seems the best one. To do so, it is more convenient
to use the following entropy generating function
1 1
Φ(y) = y 2 + (ν + 1)2 ,
2 2
where ν has been introduced in (12). With this definition Φ(y) ≥ (ν + 1)y =
Φ0 (ν)y + y, i.e.
(14) Φ(y) − Φ0 (ν)y ≥ y
and this last inequality is helpful in handling boundary terms.

4.2. Entropy inequalities. In this subsection we show in which way the

control on radiation flux come from a entropy production control.
We multiply the first equation of system (4) by Φ0 (uε ) and integrate with
respect to x and ω in Ω × S 2 . We obtain:
Z Z
2 d
ε Φ(uε ) dω dx + ε ω · nx Φ(uε ) dω ds(x)+
dt Ω×S 2 ∂Ω×S 2
Z
(15) σ(θε )(uε − θε )Φ0 (uε ) dω dx = 0.
Ω×S 2
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 19

We write now Equation (15) in the form

Z Z
d 1
Φ(uε ) dω dx + ω · nx [Φ(uε ) − Φ0 (f )uε ] dω ds(x) =
dt Ω×S 2 ε ∂Ω×S 2
Z Z
1 0 1
(16) − 2 σ(θε )(uε − θε )Φ (uε ) dω dx − ∇ · (ωΦ0 (f )uε ) dω dx.
ε Ω×S 2 ε Ω×S 2
Let us study now the various terms of the previous equation.
(i) We have that
Z Z Z
ω·nx [Φ(uε )−Φ0 (f )uε ] dω ds = ω·nx [Φ(uε )−Φ0 (f )uε ] dω ds
∂Ω×S 2 ∂Ω ω·nx >0
Z Z
+ ω · nx [Φ(uε ) − Φ0 (f )uε ] dω ds ≥
∂Ω ω·nx <0
Z Z Z Z
0
ω·nx [Φ(f )−Φ (f )f ] dω ds+ ω·nx [Φ(uε )−Φ0 (f )uε ] dω ds
∂Ω ω·nx >0 ∂Ω ω·nx <0
Z Z 
0
= (Φ(f ) − Φ (f )f ) ω · nx dω ds = 0.
∂Ω S2
Here we have used the convexity of Φ, inequality (14) and the fact that
[Φ(f ) − Φ0 (f )f ] does not depend on ω.
This term is therefore non-negative.
(ii) We perform now some manipulations of the last term in Equation (16):
Z Z
1 0 1
∇·(ωΦ (f )uε ) dω dx = ω·[uε Φ00 (f )∇x f +Φ0 (f )∇x uε ] dω dx =
ε Ω×S 2 ε Ω×S 2
Z Z
1 00
[uε Φ (f )ω · ∇x f ] dω dx + Φ0 (f )(∇x · jε ) dx.
ε Ω×S 2 Ω
We notice that
Z Z  
∂ ∂
Φ0 (f )(∇x · jε ) dx = Φ0 (f ) E(θε ) − ρε dx =
Ω Ω ∂t ∂t
Z Z
 0
Φ00 (f )E(θε )ft dx−

Φ (f )E(θε ) t dx −
Ω Z ZΩ
 0
Φ00 (f )ρε ft dx.

Φ (f )ρε t dx +
Ω Ω
Also the other term can be controlled: we have that
Z Z
1 00
[uε Φ (f )ω · ∇x f ] dω dx = [Φ00 (f )jε · ∇x f ] dω dx ≤
ε Ω×S 2 Ω×S 2
Z Z
1 0 2
|∇x Φ (f )| dx + γ |jε |2 dx.
2γ Ω Ω

(iii) Finally, we deduce that

Z
1
σ(θε )(uε − θε )Φ0 (uε ) dω dx =
ε2 Ω×S 2
20 F. GOLSE AND F. SALVARANI
Z Z
1 0 0
σ(θ ε )(u ε − θ ε )[Φ (uε ) − Φ (θ ε )] dω dx + Et (θε )Φ0 (θε ) dx.
ε2 Ω×S 2 Ω×S 2

If we remember the form of Φ, we see immediately that the first member

in the right-hand side of the previous equation controls the L2 -norm of the
flux: Z
1
σ(θε )(uε − θε )2 dω dx ≥
ε2 Ω×S 2
Z Z Z 
1 2 (i) 2
σ(θε ) (uε − θε ) dω (ω ) dω dx ≥
ε2 Ω S2 S2
Z Z 2 Z Z 2
1 (i) 1 (i)
σ(θ ε ) (uε − θ ε )ω dω dx = σ(θ ε ) uε ω dω dx =
ε2 Ω S2 ε2 Ω S 2
Z
σ(θε )(jε(i) )2 dx.
Ω
Hence
Z Z Z
1 2 1 2 m
σ(θε )(uε − θε ) dω dx ≥ σ(θε )|jε | dx ≥ |jε |2 dx.
ε2 Ω×S 2 3 Ω 3 Ω

We therefore deduce that Equation (16) implies

Z   Z
d 1 2  m
uε + f (E(θε ) − ρε ) dxdω ≤ γ − |jε |2 dx−
dt Ω×S 2 2 3 Ω
Z Z Z
1
Et (θε )θε dx − ft [ρε − E(θε )] dx + |∇x f |2 dx.
Ω Ω 2γ Ω
If we choose γ < m/3 in the previous inequality and denote by
Z y
Q(y) = E(s)s ds,
0
we deduce that
Z   Z
d 1 2 m
uε + f (E(θε ) − ρε ) + Q(θε ) dxdω + −γ |jε |2 dx ≤ C,
dt Ω×S 2 2 3 Ω
where C is a positive constant which can be computed explicitly. We have
then deduced the following result:
Proposition 4.1. Let (uε , θε ) be a solution of the scaled system (4), in
(0, τ ) × Ω × S 2 , with initial conditions
uε (0, x, ω) = uin (x, ω) ∈ L∞ (Ω × S 2 )
θε (0, x) = θin (x) ∈ L∞ (Ω)
and boundary data
uε (t, x, ω)|Γ− = ub (t, x, ω) ∈ W 1,∞ ((0, τ ) × ∂Ω × S 2 ).
Then there exists a positive constant J = J(τ, ub , uin , θin , m) such that the
current jε satisfies Z τZ
|jε (t, x)|2 dt dx ≤ J
0 Ω
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 21

and
1 τ
Z Z
σ(θε )(uε − θε )2 dt dx dω ≤ J
ε2 0 Ω×S 2
uniformly for each ε > 0.
Thanks to Proposition 4.1, we can deduce the following corollary:
Corollary 4.2. Let (uε , θε ) as in Proposition 4.1. Then, if one of the se-
quences u , or θ , or ū admits a limit, then all the three sequences go to the
same limit in strong L2 -sense.
Proof. Since σ is bounded below by a strictly positive constant, we obtain
from Proposition 4.1 that
1
kuε − θε kL2t,x,ω = O(1)
ε
and therefore
lim ku − θ kL2t,x,ω = 0.
→0
Since
kūε − θε kL2t,x ≤ kuε − θε kL2t,x,ω
and
kuε − ūε kL2t,x,ω ≤ kuε − θε kL2t,x,ω + kūε − θε kL2t,x ,
we obtain that also
lim ku − θ kL2t,x,ω = lim kθ − ū kL2t,x = 0.
→0 →0


5. Convergence to the Rosseland equation

This section is devoted to conclude the proof on the convergence of the
diffusive limit stated in Theorem 2.6. Since we will use various relatively
compact sequences, when we will indicate that a sequence converges to a
limit, we will mean that there exists a subsequences which converges to the
limit.
Let (uin , θin ) and ub be initial and boundary data which satisfy the hy-
potheses of Theorem 2.6; then, for each  > 0, let (u , θ ) be the solution to
the scaled system (4).
It follows from Propositions 3.1 and 4.1 that, for each τ > 0, one has
kρ kL∞ ((0,τ )×Ω) ≤ K
and
kj kL2 ([0,τ ]×Ω) ≤ J 1/2
for each  > 0.
By the Banach-Alaoglu theorem, for each τ > 0
(17) the family ρ is relatively compact in L∞ ((0, T ) × Ω) weak-*,
22 F. GOLSE AND F. SALVARANI

and therefore there exists ρ ∈ L∞ ((0, τ ) × Ω) such that ρ *∗ ρ, while

(18) the family j is relatively compact in L2 ([0, τ ] × Ω) weak,
which means that there exists j ∈ L2 ((0, τ ) × Ω) such that j * j.
Because of the nonlinearities, that are present both in the scaled system
(4) and in the limiting nonlinear diffusion equation, weak compactness re-
sults as above are not enough to pass to the limit as  → 0. Strong L2
compactness of the family ρ is required, which will be obtained by using
Minty’s trick [13] and compensated compactness.
Consider the vector fields
p = (ρ + E(θ ), j ) and q = (ρ , 0, 0, 0) .
By the arguments as above, both vector fields satisfy
p and q are bounded in (L2 ([0, τ ] × Ω))4 .
Moreover, if we integrate the first equation of system (4) with respect to ω
in S 2 , we obtain
∂ h ρε i
(19) + E(θε ) + ∇x · jε = 0,
∂t c
which means:
divt,x p = 0.
In order to prove that
−1
curlt,x q is relatively compact in Ht,x ,
we simply use the averaging lemma [9], which give us compactness in x,
the behavior of the time derivative being irrelevant, since the only non-zero
component of qε is the first one.
Let E∗ the weak-∗ limit of E(θε ). Since (ūε − θε ) tends strongly to zero in
L , we can deduce that also E(ρε ) converges weakly-∗ to E∗ . By compensated
2

compactness (the div-curl lemma in [16]) we find that, by subsequences,

(20) pn · qn = ρ2n + ρn E(θn ) * p · q = ρ2 + ρE∗
in the sense of Radon measures on (0, T ) × (0, 1) as n → 0.
The family
(21) ρ is relatively compact in L2 ([0, T ] × [0, 1]) strong:
indeed,
(ρε − ρ)2 + (ρε − ρ)(E(ρε ) − E(ρ)) ≥ 0
since E is monotone non-decreasing. Moreover, we deduce that, by (20),
(ρε − ρ)2 + (ρε − ρ)(E(ρε ) − E(ρ)) → 0
in D0 . A non-negative sequence which converges in D0 to zero converges also
strongly to zero. Since also the term
(ρε − ρ)(E(ρε ) − E(ρ))
is non-negative, we deduce finally that ρε → ρ (strongly) in L2 ((0, τ ) × Ω).
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 23

Because of the regularity on E (see Definition 2.4) and because (θ − ū ) →
0 strongly in L2 ((0, τ ) × Ω), then E(θε ) → E(ρ) (strongly) in L2 ((0, τ ) × Ω).
Let us now take into account the flux equation
(i) Z
∂jε 1
(22) ε2 + ω (i) ∇x · (ωuε ) dω + σ(θε ) jε(i) = 0.
∂t 4π S 2
Since σ is strictly positive, we can state that
 Z 
1/2 −1/2 2 ∂jε 1
σ(θε ) jε = −σ(θε ) ε + ∇x · (ω ⊗ ω)uε dω .
∂t 4π S2
Now, by Proposition (4.1), we have that
kσ(θε )1/2 jε k2L2 ((0,τ )×Ω) ≤ kσ(θε )1/2 (uε − θε )k2L2 ((0,τ )×Ω×S 2 ) +

kσ(θε )1/2 θε k2L2 ((0,τ )×Ω) ≤ ε2 J + Cτ,

where C is the maximum of the function σ(y)y 2 for y ∈ [0, M ].
Therefore, there exists η ∈ L2 ((0, τ ) × Ω) such that σ(θε )1/2 jε * η in
L ((0, τ ) × Ω). Since σ is strictly positive and θε → ρ strongly in L2 ((0, τ ) ×
2

Ω), we deduce that

jε * σ(θε )−1/2 η = j.
On the other hand, if we multiply the first equation of the system (4) by uε ,
we deduce:
∂u2ε
   
1 2 1/2 1 1/2
(23) ε + ω · ∇x uε = [uε σ(θε ) ] σ(θε ) (uε − θε ) .
2 ∂t 
Thanks to Proposition 4.1, there exists a function ζ ∈ L2 ((0, τ ) × Ω) such
that
1
σ(θε )1/2 (uε − θε ) * ζ

in L2 ((0, τ ) × Ω) and therefore we can pass to the limit and obtain that
1
ζ=p ω · ∇u,
σ(ρ)
since σ is bounded below by a positive constant by hypothesis.
But it is easy to see that
Z
1
η=− ω ζ, ∂ω = − p ∇ρ,
S 2 3 σ(ρ)
and finally that  
∇x ρ
(ρ + E(ρ))t − ∇x · = 0,
3σ(ρ)
which is equation (5).

In order to verify that the initial and boundary conditions are the correct
limits of the conditions imposed on the radiative transfer system, we will
use a regularity argument.
24 F. GOLSE AND F. SALVARANI

Let us now consider the continuity equation (19): since the flux jε is
bounded in L2 ((0, τ ) × Ω), we deduce that
(ρε + E(θε ))t is bounded in L2 (0, T ; H −1 (Ω)) ;
which, joint to the bound on ρ and E, implies that
(24) ρ + E(θ ) is relatively compact in C([0, τ ]; H −1 (Ω))
by Arzela-Ascoli’s theorem and therefore the target equation recovers the
correct initial data.
Because of the initial condition of the radiative transfer system (4), one
has
Z
−1 1
(25) ρ ∈ C([0, τ ]; H (Ω)) and ρ t=0 = uin (x, ω) dω.
4π S 2
Let us now care about the boundary conditions: we consider equation
(23) and perform the limit as  → 0. We note immediately, thanks to the
identity
σ(θε )(uε − θε )uε = σ(θε )(uε − θε )2 + σ(θε )uε θε − σ(θε )θε2 ,
that ω·∇x uε is bounded in L2 ((0, τ )×Ω×S 2 ), we can deduce that, thanks to
Cessenat’s results [4, 5], uε has a trace in L2 -sense and that uε * u weakly
in L2 ((0, τ ] × ∂Ω × S 2 ), and therefore the trace of the limit exists. Thus the
proof of the nonlinear diffusion limit is competed.

Acknowledgment: This paper has been written when the second author
was visiting the Centre de Mathématiques Laurent Schwartz of the Ecole
Polytechnique, France. FS is very grateful to the Ecole Polytechnique for
the invitation and the facilities provided to him.
Financial support of the Italian National Group for Mathematical Physics
(GNFM) is acknowledged.

References
[1] Alexander, D. R.; Ferguson, J. W. Low temperature Rosseland opacities. Astro-
physical Journal 437, 2 (1994) 879-891
[2] Bardos, C.; Golse, F.; Perthame, B. The Rosseland approximation for the radia-
tive transfer equations. Comm. Pure Appl. Math. 40, 6 (1987) 691–721.
[3] Bardos, C.; Golse, F.; Perthame, B.; Sentis, R. The nonaccretive radiative
transfer equations: existence of solutions and Rosseland approximation. J. Funct. Anal.
77, 2 (1988) 434–460.
[4] Cessenat, M. Théorèmes de trace Lp pour des espaces de fonctions de la neutronique.
C. R. Acad. Sc. Paris 299, 16 (1984), 831–834.
[5] Cessenat, M. Théorèmes de trace pour des espaces de fonctions de la neutronique.
C. R. Acad. Sc. Paris 300, 3 (1985), 89–92.
[6] Dogbe, C. The radiative transfer equations: diffusion approximation under accretive-
ness and compactness assumptions. Comput. Math. Appl. 42 (2001), 783–791.
[7] Golse, F. The Milne problem for the radiative transfer equations (with frequency
dependence). Trans. Amer. Math. Soc. 303, 1 (1987) 125–143.
[8] Golse, F., Perthame, B. Generalized solutions of the radiative transfer equations
in a singular case. Comm. Math. Phys. 106 (1986), 211–239.
 ROSSELAND LIMIT FOR RADIATIVE TRANSFER 25

[9] Golse, F.; Saint-Raymond, L. The Navier-Stokes limit of the Boltzmann equation
for bounded collision kernels. Invent. Math. 155 (2004), 81-161
[10] Golse, F.; Salvarani, F. The nonlinear diffusion limit for generalized Carleman
models: the initial-boundary value problem Nonlinearity 20 (2007), 927-942
[11] Goudon, T.; Poupaud, F. Approximation by homogenization and diffusion of ki-
netic equations. Comm. Partial Differential Equations 26 (2001), 537–569.
[12] Larsen, E.; Pomraning, G. ; Badham, V. Asymptotic analysis of radiative transfer
problems. J. Quant. Spectros. Radiat. Transfer 29 (1983), 285–310
[13] Marcati, P.; Milani, A. J. The one-dimensional Darcy’s law as the limit of a
compressible Euler flow. J. Diff. Eq. 84 (1990) 129–147.
[14] Mercier, B. Application of accretive operators theory to the radiative transfer equa-
tions. SIAM J. Math. Anal. 18 (1987) 393–408.
[15] Mihalas, D.; Weibel Mihalas B. Foundations of Radiation Hydrodynamics, Ox-
ford University Press, New York 1984.
[16] F. Murat. Compacité par compensation, Ann. Scuola Norm. Sup. Pisa Cl. Sci. 5
(1978) 489–507.
[17] Pazy, A. Semigroups of linear operators and applications to partial differential equa-
tions, Springer-Verlag, New York 1983
[18] Pomraning, G. The Equations of Radiation Hydrodynamics, Dover Publications,
Mineola NY, 2005.

F.G.: Centre de Mathématique Laurent Schwartz, Ecole Polytechnique,

91128 Palaiseau Cedex, France

F.S.: Dipartimento di Matematica, Università degli Studi di Pavia, Via Fer-

rata 1, 27100 Pavia, Italy