Transport Theory for Acoustic Waves with Re

ection and
Transmission at Interfaces
Guillaume Bal  Joseph B. Keller y George Papanicolaou z Leonid Ryzhik x
March 8, 1999

Abstract
Transport theoretic boundary conditions are derived for acoustic wave re ection and trans-
mission at a rough interface with small random uctuations. The Wigner distribution is used to
go from waves to energy transport in the high frequency limit, and the Born expansion is used to
calculate the eect of the random rough surface. The smoothing method is also used to remove
the grazing angle singularity due to the Born approximation. The results are presented in a form
that is convenient both for theoretical analysis and for numerical computations.

Contents
1 Introduction 2
1.1 The Radiative Transport Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Half Space Problems and Boundary Conditions . . . . . . . . . . . . . . . . . . . . . 3
2 Acoustic Wave Transport 5
3 Flat Boundary 7
3.1 Re ection and Transmission of Waves in Homogeneous Media . . . . . . . . . . . . . 7
3.2 High Frequency Limit: Re ection and Transmission of Energy . . . . . . . . . . . . . 9
4 Rough Boundaries 10
4.1 Surface Re ection Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2 Re ection and Transmission at a Rough Interface . . . . . . . . . . . . . . . . . . . . 13
4.3 Re ection and Transmission from Small Amplitude Rough Interfaces . . . . . . . . . 14
4.4 Derivation of the Dirichlet and Neumann Boundary Conditions . . . . . . . . . . . . 16
5 The Initial Boundary Value Problem for the Radiative Transport Equation 17
A Derivation of the Interface Conditions 18
B Grazing Angles and the Smoothing Method 21
B.1 Smoothing Method for Dirichlet and Neumann Boundary Conditions . . . . . . . . . 22
B.2 Extension to Interface Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Department of Mathematics, Stanford CA, 94305; bal@math.stanford.edu
y
Department of Mathematics, Stanford CA, 94305; keller@math.stanford.edu
z
Department of Mathematics, Stanford CA, 94305; papanico@math.stanford.edu
x
Department of Mathematics, University of Chicago, Chicago IL, 60637; ryzhik@math.uchicago.edu

1
1 Introduction
Wave propagation in weakly uctuating random media over distances long compared to the wave
length can be described by the energy transport equation. This is the radiative transport regime
described in the next section. Near boundaries and interfaces the waves undergo coherent or partially
coherent re ection and transmission, according as the interfaces are smooth or randomly rough.
Starting from the acoustic equations, we derive systematically, boundary conditions for the radiative
transport equation. This equation is asymptotically valid away from the boundaries and interfaces.
The asymptotic limit that leads to radiative transport corresponds to weak uctuations and high
frequency waves, with the correlation length of the inhomogeneities comparable to the wave length.
We use the Wigner distribution to analyze this limit, as we did in [23]. The main result of this work is
the derivation of the transport boundary conditions (1.3) and (1.4) for re ection and transmission at
a rough interface. The boundary conditions are intuitively clear. They combine coherent or specular
re ection and transmission, and incoherent or diuse re ection and transmission, in the form of a
linear input-output relation. If there is no dissipation, as we assume, energy ux is conserved as
(1.5) shows. We calculate the explicit form of the coherent and incoherent energy ux re ection and
transmission coecients and cross-sections when the uctuations of the interface are weak. Special
cases of these coecients have been obtained before in the Born approximation [5, 10, 29]. Our
results are summarized at the end of section 1.2.
The re ection and transmission coecients, computed by a Born expansion, are singular at
grazing angles. To eliminate the singularity, we recalculate them in Appendix B by using the
smoothing method, as was done in [31].
1.1 The Radiative Transport Equation
Radiative transport is a theory that was introduced phenomenologically to describe the propagation
of light intensity through the Earth's atmosphere. It has been applied successfully to many other
problems of wave propagation in complex media. In its simplest form, a(t; x; k) denotes the angularly
resolved energy density dened for wave vector k, position x and time t. Because of interaction with
the inhomogeneous medium through which it propagates, a wave with wave vector k may be scattered
into any other direction k^ 0 , where k^ = jkkj . Energy balance leads to the transport equation
@a(t; x; k) + r !(x; k)
r a(t; x; k) , r !(x; k)
r a(t; x; k) (1.1)
k x x k
@t Z
= n (x; k; k0 )a(t; x; k0 )dk0 , (x; k)a(t; x; k):
R
Here n is the dimension of space (n = 2 or 3), !(x; k) is the local frequency at position x of a wave
with wave vector k, the dierential scattering cross-section (x; k; k0 ) is the rate at which energy
with wave vector k0 is converted to wave energy with wave vector k at position x, and
Z
(x; k) = (x; k0 ; k)dk0 (1.2)
is the total scattering cross-section. The function (x; k; k0 ) is nonnegative and usually symmetric
in k and k0 . The left side of (1.1) is the total time derivative of a(t; x; k) at a point moving along a
trajectory in phase space (x; k) and the right side describes the eects of scattering.
The transport equation (1.1) is conservative when (1.2) holds because then the total energy is
independent of time:
ZZ
a(t; x; k)dxdk = const:

2
Absorption may be accounted for easily by letting the total scattering cross-section be the sum of
two terms
(x; k) = sc(x; k) + ab (x; k)
where sc(x; k) is the total scattering cross-section given by the right side of (1.2) and ab (x; k) is
the absorption cross section.
Derivations of (1.1) from equations for wave propagation by many authors are nicely presented
in a recent review [22]. (See [23] for references). In our work [23], radiative transport theory for
scalar and vector waves, including mode conversion and polarization, was derived in the following
regime:
 Distances of propagation L are much larger than the wavelength ,
 The medium parameters vary on a scale comparable to the wavelength,
 The mismatch between the inhomogeneities and the background medium is small,
 Absorption is small.
This regime arises in many physically important situations such as seismic wave propagation, where
teleseismic events can be modeled by radiative transport equations [1, 13].
1.2 Half Space Problems and Boundary Conditions
The radiative transport equation (1.1) has been derived in an unbounded domain. Boundary condi-
tions at a at boundary or interface were derived for time harmonic acoustic waves for the monochro-
matic transport equation without scattering. They can be used with (1.1) in a domain with curved
boundaries and interfaces varying on a scale large compared to the wave length because of the
separation of scales and weak scattering.
The problem of wave scattering from rough boundaries has been studied extensively [5, 10, 29].
We show that scattering from a rough interface can be incorporated in radiative transport theory
under the following conditions:
 Wavelength small compared to the curvature of the mean interface
 Roughness height smaller than or comparable to the wavelength
 Correlation length of the surface uctuations comparable to the wavelength.
Then the boundary conditions for (1.1) are as follows:
Let a1in (x0 ; k0 ) and a2in (x0 ; k0 ) be the energy densities of the waves coming to the interface from
above and below, respectively, with x0 being a horizontal coordinate on the rough surface, and k0
the horizontal wave vector. Then the outgoing energy densities are given by
v1 k^n1 a1out (x0 ; k0 ) = vZ1 k^n1 jR11 (k0 )j2 a1in(x0 ; k0 ) + v2 k^n2 jT12 (k0 )j2 a2in (x0 ; k0 ) (1.3)
 
+ dp0 v1 k^n1 (p0 )11 (x0 ; k0 ; p0 )a1in (x0 ; p0 ) + v2 k^n2 (p0 )12 (x0 ; k0 ; p0 )a2in (x0 ; p0 )
and
v2 k^n2 a2out (x0 ; k0 ) = vZ2 k^n2 jR22 (k0 )j2 a2in (x0 ; k0 ) + v1 k^n1 jT21 (k0 )j2 a1in(x0 ; k0 ) (1.4)
 
+ dp0 v2 k^n2 (p0 )22 (x0 ; k0 ; p0 )a2in (x0 ; p0 ) + v1 k^n1 (p0 )21 (x0 ; k0 ; p0 )a1in (x0 ; p0 ) :

3
 a 1in
a 1out

a 2out

Figure 1.1: Incident, re ected, transmitted, and diusely scattered waves at a at interface.

Rough surface, re ection Rough surface,

Flat surface
coecient correction scattering cross-section
Dirichlet 1 (4.27) entry 2; 2 (4.28) entry 2; 2
Neumann 1 (4.27) entry 1; 1 (4.28) entry 1; 1
Interface (3.11) (4.23) (4.22)
Table 1.1: References to the explicit formulas for the re ection and transmission coecients and
scattering cross-sections

The rst and second terms on the right in (1.3) and (1.4) represent coherent re ection and trans-
mission respectively, and the integrals represent diuse or incoherent scattering. The coecients in
these equations satisfy the energy ux conservation law
Z  
1 = jR11 (k0 )j2 + jT21 (k0 )j2 + dp0 11 (p0 ; k0 ) + 21 (p0 ; k0 ) ; (1.5)
with a similar relation holding for R22 , T12 , 12 and 22 . This is the analog of (1.2) for surface
scattering. Here vj are the acoustic wave speeds in medium j , and k^nj is the normal component of
the normalized wave vector k^ j = kj =jkj j where kj has tangential component k0 ; j = 1; 2.
We derive explicit formulas for the surface dierential cross sections ij and the re ection and
transmission coecients. For simplicity, we consider a homogeneous background medium. A gener-
alization to an inhomogeneous medium is presented in [2]. The results remain essentially unchanged
but their derivation is substantially more involved. The equation numbers for the explicit formulas
are given in Table 1.1.
Interface conditions (1.3) and (1.4) are derived when the adjacent media are homogeneous. We
expect that they remain valid when the media are random. This is because waves are not trapped
near the boundary when the uctuations of the interface are small. A mathematical justication of
4
this assumption is not available at present.

2 Acoustic Wave Transport

We recall how transport equations for the phase space energy densities are constructed [23, 14, 24].
For simplicity we assume here that the space domain is IR3 (n = 3). The acoustic equations for the
velocity v and pressure p are
 @@tv + rp = 0 (2.1)
 @p
@t + r
v = 0:
This system may be written as a symmetric hyperbolic system (with summation over repeated
indices):
A(x) @@tu + Dj @x
@ u = 0;
j (2.2)
where u = (v; p), and x 2 IRn . The matrix A(x) = diag(; ; ; ) is symmetric and positive
denite and the matrices Dj are symmetric and independent of x and t. We consider high frequency
solutions of (2.2). Physically this means that the typical wave length  of the initial data is much
smaller than the overall propagation distance L, so that " = =L  1. The spatial energy density
for the solutions of (2.2) is given by
E (t; x) = v2 + p2 = 21 (A(x)u(t; x)
u(t; x)) = 12 Aij (x)ui (t; x)uj (t; x)
2 2
(2.3)
and the ux F (x) by
F i(t; x) = pv = 21 (Di u(t; x)
u(t; x)): (2.4)
We have the energy conservation law
@ E + r
F = 0: (2.5)
@t
and thus the total energy is conserved:
d Z E (t; x)dx = 0: (2.6)
dt
The high frequency limit " ! 0 of the energy density E (t; x) is described in terms of the Wigner
transform, which is dened by
 1 n Z
W"(t; x; k) = 2 eik
y u" (t; x , "y=2)u" (x + "y=2)dy; (2.7)
where u" (t; x) is the solution of (2.2). The Wigner transform W" is a 4  4 Hermitian matrix. Its limit
as " ! 0 is called the Wigner distribution and is denoted by W (t; x; k). The limit Wigner matrix
is not only Hermitian but also positive denite. The limit energy density and ux are expressed in
terms of W (t; x; k) by Z
1
E (t; x) = 2 Tr(A(x)W (t; x; k))dk

5
and Z
Fi (t; x; k) = 12 Tr(DiW (t; x; k))dk:
The limit Wigner distribution may be decomposed into dierent wave modes in a way that
generalizes the plane wave decomposition in a homogeneous medium. The dispersion matrix of the
system (2.2) is dened by
0 0 0 0 k1 = 1
B 0
L(x; k) = A, (x)ki Di = B 0 0 k2 = C
C
@ k3 = A : (2.8)
1
0 0 0
k1 = k2 = k3 = 0
It has one double eigenvalue !1 = !2 = 0 and two simple eigenvalues
!f = vjkj ; !b = ,vjkj ; (2.9)
q
where jkj = k12 + k22 + k32 and v is the sound speed

v = p1 : (2.10)

The corresponding basis of eigenvectors is

b1 = p1 (z(1) (k); 0)t ; b2 = p1 (z(2) (k); 0)t ;
^ ^
bf = ( pk2 ; p12 )t ; bb = ( pk2 ; , p12 )t ; (2.11)

where the vectors k^ , z(1) (k) and z(2) (k) form an orthonormal triplet. The eigenvectors b1 (k) and
b2(k) correspond to transverse advection modes, orthogonal to the direction of propagation. These
modes do not propagate because !1 = !2 = 0. The eigenvectors bf (k) and bb (k) represent forward
and backward acoustic waves, which are longitudinal , and which propagate with the sound speed v
given by (2.10).
The limit Wigner distribution matrix W (t; x; k) has the form [23]:
X2
W (t; x; k) = Wij (t; x; k)bi (k)bj  (k) + af (t; x; k)bf (k)bf (k) + ab(t; x; k)bb (k)bb (k): (2.12)
 =1
The rst term corresponds to non-propagating modes and may be set to zero without loss of gener-
ality. The last two terms correspond to forward and backward propagating sound waves. The scalar
functions af and ab are related by af (t; x; k) = ab (t; x; ,k), and af satises the Liouville equation
@a + r !
r a , r !
r a = 0: (2.13)
k x x k
@t
The functions af and ab can be interpreted as phase space energy densities since they are non-
negative (because the matrix W (t; x; k) is non-negative) and
1 Z Z
E (x) = 2 dk[af (t; x; k) + ab(t; x; k)] = dkaf (t; x; k):

6
The ux is given by
Z Z
F = v2 dk[k^ af (t; x; k) , k^ ab (t; x; k)] = v dkk^ af (t; x; k): (2.14)
The radiative transport equation (1.1) arises when the density  and compressibility  are random
and varying on the scale of the wave length, so we assume they have the forms
p p
 ! (1 + "1 ( x" ));  ! (1 + "1 ( x" )):
The random processes 1 and 1 are mean zero space homogeneous with power spectral densities
R^  , R^ , and cross spectral density R^  . The radiative transport equation for a(t; x; k) = af (t; x; k)
has the form
@a +vk^
r a , jkjr v
r a = v2 jkj2 Z (vjkj , vjk0 j)[a(k0 ) , a(k)]
@t n x x k
2 o

(k^
k^ ) R^ (k , k ) + 2(k^
k^ 0 )R^(k , k0) + R^ (k , k0) dk0 :
0 2 0 (2.15)
This equation is of the form (1.1). A systematic derivation using the Wigner distribution is given
in [23].
Equation (1.1) has been derived from equations governing particular wave motions by various
authors, such as Stott [26], Watson et.al. [32, 33, 19], Barabanenkov et.al. [3], Besieris and Tappert
[6], Howe [16], Ishimaru [17] and Besieris et. al. [7], with a recent survey presented in [4]. See also
[12] for similar results. These derivations also determine the functions !(x; k) and (x; k; k0 ) and
show how a is related to the wave eld. In [23], (1.1) and these functions are derived as a special
case of a more general theory.

3 Flat Boundary
3.1 Re ection and Transmission of Waves in Homogeneous Media
We will derive transport theoretic boundary conditions assuming that the media adjacent to the
interface are homogeneous, with the interface at in this section, and random, in later sections.
In this section we recall brie y the transmission and re ection of time harmonic, acoustic plane
waves at a plane interface. The time harmonic solutions of (2.1) with frequency ! satisfy
rp , i!v = 0 (3.1)
r
v , ip = 0:
We consider the high frequency regime, and replace accordingly ! ! !=" with "  1. Then (3.1)
becomes
"rp , i!v = 0 (3.2)
"r
v , ip = 0:
Consider a medium composed of two homogeneous half spaces in IRn , characterized by constant
densities and compressibilities i and i , i = 1; 2. Here i = 1 in thepupper half space xn > 0 and
i = 2 in the lower one xn < 0. The velocity of propagation is vi = 1= ii in each half space.
The wavenumber jki j of a plane wave is related to the frequency ! by the dispersion relation
! = vi jki j for i = 1; 2: (3.3)
7
Any solution of (3.2) in xn > 0 can be decomposed into a sum of incoming and outgoing waves,
ui"(x) = uiI;"(x) + uiR;"(x): (3.4)
When the medium is homogeneous
Z
uiI;"(x) = (2"d) kn, exp [i k "
x ] exp [,i  k"n xn ] i" (k0 ) bif (ki, )
0 0 0 i i
( 1)=2
(3.5)
Z
uR;"(x) = (2"d) kn, exp [i k "
x ] exp [i  k"n xn ] "i (k0 ) bif (ki+ ):
0 0 0 i i
i (3.6)
( 1)=2

The wave vector k0 is the horizontal component of k and kni , i = 1; 2 is its normal component:
q
ki = (k0 ; ikni ); kni (k0 ) = jkij2 , jk0 j2 ; 1 = 1; 2 = ,1 (3.7)
We consider only propagating waves because the energy in evanescent waves is exponentially small in
the high frequency limit, and vanishes as " ! 0. Therefore we assume that the normal components
of the wave vectors kn1 and kn2 are real. Then the support of the amplitudes i" and "i is uniformly
inside the ball fjk0 j < !=vi g. Then u^ iI;", the Fourier transform of uiI;" , and the corresponding
amplitude i" (k0 ) are related by
u^ iI;"( k" ) = (2") n( +1) 2= i" (k0 ) bi (ki,) (kn + i kni ): (3.8)
An analogous relation holds for the re ected waves.
The amplitudes i" and "i are not independent of one another. They are related through the
interface conditions at xn = 0, expressed by the continuity of the normal velocity and pressure
u1n (x0 ; 0) = u2n(x0 ; 0)
(3.9)
p1 (x0 ; 0) = p2 (x0 ; 0):
Equations (2.11) and (3.5-3.6) give the following relations for the wave amplitudes (for i 6= j )
(kni )1=2 (ki ),1 "i (k0 ) = Ri (k0 ) (kni )1=2 (ki ),1 i" (k0 ) + T j (k0 ) (knj )1=2 (kj ),1 j" (k0 ): (3.10)
Here Ri is the Fresnel re ection coecient for medium i, and T j is the Fresnel transmission coecient
from medium j into medium i. They are given by (see e.g. [18, 8])

Ri(k0 ) = j kni , iknj ;

j i i j i j i j =
T i(k0 ) = 2(j i knkni ) j ; i 6= j:
1 2
(3.11)
 kn +  kn  kn +  kn
Thus R2 (k0 ) = ,R1 (k0 ) and T 1 (k0 ) = T 2 (k0 ). Moreover, the conservation of energy ux implies
that
(Ri (k0 ))2 + (T i (k0 ))2 = 1 ; i = 1; 2: (3.12)
The angles of incidence and refraction i and j are related by Snell's law
sin i = sin j : (3.13)
vi vj
8
3.2 High Frequency Limit: Re ection and Transmission of Energy
To derive interface conditions on a at a at boundary we consider the Wigner transform of the
incoming and outgoing waves. We denote by W [u] the limit as " ! 0 of W"[u" ], the Wigner transform
of u" dened by (2.7). The directions of the incident and re ected waves are not overlapping. Then
by the general theory [14, 24], they decouple in the high frequency limit " ! 0 so that
W [ui ](x; k) = W [uiI ](x; k) + W [uiR](x; k): (3.14)
This also may be checked directly using (3.5) and (3.6). We will now show that
W [uiI ](x; k) = W n,1[i ] (k0 ; ,[x0 , xkn k0 ]) (kn + i kni ) bi (ki, )bi (ki, )
n (3.15)
i n , 1 i 0 0 x
W [uR](x; k) = W [ ] (k ; ,[x , kn k0 ]) (kn , i kni ) bi(ki+ )bi (ki+);
n
where W n,1[i" ] is the limit as " ! 0 of the (n , 1)-dimensional Wigner transform of i" .
Let us compute W [uiI ](x; k). By virtue of the relation [14, 24]
W"[f"(:)](x; k) = W" [ (2"1)n=2 f^"( ": )](k; ,x);
and (3.8), we have

W" [uiI;"](x; k) = 2" Z dp i (k0 + "p0 )i (k0 , "p0 )(k + " pn + i ki (k0 + " p0 )) (3.16)
(2)n " 2 " 2 n 2 n 2
(kn , " p2n + ikni (k0 , " p2 )) exp (ip
x)bi((k + " p2 )i, )bi((k , " p2 )i,):
0

From the dispersion relation (3.3) we deduce

(kn + " p2n )2 = (v!i )2 , (k0 + " p2 )2

2 0

(kn , " p2n )2 = (v!i )2 , (k0 , " p2 )2 :

2 0

Then
jkj + j "2p j = (v!i) ;
2
2 2
2
kn pn = k0
p0 :
On the other hand one readily checks that for constants c1 and c2 we have
(kn + "p2n + c1 ) (kn , "p2n + c2) = (kn + c1 +2 c2 ) (pn , c2 ," c1 ):
By using this relation and the change of variables pn ! "pn in (3.16), we get for " small,
Z
W [uiI ](x; k)  (21)n,1 dp0 i" (k0 + "2p )i" (k0 , "2p ) exp i[p0
(x0 , xkn k0 )]
0 0
n (3.17)
b (k, )b (k,) (kn +  kn):
i i i  i i i

When " ! 0 in (3.17) it yields the rst equation in (3.15), and the second equation in (3.15) can be
derived similarly.

9
 The explicit representation (3.15) implies that the Wigner matrix of the incoming waves has the
form
WIi = aiI (x; k)bi (k)bi (k)
with aiI supported on the wave vectors k with kn = ,i kni . In the interior, the function aiI satises
the transport equation
k^
rxaiI = 0: (3.18)
Similarly
WRi = aiR (x; k)bi (k)bi (k):
The function aiR is supported on the hemisphere of outgoing wave vectors with kn = i kni and it
satises the same transport equation (3.18).
Since the coecients Ri and T i do not depend on ", we deduce from (3.10) that
kni W n,1[ i ] = kni jRi (k0 )j2 W n,1[i ] + knj jT j (k0 )j2 W n,1[j ]
(ki )2 (kq
i )2
qj (kj )2 (3.19)
kni kn i 0 j 0
+ ki kj [R (k ) T (k ) + Ri (k0 ) T j (k0 )] W n,1[i ; j ]:
By virtue of (3.15), the rst two terms on the right side of (3.19) are related to the incident energy
W [uiI ]. The third term, involving W n,1[i ; j ], corresponds to the correlation between amplitudes
of waves coming to the surface from opposite sides. We assume that they are uncorrelated, so that
W n,1 [i ; j ] = 0: (H1)
Condition (H1) is not satised in general; in particular not for two incoming plane waves. How-
ever it is satised when the wave packets arriving at the interface from dierent sides have uncorre-
lated random phases. This is the case in the radiative transport regime when waves arriving at the
interface have undergone independent multiple scattering.
When (H1) holds, (3.19) shows that the ai satisfy the interface conditions
vi k^ni ai(x0 ; 0; k0 ; ikni ) = jRi (k0 )j2 vi k^ni ai(x0 ; 0; k0 ; ,i kni ) + jT j (k0 )j2 vj k^nj aj (x0 ; 0; k0 ; ,j knj ): (3.20)
Note that evanescent waves disappear in the high frequency regime. Therefore aj (x0 ; 0; k0 ; knj ) = 0
when knj is imaginary. For incident angles below the critical angle (obtained through Snell's law
(3.13)), the re ection coecient jRi (k0 )j2 = 1 and no energy is transmitted. These relations also
hold if the two adjacent media are deterministic but inhomogeneous, as is shown in [24]. We expect
that they hold when the adjacent media are random with weak uctuations.

4 Rough Boundaries
4.1 Surface Re ection Operator
Let us assume that the boundary of the domain or the interface between two subdomains is a surface
given by (x0 ; h(x0 )). First we describe the re ected and transmitted waves generated by an incident
wave, and then derive boundary conditions for the phase space energy density in the high frequency
limit.
The rst part of this problem has been studied at length in the past. (See [5, 11, 29].) For
simplicity we deal rst with re ected waves without transmission, so the superscript corresponding
10
to the medium can be dropped. Consider an incident plane wave, with k00 the horizontal component
of the wave vector and with amplitude " = 1:
in (x) = exp [ik00
x0 , ikn (k00 )xn ] b(k, ):
In the region xn > hmax , the scattered wave can be decomposed into a sum of outgoing plane waves:
Z
sc(x) = dk0 S (k0 ; k00 ) exp [ik0
x0 + ikn (k0 )xn ] b(k+ ):
The kernel S (k0 ; k00 ) is the scattering amplitude for waves of incident horizontal wave number k00
scattered to waves of horizontal wave number k0 . It is determined by the height h(x0 ) of the surface.
S can be obtained by solving an integral equation that follows from the boundary conditions.
Let the boundary vary on the scale of the wavelength, let " be proportional to the wave length,
and let the scaled boundary h" be given by
h" (x0 ) = "h( x" ):
0
(4.1)
Here h is a bounded function. This scaling implies that the roughness height, the scale length of
the boundary, and the wavelength are all of the same order. Then S is independent of ". In the
high frequency limit an incident wave has scaled horizontal wave vector k0 =". Then the change of
variables x ! x=" yields the same equation for S as when " = 1. However, the incoming wave still
depends on " through the amplitude " .
In order to express conservation of energy ux at the boundary in a natural way we introduce
the amplitudes for energy uxes:
q q
~ " = vk^n " ; ~" = vk^n " :
These are energy ux densities, as can be seen from the expression (2.14) for the total ux.
Consider the general incoming and outgoing plane waves given by (3.5), (3.6) with the incident
and re ected amplitudes related by
Z
~"(k0 ) = dk00 S (k0 ; k00 )~" (k00 ): (4.2)
The incoming and outgoing Wigner transforms are given in (3.15) with W n,1 [~" ] being the known
incoming energy ux. The unknown W n,1[~] is given by
Z Z Z
0 exp [ip0
x0 ] dk0 S (k0 , " p ; k0 )~" (k0 ) dk~ 0 S (k0 + " p ; k~ 0 )~" (k~ 0 )
0
W n,1 [~](k0 ; ,x0 ) = "lim d p 2 20 2 0
Z!0 Z 0


0 0 0

= dp0 exp [ip0
x0 ]  dk00 S (k0 ; k00 )~" (k00 ) + O(");

since S does not depend on ". It is clear from this expression that we cannot write W n,1 [~] only in
terms of W n,1 [~] for a general scattering kernel S .
We consider random boundaries that are stationary, i.e. statistically invariant under translation.
The physical problem is not changed by translation and the incident and re ected waves change phase
only. Then the scattering operator for a surface shifted by a vector d0 must satisfy the relation
Sh(x ,d ) (k0; k00 ) = exp [i(k0 , k00 )
d0 ]Sh(x ) (k0 ; k00 ):
0 0 0 (4.3)

11
Since hh(x0 )i, the mean over all realizations, is a constant, we conclude from (4.3) that
hS i(k0 ; k0 ) = exp [i(k0 , k0 )
d0]hS i(k0 ; k0 ):
0 0 0

Because this relation is true for all shifts d0 we conclude that hS i has the form
hS i(k0 ; k0 ) = R(k0 )(k0 , k0 ):
0 0 (4.4)
Here R(k0 ) is the coherent re ection coecient.
Consider now the uctuation
S of the scattering amplitude:
S (k0 ; k0 ) = R(k0 )(k0 , k00 ) +
S (k0 ; k00 ): (4.5)
The product of two uctuations are given by
 0  a0 0
k0 , a ; k0 a a
0
(k0 ; k00 ; a0 ; a00 ) =
S 2 0 2 2 0 2
00 0
,
S k + ; k + : 0

Again, because of translational invariance, the average of  is proportional to (a0 , a00). Thus there
exists a kernel
E such that:
hi(k0 ; k0 ; a0; a0 ) =
E (k0 ; k0 ; a0)(a0 , a0 ):
0 0 0 0 (4.6)
Conservation of energy means that for every incident eld ~ (k ), 0
Z Z
j~(k0 )j dk0 = j~ (k )j dk :
2
0
2
0

Squaring and averaging (4.2), using (4.5), yield then

Z
jR(k0 )j +
2
 (k00 ; k0 ) dk00 = 1; jk0j < !=v: (4.7)
jk0 j<!=v
0

Here
 (k00 ; k0 ) =
E (k00 ; k0 ; 0) (4.8)
is the dierential cross section for the incoherent or diuse surface scattering.
We are now in a position to calculate the average energy associated with the re ected waves in
terms of the incident energy. We have
hW nZ, [~]i(k0 ; ,x0 ) Z
1
Z
= "lim d p0 exp [ i p 0
x 0 ] h dk 0 S ( k0 , " p 0
; k0 
)~ (k 0 ) d ~0 S (k0 + " p ; k~0 )~" (k~0 )i
k
0
! Z 2 " 2
0
Z Z 0 0

p 0
0

p0
0 0 0

= "lim
! 0
dp exp [ip
x ] dk dk hS (k , " 2 ; k )S (k + " 2 ; k )i~ " (k0 )~" (k~0 ):
0 0 0 0
0
~0 0
0 0 0
0
~0 0 0 0

Because of (4.4){(4.6) we have

< S (k0 , " p2 ; k00 )S (k0 + " p2 ; k~0 0 )i = jR(k0 )j2 (k0 , " p2 , k00 )(k0 + " p2 , k~0 0 )
0 0 0 0

+
E (k0 ; k0 +2k 0 ; "p0 ) (k00 , k~0 0 , "p0 ) + O(")

~ 0 0

12
since R is independent of ". Since
E is also independent of ", we get
hW n, [~]i(k0 ; ,x0 ) = jRZ (k0 )j hW n, [~]i(kZ0 ; ,x0 )
1 2 1

+ dp0 exp [ip0
x0 ] dk0
E (k0 ; k0 + "p0 ; 0)~" (k0 )~" (k0 , "p0 ) + O(")
Z 0 0 0 0

= jR(k0 )j hW n, [~]i(k0 ; ,x0 ) + dk0  (k0 ; k0 )hW n, [~]i(k0 ; ,x0 ):

2 1
0 0
1
0

This implies the following boundary condition for the average scalar energy a(x; k) at xn = 0:
vk^n (k0 ) a(x0 ; 0; k0 ; kn ) = jZR(k0)j2 vk^n(k0 ) a(x0 ; 0; k0 ; ,kn )
(4.9)
+ dk00  (k0 ; k00 ) vk^n (k00 ) a(x0 ; 0; k00 ; ,kn (k00 )); kn > 0:
Here R(k) dened by (4.4) is the re ection coecient and  (k; k0 ) dened by (4.8) is the dierential
scattering cross-section of the boundary.
4.2 Re ection and Transmission at a Rough Interface
Let us now consider two homogeneous half spaces separated by a rough interface h. The scalar
scattering operator S must be replaced by a two by two matrix-valued operator (S ij )1i;j 2 . When
a wave of unit amplitude is incident in medium j , S ij determines the amplitude of the wave scattered
into medium i. As before, we introduce the ux amplitudes ~ i" and ~"i :
q q p! q
~i" = vi k^ni i";
= ki kni "i : ~"i = vi k^ni "i
(4.10)
We consider general incoming and outgoing plane waves given by (3.5) and (3.6), with the incident
and re ected amplitudes related by
Z Z
~"i (k0 ) = dk00 S ii (k0 ; k00 )~i" (k00 ) + dk00 S ij (k0 ; k00 )~j" (k00 ):
Again we deduce from translational invariance that
S ij (k0; k00 ) = Rij (k0 )(k0 , k00 ) +
S ij (k0 ; k00 ): (4.11)
Moreover,
hi kl i(k0 ; k0 ; a0; a0 ) =
E i kl(k0 ; k0 ; a0 )(a0 , a0 );
;
0 0
;
0 0 (4.12)
where  a0 0  a0 0
0 0 0 0 a
 (k ; k ; a; a ) =
S k , 2 ; k , 2
S k + 2 ; k + 2 :
i kl
; ik il 0 0 a 0 0
0 0 0 0

Therefore we have
hS ik (k0 , " p ; k0 )S il (k0 + " p ; k~0 )i = Rik (k0 )Ril (k0 )(k0 , " p , k0 )(k0 + " p , k~0 )
2
0

0 2
0

0 2
0

0 2
0

+
E i kl (k0 ; k0 k 0 ; "p0 ) (k0 , k~0 , "p0 ) + O(")

+ ~0
0
;
2 0 0

Here Rik (k0 ) are the re ection coecients.

We denote the dierential scattering cross-sections by
 i;kl(k0 ; k00 ) =
E i;kl (k0 ; k00 ; 0): (4.13)
13
Then as above, we obtain
X 2
hW n,1[~i ]i(k0 ; ,x0) = Rik (k0 )Ril (k0 )W n,1 [~k ; ~l ](k0 ; ,x0 )
k;l=1
XZ 2
+ dk00  i;kl (k0 ; k00 )W n,1 [~k ; ~l ](k00 ; ,x0 ):
k;l=1
When the decorrelation condition (H1) stated in the previous section holds, we let  ij =  i;jj and
we get the incoming-outgoing energy ux relations
X2
vik^ni (k0 ) ai (x0 ; 0; k0 ; ikni ) = jRij (k0 )j vj k^nj (k0 ) aj (x0 ; 0; k0 ; ,j knj )
2

j
X Z 0 ij 0 0 j ^j 0 j 0 0 j j 0 (4.14)
=1
2
+ dk  (k ; k ) v kn (k ) a (x ; 0; k ; , kn (k )):
0 0 0 0 0
j =1
Here Rij is dened by (4.11) and  ij by (4.13).
4.3 Re ection and Transmission from Small Amplitude Rough Interfaces
We now specialize to re ection and transmission at a rough interface with small random uctuations.
The interface is given by xn = "h( x" ), where  is another small parameter. The random process
0

h(y0 ) has mean zero and is stationary, with covariance function Q(y0 ) dened by
hh(x0 + y0)h(x0 )i = Q(y0 ):
The power spectral density Q^ (p) of the surface uctuations is the Fourier transform of Q:
Z dp0 eip
y Q^ (p0 ):
Q( y 0 ) = (2)n,1
0 0
(4.15)
To simplify the calculations, we assume that the scattering operator S , a 2  2 matrix, has an
expansion of the form
S = S0 +  S1 + 2 S2 + o(2 ): (4.16)
Here S0 (k0 ; k00 ) = R(k0 )(k0 , k00 ) is the scattering operator of the unperturbed at surface. Clearly
S1 has mean zero because it is linear in h. Then
R(k0 ) = R0 (k0)(k0 , k00) + 2 R2 (k0 )(k0 , k00 ) + o(2 )
(4.17)

S = S1 + O(2 ):
Then (4.13) implies that  depends only on S1 up to o(2 ). Moreover R depends only on S2 through
its statistical average R2 .
The details of the small  computation of the re ection and transmission coecients is given in
Appendix A. The results are as follows. The leading order term is
0 1
R T C
R = B
@ A: (4.18)
T ,R
0

14
The Fresnel coecients R and T are given in (3.11). They can also be written in the form

R =  +,  ; T = 2+ ;
2 2
1
2
2
2 2
1 2
2
(4.19)
1 2 1 2

where s
i = kni :
i
Now we dene the auxiliary matrices P , M and N by
0 1 0 1
( k ) , p0
k0
0 ( k )
1 2
, p 0
k0
0
2 2

P =B

@ , 1  C
2
A ; M ( p B@ , kn (k0)  (k )
0 ; k0 ) = B kn (k0 )  (k ) C
1 CA (4.20)
1 2 2

1 1 , , , ,
2
1
kn(k ) (k )
0 0 kn (k ) (k )
1 0 0
1
1 2
2
1

and N (p0 ; k0 ) = P ,1 (p0 )V0 M (p0 ; k0 ); where

0 1
1 0
V = B
@ CA :
0 ,1
0

Then Bij is given by B (p0 ; k0 ) = 12 [R0 (k0 )N (p0 ; k0 ) + N (p0 ; k0 )R0 (k0 )].

Bij = ,(p
k )i (p )j (k )(+2 + 2 )((p10, 2 )1 2 (p0 )1 2 (k0 )i (k0 )j (p0 ) ;
0 0 0 0 ij
(4.21)
0 0

1 )(2 + 2 )(k0 )
2 1 2

where i0 = 2 if i = 1 and vice-versa, and

,(p0
k0 ) = (k ) ,
2 2
k0
p0 , (k1 )2 , k0
p0 :
2 1
The interface conditions (3.9) are continuity of the normal velocity and pressure. Then from
(4.13) and (4.17) we nd, after a long computation given in Appendix A, that the scattering cross-
section tensor is
 ij (p0 ; k0 ) = 4Bij2 (p0 ; k0 )Q^ (p0 , k0 ): (4.22)
The part R2 of the coherent re ection coecient may be split into a sum of two terms
R2 = R2;0 + R2;1 ;
which are given by
0 1
R ; = T2 ((k ) , (k ) ) Q(0) B
0 1C
1 2
@ 2 2
A
,
20

Z 1 0 (4.23)
R ; (p0) = ,2 dk0 N (p0 ; k0 ) B (k0 ; p0 )Q^ (p0 , k0):
21

Conservation of energy ux (4.7) has to be satised at the interface. This means that energy
incident upon the interface is either re ected, transmitted, or scattered. From (4.17)
jRij j = jRij +  Rij j = (Rij ) + 2 Rij Rij + o( ):
2
0
2
2
2
0
2 2
0 2
2
(4.24)
15
Therefore up to o(2 ) conservation of energy ux for each p0 is given by
Z
1 = jR11 (p0 )j2 + jR21 (p0 )j2 + 2 dk0 [ 11 (k0 ; p0 ) +  21 (k0 ; p0 )] (4.25)
Z
1 = jR22 (p0 )j2 + jR12 (p0 )j2 + 2 dk0 [ 22 (k0 ; p0 ) +  12 (k0 ; p0 )]:
In the limit  = 0 this simplies to (3.12). The 2 terms in (4.25) give the identity
Z
0 = 2(,T R221;0 + RR211;1 + T R221;1 ) + dk0 [ 11 (k0 ; p0 ) +  21 (k0 ; p0 )]
Z
0 = 2(T R2;0 , RR2;1 + T R2;1 ) + dk0 [ 22 (k0 ; p0 ) +  12 (k0 ; p0 )]:
12 22 12

A straightforward though tedious computation shows that this is indeed true when we use the
formulas given above. Note that the scattering cross section (4.22) is positive for every incident and
re ected or transmitted angle. Therefore, when the incident angle is smaller than the angle of total
re ection for the at interface, only diuse energy is transmitted.
To summarize, we have derived boundary conditions for the transport equation (3.18) in two
adjacent homogeneous half spaces from the plane wave decomposition (3.5) and (3.6) for both at
and rough interfaces, with small uctuations. For slowly varying heterogeneous media with at or
smoothly varying interfaces such conditions are derived in detail in [24]. For weakly uctuating
random media or heterogeneous media with random rough interfaces no such derivation has been
given. In the high frequency limit the boundary conditions for the transport equation must remain
valid, but this is dicult to justify. When the uctuations of the interface are small, the interaction of
the volume scattering with the surface uctuations must be of higher order and therefore negligible.
4.4 Derivation of the Dirichlet and Neumann Boundary Conditions
When 1 =2 ! 0, the interface corresponds to a hard boundary with vanishing normal velocity
and Neumann boundary condition for waves incident from medium 1, and to a soft boundary with
vanishing pressure and Dirichlet boundary condition for waves incident from medium 2. Then R ! 1,
T ! 0, and 0 1
1 0 C
R0 = B@ A:
0 ,1
Moreover
0 1
(k )
1 2
, p 0
k0
^ 0
B ,2i (k1 (p0)k1 (k0 ))1=2 h(p , k )0 0 CC : (4.26)
S1(p0; k0 ) = B@ n n A
0 ,2i(kn2 (p0 ) kn2 (k0 ))1=2 h^ (p0 , k0)
As is to be expected in this limit, R2;0 = 0 and R2 = R2;1 is given by
0 Z ((k1 )2 , k0
p0)2 1
B ,2 dk kn1 (k0 )kn1 (p0 ) Q^ (p , k ) Z
R2 (p0 ) = B
0 0 0 0 CC : (4.27)
@ A
0 2 dk0 kn2 (k0 )kn2 (p0 )Q^ (p0 , k0 )
The dierential scattering tensor  , deduced from the expression for S1 , is
0 12 0 02 1
((k ) , k
p ) Q^ (p0 , k0 )
 (p ; k ) = B
B 4 0 CC :
@ kn (k )kn (p )
0 0 0 1 0 (4.28)
A
1

0 4kn2 (k0 )kn2 (p0 )Q^ (p0 , k0 )

16
 γ0

γ1 Ω1

γ2 Ω2

Figure 5.1: Geometry of the media

Here 11 is the dierential scattering cross-section for the Neumann problem in the rst medium,
and 22 is the dierential scattering cross-section for the Dirichlet problem in the second medium.
Conservation of energy ux also holds:
Z
0 = 2R0ii R2ii;1 (p0 ) + dk0  ii(p0 ; k0 ); i = 1; 2:

5 The Initial Boundary Value Problem for the Radiative Transport

Equation
Now we summarize the initial, boundary, and interface conditions for the radiative transport equation
in two adjacent plane layers
1 and
2 with rough boundaries, separated by a rough interface, as is
shown in Figure 5.1. Radiative transport theory is valid in the interior of each layer. The boundary
and interface conditions are those for high frequency time harmonic waves derived in the previous
sections. We now use them in the time domain, assuming that time variations in the transport
regime are slow compared to the oscillation frequency ",1 . This is so if the frequency spread of the
waves about a carrier frequency is narrow. At the outer boundaries we impose a Neumann condition
at 0 and a Dirichlet condition at 2 .
The transport equation in
i is
@ai (t; x; k) + vi k^ i
r ai (t; x; k) = Z i (x; k; k0 ) hai (t; x; k0 ) , ai (t; x; k)i dk0 ; i = 1; 2: (5.1)
x
@t R3
The scattering cross-sections are given by
i (x; k; k0 ) = M (v 2) jkj (vi jkj , vi jk0 j)
i2 2
(5.2)
n o

(k^
k^ 0)2 R^
i (k , k0 ) + 2(k^
k^ 0 )R^ i (k , k0 ) + R^ i (k , k0 ) :
 
The mean boundaries 0 , 1 and 2 are the at surfaces xn = L0 , xn = 0 and xn = ,L2
respectively. At 0 we require p = 0 and at 2 , vn = 0. Then the conditions at 0 and 2 are given
respectively by
h i
a1(t; x0 ; L0 ; k0 ; kn ) = 1 Z, 2R211;1(k0 )2 a1(t; x0 ; L0 ; k0 ; ,kn)
(5.3)
+ 2 dp0  11 (k0 ; p0 )a1 (t; x0 ; L0 ; p0 ; p1n )

17
 h i
a2 (t; x0 ; ,L2 ; k0 ; kn ) = 1 Z, 2R222;1 (k0 )2 a2 (t; x0 ; ,L2; k0 ; ,kn)
(5.4)
+ 2 dp0  22 (k0 ; p0 )a2 (t; x0 ; ,L2 ; p0 ; p2n ):
Here kn < 0, pn > 0 at 1 , and kn > 0, pn < 0 at 2 , so that the part on the left represents
the outgoing or scattered wave, and the part on the right corresponds to the incoming waves. The
matrices R2;1 and  appearing in (5.3) and (5.4) are given by (4.27) and (4.28).
The interface condition at 1 for the outgoing eld a1 (t; x0 ; 0; k0 ; kn1 ) with kn > 0 is
,
,

a1 (t; x0 ; 0; k0 ; kn1 )= jRj2 + 2RR211;1 2 a1 (t; x0 ; 0; k0 ; ,kn1 ) + jT j2 + 2T (R212;0 + R212;1 )2 a2 (t; x0 ; 0; k0 ; kn2 )
Z Z
+2 dp0  11 (k0 ; p0 )a1 (t; x0 ; 0; p0; ,p1n ) + 2 dp0  12 (k0 ; p0 )a2 (t; x0 ; 0; p0; p2n ): (5.5)
Here ,p1n < 0 and p2n > 0 correspond to incoming waves. The condition on a2 is similar to (5.5).
The matrices R2;0 , R2;1 and  are given by (4.23) and (4.22). These interface conditions are valid
under the assumption that waves incident on the interface from the two sides are uncorrelated.
This assumption is valid provided that the widths of the layers L1 , L2 are much larger than the
corresponding scattering mean free paths
j
lj = v j ; j = 1; 2:
The total scattering cross-section j is dened by (1.2).

Acknowledgement
We thank John G. Watson for the use of his calculations in Appendix B.2. The work of Bal,
Keller and Papanicolaou was supported by AFOSR grant F49620-98-1-0211 and by NSF grant
DMS-9622854. The work of Ryzhik was supported in part by the MRSEC Program of the National
Science Foundation under Award Number DMR-9400379.

A Derivation of the Interface Conditions

We shall now calculate the re ection coecients (4.18) and (4.23), and the scattering cross section
tensor (4.22). We note rst that the normal vector to the surface [ x" ; h( x" )] is 0 0

 0 0  0   0 
 x" ; h( x" ) = C ez , rh" ( x" ) = C ez , (rh)( x" ) ; (A.1)
where C is a normalization constant. Using the plane wave decomposition (3.5) and (3.6), the
denition (2.11) of the eigenvector bf , and the ux amplitudes (4.10), we obtain
Z Z
dk0 ei k "x e,i1kn1 h( x" ) k
 ,  kn n  ~ 1 + dk0 ei k "x ei1 kn1 h( x" ) k
 +  knn  ~1
0


0 0 0 1 1
0 0 0 1 1
0


0 0

k " 1
k 1 "
Z 1
n Z 1
n
dk0 ei k "x e,i2kn2 h( x" ) k
 ,k2 kn n 2 ~ 2" + dk0 ei k "x ei2 kn2 h( x" ) k
 +k2 knn 2 ~"2
0 0 0 0 2 2
0 0 0 0 2 2
0 0

=

n n
(A.2)
Z Z
0
dk e e
i k x ,i1 k1 h x
 ~ " + dk0 ei k x ei1 k1 h x , ~"
,
0 0 0 0 0 0



" n ( " ) 1 1 " n ( " ) 1 1

Z Z 1 1

0
= dk e e
i k x ,i2 k2 h x
 ~ " + dk0 ei k x ei2 k2 h x , ~" :
,
0 0 0 0 0 0



n ( " ) 1 2 n ( " ) 1 2
" "
2 2

18
Here i is dened by s
i = kni :
i
The incoming energy ux amplitudes ~i" are not random.
To calculate the re ected wave amplitudes ~"i using a Born expansion in the height  of the rough
boundary, we need the following intermediate expansions:
exp [ikn h( x" )] = 1 + ikn h( x" )  , 21 (kn )2 h2 ( x" ) 2 + o(2 )
0 0 0

x 0  x 0 x0 
exp [ikn h( " )](k
n)C = kn + i(kn ) h( " ) , k
rh( " ) 
, 1 2 0
 0  (A.3)
+ , 21 (kn )3 h2 ( x" ) , ikn h( x" )k0
rh( x" ) 2 + o(2 )
0 0

~"i = ~";0 +  ~";1 + 2 ~";2 + o(2 ):

Using (A.3) in (A.2), we nd at zeroth order in 
~";1 0 1 , ~ 1" 1 = ,~";2 0 2 + ~ 2" 2
(A.4)
~";1 01,1 + ~ 1" 1,1 = ~";2 02,1 + ~ 2" 2,1 :
We can now invert (A.4) to obtain
0 1
R T C
~"; = R ~ " = B
@ A ~": (A.5)
T ,R
0 0

Here ~ " = (~i" ) and ~";k = (~";k

i ), with k = 0; 1; 2, are two-component vectors, and R and T are
given by
R = 12 +, 22 ; T = 22+1 22 :
2 2
(A.6)
1 2 1 2

At the rst order in  in equation (A.2) we obtain

Z (0
(rh)( x ) ! )
dk0 exp [i k x
0
x0 0 k ~ ~
0

" ]( iknh( " ) ,

"
kn !  (~" + "; ) +  "; )
1 1 1 1
1 1 0 1 1

Z k0
(rh)( x" )  (~ + ~ ) ,  ~
= dk0 exp [i k "
x ] ikn h( x" ) ,
0 0 0 0

2 2 2 2
k " "; "; 2 0 2 1
n2
(A.7)
Z k 0
x0  x 0 
0 , 1 ~
dk exp [i " ] ikn h( " ) (,~ " + "; ) +  "; , ~ 1 1 1 1 1

Z 0 0  
1 0 1 1

= dk0 exp [i k "
x ] ikn h( x" ), (~" , ~"; ) + , ~"; :

0
2 1 2 2 1 2
2 0 2 1

Denoting by ^h the Fourier transform of h, we have

Z 0 0 Z 0 0
h( x" ) = exp (i  "
x )h^ ( 0 )d 0 ; (rh)( x" ) = i 0 exp (i  "
x )h^ ( 0 )d 0 :
0 0

19
Now we replace ~";0 in (A.7) by its expression (A.5), and compute the Fourier transform of (A.7)
with respect to x0 , denoting the conjugate variable p0 :
Z
 ~1 (p0 ) + dk0 h^ (p0 , k0)  (ik1 , i k
(p , k ) ) ((1 + R)~1 + T ~ 2 )
0 0 0
";1 n k0 n1 0 0 " "
Z
1 1

= ,2~";2 1(p0 ) + dk0h^ (p0 , k0) 2 (ikn2 , i k
(pk2, k ) ) (T ~1" + (1 , R)~2" )
n
(A.8)
Z
1,1 ~";1 1 (p0 ) + dk0 h^ (p0 , k0 ) 1,1 ikn1 ((,1 + R)~1" + T ~ 2" )
Z
= 2,1 ~";2 1 (p0 ) + dk0 h^ (p0 , k0 ) 2,1 ikn2 (,T ~ 1" + (1 + R)~2" ):
For m = 1; 2 we have
iknm , i k
(pkm, k ) = i (k ) k,mk
p :
0 0 0 m2 0 0
n n
Using the denition of M given in (4.20) we can rewrite (A.8) in vector form:
Z
P (p0)~";1 (p0) = dk0 ih^ (p0 , k0)(M (p0 ; k0 )~" (k0 ) + V0 M (p0 ; k0 )~";0 )(k0 ): (A.9)
From (A.6) we have

1 (1 + R) = 22+12 ; 2 (1 , R) = 22+22
3 3

1 2 1 2

along with similar expressions for T . Therefore, from (A.8) we get

Z
(1 ";1 + 2 ";1 )(p ) = 2i dk0 h^ (p0 , k0 ),(p0
k0 ) (1(~2" + + 2 ~2" ) (k0 )
1
~ 1 ~ 2 0
Z 22 ) (A.10)
, 2 12 ~2" ) (k0)
1

(1,1 ~";1 1 , 2,1 ~";2 1 )(p0 ) = 2i dk0 h^ (p0 , k0 )(1 , 2 ) (1 2(~2" +
2 1

2 )
1 2

where
,(p0
k0 ) = (k ) ,2k
p ) , ((k ) , k0
p0) :
2 2 0 0 1 2

1
Using the solution (A.6) of (A.5) we obtain in vector form
Z Z
(~";1 )(p0 ) = dk0 S1 (p0 ; k0 )(~" )(k0 ) = 2i dk0 h^ (p0 , k0 )B (p0 ; k0 )(~" )(k0 ) (A.11)
where B is given by (4.21).
The terms of order 2 in (A.2) yield
Z Z
k
exp ( " ) dk = exp ( k "
x ) dk0
0
x0 0 0
0 1 2
(A.12)
where the vectors m are
1 x 0 x x 0 0 x k
rh( " ) )~m + m~m ]
1m = m [,( (knm )2 h2 ( ) + ih( )k0
rh( ))m (,~" + ~";m0 ) + (iknm h( ) ,
0 0 x0
2 "0 " " " knm ";1 ";2
(A.13)
, 1 m x m ~m m m x 0
~ m ~
2 = m [, (kn ) h ( )(~ + ";0 ) + (i kn h( ))";1 + ";2 ]:
m 1 2 2 m
2 " "

20
We average the interface conditions (A.12) before inverting them.
We note rst that
x 0 Z
hh ( )i = Q^ (0 )d0 = Q(0)
2
"0
x 0 x Z
hh( " ) k
rh( " )i = ik0
0 Q^ (0 )d0 = k0
rQ(0):
0

We also note the following Fourier transforms for m = 1; 2

0
rh( x )
0
Z x0 k
Fx,!p h dk0 m (k0 ) exp [ik0
x0 ="](iknm h( " ) , km " )~";m i
1

Z
0 0 1
m ) , p0
k0 n
0 0 ( k 2
^
= ,2 dk m (k ) km (k0 ) Q(k , p )(B (k ; p )~" (p0 ))m
0 0 0 0
Z n
Fx !p h dk m (k ) exp [ik0
x0 ="]iknm h( x" )~";m i
0
, 1 0 0
0
Z 0 1

= ,2 dk0 m (k0 )knm (k0 )Q^ (k0 , p0 )(B (k0 ; p0 )~" (p0 ))m
Here Fx,1!p
0 0 is the inverse Fourier transform with conjugate variables x0 and p0 .
Taking the Fourier transform of the interface conditions (A.12), we get the following 2  2 system
1 (p0 )[,( 21 (kn1 )2 Q(0) + ip0
rQ(0))(,~ 1" + ~";1 0 )(p0 )]
Z
,2 dk01 (k0 ) (k )k1,(kp0 )
k Q^ (p0 , k0)(B (k0 ; p0 )~"(p0 ))1 + 1 (p0 )h~";1 2 i
1 2 0 0
n
0 1
= 2 (p )[,( 2 (kn ) Q(0) + ip0
rQ(0))(~2" , ~";2 0 )(p0 )]
2 2

Z
,2 dk02 (k0 ) (k )k2,(kp0 )
k Q^ (p0 , k0)(B (k0 ; p0 )~"(p0 ))2 , 2 (p0 )h~";2 2 i
2 2 0 0
n
(A.14)
1,1Z(p0 )[, 12 (kn1 )2 Q(0)(~1" + ~";1 0 )]
,2 dk01,1 (k0)kn1 (k0 ) Q^ (p0 , k0 )(B (k0 ; p0)~" (p0))1 + 1,1 (p0 )h~";1 2i
= 2,1 (p0 )[, 21 (kn2 )2 Q(0)(~2" + ~";2 0 )]
Z
+2 dk0 2,1 (k0 )kn2 (k0 ) Q^ (p0 , k0 )(B (k0 ; p0 )~" (p0 ))2 + 2,1 (p0 )h~";2 2 i:
We solve (A.14) for h~";2 i and write the result in terms of the tensor R2 . Then we split the tensor
R2 into two parts, R2 = R2;0 + R2;1, and we obtain the formulas (4.23).

B Grazing Angles and the Smoothing Method

As has been pointed out by several authors [28, 10, 30, 31] the re ection coecient blows up as
the incidence direction becomes parallel to the mean surface in the case of the Neumann boundary
condition. As can be seen from (4.27), the term R211;1 is innite for grazing angles, since kn (p0 ) = 0.
This shortcoming can be avoided by using the smoothing method instead of the Born expansion.
For the Neumann boundary condition, the smoothing method gives the re ection coecient [31]
R(k) = 11 , + 2 Q (B.1)
2 Q
21
where Z
Q = , dk0 ((kk )(k,0 )kk (
pp0 )) Q^ (p0 , k0):
0 01 2 2
(B.2)
n1 n1
The Born expansion for the re ection coecient gives, when (4.27) is used,
R(k) = 1 + 22 Q + O(3 ):
Thus, (B.1) agrees with the Born expansion to O(3 ), but it is not singular at grazing angles. Note
that Q is negative since it involves only propagating waves for which kni is real. Therefore the
re ection coecient R given by (B.1) ranges between 0 and 1.
The smoothing method gives for the scattering cross-section [31]

ren
11
(p0 ; k0 ) = (1 , 24Q(k0 ))2 ((kk1)(k,
1 2
k0
p0)2 Q^ (p0 , k0): (B.3)
n )kn1 (p0 )
0
This coecient is nonsingular at grazing angles.
The reason why the second Born approximation (4.27) is singular at grazing incidence is that
it includes only single and double scattering. At grazing incidence, all the singly scattered waves
combine in phase in the direction of the re ected wave, and their sum is innite. Double scattering
alone does not reduce these waves enough to make the sum converge. The smoothing method takes
into account some multiple scattering of all orders|that corresponding to the ladder diagrams of
diagrammatic resummation methods [25, 27]|and this reduces the scattered waves enough to make
their sum converge.
B.1 Smoothing Method for Dirichlet and Neumann Boundary Conditions
The smoothing method is an alternative to the Born expansion which includes ladder diagrams. We
will present it in a form suitable for re ection from a random surface with a Dirichlet or Neumann
boundary condition, as in [31].
Vanishing pressure (soft boundary condition) and vanishing normal velocity at the rough bound-
ary (hard boundary condition) give equations of the form
K (,)~ = K ()~: (B.4)
Here ~ and ~ are the incident and re ected wave amplitudes and the sign is , for the Dirichlet and
+ for the Neumann boundary condition. The operator K for the Dirichlet and Neumann conditions
is
Z Z
[KD ()~g ] (p0 ) = (2M ),n dx0 e,ip
x0 0
dk0 eik
x e,iknh(x ) ,1 (k0 )~g(k0 )
0 0 0

Z Z
dk0 eik
x e,iknh(x ) (,1 +  k
rk h(x ) )(k0 )~g(k0 ):
0 0
[KN ()~g ] (p0 ) = (2M ),n dx0 e,ip
x0 0 0 0 0

n
Expansion in powers of  gives
K () = 1 + K1 + 2 K2 + O(3 ): (B.5)
Plugging (B.5) into (B.4) yields
(I , K1 + 2 K2 )~ = (I + K1 + 2 K2 )~ + O(3 ): (B.6)

22
 We denote by < f > the average over realizations of a random quantity f . We denote the
uctuation by f 0 = f , < f >. We now use this notation to invert (B.6) approximately. Note that
the incoming energy ux amplitude ~ is independent of h. We also note from (4.26) that < K1 >= 0.
Averaging (B.6) over the realizations yields
(I + 2 < K2 >) < ~ > , < K10 ~0 > +2 < K20 ~0 >= (I + 2 < K2 >)~ + O(3 ): (B.7)
Subtracting (B.7) from (B.6) and keeping terms through O() we get an expression for the uctuating
part of the re ected wave amplitude
~0 = (K10 < ~ > K10 ~ ) + O(2 ): (B.8)
Finally we use (B.8) for ~0 in (B.7) and obtain
([I + 2 < K2 >] , 2 < K10 K10 >) < ~ > = ([I + 2 < K2 >] + 2 < K10 K10 >)~ + O(3 ): (B.9)
Dropping the error and noting that < K2 > and < K10 K10 > are scalars, we get for the mean re ected
amplitude
0 0
< ~ >= 1(1++2 <<KK2>>,+2 <<KK0 1KK0 1>>) ~ :
2 2

2 1 1

The factor < K2 > is not singular at grazing angles, while < K10 K10 > is singular in the Neumann
case. We can therefore divide both the numerator and denominator in (B.9) by 1 + 2 < K2 >, use
the approximation [1 + 2 < K2 >],1  1 , 2 < K2 > and obtain to order 3 ,
0 0
< ~ >= 1(1,+2 <<KK0 1KK0 1>>) ~ :
2
(B.10)
1 1

Upon dening Q =< K10 K10 >, (B.10) becomes (B.2) for the Neumann boundary condition. For the
Dirichlet condition the smoothing method is not necessary since the Born approximation is bounded
independently of the angle of incidence.
Using (B.8) for ~0 and (B.10) for < ~ > allows us to recompute the correlation function
< W n,1[~] >. By doing so we check that the recomputed scattering cross-section for the Neumann
boundary condition is (B.3). We can interpret the smoothing method as a way of renormalizing the
results of the Born approximation.
B.2 Extension to Interface Problems
We will now use the smoothing method to obtain nonsingular interface re ection and transmission
coecients, following unpublished calculations of John G. Watson. Continuity of pressure and
normal velocity give interfacial conditions of the form
0 1 0 1
B K ( ) 
11
K 12 (,)2 C ~ B K 11 (,)1 K 12 ()2 C
@ ,
1
A  = @ 21 A ~ : (B.11)
K ()1 1
21
,K 22(,)2,1 ,K (,)1,1 K 22()2,1
Here ~ = (~i ), ~ = (~i ), i = 1; 2, and we have used the following denitions:
K ij () = 1Z + K1ij + 2 K2ij + O(3 )
h i
K1ij () (p0 ) = h^ (p0 , k0 )mji (p0 ; k0 ) (p0 )dp0 (B.12)
h ij i
< K2 () > ( ) (k0 ) = 21 (mj1)2 Q(0) (k0 ):

23
Only the real part of K 1j has been written since the imaginary part does not contribute, and

mj2 (p0 ; k0 ) = i (k )j ,j p0
k :
j 2 0 0
mj1 = ij knj ;
 kn (k )
These expressions are obtained from (A.2).
We note that R0 = P ,1 V0 P where P is dened in (4.20). Therefore, if we multiply both sides of
(B.11) by P ,1 , we can write it in the form
(I + R1 + 2 R2 ) ~ = R0 (I , R1 + 2 R2 ) ~ ; (B.13)
where R1 and R2 are dened by using (B.12) in (B.11) and adopting matrix notation.
We can now use the smoothing method, as in the previous section. With the notation of the
previous section for average and uctuation, we have
~0 = ,[R10 < ~ > +R0 R10 ~ ] + O(2 ); (B.14)
and then up to terms of order 3
< I + 2 R2 >< ~ > ,2 < R10 R10 >< ~ > = R0 < I + 2 R2 > ~ + 2 < R10 R0 R10 > ~ : (B.15)
Again < R2 > is a local multiplication operator and we can multiply both sides of (B.15) by
(I + 2 R2 ),1 . This yields, to order 3
< ~ > = [I , 2 < R10 R10 >],1 [R0 + 2 < R10 R0R10 > +2 (R0 < R2 > , < R2 > R0)]~ (B.16)
uniformly in angles, including grazing angles.
The matrices appearing in (B.16) can be related to the quantities already computed in section
4.3 as follows. We deduce from (A.9) given in Appendix A that the operator R1 = R10 has the form
Z
(R10 g)(p0 ) = dk0 ih^ (p0 , k0 )N (p0 ; k0 )g(k0 ):
Therefore we have
(I , 2 < R10 R10 >) = I + 2 W
(R0 + 2 < R10 R0 R10 >) = R0 (I , 2 V );
where Z
V (p0 ) = 0 , 0 0 0 , 0 0 0 ^ 0 0
Z dk P (p )M (p ; k )P (k )M (k ; p )Q(p , k )
1 1

W (p0 ) = dk0 N (p0 ; k0)N (k0 ; p0 )Q^ (p0 , k0 ):

We rewrite expression (B.16) for the coherent re ection coecient in the form
< ~ > = [R~0 + 2 (R0 < R2 > , < R2 > R0 ) ] ~ + O(3 ); (B.17)
where
R~0 = (I + 2 W ),1 R0 (I , 2 V ): (B.18)
The dierential scattering cross-sections (4.22) also need to be recomputed. From (B.14) and
(B.17) we nd that
~0 = ,[R10 R~ 0 + R0 R10 ] + O(2 ):

24
 Born expansion Smoothing Method
(4.27) entry 2; 2 (4.27) entry 2; 2
Dirichlet B.C.
(4.28) entry 2; 2 (4.28) entry 2; 2
(4.27) entry 1; 1 (B.1)
Neumann B.C.
(4.28) entry 1; 1 (B.3)
(4.23) (B.17)
Interface Condition
(4.22) (B.20)
Table B.1: Reference to the explicit formulas for the Smoothing Method

Therefore replacing B by
B~ (p0 ; k0 ) = 12 (N (p0 ; k0 )R~ 0 (k0 ) + R0 (k0 )N (p0 ; k0 )); (B.19)
we get for the cross-sections the nonsingular result
ij (p0 ; k0 ) = = 4B~ 2 (p0 ; k0 )Q^ (p0 , k0 ):
ren (B.20)
ij
If we expand the inverse in (B.16) to order 3 we get the standard Born expansion
< ~ > = [R0 + 2 ((R0 < R2 > , < R2 > R0)+ < R10 R0R10 > + < R10 R10 > R0 )] ~:
The term < R10 R0 R10 > + < R10 R10 R0 > corresponds to R2;1 in (4.23). These coecients blow
up at grazing angles, as in the case of the Neumann boundary 0 condition1 We note that the term
0 1C
(R0 < R2 > , < R2 > R0 ), which is proportional to Q(0) B @ A and corresponds to the
,1 0
term R2;0 , does not vanish. Since it does not depend on the angle of incidence, this term is always
bounded and does not need to be renormalized.
To summarize, the re ection and transmission coecients given in Table 1.1, valid in the Born
expansion, are now replaced by those in Table B.1 given by the smoothing method.

References
[1] K.Aki and R.S.Wu, Multiple scattering and energy transfer of seismic waves { separation of
scattering eect from intrinsic attenuation {II. Application of the theory to Hindu Kush region,
in Seismic wave scattering and attenuation, vol. I (R.S.Wu and K.Aki, eds), 1988, 49-80
[2] G.Bal, G.Papanicolaou and L.Ryzhik, Diuse scattering from a random surface, preprint, 1998.
[3] Yu.Barabanenkov, A.Vinogradov, Yu.Kravtsov and V.Tatarskii, Application of the theory of
multiple scattering of waves to the derivation of the radiative transfer equation for a statistically
inhomogeneous medium, Radiozika, 15, 1972, 1852-1860. English translation 1420-1425.

25
 [4] Yu.Barabanenkov, Yu.Kravtsov, V.Ozrin and A.Saichev, Enhanced backscattering in optics,
Progress in Optics 29, 1991, 67-190.
[5] F.G.Bass and I.M.Fuks, Wave Scattering from Statistically Rough Surfaces, Pergamon, NY,
1978.
[6] I.M.Besieris and F.D.Tappert, Propagation of frequency modulated pulses in a randomly strat-
ied plasma, Jour. Math. Phys. 14, 1973, 704-707.
[7] I.M.Besieris, W.Kohler and H.Freese, A transport-theoretic analysis of pulse propagation
through ocean sediments, J. Acoust. Soc. Am., 72, 1982, 937-946.
[8] L.Brekhovskikh, Waves in Layered Media, Academic Press, 1980.
[9] S.Chandrasekhar, Radiative Transfer, Dover, New York, 1960.
[10] J.A.DeSanto, Theoretical methods in ocean acoustics, in Topics in current physics, Vol 8,
Chap.2, Springer, New York, 1979, 7-77.
[11] J.A.DeSanto and G.S. Brown, Analytical techniques for multiple scattering from rough surfaces,
in Progress in Optics XXIII (E. Wolf, ed.), Elsevier, Amsterdam, 1986, 1-62.
[12] A.Fannjiang and L.Ryzhik, Phase space transport theory for sound waves in random ows,
submitted to SIAM, J. App. Math, 1998.
[13] M.C.Fehler and H.Sato, Seismic Wave Propagation and Scattering in the Heterogeneous Earth,
AIP Press, 1998.
[14] P.Gerard, P.Markowich, N.Mauser and F.Poupaud, Homogenization limits and Wigner trans-
forms, Comm. Pure Appl. Math., 50, 1997, 323-380.
[15] J.J.Greet and M.Nieto{Vesperinas, Field theory for the generalized bidirectional re ectivity:
derivation of Helmholtz's reciprocity principle and Kirchho's law, submitted to Jour. Acoust.
Soc. Am., 1998.
[16] M.S.Howe, On the kinetic theory of wave propagation in random media, Phil. Trans. Roy. Soc.
Lond. 274, 1973, 523-549.
[17] A.Ishimaru, Wave propagation and scattering in random media, vol. II, Academic Press, New
York, 1978.
[18] L.D. Landau and E.M. Lifschitz, Fluid mechanics; Pergamon Press, NY, 1987.
[19] C.W.Law and K.Watson, Radiation transport along curved ray paths, Jour. Math. Phys., 11,
1970, 3125-3137.
[20] L.Margerin, M.Campillo and B. van Tiggelen, Radiative transfer and diusion of waves in a
layered medium: a new insight into coda Q, to appear in Geophys. Jour. Int., 1998.
[21] M.Nieto-Vesperinas, Scattering and diraction in physical optics, Wiley, New York, 1991.
[22] M.C.W.van Rossum and Th.M.Nieuwenhuizen, Multiple scattering of classical waves: from
microscopy to mesoscopy and diusion, to appear in Phys. Rev. E.

26
[23] L.Ryzhik, G.Papanicolaou and J.B. Keller, Transport Equations for Elastic and Other Waves
in Random Media, Wave Motion 24, 1996, 327-370.
[24] L.Ryzhik, G.Papanicolaou and J.B. Keller, Transport equations in a half space, Comm. PDE's,
22, 1997, 1869-1911.
[25] P.Sheng, Introduction to wave scattering, localization, and mesoscopic phenomena, Academic
Press, 1995.
[26] P.Stott, A transport equation for the multiple scattering of electromagnetic waves by a turbulent
plasma, Jour. Phys. A, 1, 1968, 675-689.
[27] B.A.van Tiggelen and R.Maynard, Reciprocity and coherent backscattering of light, in `Waves
in Complex Media', G. Papanicolaou editor, IMA Volume 96, Springer Verlag, 1997, 247-271.
[28] V.Twersky, On scattering and re ection of sound by rough surfaces, J. Acoust. Soc. Am. 29,
1957, 209-225.
[29] A.Voronovich, Wave Scattering from Rough Surfaces, Springer Verlag, 1991.
[30] J.G.Watson and J.B.Keller, Re ection, scattering, and absorption of acoustic waves by rough
surfaces, J. Acoust. Soc. Am. 74 (6), 1983, 1887-1894.
[31] J.G.Watson and J.B.Keller, Rough surface scattering via the smoothing method, J. Acoust.
Soc. Am. 75 (6), 1984, 1705-1708.
[32] K.Watson and J.L.Peacher, Doppler shift in frequency in the transport of electromagnetic waves
in an underdense plasma, Jour. Math. Phys11, 1970, 1496-1504.
[33] K.Watson, Multiple scattering of electromagnetic waves in an underdense plasma, Jour. Math.
Phys., 10, 1969, 688-702.

27