Multi-scaled diffusion-approximation. Applications to wave propagation in random media

MULTI-SCALED DIFFUSION-APPROXIMATION.

APPLICATIONS TO WAVE PROPAGATION

IN RANDOM MEDIA

JOSSELIN GARNIER

Abstract. In this paper a multi-scaled diusion-approximation theo- rem is proved so as to unify various applications in wave propagation in random media: transmission of optical modes through random planar waveguides time delay in scattering for the linear wave equation decay of the transmission coecient for large lengths with xed output and phase dierence in weakly nonlinear random media.

1. Introduction

Wave propagation in random media has become an extensively studied subject. In one-dimensional linear media with random inhomogeneities, lo- calization occurs, which means in particular that the transmitted intensity decays exponentially as a function of the size of the medium. This problem has been analyzed in detail by Carmona et al. (1990).

In our paper, we consider wave reection and transmission from a one- dimensional random slab. Several quantities characterize the reected wave here we focus on the reection coecient, the phase dierence and the time delay. The analysis puts into evidence the usual scales (see Knapp et al. (1989) and Papanicolaou (1988)): length of the slab, wavelength, amplitude and correlation radius of the random perturbations. We study the asymptotic behavior of the scattered wave in the framework introduced by Papanicolaou based on the separation of these scales. The uctuations of the random coecients are on a small scale so that we actually deal with diusion-approximation problems. However we consider here situa- tions where many scaled quantities play a role, so that we need to prove, then to use general multi-scaled diusion-approximation theorems. Indeed, the study of planar waveguides depends on the above quantities, but also on the thickness of the core. The time delay depends not only on the high carrier frequency of the wave packet, but also on its bandwidth. Finally the amplitude of the nonlinear term is an essential requirement for the study of the behavior of the transmittivity for the nonlinear wave equation. The action of nonlinearity seems to be opposite to that of disorder. Nonlinearity may change the dependence of the transmission coecient on the length, that still tends to zero as the size of the medium increases, but following a

URL address of the journal: http://www.emath.fr/ps/

Received by the journal March 18, 1996. Revised January 23, 1997. Accepted for publication February 14, 1997.

c Societe de Mathematiques Appliquees et Industrielles. Typeset by L^ATEX.

power law (see Desvillard et al. (1986) and Knapp et al. (1989)) instead of the exponential behavior observed in the linear case.

The organization of the paper is as follows: The rst section is devoted to the transmission of optical modes through a random metallic planar wave- guide. In Section 3 we rene some results of Faris et al. (1994) concerning the time delay in scattering for the linear wave equation with a random index of refraction. The last sections deal with the random nonlinear wave equation. In Section 4 we study the nonlinear xed output problem, i.e. the transmission problem with a xed outgoing intensity, which is simpler than the xed input problem, since there are unique values of input intensity and transmittivity for a given output intensity. In Section 5 we deal with the phase dierence of the reected wave. Finally the appendix is devoted to the statements and the proofs of multi-scaled diusion-approximation theorems.

2. Propagation in metallic planar waveguides

In this section we study propagation of optical modes in dielectric lms with thicknesses comparable to the wavelength. The main idea of a wave- guide is to guide a beam of light by employing a variation of the index of refraction in the transverse direction so as to cause the light to travel along a well-dened channel. The dependence of the index of refraction on the transverse direction may be continuous or discontinuous (we shall consider a particular case of the discontinuous situation), but the essential element is that the index of refraction is maximal in the channel along which one whishes to guide the light.

We shall consider the basic problem of TE (transversal electric) mode propagation in slab dielectric waveguides (see Collins (1960)). Indeed the basic features of the behavior of dielectric waveguide can be extracted from a planar model in which no variation exists in one direction (say

y

). Channel waveguides with axis

z

, in which the waveguide dimensions are nite in both the

x

and

y

directions, approach the behavior of a planar waveguide when one dimension is much larger than the other (see Groell (1968)). Even when this is not the case, most of the phenomena of interest are only modied in a simply quantitative way when going from a planar waveguide to a channel waveguide. This fact combined with the great mathematical simplication of the one-dimensional case has led us to consider here planar waveguides.

We assume that the slab waveguide has thickness 2

a

and is located in the region

x

^2;

aa

]. Its axis is

z

and the slab is innite in the

y

-direction. The slab is switched between two metallic slabs for

x > a

and for

x <

^;

a

. We shall restrict ourselves to the

y

-independent case and consider waves which depend only on

x

and

z

.

We begin by studying the properties of guided modes in a perfect wave- guide, whose core has a homogeneous index of refraction equal to

n

⁰. A mode of a dielectric waveguide is a monochromatic wave^E(

t

^r) =^E(^r)

e

^;i!t solution of the wave equation

^E(^r) +

k

²⁰

n

²(^r)^E(^r) = 0

(2.1)

where

k

⁰ =

!=c

is the vacuum wavenumber. The solution satises continuity conditions of the tangential components of the eld at the dielectric inter- faces. Limiting ourselves to waves with phase front normal to the waveguide axis

z

, we have^E(^r) =^E(

x

)

e

^iz.

We are looking for TE modes ^E = (0

E

y

0) with eld component

E

y =

E

(

x

)

e

^iz. The scalar eld

E

satises

@

²

E

@x

^{2 + (}

k

^2;

²)

E

= 0

(2.2) where

k

is the homogeneous wavenumber

k

=

n

⁰

k

⁰. In the metallic slabs

x > a

and

x <

^;

a

the electric eld is zero. Because of the need to match

E

y at

x

=^;

a

and

x

=

a

the eld

E

solution of (2.2) satises the boundary conditions

E

(^;

az

) =

E

(

az

) = 0. There exists solutions only for some values of

. Thus the metallic planar waveguide can only support a nite number of conned TE modes,

E

j(

xz

) =

R

j(

x

)

e

^ijz, where

R

j(

x

) =

8

>

>

<

>

>

:

cos

j x

2

a

if

j

is odd, sin

j x

2

a

if

j

is even, (2.3)

and

j satises the dispersion relation

_j²+ ²

j

²

4

a

^{2 =}

k

²

:

(2.4)

There exists

N

guided modes, where

N

is the integer that satises 2

ak

;1

N <

²

ak

:

^(2.5)

We shall assume that a monochromatic guided wave is incoming from the left through a perfect waveguide and has the form (

x <

0):

E

in(

xz

) =^XN

j⁼¹

R

j(

x

)

E

j(

z

)

E

j(

z

) =

B

j⁰

e

^ijz

(2.6) where

j is the positive solution of (2.4) and

B

j⁰is the decomposition of the incident wave

E

in on the

j

-th mode. This wave is scattered by a perturbed waveguide occupying the interval 0

L

^"], so that the total eld is consti- tuted of the sum of the incident wave (2.6) and the reected wave in the region

x <

0, and of the transmitted wave in the region

x > L

^", where the waveguide is unperturbed. We shall consider here that the wavelength of the incident wave and the thickness of the waveguide 2

a

are of order 1. We assume that the medium inside the waveguide is aected by small random inhomogeneities for

z

² 0

L

^"], so that its index of refraction admits the representation:

n

^"2(

z

) =

n

²⁰1 +

"m

(

z

"

^r)

(2.7)

where

"

is a small parameter which characterizes the amplitude of the ran- dom inhomogeneities. The random coecient

m

which describes the in- homogeneities is assumed to be an ergodic Markov process. More exactly, we consider that the process

m

has a unique invariant probability, under

which it is ergodic and that it satises the Fredholm alternative. As a con- sequence, its innitesimal generator admits an inverse on the subspace of functions centered under the invariant probability of

m

. In our paper we think at Markov processes on a compact space satisfying the Doeblin condi- tion, however some Markov diusion processes studied by Bouc et al. (1984) are suitable. We refer in particular to Kesten et al. (1979), Kushner (1984) and Papanicolaou et al. (1976) for sharp conditions. Throughout the pa- per expectations will be taken under the invariant probability of

m

and we denote:

c(

k

) =

Z

1

0

cos(2

ks

)^E

m

(0)

m

(

s

)]

ds:

(2.8) Three cases appear. If ^;1

r <

0 (resp.

r

= 0,

r >

0), the correlation radius of the inhomogeneities is much larger (resp. of the same order, much shorter) than the wavelength. The rst case corresponds to a very high frequency regime, and the third one to a low frequency regime. The case

r <

^;1 does not provide us with any asymptotic regime.

We shall consider a perturbed slab waveguide of length

L

^" =

L="

^2+r located in the region

z

²0

L

^"]. It will appear in the following that this is the judicious scale to put into evidence a macroscopic eect of the perturbations.

We shall see that the uctuations of the index of refraction induce a coupling between forward and backward modes when

r

0 and only aect the phases of the modes when

r <

0. We can now state our main result:

Proposition 2.1. The transmitted wave has the following form, for

z

L="

^2+r:

E

tr(

xz

) =^XN

j⁼¹

R

j(

x

)

E

j(

z L "

^2+r)

E

j(

z L "

^{2+r) =}

B

j(

L

"

^2+r)

e

^ijz

(2.9) 1. If ^;1

r <

0, then the processes ^;

B

j(

L="

^2+r)_L⁰,

j

= 1

::: N

converge weakly to

B

j⁰

e

^ij(L) as

"

^!0,

j(

L

) =

8

>

>

<

>

>

:

k

²

2

j

w

L, if^;1

< r <

0,

k

²

2

j

Z L

0

m

(

z

)

dz

, if

r

=^;1, (2.10) where

w

is a standard Brownian motion independent of

j

and

² =

c(0).

2. If

r

0, then the processes^;j

B

j(

L="

^2+r)^j2_L⁰,

j

= 1

::: N

converge weakly to independent Markov processes^Ij whose generators are

Lj = 1

j^I j2

;

@ @

^Ij + (1^;Ij)

@

²

@

^I_j²

!

(2.11)

where 1

j =

8

>

>

>

<

>

>

>

:

k

⁴

c(

j)

2

_j^{2 , if}

r

= 0,

k

⁴

c(0)

2

_j^{2 , if}

r >

0. (2.12)

In particular, the transmitted intensities decay exponentially with the length of the random waveguide:

Llim^!1 1

L

^lnEIj(

L

)] =^; 1

4

j

:

(2.13)

In the very high frequency regime, the perturbations only aect the phases of the optical modes by adding random phases on every mode. In the other regimes, we can observe an exponential localization of the modes, but the striking point is that each mode has its own localization length. The smaller the eective wavenumber

is, the less the corresponding mode can penetrate in the random waveguide.

Proof. Inside the perturbed slab we expand the total eld

E

in the form

E

(

xz

) =^XN

j⁼¹

R

j(

x

)

E

j(

z

)

E

j(

z

) =

A

j(

z

)

e

^;ijz+

B

j(

z

)

e

^ijz

(2.14) where

A

j and

B

j are respectively backward (going to the right) and forward (going tot the left) optical modes.

E

satises the boundary conditions at the dielectric interfaces

x

=^;

a

and

x

=

a

and the evolution equation (2.1) which writes:

N

X

j⁼¹

R

j(

x

)_"j(

zA

j

B

j) = 0

(2.15)

_"j(

zA

j

B

j) =

d

²

A

j

dz

²

e

^;ijz+

d

²

B

j

dz

²

e

^ijz;2

i

j

dA

j

dz e

^;ijz;

dB

j

dz e

^ijz

+

k

²

"m

(

z

"

^r)

A

j

e

^;ijz+

B

j

e

^ijz

:

Integrating (2.15) with respect to

R

l(

x

)

dx

, we can deduce from the orthog- onality of the family (

R

j)j⁼¹:::N that

_"j(

zA

j

B

j) = 0

for every

j

= 1

::: N

. (2.16)

The orthogonality of the modes is of great importance. It will insure that there exists only coupling between forward and backward modes, and not between

j

and

j

⁰ modes.

Let us study the equation (2.16). If the couple ( ~

A

j

B

^~j) is a solution, then the couple (

A

j

B

j) dened by

A

j(

z

) = ~

A

j(

z

) +

f

(

z

)

e

^ijz,

B

j(

z

) = ~

B

j(

z

)^;

f

(

z

)

e

^;ijz, is another solution of (2.16), whatever

f

is. Choosing

f

(

z

) =

i

2^j

dA^~j

dz

e

^;ijz+^d_dz^B~j

e

^ijz, we come to the conclusion that (

A

j

B

j) satises both (2.16) and the relation

dA

j

dz e

^;ijz+

dB

j

dz e

^{ijz = 0}

:

(2.17) We actually do not look for every solution of (2.16) since we aim at studying

E

j dened in (2.14). So we can restrict ourselves to (

A

j

B

j) which satises both (2.16) and (2.17). Finally, injecting (2.17) in (2.16), the normalized processes

A

_"j,

B

_"j,

j

= 1

::: N

given by

A

_"j(

z

) =

A

j(

z

"

^2+r)

B

_"j(

z

) =

B

j(

z

"

^2+r)

(2.18)

are solutions of

dz d

A

_"j(

z

)

B

_"j(

z

)

=

P

_"j(

z

)

A

_"j(

z

)

B

_"j(

z

)

(2.19)

P

_"j(

z

) =

ik

²

2

j

"

^1+r

m

(

z

"

^2+2r)

0

@

;1 ^;

e

^i"22+rjz

e

^;i"22+rjz 1

1

A

:

(2.20) There is no wave entering the perturbed slab at

z

=

L

(in the normalized scale) and the incident wave at

z

= 0 has the form (2.6), so

A

_"j and

B

_"j satisfy the boundary conditions

A

_"j(

L

) = 0

B

_"j(0) =

B

j⁰

:

(2.21)

We aim at proving an asymptotic theorem for

B

_"j(

L

) as

"

goes to 0. Instead of working with

A

_"j and

B

_"j, we shall use the propagator

Y

_"j, i.e. the matrix which satises:

A

_"j(

L

)

B

_"j(

L

)

=

Y

_"j(

L

)

A

_"j(0)

B

_"j(0)

:

(2.22)

The matrix

Y

_"jis solution of the linear dierential equation:

dY

_"j

dz

⁽

z

) =

P

_"j(

z

)

Y

_"j(

z

)

Y

_"j(0) =

I

d

:

(2.23)

If (

a

_"j

b

_"j) is a solution of (2.19) with the initial conditions

a

_"j(0) = 1

b

_"j(0) = 0

(2.24)

then it can be readily checked that (

b

_"j

a

_"j) is another solution of (2.19) linearly independent of (

a

_"j

b

_"j), so we can write

Y

_"j(

z

) =

a

_"j(

z

)

b

_"j(

z

)

b

_"j(

z

)

a

_"j(

z

)

:

(2.25)

From

0

B

_"j(

L

)

=

Y

_"j(

L

)

A

j"(0)

B

j⁰

(2.26) we can deduce that

A

_"j(0) =^;

b

_"j(

L

)

B

j⁰

a

_"j(

L

)

B

_"j(

L

) =

B

j⁰

a

_"j(

L

)

:

(2.27)

Since the matrix

P

_"j has trace zero, the determinant of the matrix

Y

_"j is constant, i.e. ^j

a

_"j(

z

)^j2;j

b

_"j(

z

)^j2= 1, so that we get the energy conservation relation:

j

A

_"j(0)^j2+^j

B

_"j(

L

)^j2=^j

B

j^0j2

(2.28)

which means that the intensity of each incident mode has split into a trans- mitted intensity ^j

B

_"j(

L

)^j2 and a reected intensity^j

A

_"j(0)^j2.

So we have transformed the boundary value problem (2.19), (2.21) into an initial value problem (2.19), (2.24). In order to be allowed to apply the diusion-approximation theorems, we have to take care to consider sepa- rately the real and imaginary parts of each coecient

a

_"j and

b

_"j, so that we actually deal with a system with 4

N

linear dierential equations. Denoting

X

^4"_j⁺¹ = Re(

a

_"j),

X

^4"_j⁺² = Im(

a

_"j),

X

^4"_j⁺³ = Re(

b

_"j) and

X

^4"_j⁺⁴ = Im(

b

_"j),

j

= 1

::: N

, the ^R4N-valued process

X

^" satises the linear dierential equation

dX

^"(

z

)

dz

^{= 1}

"

^1+r

F

m

(

z

"

^2+2r)

z "

^2+r

X

^"(

z

)

(2.29)

with the initial conditions

X

^4"_j⁺_j⁰(0) = 1 if

j

⁰ = 1,

X

^4"_j⁺_j⁰(0) = 0 if

j

⁰ = 2

3

4, where

F

(

mh

) =_Nj⁼¹

k

²

m

2

j

0

B

B

@

0 1 sin(2

j

h

) cos(2

j

h

)

;1 0 ^;cos(2

j

h

) sin(2

j

h

) sin(2

j

h

) ^;cos(2

j

h

) 0 ^;1 cos(2

j

h

) sin(2

j

h

) 1 0

1

C

C

A

Then we transform

"

^{1+r 7!}

"

, so that the system writes now as an usual diusion-approximation problem. The application of Theorem 2-7 Papani- colaou et al. (1976) (resp. Theorem 2-8 Papanicolaou et al. (1976)) provides us with the desired result for the case

r

= 0 (resp.

r >

0). The case

r <

0 can be deduced from Theorem 6.2 in the appendix.

3. Time delay

We aim at studying the time delay of a wave packet scattered by a one- dimensional random medium. The time delay is the dierence between times spent by the wave packet in the perturbed region and in the homogeneous space. For instance, a positive time delay means that the wave packet has spent more time in the scattering region than the one it would have spent in homogeneous space.

We consider a wave packet reected by a one-dimensional random medium located in the region 0

L

]. The wave packet is assumed to satisfy the linear wave equation:

n

^"2(

x

)

c

²

@

²

E

@t

^{2 =}

@

²

E

@x

²

(3.1)

where the index of refraction

n

^" admits the representation

n

^"2(

x

) =

n

²⁰(1 +

m

^"(

x

))

m

^"(

x

) =

"m

(

x

"

²⁾

:

(3.2)

m

is an ergodic Markov process, which describes the perturbations of the index of refraction. Their amplitude is of order

"

and their correlation radius is of order

"

², which is small compared to the length

L

of the slab.

We consider the matched medium boundary condition. It is assumed that a wave packet is incident on the random slab 0

L

] from a homogeneous medium with index

n

⁰ occupying

x <

0:

u

i(

tx

) =

Z

1

0

e

^i(kx;ckt=n0)

u

^i(

k

)

dk

(3.3)

where ^

u

i is the spectrum of the incident wave packet. The boundary condi- tion at

x

=

L

is for termination of the random slab by a uniform medium.

We will have total reection for large length

L

because the wave cannot penetrate to innite depth. Indeed the reection coecient converges expo- nentially fast to 1 which follows from Furstenberg's theorem (see Carmona (1985)). So it is convenient to analyze the problem with a totally reecting

termination at

x

=

L

, so that the number of degrees of freedom is reduced by one. Thus the wave packet is scattered and the reected wave has the following form for large

t

:

u

_"r(

tx

) =

Z

1

0

R

^"(

k

)

e

^{i(;kx;ckt=n0)}

u

^i(

k

)

dk

(3.4)

where

R

^"(

k

) is the reection coecient associated to the homogeneous wave- number

k

. Because of the reecting boundary condition, the modulus of

R

^"(

k

) is equal to 1 so that

R

^"(

k

) =

e

^{i "(k)+i 0(k)}. The phase

⁰(

k

) = 2

kL

+

is related to the homogeneous part of the index of refraction. If the slab 0

L

] were occupied by a homogeneous medium, the reected wave would be exactly:

u

r⁰(

tx

) =

Z

1

0

e

^{i(;kx;ckt=n0)}

e

^{i 0(k)}

u

^i(

k

)

dk:

(3.5)

The phase

^"(

k

) characterizes the inuence of the uctuations of the index of refraction onto the reected wave. In particular the time delay is given by

T

^"(

u

i) =

n

⁰

c

Z

1

0

^"(

k

)^j

u

^i^j2(

k

)

dk

(3.6)

where

^"(

k

) := ^d _dk^"(

k

). A rigorous proof may be derived from Jauch et al.

(1972), but we can heuristically put into evidence (3.6) (cf Faris et al.

(1994)). Assume that ^

u

i is concentrated near some wavenumber

k

c. By expanding

R

^"(

k

) about

k

c in (3.4), we nd that the reected wave is, up to a multiplicative constant:

u

_"r(

tx

)^'

Z

1

0

e

^{i(;kx;ck(t;n0"(kc)=c)=n0)}

e

^{i 0(k)}

u

^i(

k

)

dk:

(3.7) By such a way we can understand why

n

⁰

^"(

k

c)

=c

represents the time delay at frequency

k

c.

Faris et al. (1994) studied the one-dimensional distribution of

^"(

k

) for some

k

and considered the limit case

"

^! 0, and then

k

^{! 1}. We rst prove that we can deal with simultaneous limits when

k

is of order

"

^;2, i.e.

when the wavelength and the correlation radius of the perturbations are of the same order.

Proposition 3.1. The process (

^"(

k="

²

L

))L⁰ converges in distribution in ^C0(0

¹)

^R) to the inhomogeneous Markov process (

(

L

))L⁰ whose in- nitesimal generator is:

Lk = 12

c(

k

)

k

²(

+ 2

L

)²

@

²

@

²

(3.8)

where

c(

k

) has been dened by (2.8). In particular

(

L

) is a centered process, whose variance is:

E

(

L

)²] = 8

c(

k

)²

k

⁴

e

^c(k)k2L;1^;

c(

k

)

k

²

L

^;1₂^;

c(

k

)

k

²

L

²

:

(3.9) Its distribution can be represented as follows:

(

L

) =

X

(

L

)^;2

L X

(

L

) = 2

c(

k

)

k

²

Z ^c(k⁾k²L

0

e

^ws;12s

ds

(3.10)

where

w

is a standard Brownian motion.

Proof. Following Papanicolaou (1988), we nd that the phase

^" satises:

d

^"

dL

⁼

k

^"(1^;cos(

^"+ 2

k

^"

L

))

m

^"(

L

) ^"(0) = 0

:

Dierentiating with respect to the wavenumber yields the equation which governs the evolution of

^":

d

^"

dL

^{= (1;cos(}

^"+ 2

k

^"

L

))

m

^"(

L

) +

k

^"sin(

^"+ 2

k

^"

L

)(

^"+ 2

L

)

m

^"(

L

)

:

Introducing the dierent scales in the evolution equation, the process (

^"

^") satises, up to negligible terms of order

"

:

8

>

>

<

>

>

:

d

^"

dL

⁼

k

"

1^;cos

^"+ 2

kL

"

²

m

(

L

"

²⁾

^"(0) =

d

^"

dL

⁼

k

"

^sin

^"+ 2

kL

"

²

(

^"+ 2

L

)

m

(

L

"

^{2) "(0) = 0}

:

The application of Theorem 2-7 Papanicolaou et al. (1976) completes the proof.

The above analysis leads to the asymptotic distribution of

^"(

k

) for a xed wavenumber. However it is necessary to nd the joint distribution of

^"(

k

) for several wavenumbers

k

in order to study the behavior of the time

delay

T

^"(

u

i) for an incident wave packet

u

i. For instance the estimation of the

p

-th moment of

T

^"(

u

i) requires the computation of the joint distribution of the

p

-uplet (

^"(

k

¹)

:::

^"(

k

p)), since

E

T

^"(

u

i)^p] =

n

^p0

c

^p

Z

E

^"(

k

¹)

:::

^"(

k

p)]^j

u

^i^j2(

k

¹)

:::

^j

u

^i^j2(

k

p)

dk

¹

:::dk

p

:

We now state the main result of this section about the convergence of the time delay. We assume that an incident wave (3.3) with a high carrier frequency

k

c

="

² and a bandwidth of order

"

^a;2 is incoming from the right.

Proposition 3.2. Let us assume that the Fourier transform of the incident wave ^

u

_"i has the form

u

^_"i(

k

) = ^

"

^a

k

^;

k

c

"

²

(3.11)

where ^

is a continuous function with compact support.

If 0

a <

2, then the time delay (

T

^"(

u

_"i)(

L

))L⁰ converges weakly to the process 0.

If

a

= 2, then (

T

^"(

u

_"i)(

L

))L⁰ converges weakly to a centered and con- tinuous process with variance:

E

T

²(

L

)] =

n

²⁰

c

^{2 8}

e

^c(kc)k2cL

c(

k

c)²

k

_c⁴

L

Z

j

^(

k

)^j4

dk

+ _L

o

!1

e

^c(kc)k2cL

c(

k

c)

k

²_c

L

!

:

(3.12)

If

a >

2, then (

"

^a;2

T

^"(

u

_"i)(

L

))L⁰ converges weakly to a centered and continuous process with variance:

E

T

²(

L

)] =

n

²⁰

c

²⁸

e

^c(kc)k2cL

c(

k

c)²

k

_c⁴

Z

j

^(

k

)^j2

dk

²+ _L

o

!1

e

^c(kc)kc2L

:

(3.13) Proof. We assume that the support of ^

is contained in (^;

NN

). We denote

by

^"(

Lh

) the process

^"(

Lk

^") ^j_k^"=(_k^c+_"^a_h⁾_="². From the expression (3.6)

of the time delay, if we prove that

^"(

Lh

) converges weakly as a process in

C

0(0

¹)

L

²(^;

NN

)), then it will follow that

T

^"(

u

_"i) converges weakly in

C

0(0

¹)

^R).

Step 1. Convergence of the finite-dimensional distributions

of

^". First we deal with the two-dimensional distributions of the pro-

cess

^". We aim at nding the asymptotic joint distribution of the couple

(

^1"(

L

)

^2"(

L

))L⁰, where

_"i =

^"(

Lk

_"i) and

k

_"i are wavenumbers given by

k

_"i = (

k

c+

"

^a

h

i)

="

². If we denote

(

kL

) ^jk⁼k^i" by

_"i then we nd that

(

^"1(

L

)

^"2(

L

)

^1"(

L

)

^2"(

L

))L⁰ satises:

8

>

>

>

>

>

>

>

>

>

>

<

>

>

>

>

>

>

>

>

>

>

:

d

^"1

dL

⁼

k

c+

"

^a

h

¹

"

1^;cos

^1"+ 2

k

c

L

"

^{2 + 2}

h

¹

L

"

^2;a

m

(

L

"

²⁾

^1"(0) = 0

d

^"2

dL

⁼

k

c+

"

^a

h

²

"

1^;cos

^2"+ 2

k

c

L

"

^{2 + 2}

h

²

L

"

^2;a

m

(

L

"

²⁾

^2"(0) = 0

d

^1"

dL

⁼

k

c+

"

^a

h

¹

"

^sin

^1"+ 2

k

c

L

"

^{2 + 2}

h

¹

L

"

^2;a

(

^1"+ 2

L

)

m

(

L

"

^{2) 1"(0) = 0}

d

^2"

dL

⁼

k

c+

"

^a

h

²

"

^sin

^2"+ 2

k

c

L

"

^{2 + 2}

h

²

L

"

^2;a

(

^2"+ 2

L

)

m

(

L

"

^{2) 2"(0) = 0}

:

If 0

< a <

2, then we get by applying Theorem 6.1 in the appendix in the case

A

= 2

k

c,

B

¹ = 1 and

B

j = 0,

j

2 that the process (

^1"

^2") converges to (

¹

²), where

¹ and

² are independent Markov process with the same generator^Lk^c given by (3.8).

If

a

= 0, then by the standard Theorem 2-7 Papanicolaou et al. (1976) the process (

^1"

^2") converges to (

¹

²), where

¹ and

² are independent Markov process with generators ^Lk^c+h¹ and ^Lk^c+h² respectively.

If

a >

2, then (

^1"

^2") converges to (

), where

is a Markov process with generator ^Lk^c.

If

a

= 2, then (

^"1(

L

)

^2"(

L

)

^1"(

L

)

^2"(

L

))L⁰converges in distribution in

C

0(0

¹)

^R4) to the Markov process (

¹(

L

)

²(

L

)

¹(

L

)

²(

L

))L⁰ whose innitesimal generator is (

h

=

h

^2;

h

¹):

L=

c(0)

k

²_c

@

@

^{1 +}

@

@

²

2

+12

c(

k

c)

k

²_c

@

²

@

^{21 +}

@

²

@

^{22 + 2cos(}

^2;

¹+ 2

hL

)

@

²

@

¹

²

+12

c(

k

c)

k

²_c

(

¹+ 2

L

)²

@

²

@

^{12 + (}

²+ 2

L

)²

@

²

@

²²

+2(

¹+ 2

L

)(

²+ 2

L

)cos(

^2;

¹+ 2

hL

)

@

²

@

¹

²

:

Papanicolaou (1988) proved the convergence of the couple (

^"1(

L

)

^2"(

L

))L⁰, and also that the process (

²(

L

)^;

¹(

L

) + 2

hL

mod 2 )L⁰ admits an in- variant probability measure

m

h which satises:

hcos

ⁱ_m^h = 2^j

h

^j

c(

k

c)

k

²_c

Z

1

0

exp

;

2^j

h

^j

z

c(

k

c)

k

²_c

z

²

z

²+ 4

dz:

(3.14) By a straightforward generalization of the above results we can show that the nite-dimensional distributions (

^"(

Lh

¹)

:::

^"(

Lh

n))L⁰ converge in ^C0(0

¹)

^Rn).

It remains to prove the tightness of

^"(

Lh

) in ^C0(0

¹)

L

²(^;

NN

)).

We will only deal with the most delicate case

a

= 2. In fact we shall show

that

^"(

Lh

) is tight in ^D(0

¹)

L

²(^;

NN

)) and we shall conclude by

checking that the weak limit is unique and belongs to^C0(0

¹)

L

²(^;

NN

)).

We begin by stating some standard tightness criteria (see Metivier (1984)).

Lemma 3.3. Let (

Ed

) be a metric space, and

X

^" a process with paths in

D(0

¹)

E

). If for every

t

in a dense subset of^R+the family(

X

^"(

t

))"^2(01]

is tight in

E

and

X

^" satises the Aldous property:

A

] For any

M >

0

>

0

>

0

there exists

>

0 such that limsup_"

!0

sup_T sup

0< <^P(^jj

X

^"(

T

+

)^;

X

^"(

T

)^jj

>

)

<

where

T

is a stopping time and sup_T is the sup over all such

T

M

then the family (

X

^")_"^2(01]is tight in ^D(0

¹)

E

).

Lemma 3.4. Let

H

be a Hilbert space and

H

n be an increasing sequence of nite-dimensional spaces in

H

such that, for any

h

²

H

,limn^!1 Hⁿ

h

=

h

.

Let

Y

^" be a

H

-valued process.

Y

^" is tight if and only if for any

>

0 and

>

0, there exists

and a

subspace

H

such that

"²⁽⁰sup^1]P(^jj

Y

^"jj

)

and sup

"^2(01]P(

d

(

Y

^"

H

)

>

)

:

(3.15)

Here we will take

H

=

L

²(^;

NN

) and

H

n the set of all the simple functions of the form

h

(

t

) = ^NnX;1

k^=;Nn

k1It²k=nk⁺¹=n⁾

(

k)k^=;Nn:::Nn^;12R2Nn

:

If

g

^2C0(^;

NN

]

^R), we denote by n(

g

) the simple function in

H

ndened by

g

(^nt_n^]).

Step 2. (

^"(

Lh

))h^2(;NN⁾ is tight in

L

²(^;

NN

) for every

L

. By Lemma 3.4 it is sucient to check that (

^"(

Lh

))_h^2(;_NN⁾fullls (3.15). The following criteria (3.16, 3.17) insures that (

^"(

Lh

))_h^2(;_NN⁾fullls (3.15):

lim^!0limsup_"

!0 E

"

sup

jh^;h^0jjh^jN^jh^0jN^j

^"(

Lh

)^;

^"(

Lh

⁰)^j2

#

= 0

(3.16) limsup_"

!0

sup

jh^jN^E

j

^"(

Lh

)^j2

<

¹

:

(3.17)