on his 70th birthday
DETERMINISTIC EQUIVALENT OF A RANDOM RIEMANNIAN METRIC DEFINED BY GEODESICS
OANA IOSIFESCU and GÉRARD MICHAILLE
We give a deterministic equivalent in probability of a rapidly oscillating random Riemannian metric. This equivalent metric is obtained by geodesics dened in the space of functions of bounded variation via suitable ergodic theorems.
AMS 2000 Subject Classication: Primary 60F10;
Secondary 34F05, 60F15, 73B27.
Key words: stochastic homogenization, Poisson process, Gamma-convergence, function of bounded variation, ergodic theorem.
1. INTRODUCTION
LetΩbe the set of locally nite sequences(ωi)i∈N inRN andMthe set of all countable and locally nite sums of Dirac measures, equipped with their standard σ-algebra. Given a positive numberλand ε >0 intended to tend to 0, we consider the Poisson point process
Ω→ M, ω7→ Nε(ω,·) =X
i∈N
δωi
with intensityλε=λ/εNLN, whereLN denotes the Lebesgue measure onRN, and dene the heterogeneous random Riemannian metric on RN by
dsε(ω, x, ξ) :=
( ds2(ξ) if x∈S
i∈NB(ωi, εr), ds1(ξ) otherwise.
For every iin N, B(ωi, εr) is the open ball in RN with radius εr centered at ωi and dsk(ξ) := (P
i,jaki,jξiξj)1/2 for all ξ ∈ RN with (aki,j)i=1,...,N, j=1,...,N, k= 1,2, two symmetric positive matrices (see Section 3 for more details).
The spaceRN is then equipped with the highly random oscillating geo- desic metric (a, b)7→dε(ω)(a, b) dened for any pointsa, b inRN by
inf Z 1
0
dsε(ω, γ(t), γ0(t)) dt:γ ∈W1,1((0,1),RN), γ(0) =a, γ(1) =b
.
REV. ROUMAINE MATH. PURES APPL., 53 (2008), 56, 479498
Classically, there exists a probability measurePon Ωsatisfying P([Nε(·, A) =k]) =λkεLN(A)k exp(−λeLN(A))
k!
for every bounded Borel set A in RN, so that the expectation of the random variable Nε(·, A) isE Nε(·, A)
=λεLN(A). Consequently, dε is statistically like a geodesic distance in a Riemaniann space with small balls as hetero- geneities, and whose number is of orderλε=λ/εN per unit of volume. It will be noted that the balls can interpenetrate.
It is easy to see that the random distanceω 7→dε(ω)(a, b) has the same law as (i.e., is statistically equivalent to) the random distance given by
inf Z 1
0
ds(ω,γ(t)
ε , γ0(t)) dt:γ ∈W1,1((0,1),RN), γ(0) =a, γ(1) =b
,
where ds(ω, x, ξ) :=ds1(ξ) + ds2(ξ)−ds1(ξ)
min(1,N(ω, B(x, r)) and N is the Poisson point process
Ω→ M, ω7→ N(ω,·) =X
i∈N
δωi
with intensity λLN. The scaling 1/ε on the space variable, well known in homogenization theory, is more adapted to a mathematical processing. In this paper, we look at this distance that we still denote by dε(ω). We would like to stress that we can treat the problem, without changing proofs, for random Riemannian metrics in general stationary ergodic media.
Since we only have a statistical knowledge of dε(ω) which is moreover highly oscillating, for obtaining information on dε(ω) it is natural to identify its a.s. variational limit as ε → 0 and take advantage of the stationarity of the Poisson point process for proving the deterministic nature of the limit. In fact, we have been unable to fully carry out this program: we only establish the variational convergence in law and in probability of dε to a deterministic distanced0. Our proof is based on the following compactness result established in Section 3: there exists a subsequence of the sequence of integral functionals Fε deningdε, which almost surely converges to a deterministic functional F0
in the sense of Gamma-convergence. We deduce the convergence in law of any sequence of relaxed functionalsF¯εto the Dirac probability measure δF0. More precisely, we know that Gamma-convergence induces a metric on the set SCof all semicontinous functionsG:BV(0,1),RN →R∪{+∞}. Then, considering the probability images PFε−1 and PF0−1 =δF0 of
F¯ε, F0 : Ω→ SC
we obtain PFε−1 * δF0. Denoting bydΓthe metric induced inSC by Gamma- convergence, we obtain as a corollary the convergence in probability of all
sequences, (Fε)ε>0, i.e.,
P{ω:dΓ( ¯Fε(ω), F0)> η} →0asε→0.
Finally, using the continuity of the map G 7→ minG in the metric space (SC, dΓ), we deduce our main result, the convergence in probability
P{ω:|dε(ω)(a, b)−d0(a, b)|> η} →0 asε→0.
We would like to stress that we could treat this analysis for random Riemannian metrics in general stationary ergodic media, without changing proofs.
This problem has already been studied in the deterministic case when the Riemannian metric is periodically distributed. Precisely, it has been solved in [2] when p > 1 and in [4] when p = 1 by using an integral representation theorem for functionals dened inW1,p((0,1),RN) or BV((0,1),RN).
The paper is organized as follows. The next section contains a brief summary of some results related to the ergodic theory of random subadditive processes, to Gamma-convergence and to the BV((0,1),RN) space. Section 3 is devoted to the study of two suitable random processes and to the deni- tion of the deterministic limit distance. In Subsection 4.1 we establish our compactness result using both random processes. Finally, in the last section, we conclude as to the convergence in probability of all sequences of random geodesic metrics to the deterministic metric dened in Section 3.
2. BASIC NOTIONS
2.1. Basic notions on Gamma-convergence
Let(X, d)be a metric space and{(Fn)n, F, n→+∞}functionals map- pingX intoR∪ {+∞}. The notion of convergence below is equivalent to con- vergence of the epigraph ofFnto the epigraph ofF in the Kuratovski-Painlevé sense and has been rst introduced in [11]. We give the denition in the frame- work of metric spaces. Indeed, in the sequel, (X, d) will be L1((0,1),RN) or BV((0,1),RN) equipped with the strong convergence of L1((0,1),RN). For details, we refer the reader to [7] and [9].
Denition 1. A sequence(Fn)n∈N Gamma-converges toF atx in X i i) there exists a sequence (xn)n∈N converging to xinX such that
F(x)≥lim sup
n→+∞ Fn(xn),
ii) for every sequence(yn)n∈N converging to xinX we have F(x)≤lim inf
n→+∞Fn(yn).
When this property holds for everyxinX, we say that(Fn)nGamma-converges to F in (X, d) and writeF = Γ limFn.
It is straightforward to show that this convergence is equivalent to Γ lim sup
n→+∞
Fn(x)≤F(x)≤Γ lim inf
n→+∞Fn(x) for any x∈X, where
Γ lim sup
n→+∞ Fn(x) := min n
lim sup
n→+∞ Fn(xn) :xn→x o
,
Γ lim inf
n→+∞Fn(xn) := minn lim inf
n→+∞Fn(xn) :xn→xo .
The main interest of this concept lies in the following variational property.
Proposition 1. Assume that (Fn) Gamma-converges to F and let xn∈ X be such that
Fn(xn)≤inf{ Fn(x) :x∈X }+εn,
where εn > 0, εn → 0. Assume furthermore that {xn, n ∈ N} is relatively compact. Then any cluster point x of {xn, n∈N} is a minimizer of F and
n→+∞lim inf{Fn(x) :x∈X}=F(x).
2.2. Basic notions on BV-functions
The setBV(I,RN)is the space of all functionsu∈L1(I,RN)such that the distributional derivative u0 iof u is a bounded Borel measure with values in RN. The nite total variation of u0 is denoted byR
I|u0|and BV(I,RN) is equipped with the norm
|u|BV(I,RN)=|u|L1((0,1),RN)+ Z
I
|u0|.
The unit ball of BV(I,RN) is weakly sequentially compact in the follow- ing sense: if |un|BV(I,RN) ≤ 1, there exist a subsequence (unk)k∈N and u in BV(I,RN) with |u|BV(I,RN) ≤ 1 such that unk weakly converges to u in BV(I,RN), that is,
( unk →u strongly in L1(I,RN),
u0nk * u0 in the sense of weak convergence of bounded measures.
In the sequel, u0a and u0s will denote the density of the regular part of the measure u0 and its singular part in the Lebesgue decomposition, i.e., u0 = u0aL1bI+u0s. The functionuais actually the approximate dierential ofuinI. More precisely,
Proposition 2. Letu be any element of BV(I,RN). Then
ρ→0lim 1 ρ
Z
Iρ(t0)
|u(t)−u(t0)−(t−t0) u0a(t0)|
|t−t0| dt= 0
for almost all t0 in I, where Iρ(t0) denotes the interval centered at t0 of length 2ρ.
For a general study ofBV-functions we refer the reader to [6].
2.3. Basic notions on subadditive random processes
We consider a probability space (Ω,T,P) and a group (τs)s∈R of P- preserving transformations of (Ω,T). The group (τs)s∈R is said to be ergodic if every set E inT such thatτsE=E,s∈R, has probability equal to0 or 1. A sucient condition to ensure ergodicity of (τs)s∈R is the mixing condition:
for every E andF inT,
|s|→+∞lim P(τsE∩F) =P(E)P(F) which expresses an asymptotic independence.
LetEdenote the expectation operator andJ the set of half open intervals [a, b) of R. A random subadditive process with respect to (τs)s∈R is a set function S :J →L1(Ω,T,P) satisfying the conditions below:
(i) for everyI ∈ J such that there exists a nite family(Ij)j∈J of pairwise disjoint intervals in J withI =S
j∈JIj, we haveSI(·)≤P
j∈JSIj(·);
(ii)∀I ∈ J, ∀s∈R,Ss+I=SI◦τs.
The subadditive ergodic Theorem 1 below is due to Ackoglu-Krengel. It was rst applied in the context of stochastic homogenization in [10]. In order to dene the limit metric, it will be applied in the next section with d= 1.
Theorem 1. LetS be a subadditive process with respect to (τs)z∈R that is assumed to be ergodic. Assume that
inf Z
Ω
SI(ω)
|I| P(dω) :|I| 6= 0
>−∞.
Let I be any xed interval in I. Then, almost surely,
t→+∞lim StI(ω)
|tI| = inf
m∈N∗
ES[0,m[d
md
.
For a proof see [3] and, for some extensions, [12], [13].
3. IDENTIFYING THE LIMIT METRIC
Let us give more details on the probabilistic setting. We consider two symmetric N×N-matrices(aki,j)i=1,...,N, j=1,...,N, satisfying
ν|ξ|2≤X
i,j
aki,jξiξj ≤Λ|ξ|2, k= 1,2,
for every ξ ∈ RN, where ν, Λ are two given positive constants. Denote by ds1 and ds2 the two corresponding homogenous Riemannian metrics. On the other hand, we consider a Poisson point process ω 7→ N(ω,·) with intensity λLN,λ >0, from the probability space(Ω,T, P)intoNB(RN)equipped with the standard productσ-algebra, which classically satises:
i) for every bounded Borel setA inRN, we have N(ω, A) =X
i∈N
δωi(A);
ii) for every nite family(Ai)i∈I of pairwise disjoint bounded Borel sets inRN, the random variablesN(·, Ai), i∈I, are independent;
iii) for every bounded Borel setA and everyk∈N, P([N(·, A) =k]) =λkLN(A)k exp(−λLN(A))
k! .
Note that N(ω, A) = card(A∩Ω) and that E(N(·, A)) =λLN(A) for every bounded Borel set of RN.
For a givenr >0, we dene a random Riemannian metric by setting ds(ω, x, ξ) =
( ds2 if x∈S
i∈NB(ωi, r), ds1 otherwise.
It may be interesting to note that
(1) ds(ω, x, ξ) :=ds1(ξ) + ds2(ξ)−ds1(ξ)
min(1,N(ω, B(x, r)).
For all ω in Ω, ds takes on values in the class Fν,Λ of the functions f fromRN×RN intoRthat are convex with respect to the second variable and satisfy ν|ξ| ≤ f(x, ξ) ≤Λ(1 +|ξ|). This class will be equipped with the trace denoted by T˜ of the Borel σ-eld of RRN×RN, that is the smallest σ-algebra on Fν,Λ such that all evaluation maps f 7→ f(x, ξ), x ∈ RN, ξ ∈ RN, are measurable. According to (1), it is readily seen that the map ds : Ω→ Fν,Λ, ω 7→ ω˜ := ds(ω) is (T,T˜)-measurable. Throughout the paper we will reason in the probability space( ˜Ω,T˜,P) := (F˜ ν,Λ,T˜,P˜)whereP˜ is the law ofds, i.e.
the probability image of Pby the measurable map ds.
For every xed point a in RN we introduce the ergodic group (Ts)s∈R
of P˜-preserving transformations by setting Tsω(x, ξ) = ˜˜ ω(x+sa, ξ)and dene the subadditive random process with respect to (Ts)s∈R by setting
SI(˜ω, a) = inf Z
I
˜
ω(γ(t), γ0(t)) dt:γ ∈W1,1(I,◦ RN), γ(α) =αa, γ(β) =βa
.
for every open half interval I = [α, β) in R. Conditions (i) and (ii) dening a subadditive random process are easily checked as well as measurability and integrability ofSI(·, a). Ergodicity of(Ts)s∈Ris a straightforward consequence of condition (ii) in the denition of the Poisson point process which ensure the mixing condition introduced in Subsection 2.3. In what follows, ε denotes a sequence of positive numbers going to 0. Applying Theorem 1, we obtain
Theorem 2. Let I be an arbitrary interval of R. Then there exists a subset Ω˜0 of full probability such that for all ω˜ in Ω˜0 and for all a ∈ QN the limit
ds0(a) = lim
ε→0
S1 εI(˜ω, a)
|1εI| = lim
n→+∞
S]0,n[(˜ω, a)
n = inf
n∈N∗
ES]0,n[(·, a) n
exists.
Proof. For every xedainQN, using Theorem 1 we deduce the existence of Ω˜0a of full probability such that for everyω˜ ∈Ω˜0a the limit
ε→0lim S1
εI(˜ω, a)
|1εI| = lim
n→+∞
S]0,n[(˜ω, a) n exists, and it is sucient to set Ω˜0=T
a∈QNΩ˜0a.
We are going to establish the existence of the limit for every a ∈ RN and every ω˜ ∈ Ω˜0. This will be an easy consequence of the equilipschitzian property below.
Proposition 3. The above normalized subadditive random process satis- es
| S1
εI(˜ω, a)
|1εI| − S1
εI(˜ω, b)
|1εI| | ≤Λ|a−b|.
for every ω˜ and everya andb in QN. Consequently,
|ds0(a)−ds0(b)| ≤Λ|a−b|
for every ω˜ in Ω˜0, andds0 may be extended to RN by a function still denoted ds0, satisfying the same lipschitzian property and which is still the almost sure limit of the previous normalized subadditive process.
Proof. We assume without loss of generality that ε = (1/n)n∈N∗ and I = (0,1). Let s∈R, s >0 intended to go to 0and γ ∈W1,1((0, n),RN) an admissible function in the denition of S]0,n[(˜ω, a)/n, which requires γ(0) = 0 and γ(n) = na. We modify the trajectory γ in [n−s, n) by the segment [γ(n−s), nb] in RN so that the new trajectory γs satises γs(0) = 0 and γs(n) = nb and is an admissible function in the denition of S]0,n[(˜ω, b)/n. More precisely, we set
γs(t) =
( γ(t) in[0, n−s]
nb+t−ns nb−γ(n−s)
in[n−s, n].
According to the growth condition, we have S]0,n[(˜ω, b)
n ≤ 1
n Z n
0
˜
ω(γs(t), γs0(t)) dt
≤ 1 n
Z n 0
˜
ω(γ(t), γ0(t)) dt+Λ
n|nb−γ(n−s)|.
Letting s→ 0 andn→+∞ in this order, and noticing that every element of W1,1((0, n),RN]is continuous in [0, n], we obtain
S]0,n[(˜ω, b)
n ≤ S]0,n[(˜ω, a)
n + Λ|b−a|, which completes the proof.
In what follows, the same notation Ω˜0 denotes various subsets of full probability. In the proofs below we will consider theP˜-preserving transforma- tion τξ on Ω˜ dened for every ξ ∈ RN by τξω(x, ζ) = ˜˜ ω(x+ξ, ζ). We also introduce a second process S˜from J intoL1( ˜Ω,T,P˜) dened by
S˜I(˜ω, a) = inf Z
I
˜
ω(u(t), u0(t)) dt:u∈W1,1(I,◦ RN), u(α)−u(β) = (α−β)a
.
Clearly, S˜ is invariant under translations of R and satises SI ≥ S˜I for all I ∈ I. Although S˜is not subadditive we have
Theorem 3. For everyt0 ∈R,ρ >0anda∈RN,S˜1
εIρ(t0)(·, a)/|1εIρ(t0)|
tends in probability to ds0(a) as ε→0.
Therefore, if a belongs to QN and ρ to Q+, there exists a subset Ω˜0 of Ω˜ of full probability which does not depend on a, ρ and t0, and a subsequence, not relabelled, possibly depending on a andρ, such that
ε→0lim S˜1
εIρ(t0)(˜ω, a)
|1εIρ(t0)| =ds0(a) for all ω˜ in Ω˜0.
Proof. AsS˜I+s(˜ω, a) = ˜SI(˜ω, a)for everysinR, it is sucient to establish the proof for the interval Iρ(0). To simplify notation, we denote by I this interval. Let γε be an ε-minimizer of S˜1
εI, i.e., S˜1
εI(˜ω, a)
|1εI| ≥ 1
|1εI|
Z
1 εI
˜ ω
γε
ε, γε0
dt−ε,
and set ˜γε =γε−bεwherebε=γε(αε)−αεaandαε is the lower bound of the interval (1/ε)I. Since ˜γε is an admissible function for S1
εI, we have S˜1
εI(˜ω, a)
|1εI| ≥ 1
|1εI|
Z
1 εI
τbεω˜ γ˜ε
ε,γ˜ε0
dt−ε≥ S1
εI(τbεω, a)˜
|1εI| −ε.
Thus
S1 εI(˜ω, a)
|1εI| ≥ S˜1
εI(˜ω, a)
|1εI| ≥ S1
εI(τbεω, a)˜
|1εI| −ε, so that, for every η >0,
S˜1 εI(·, a)
|1εI| −ds0(a) > η
⊆
S˜1 εI(·, a)
|1εI| − S1
εI(τbε·, a)
|1εI| +ε > η
2
∪
S1
εI(τbε·, a)
|1εI| −ε−ds0(a) > η
2
⊆
S1 εI(·, a)
|1εI| − S1
εI(τbε·, a)
|1εI| +ε > η
2
∪
S1
εI(τbε·, a)
|1εI| −ε−ds0(a) > η
2
.
Sinceτbεis aP˜ preserving transformation, Theorem 2 implies thatS˜1
εI(·, a)/|1εI|
tends in probability to ds0(a) (use the Markov inequality P˜
|fε| > η2
≤
2 η
R
Ω|fε|dP˜).
4. THE MAIN COMPACTNESS RESULT
As noted in the introduction,ds(ω, x/ε, ξ)denes a random metric which has the same law as the strongly oscillating random Riemannian metric de- ned by
dsε(ω, x, ξ) :=ds1(ξ) + ds2(ξ)−ds1(ξ)
min(1,Nε(ω, B(x, εr)) where Nε is a Poisson process with intensity ελN (reason by using the Laplace transform of N). The next two subsections are devoted to the proof of the following compactness theorem and its corollary.
Theorem 4. There exists a subsequence ofεsuch that the length dened in L1((0,1),RN) by
Fε(˜ω)(γ) :=
Z 1
0
˜ ω
γ(t) ε , γ0(t)
dt if γ ∈ W1,1((0,1),RN),
+∞ otherwise,
Gamma-converges almost surely to the deterministic function dened in L1((0,1),RN) by
F0(γ) =
Z 1
0
ds0(γa0(t)) dt+ Z 1
0
ds∞0 γs0
|γs0|
|γs0| ifγ ∈BV((0,1),RN),
+∞ otherwise,
where γa0 is the density of the regular part of γ0 in its Lebesgue decomposition and ds∞0 is the recession function dened for every a∈RN by
ds∞0 (a) = lim
t→+∞
ds0(ta) t .
Before proving Theorem 4, let us take into account the boundary condi- tions. For every (a, b) inR2N, consider the functionals
Ia,b,I¯a,b:L1((0,1),RN)→R+∪ {+∞}
dened by
Ia,b(γ) =
(0 if γ(0) =aand γ(1) =b, +∞ otherwise,
I¯a,b(γ) =
(ds∞0 (γ(0)−a) +ds∞0 (b−γ(1)) if γ ∈BV((0,1),RN),
+∞ otherwise.
It is well known that I¯a,b is the relaxed functional of Ia,b. As a consequence of Theorem 4, we obtain the following result whose proof is a straightforward adaptation of that of Theorem 4 (see [8]).
Corollary 1. There exists a subsequence ofεsuch thatFε+Ia,bGamma- converges almost surely to the deterministic function F0+ ¯Ia,b.
We are going to establish Theorem 4. For proving the upper bound, we use the subadditive ergodic process. The lower bound will be established using the second ergodic process.
4.1. The upper bound
With the notation and denitions of previous sections, we establish, for a subsequence, the upper bound in the denition of Gamma-convergence. More precisely, we have
Proposition 4. There exists a subset Ω˜0 of Ω˜ of full probability such that for all ω˜ in Ω˜0 there exists a subsequence still denoted ε such that
Γ lim sup
ε→0
Fε(˜ω)≤F0. Proof. We proceed in two steps.
Step 1. We prove the result in a dense countable subsetDofC1((0,1),RN) for the strong topology of L1((0,1),RN). Let γ ∈ D. Subdivise (0,1)in in- tervals Ii = (ti−1, ti), i = 1, . . . , m, t0 = 0, tm = 1, of length h intended to go to 0, and consider the continuous piecewise linear function γh taking the values γ(ti)for every i= 0, . . . , m. Denote byai the gradient ofγh inIi. It is easily seen that γh strongly converges to γ inW1,1((0,1),RN). According to Theorem 2 and Proposition 3, there exists Ω˜0 of full probability such that, for all ω˜ ∈Ω˜0 and i= 1, . . . , m,
ds0(ai) = lim
ε→0
S1
εIi(˜ω, ai)
|1εIi| = lim
ε→0
1
|Ii| Z
Ii
˜
ω(γh,εi (˜ω, t)
ε , γh,ε0i (˜ω, t)) dt, where we performed a change of scale andγih,ε(˜ω,·)is anε/m-minimizer of the problem
(2) inf
v∈W1,1(Ii,RN)
Z
Ii
˜ ω
v(t) ε , v0(t)
dt:v(ti−1) =ti−1ai, v(ti) =tiai
.
Set bi = γ(ti)−tiai for i = 1, . . . , m and denote by χIi the characteristic function of the interval Ii. The function
γh,ε(˜ω,·) =
m
X
i=1
γh,εi (τbi
ε
˜
ω,·) +bi χIi
belongs to W1,1((0,1),RN). The measure preserving transformation τbi ε
is introduced for technical reasons which will appear below. We have
lim sup
ε→0
E
Z 1 0
ds0(γh0) dt− Z 1
0
˜ ω
γh,ε(˜ω, t)
ε , γh,ε0 (˜ω, t)
dt
≤lim sup
ε→0 m
X
i=1
E
ds0(ai)|Ii| − Z
Ii
˜ ω
γh,εi (τbi
ε
˜
ω, t) +bi
ε , γh,ε0i (τbi
ε
˜ ω, t)
dt
= lim sup
ε→0 m
X
i=1
E
ds0(ai)|Ii| − Z
Ii
τbi
ε
˜ ω
γh,εi (τbi
ε
˜ ω, t)
ε , γh,ε0i (τbi
ε
˜ ω, t)
dt
= lim sup
ε→0 m
X
i=1
E
ds0(ai)|Ii| − Z
Ii
˜
ω γh,εi (˜ω, t)
ε , γh,ε0i (˜ω, t)
! dt
= 0, where we have used the Lebesgue dominated convergence theorem. According to Proposition 3, we obtain
(3) lim
h→0lim
ε→0 E
Z 1 0
ds0(γ0(t)) dt− Z 1
0
˜ ω
γh,ε(˜ω, t)
ε , γh,ε0 (˜ω, t)
dt
= 0.
On the other hand, using Poincaré's inequality (note that γ(ti) =γh,εi (˜ω, ti) + bi), we obtain
E Z 1
0
|γ(t)−γh,ε(˜ω, t)|dt≤
m
X
i=1
E Z
Ii
|γ(t)−γh,εi (τbi
ε
˜
ω, t)−bi|dt
=
m
X
i=1
E Z
Ii
|γ(t)−γh,εi (˜ω, t)−bi|dt≤ 1 2h
m
X
i=1
E Z
Ii
|γ0(t)−γh,ε0i (˜ω, t)|dt.
From coercivity, growth condition, and since γh,εi (˜ω,·) is an ε/m-minimizing sequence of (2), we have
Z
Ii
|γh,ε0i (˜ω, t)|dt≤ 1 ν
Z
Ii
˜ ω
t εai, ai
dt+ ε mν ≤ Λ
ν|ai| |Ii|+ ε mν, so that the previous inequality yields
(4) E
Z 1 0
|γ(t)−γh,ε(˜ω, t)|dt≤Ch+ ε ν, whereCdepends only on Λ, ν andsup
[0,1]
|γ0|. Letε→0andh→0in this order in (3), (4). Using a simultaneous diagonalisation argument, we deduce that there exists a map ε7→h(ε) such that
ε→0limE
Z 1 0
ds0(γ) dt− Z 1
0
˜ ω
γh(ε),ε(˜ω, t)
ε , γh(ε),ε)0 (˜ω, t)
dt
= 0,
ε→0limE Z 1
0
|γ(t)−γh(ε),ε(˜ω, t)|dt= 0.
Set γε(˜ω,·) = γh(ε),ε(˜ω,·). By the above, there exist a set Ω˜0 of full probability that we may assume independent on γ and, by a classical Can- tor diagonalization argument, a subsequence of εindependent onγ such that
(γε(˜ω,·))ε strongly converges inL1((0,1),RN) to γ with lim sup
ε→0
Fε(˜ω)(γε(˜ω,·))≤F0(γ), so that
Γ lim sup
ε→0
Fε(˜ω)≤F0 inD.
Step 2. Let us consider the functional dened inL1((0,1),RN)by F˜0(γ) =
Z 1
0
ds0(γ0(t)) dt if γ ∈D,
+∞ otherwise.
According to Step 1, the estimate Γ lim sup
ε→0
Fε(˜ω)≤F˜0
holds in L1((0,1),RN) for every ω˜ in Ω˜0. We conclude by taking the lower semicontinuous envelope for the strong convergence in L1((0,1),RN) of the two functionals in the inequality above. Indeed, from the semicontinuity of Γ lim sup
ε→0
Fε(˜ω) (see [7]) and classical integral representation results on the lower semicontinuous envelope of convex functions with linear growth (see [8]
or [5] in the quasi-convex case), we obtain Γ lim sup
ε→0
Fε(˜ω)≤F0.
4.2. The lower bound
We establish now the lower bound in the denition of Gamma-convergence.
More precisely, we have
Proposition 5. There exists a subsetΩ˜0 ofΩ˜ of full probability such that Γ lim inf
ε→0 Fε(˜ω)≥F0. for all ω˜ in Ω˜0.
Proof. It is enough to prove that there existsΩ˜0of full probability such that F0(γ)≤lim inf
ε→0 Fε(˜ω)(γε) for every γε strongly converging toγ inL1((0,1),RN).
Fixω˜ in the set Ω˜0 of full probability obtained in Theorem 3 and assume thatlim inf
ε→0 Fε(˜ω)(γε)<+∞. There exists a (not relabelled) subsequence such
that the bounded Radon measure µε(˜ω) := ˜ω
γε(·) ε , γε0
L1b(0,1)
weakly converges to a radon measure µ(˜ω) for any ω˜ ∈ Ω˜0. Our method consists in analyzing the Lebesgue decomposition µ(˜ω) = µa(˜ω) +µs(˜ω) of the limit measure, where µa(˜ω) and µs(˜ω) are respectively the regular and the singular part of µ(˜ω) with respect to L1b(0,1). More precisely, using the ergodic Theorem 3, we establish the existence of a subset of full probability still denoted Ω˜0 such that
(5) µa(˜ω)≥ds0(γa0(·))L1b(0,1), µs(˜ω)≥ds∞0 γs0
|γs0|(·)
|γs0|
for all ω˜ ∈Ω˜0, and the conclusion of Theorem 5 will follow from the classical Alexandro inequality
µ(˜ω)(O)≤lim inf
ε→0 µε(˜ω)(O)
for every open subset of(0,1), applied here forO= (0,1). To simplify notation we drop the xed variableω˜ ∈Ω˜0 appearing inµεandµ. The next two lemmas are devoted to the proof of the two estimates in (5).
Lemma 1. For L1b(0,1)-almost all t0 in (0,1), we have
ρ→0lim
µ(Iρ(t0))
|Iρ(t0)| ≥ds0(γa0(t0)).
Proof. By the classical dierentiation theorem, the limit lim
ρ→0µ(Iρ(t0))/
|Iρ(t0)| exists for L1b(0,1)-almost allt0 in (0,1). We x such a t0. Clearly, for ρ∈(t0,1−t0)\D, whereDis a countable set, we have
ρ→0lim
µ(Iρ(t0))
|Iρ(t0)| = lim
ρ→0lim
ε→0
µε(Iρ(t0))
|Iρ(t0)| = lim
ρ→0lim
ε→0
1
|Iρ(t0)|
Z
Iρ(t0)
˜ ω
γε(t) ε , γε0(t)
dt.
Assume for a moment that the trace of γε agrees with the ane function γ0(t) =γ(t0) + (t−t0)awitha∈QN,|γa0(t0)−a| ≤η, whereη >0is intended to go to 0. It follows that
ρ→0lim
µ(Iρ(t0))
|Iρ(t0)| = lim
ρ→0lim
ε→0
µε(Iρ(t0))
|Iρ(t0)|
≥lim sup
ρ→0
lim sup
ε→0
inf 1
|Iρ(t0)|
Z
Iρ(t0)
˜ ω
v(t) ε , v0(t)
dt: v∈W1,1(Iρ(t0),RN), v(t0−ρ/2)−v(t0+ρ/2) =ρa
≥lim sup
ρ→0
lim sup
ε→0
S˜1
εIρ(t0)(˜ω, a)
|1εIρ(t0)| =ds0(a)≥ds0(γa0(t0))−Λη,
where we have used Theorem 3 giving the expression of ds0. The conclu- sion follows by letting η → 0. The idea consists now in modifying γε in a neighborhood of the boundary of Iρ(t0) by a function ˜γε which belongs to γ0+W01,1(Iρ(t0),RN) and to follow the above procedure. We obtain (for the details, we refer the reader to [1])
(6) µ(Iρ(t0))
|Iρ(t0)| ≥ds0(a)− Λ ρ2
Z
Iρ(t0)
|γ(t)−γ0(t)|dt.
By the denition of γ0, (6) yields µ(Iρ(t0))
|Iρ(t0)| ≥ds0(γa0(t0))−Λ ρ
Z
Iρ(t0)
|γ(t)−γ(t0)−(t−t0)γa0(t0)|
|t−t0| dt−2Λη.
Letting ρ→0, according to Proposition 2 we conclude as previously.
We establish now the second estimateµs≥ds∞0 γs0
|γs0|(·)
|γs0|. Lemma 2. For γs0-almost all t0 in (0,1), we have
ρ→0lim
µ(Iρ(t0))
|γs0|(Iρ(t0)) ≥ds∞0 γ0
|γ0|(·) .
Proof. We rst introduce some notation and state a few technical results established in [5]. Dene in BV(I,RN) and W1,1(I,RN), respectively, the rescaled functions
γρ= 1
|γ0|(Iρ(t0))|
γ(t0+ρt)− 1
|Iρ(t0)|
Z
Iρ(t0)
γ(s) ds
,
γε,ρ= 1
|γ0|(Iρ(t0))|
γε(t0+ρt)− 1
|Iρ(t0)|
Z
Iρ(t0)
γε(s) ds ,
whereI is the interval(−1/2,1/2). In the sequel, we xt0 in the Borel subset E of(0,1)satisfying |γs0| (0,1)\E
= 0, for which the limit
ρ→0lim
γ0(Iρ(t0))
|γ0|(Iρ(t0)) = γ0
|γ0|(t0) exists and lim
ρ→0 ρ
|γ0|(Iρ(t0)) = 0. Set tρ = |γ0|(Iρρ(t0)), which thus tends to +∞. From now on, ρ denotes a sequence of positive rational numbers. Note that γε,ρ tends toγρstrongly inL1((0,1),RN)asρ→0 and that, according to the Poincaré-Wirtinger inequality,
Z
I
|γρ|dt≤C Z
I
|γ0ρ|.
SinceR
I|γρ0|dt≤1, there exists a function˜γ inBV(I,RN)and a subsequence inρthat we shall not relabel such that γρtends toγ˜ weakly inBV(I,RN). A carefull analysis of this convergence leads to the following property ofγ˜: there exists a nondecreasing scalar function ψ in BV(I) such that ˜γ admits the representation ˜γ(t) =ψ(t) γ0/|γ0|(t0). Moreover, for every δ >0,δ ∈]0,1[\D, where Dis a countable subset of ]0,1[,γρ0(δI) →γ˜(δI). For a proof, we refer the reader to [5], Theorem 2.3.
We are now in a position to complete the proof of Lemma 2. We want to estimate from below lim
ρ→0
µ(Iρ(t0))
|γ0|(Iρ(t0)) for the subsequence invoked above. We have
ρ→0lim
µ(Iρ(t0))
|γ0|(Iρ(t0)) = lim
ρ→0lim
ε→0
µε(Iρ(t0))
|γ0|(Iρ(t0))
= lim
ρ→0lim
ε→0
1
|γ0|(Iρ(t0)) Z
Iρ(t0)
˜ ω
γε(t) ε , γε0(t)
dt.
In order to reason on the xed space BV(I), we change the scale and set vε,ρ= 1
ρ γε(t0+tρ)−rρ ,
whererρapproximates inQN the mean valuemρofγρinIρ(t0). More precisely,
|mρ−rρ| ≤η, where η is a positive number intended to go to 0. The reason for which we substitute rρ to this mean value will be explained later. Note thatvε,ρ/tρapproximates the functionγε,ρ previously dened. More precisely,
we have
vε,ρ tρ −γε,ρ
≤ |mε,ρ−mρ|
|γ0|(Iρ(t0)) + η
|γ0|(Iρ(t0)) and we obtain
(7) lim
ρ→0
µ(Iρ(t0))
|γ0|(Iρ(t0)) = lim
ρ→0lim
ε→0
1 tρ
Z
I
τrρ
ε ω˜ρvε,ρ ε , vε,ρ0
dt.
In order to apply the ergodic Theorem 3 witha=tρ1δ˜γ0(δI), we modify vε,ρ near the boundary of I and adapt an argument of [1]. More precisely, we slice I near its boundary by small intervals of length αρ/n, where αρ =
|γρ−˜γ|1/2
L1(I,RN) and n∈Nis intended to go to +∞: I0 = (1−αρ)I, . . . , Ii =
1−αρ+iαρ
n
I, i= 1, . . . , n, and consider the cut-o functions
ϕi ∈ C0∞(Ii), 0≤ϕ≤1, ϕ= 1 inIi−1, |ϕ0i|L∞(Ii)≤ n αρ
, i= 1, . . . , n.
We nally dene
vε,ρ,i=tργ0+ϕi(vε,ρ−tργ0),
whereγ0 is the ane function γ0(t) = tρ
δγ˜0(δI) +ψ(δ/2−) +ψ(δ/2+) 2
γ0
|γ0|(t0)
with gradient 1δγ˜0(δI). The constant in the denition of γ0 will guarantee the control of the defect at the end of the proof. We then obtain for i= 1, . . . , n the estimate
(8) 1 tρ
Z
I
τrρ
ε ω˜ vε,ρ
ε/ρ, vε,ρ0
dt≥ 1 tρ
Z
I
τrρ
ε ω˜ vε,ρ,i
ε/ρ , vε,ρ,i0
dt−Ri(ρ, ε, n, η), whose proof is an easy consequence of the growth condition (see [1]), where
Ri(ρ, ε, n, η) =O(ρ) + Λ tρ
Z
Ii\Ii−1
|v0ε,ρ−tργ00|dt+ nΛ tραρ
Z
Ii\Ii−1
|vε,ρ−tργ0|dt and O(ρ) does not depend onε, n, η and tends to0 asρ→0.
As vε,ρ,i are admissible functions in the denition of S˜, for i= 1, . . . , n we have
1 tρ
Z
I
τrρ
ε ω˜ vε,ρ,i
ε/ρ , v0ε,ρ,i
dt≥ 1 tρ
S˜ 1 ε/ρI
τrρ
ε ω, t˜ ρ1 δ˜γ0(δI)
|ε/ρ1 I| . Averaging these ninequalities, from (8) we obtain
1 tρ
Z
I
τrρ
ε ω˜ ve,ρ
ε/ρ, vε,ρ0
dt≥ 1 tρ
S˜1 ε/ρI
τrρ
ε ω,˜ tδργ˜0(δI)
|ε/ρ1 I| −O(ρ)−O(1/n)
−Λ αρ
Z
I\(1−αρ)I
vε,ρ tρ
−γ0
dt≥ 1 tρ
S˜ 1 ε/ρI
τrρ
ε ω,˜ tδργ˜0(δI)
|ε/ρ1 I| −O(ρ)−O(1/n)
− Λη
|γ0|(Iρ(t0))− Λ|mε,ρ−mρ|
|γ0|(Iρ(t0)) − Λ αρ
Z
I\(1−αρ)I
|γε,ρ−γ0|dt,
where O(1/n) tends to 0 as n goes to +∞. We let successively n → +∞, ε→0,η→0. According to Theorem 3, there exists a setΩ˜0 of full probability such that
lim sup
ε→0
1 tρ
Z
I
τrρ
ε ω˜ ve,ρ
ε/ρ, vε,ρ0
dt
≥ 1 tρ
ds0 tρ
δγ˜0(δI)
− Λ αρ
Z
I\(1−αρ)I
|γρ−γ0|dt−O(ρ)
≥ 1 tρ
ds0
tρ δγ˜0(δI)
−Λ√
αρ− Λ αρ
Z
I\(1−αρ)I
|˜γ−γ0|dt−O(ρ)
for everyω˜ ∈Ω˜0. Note thatΩ˜0 a priori depends onrρandtρ/δγ˜0(δI). Actually, we consider the set of full probability Ω˜0 = TΩ˜r,a, where (r, a) belongs to QN ×QN. Without any loss of generality, we suppose here that tρ/δ ˜γ0(δI) belongs to QN. If not, is sucient to reason as in the previous subsection.
Recalling (7), we nally obtain
ρ→0lim
µ(Iρ(t0))
|γ0|(Iρ(t0)) ≥lim sup
ρ→0
1 tρds0
tρ
δ ˜γ0(δI)
−lim sup
ρ→0
Λ αρ
Z
I\(1−αρ)I
|˜γ−γ0|dt.
By a straightforward computation, it is easily seen thatγ˜−γ0 has zero trace on
∂δI. This is the reason for which we chose the constant ψ(δ/2−)+ψ(δ/22 +)|γγ00|(t0) in the denition ofγ0. Therefore, by the properties of trace inBV(I), the last term on the right hand side tends to 0. For the rst term we have
lim sup
ρ→0
1 tρds0
tρ
δγ˜0(δI)
= lim sup
ρ→0
1 δ
δ tρds0
tρ
δ ˜γρ0(δI)
= lim sup
ρ→0
1 δ
δ tρds0
tρ
δ γ0
|γ0|(t0)
−lim inf
ρ→0 Λ
1−|γ0|(Iδρ(t0)
|γ0|(Iρ(t0)
≥ 1 δ ds∞0
γ0
|γ0|(t0)
−(1−δ).
We conclude after letting δ →1.
The proof of Proposition 2 is now complete.
5. CONVERGENCE IN PROBABILITY
In order to simplify notation, we writeFεandF0to denote the restrictions to BV((0,1),RN)ofFε+Ia,bandF0+ ¯Ia,b, respectively. We want to establish the convergence in probability of all sequences (Fε)ε>0 of random functionals in a suitable metric space. We equip the spaceBV((0,1),RN)with the strong topology of L1((0,1),RN) and denote by SC the space of real valued lower semicontinuous functions G dened on BV((0,1),RN) that satisfy G(u) ≥ νR
Ω|Du|. It is well known (see for instance [9]) that SC is metrizable by the Gamma-convergence. We denote by dΓ a metric induced by the latter in SC and by BΓ the Borel eld of the metric space (SC, dΓ). Recall that the law of a measurable mapG: ( ˜Ω,T˜)→(SC,BΓ)is the image ofP˜ byGon the metric space (SC, dΓ).
Theorem 5. For everyω˜ ∈Ω˜ letF¯ε(˜ω)denote the lower semicontinuous envelope of Fε(˜ω). Then all sequences F¯ε converge in law and in probability to F0 in the metric space(SC, dΓ).
Proof. Clearly (see for instance [7]), for a subsequenceFε that Gamma- converges almost surely to F0, the corresponding subsequence of F¯ε Gamma- converges almost surely to the same limitF0 which belongs to SC. Therefore, any sequence of laws of random functions F¯ε weakly converges to the Dirac measureδF0. Then, by classical probabilistic arguments, any sequenceF¯εcon- verges in probability to the constant F0 inSC, i.e.,P{˜ ω˜ ∈Ω :˜ dΓ( ¯Fε(˜ω), F0)>
η} →0asε→0 for all η >0.
Corollary 2. The random geodesic metric dε converges in probability to the deterministic metric d0 in the following sense: for all η > 0 and all points (a, b) in RN,
P{ω :|dε(ω)(a, b)−d0(a, b)|> η} →0 asε→0.
Proof. It is equivalent to establish thatP{˜ ω˜ :|dε(˜ω)(a, b)−d0(a, b)|> η}
tends to 0asε→0, which is a straightforward consequence of P{˜ ω˜ ∈Ω :˜ dΓ( ¯Fε(˜ω), F0)> η} →0
and the continuity of the real valued mapG7→minGdened in(SC, dΓ).
REFERENCES
[1] Y. Abddaimi, G. Michaille and C. Licht, Stochastic homogenization for an integral functional of a quasiconvex function with linear growth. Asymptotic Anal. 15(2) (1997), 183+.
[2] E. Acerbi and G. Buttazzo, On the limit of periodic Riemannian metrics. J. Anal. Math.
43(4) (1983), 183201.
[3] M.A. Ackoglu and U. Krengel, Ergodic theorem for superadditive processes. J. Reine Angew. Math. 323 (1981), 5367.
[4] M. Amar and E. Vitali, Homogenization of periodic Finsler metric. J. Convex Anal.
5(1) (1998), 171186.
[5] L. Ambrosio and G. Dal Maso, On the relaxation inBV(Ω;Rm)of quasi-convex inte- grals. J. Funct. Anal. 109 (1992), 7697.
[6] L. Ambrosio, N. Fusco and D. Pallara, Functions of Bounded Variation and Free Dis- continuity Problems. Oxford Univ. Press, New York, 2000.
[7] H. Attouch, Variational Convergence for Functions and Operators. Applicable Mathe- matics Series. Pitman (Advanced Publishing Program), Boston, MA, 1984.
[8] G. Bouchitté, Convergence et relaxation des fonctionnelles du calcul des variations á croissance linéaire. Application à l'homogénéisation en plasticité. Ann. Fac. Sci.
Toulouse Math (5) 8 (1986/87), 1, 736.
[9] G. Dal Maso, An Introduction toΓ-convergence. Birkhäuser, Boston, 1993.
[10] G. Dal Maso and L. Modica, Non linear stochastic homogenization and ergodic theory.
J. Reine Angew. Math. 363 (1986), 2743.
[11] E. De Giorgi and T. Franzoni, Su un tipo di convergenza variazionale. Atti Accad. Naz.
Lincei Rend. Cl. Sci. Fis. Mat. Natur. (8) 58 (1975), 842850.