Annealed and quenched fluctuations of transient random walks in random environment
on Z
Laurent Tournier
UMPA, ENS Lyon – IMPA, Rio de Janeiro Joint work with N. Enriquez, C. Sabot and O. Zindy
June 16, 2011. LaPietra Week in Probability, Florence
RWRE on Z : definitions
Letµbe a law on(0,1).
Define an i.i.d. sequenceω= (ωx)x∈Zwith lawµ(“environment”).
ω0 ω1
ω−1
1−ω−1
0 1 2 3
− 1
− 2
− 3 · · · · · ·
ex.:µ=pδα+ (1−p)δβ, µ=B(a,b)(Beta distribution) (Givenω) QuenchedlawPωof Markov chain of transitionω (Randomω) AnnealedlawPof RWRE:
P(·) =E[Pω(·)] = Z
Pω(·)dµ⊗Z(ω) UnderP, the RW “learns” aboutω; transitions arereinforced.
RWRE on Z : definitions
Letµbe a law on(0,1).
Define an i.i.d. sequenceω= (ωx)x∈Zwith lawµ(“environment”).
ω0 ω1
ω−1
1−ω1
1−ω0
1−ω−1
0 1 2 3
− 1
− 2
− 3 · · · · · ·
· · ·
· · ·
ex.:µ=pδα+ (1−p)δβ, µ=B(a,b)(Beta distribution) (Givenω) QuenchedlawPωof Markov chain of transitionω (Randomω) AnnealedlawPof RWRE:
P(·) =E[Pω(·)] = Z
Pω(·)dµ⊗Z(ω) UnderP, the RW “learns” aboutω; transitions arereinforced.
RWRE on Z : definitions
Letµbe a law on(0,1).
Define an i.i.d. sequenceω= (ωx)x∈Zwith lawµ(“environment”).
ω0 ω1
ω−1
1−ω1
1−ω0
1−ω−1
0 1 2 3
− 1
− 2
− 3 · · · · · ·
· · ·
· · ·
ex.:µ=pδα+ (1−p)δβ, µ=B(a,b)(Beta distribution)
(Givenω) QuenchedlawPωof Markov chain of transitionω (Randomω) AnnealedlawPof RWRE:
P(·) =E[Pω(·)] = Z
Pω(·)dµ⊗Z(ω) UnderP, the RW “learns” aboutω; transitions arereinforced.
RWRE on Z : definitions
Letµbe a law on(0,1).
Define an i.i.d. sequenceω= (ωx)x∈Zwith lawµ(“environment”).
ω0 ω1
ω−1
1−ω1
1−ω0
1−ω−1
0 1 2 3
− 1
− 2
− 3 · · · · · ·
· · ·
· · ·
ex.:µ=pδα+ (1−p)δβ, µ=B(a,b)(Beta distribution) (Givenω) QuenchedlawPωof Markov chain of transitionω (Randomω) AnnealedlawPof RWRE:
P(·) =E[Pω(·)] = Z
Pω(·)dµ⊗Z(ω)
UnderP, the RW “learns” aboutω; transitions arereinforced.
RWRE on Z : definitions
Letµbe a law on(0,1).
Define an i.i.d. sequenceω= (ωx)x∈Zwith lawµ(“environment”).
ω0 ω1
ω−1
1−ω1
1−ω0
1−ω−1
0 1 2 3
− 1
− 2
− 3 · · · · · ·
· · ·
· · ·
ex.:µ=pδα+ (1−p)δβ, µ=B(a,b)(Beta distribution) (Givenω) QuenchedlawPωof Markov chain of transitionω (Randomω) AnnealedlawPof RWRE:
P(·) =E[Pω(·)] = Z
Pω(·)dµ⊗Z(ω) UnderP, the RW “learns” aboutω; transitions arereinforced.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 ω
DefineVbyV(0) =0 andωx= e−V(x) e−V(x)+e−V(x−1),
e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx= 1−ωx
ωx
Theorem (Solomon 1975)
( Xn→+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 V ω
V (x) < V (x − 1) ⇔ ω
x>
12DefineVbyV(0) =0 andωx= e−V(x) e−V(x)+e−V(x−1),
e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx= 1−ωx
ωx
Theorem (Solomon 1975)
( Xn→+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 V ω
V (x) < V (x − 1) ⇔ ω
x>
12DefineVbyV(0) =0 andωx= e−V(x) e−V(x)+e−V(x−1),
e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx=1−ωx
ωx
Theorem (Solomon 1975)
( Xn→+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 V ω
V (x) < V (x − 1) ⇔ ω
x>
12DefineVbyV(0) =0 andωx= e−V(x)
e−V(x)+e−V(x−1),e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx=1−ωx
ωx
Theorem (Solomon 1975)
( Xn→+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 V ω
V (x) < V (x − 1) ⇔ ω
x>
12DefineVbyV(0) =0 andωx= e−V(x)
e−V(x)+e−V(x−1),e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx=1−ωx
ωx
Theorem (Solomon 1975)
P-a.s.,
( Xn→+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 V ω
V (x) < V (x − 1) ⇔ ω
x>
12DefineVbyV(0) =0 andωx= e−V(x)
e−V(x)+e−V(x−1),e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx=1−ωx
ωx
Theorem (Solomon 1975)
Pω-a.s.
for P-a.e.ω,
( Xn →+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Potential – Transience and speed
0 1 2 3
− 1
− 2
− 3 V ω
V (x) < V (x − 1) ⇔ ω
x>
12DefineVbyV(0) =0 andωx= e−V(x)
e−V(x)+e−V(x−1),e−V(x)=Cx,x+1
i.e.V(x) =logρ1+· · ·+logρxforx≥0, whereρx=1−ωx
ωx
Theorem (Solomon 1975)
Pω-a.s.
for P-a.e.ω,
( Xn →+∞ iff E[logρ0]<0(⇔V(x)x→∞−→ −∞) (Xn)nrecurrent iff E[logρ0] =0
AssumeE[logρ0]<0.P-a.s.,Xn
n →vwhere
v>0 if E[ρ0]<1 v=0 else.
Fluctuations: KKS theorem
Hypotheses
(a)∃κ >0 such thatE[ρκ0] =1, andE[ρκ0(logρ0)+]<∞ (b) The law of logρ0is non-arithmetic.
Ex. Forµ=B(a,b),κ=b−a.
Theorem (Kesten-Kozlov-Spitzer 1975) Assume (a)-(b). Then, underP,
If 0< κ <1, Xn
nκ
(law)
−→n (AκSκ)−1/κ E[eitSκ] =e−(−it)κ If 1< κ <2, Xn−vn
n1/κ
(law)
−→n −v1+κ1AκSκ E[eitSκ] =e(−it)κ Ifκ >2, Xn−vn
√n
(law)
−→n N(0, σ2)
whereAκ>0 (Sκis a totally asymmetricκ-stable r.v.)
Fluctuations: KKS theorem
Hypotheses
(a)∃κ >0 such thatE[eκlogρ0] =1, andE[ρκ0(logρ0)+]<∞ (b) The law of logρ0is non-arithmetic.
Ex. Forµ=B(a,b),κ=b−a.
Theorem (Kesten-Kozlov-Spitzer 1975) Assume (a)-(b). Then, underP,
If 0< κ <1, Xn
nκ
(law)
−→n (AκSκ)−1/κ E[eitSκ] =e−(−it)κ If 1< κ <2, Xn−vn
n1/κ
(law)
−→n −v1+κ1AκSκ E[eitSκ] =e(−it)κ Ifκ >2, Xn−vn
√n
(law)
−→n N(0, σ2)
whereAκ>0 (Sκis a totally asymmetricκ-stable r.v.)
Fluctuations: KKS theorem
Hypotheses
(a)∃κ >0 such thatE[eκlogρ0] =1, andE[ρκ0(logρ0)+]<∞ (b) The law of logρ0is non-arithmetic.
Ex. Forµ=B(a,b),κ=b−a.
Theorem (Kesten-Kozlov-Spitzer 1975)
Assume (a)-(b). Then, underP, (letτ(x) =inf{n:Xn=x}) If 0< κ <1, Xn
nκ
(law)
−→n (AκSκ)−1/κ τ(x) x1/κ
(law)
−→n AκSκ If 1< κ <2, Xn−vn
n1/κ
(law)
−→n −v1+κ1AκSκ τ(x)−xv
x1/κ
(law)
−→x AκSκ Ifκ >2, Xn−vn
√n
(law)
−→n N(0, σ2) τ(x)√−xv
x
(law)
−→x N(0, σ02) whereAκ>0 (Sκis a totally asymmetricκ-stable r.v.)
Main result
Theorem (Enriquez-Sabot-T.-Zindy 2010) For 0< κ <2 (κ6=1),
Aκ=2
πκ2
|sin(πκ)|(CK)2E[ρκ0logρ0] 1/κ
whereCKis Kesten’s renewal constant:P(R>t)∼CKt−κwith R=1+ρ1+ρ1ρ2+· · ·.
Description of thequenchedbehaviour for 0< κ <2
Main result
Theorem (Enriquez-Sabot-T.-Zindy 2010) For 0< κ <2 (κ6=1),
Aκ=2
πκ2
|sin(πκ)|(CK)2E[ρκ0logρ0] 1/κ
whereCKis Kesten’s renewal constant:P(R>t)∼CKt−κwith R=1+ρ1+ρ1ρ2+· · ·.
Description of thequenchedbehaviour for 0< κ <2
Sample trajectory (0 < κ < 1)
150
40000 100
20000 10000
0 50000 60000
0 50
30000 70000
200 250
-50
Sample trajectory (1 < κ < 2)
Sample trajectory (1 < κ < 2)
General scheme of proof
Excursions ofV:e0=0,en+1 =inf{k≥en|V(k)≤V(en)}. V
τ(en) =X
k
τ(ek+1)−τ(ek)
=
small exc. H<hn
+
high exc. H≥hn
•There are very few large excursions (⇒crossing times almost i.i.d.)
•
Crossings of small excursions iso(n1/κ) (0< κ <1)
Fluctuation of crossings of small excursions iso(n1/κ) (1< κ <2)
General scheme of proof
Excursions ofV:e0=0,en+1 =inf{k≥en|V(k)≤V(en)}. V
Hk= max
ek≤x<ek+1V(x)−V(ek) ek
τ(en) =X
k
τ(ek+1)−τ(ek)
=
small exc. H<hn
+
high exc. H≥hn
•There are very few large excursions (⇒crossing times almost i.i.d.)
•
Crossings of small excursions iso(n1/κ) (0< κ <1)
Fluctuation of crossings of small excursions iso(n1/κ) (1< κ <2)
General scheme of proof
Excursions ofV:e0=0,en+1 =inf{k≥en|V(k)≤V(en)}. V
Hk= max
ek≤x<ek+1V(x)−V(ek) ek
τ(en) =X
k
τ(ek+1)−τ(ek)
=
small exc.
H<hn
+
high exc.
H≥hn
•There are very few large excursions (⇒crossing times almost i.i.d.)
•
Crossings of small excursions iso(n1/κ) (0< κ <1)
Fluctuation of crossings of small excursions iso(n1/κ) (1< κ <2)
Crossing time of a (high) excursion
Goal: estimateP(τ(e1)>t)ast→ ∞.
P(τ(e1)>t) =P(τ(e1)>t,H>logt−log logt) +o(t−κ)
F
iV
H 0
S
e
1•M1,M2,Hare almost independent on{H>h}withhlarge
•Property (Feller – Iglehart):P(eH>u)∼CIu−κ Thus,P(τ(e1)>t)∼Ct−κfor some (rather explicit)C.
Crossing time of a (high) excursion
Goal: estimateP(τ(e1)>t)ast→ ∞.
P(τ(e1)>t) =P(τ(e1)>t,H>logt−log logt) +o(t−κ)
F
iV
H 0
S
e
1τ(e1) =F1+· · ·+FN
'M1M2eHe e∼ E(1),
M1=PTH
−∞e−V(x),M2=P∞ 0 eV(x)−H
•M1,M2,Hare almost independent on{H>h}withhlarge
•Property (Feller – Iglehart):P(eH>u)∼CIu−κ Thus,P(τ(e1)>t)∼Ct−κfor some (rather explicit)C.
Crossing time of a (high) excursion
Goal: estimateP(τ(e1)>t)ast→ ∞.
P(τ(e1)>t) =P(τ(e1)>t,H>logt−log logt) +o(t−κ)
F
iV
H 0
S
e
1τ(e1) =F1+· · ·+FN+S
'M1M2eHe e∼ E(1),
M1=PTH
−∞e−V(x),M2=P∞ 0 eV(x)−H
•M1,M2,Hare almost independent on{H>h}withhlarge
•Property (Feller – Iglehart):P(eH>u)∼CIu−κ Thus,P(τ(e1)>t)∼Ct−κfor some (rather explicit)C.
Crossing time of a (high) excursion
Goal: estimateP(τ(e1)>t)ast→ ∞.
P(τ(e1)>t) =P(τ(e1)>t,H>logt−log logt) +o(t−κ)
F
iV
H 0
S
e
1τ(e1) =F1+· · ·+FN+S 'Eω[F]N
'M1M2eHe e∼ E(1),
M1=PTH
−∞e−V(x),M2=P∞ 0 eV(x)−H
•M1,M2,Hare almost independent on{H>h}withhlarge
•Property (Feller – Iglehart):P(eH>u)∼CIu−κ Thus,P(τ(e1)>t)∼Ct−κfor some (rather explicit)C.
Crossing time of a (high) excursion
Goal: estimateP(τ(e1)>t)ast→ ∞.
P(τ(e1)>t) =P(τ(e1)>t,H>logt−log logt) +o(t−κ)
F
iV
H 0
S
e
1τ(e1) =F1+· · ·+FN+S 'Eω[F]N
'Eω[F]Eω[N]e
'M1M2eHe
e∼ E(1),
M1=PTH
−∞e−V(x),M2=P∞ 0 eV(x)−H
•M1,M2,Hare almost independent on{H>h}withhlarge
•Property (Feller – Iglehart):P(eH>u)∼CIu−κ Thus,P(τ(e1)>t)∼Ct−κfor some (rather explicit)C.
Crossing time of a (high) excursion
Goal: estimateP(τ(e1)>t)ast→ ∞.
P(τ(e1)>t) =P(τ(e1)>t,H>logt−log logt) +o(t−κ)
F
iV
H 0
S
e
1M
1M
2T
Hτ(e1) =F1+· · ·+FN+S 'Eω[F]N
'Eω[F]Eω[N]e 'M1M2eHe e∼ E(1),
M1=PTH
−∞e−V(x),M2=P∞ 0 eV(x)−H
•M1,M2,Hare almost independent on{H>h}withhlarge
•Property (Feller – Iglehart):P(eH>u)∼CIu−κ Thus,P(τ(e1)>t)∼Ct−κfor some (rather explicit)C.
Interlude on heavy-tailed distributions
LetT1,T2, . . .be i.i.d. r.v.≥0 such that P(Ti>t)∼Ct−κ. Then, if 0< κ <1, T1+· · ·+Tn
n1/κ
(law)
−→n (CΓ(1−κ))1/κSκ and, if 1< κ <2, T1+· · ·+Tn−nE[Ti]
n1/κ
(law)
−→n (−CΓ(1−κ))1/κSκ. Forκ >2, CLT.
Heavy-tail phenomenon:
•#{1≤i≤n:Ti≥εn1/κ}(law)−→n P(Cε−κ)
•(for 0< κ <1)E h X
1≤i≤n
Ti1{T
i<εn1/κ}
i∼Cε1−κn1/κ
⇒Up to an error of orderε1−κ,T1+···+Tn1/κ n is given by theP(Cε−κ)terms larger thanεn1/κ.
Interlude on heavy-tailed distributions
LetT1,T2, . . .be i.i.d. r.v.≥0 such that P(Ti>t)∼Ct−κ. Then, if 0< κ <1, T1+· · ·+Tn
n1/κ
(law)
−→n (CΓ(1−κ))1/κSκ and, if 1< κ <2, T1+· · ·+Tn−nE[Ti]
n1/κ
(law)
−→n (−CΓ(1−κ))1/κSκ. Forκ >2, CLT.
Heavy-tail phenomenon:
•#{1≤i≤n:Ti≥εn1/κ}(law)−→n P(Cε−κ)
•(for 1< κ <2) Var X
1≤i≤n
Ti1{Ti<εn1/κ}
∼Cε2−κn2/κ
⇒Up to an error of orderε1−κ/2, T1+···+Tn1/κn−nE[Ti] is given by the (centered) P(Cε−κ)terms larger thanεn1/κ.
Next steps of the proof of the main result (KKS)
Lethn=κ1 logn−log logn(henceehn=εnn1/κ,εn= log1n, cf.τ'MeH)
τ(en) =X
k
τ(ek+1)−τ(ek)
=
small exc.
H<hn
+
high exc.
H≥hn
Neglect (fluctuations of) crossing times of small excursions Ensure large excursions are way appart of each other w.h.p.
Neglect time spent “backtracking” far away to the left of a high excursion before crossing it
Replace neglected parts by independent versions of them (which are negligible as well)
⇒Reduction to i.i.d. copies ofτ(e1), hence the (annealed) limit theorem.
Next steps of the proof of the main result (KKS)
Lethn=κ1 logn−log logn(henceehn=εnn1/κ,εn= log1n, cf.τ'MeH)
τ(en) =X
k
τ(ek+1)−τ(ek)
=
small exc.
H<hn
+
high exc.
H≥hn
Neglect (fluctuations of) crossing times of small excursions Ensure large excursions are way appart of each other w.h.p.
Neglect time spent “backtracking” far away to the left of a high excursion before crossing it
Replace neglected parts by independent versions of them (which are negligible as well)
⇒Reduction to i.i.d. copies ofτ(e1), hence the (annealed) limit theorem.
Consequences of the quenched description
0< κ <1
τ(en)is mainly given by a few termsMeHeattached to deep valleys.
The time spent in-between is negligible in comparison to them.
⇒Localizationin a (random) deep valley. (cf. Enriquez-Sabot-Zindy) 1< κ <2
The fluctuations ofτ(en)are mainly given by a few termsMeH(e−1) attached to deep valleys.
⇒Practical interest(explicit confidence intervals,. . . )
The fluctuations ofτ(en)are almost a sum of i.i.d. terms likeMeH(e−1) (re-introducing independent small valleys).
L
τ(en)−Eω[τ(en)]
n1/κ ω
' L
n
X
i=1
MbiebHi
n1/κ (ei−1)
M,b Hb
!
⇒Limit theoremfor the law of the quenched law of fluctuations IfT1,T2, . . .are i.i.d. withP(T >t)∼Ct−κand 0< κ <2,
Ti n1/κ
1≤i≤n (law)
−→ {ξi|i≥1} whereξis a PPP of intensityλκt−(κ+1)dt.
Consequences of the quenched description
0< κ <1
τ(en)is mainly given by a few termsMeHeattached to deep valleys.
The time spent in-between is negligible in comparison to them.
⇒Localizationin a (random) deep valley. (cf. Enriquez-Sabot-Zindy) 1< κ <2
The fluctuations ofτ(en)are mainly given by a few termsMeH(e−1) attached to deep valleys.
⇒Practical interest(explicit confidence intervals,. . . )
The fluctuations ofτ(en)are almost a sum of i.i.d. terms likeMeH(e−1) (re-introducing independent small valleys).
L
τ(en)−Eω[τ(en)]
n1/κ ω
' L
n
X
i=1
MbiebHi
n1/κ (ei−1)
M,b Hb
!
⇒Limit theoremfor the law of the quenched law of fluctuations L
τ(en)−Eω[τ(en)]
n1/κ ω
(law,W1)
−→n L
∞
X
i=1
ξi(ei−1) ξ
!
whereξis a PPP of intensityλκt−(κ+1)dt. AndW1is Wasserstein distance.