Local limit theorems on some non unimodular groups
Emile Le Page and Marc Peigne Abstract.
Let Gd be the semi-direct product of R + and Rd, d 1 and let us consider the product group Gd N = Gd RN, N 1. For a large class of probability measures on Gd N, one proves that there exists( )2]01] such that the sequence of nite measuresnn(N+3)=2 ( )n n
o
n1
converges weakly to a non-degenerate measure.
Resume.
Soit Gd le produit semi-direct de R + et de Rd et Gd N le groupe produit GdRN, N 0. Pour une large classe de mesures de probabilite sur Gd N nous montrons qu'il existe ( )2]01] tel que la suite de mesures niesnn(N+3)=2 ( )n n
o
n1
converge vaguement vers une mesure non nulle.
1. Introduction.
Fix two integers d 1, N 0 and choose a norm kk on Rd and
RN (when N 1). Let Gd N be the connected group R + Rd RN 117
with the composition law
for all g= (aub) for all g0 = (a0u0b0)2G gg0 = (aa0au0+ub+b0):
We will note g = (a(g)u(g)b(g)) (or g = (aub) when there is no ambiguity). The group (Gd N) is a non unimodular solvable group with exponential growth and the right Haar measuremD on Gd N is
mD(dadudb1dbN) = dadudb1dbN
a :
Note that Gd0 is the semi-direct product of R + and Rd in particular G1 0 is the ane group of the real line.
We consider a probability measure on G we denote by n its nth power of convolution. Under quite general assumptions on we show that there exists ( )2]01 such that the sequence
nn(N+3)=2 ( )n n
o
n0
converges weakly to a non-degenerate measure. This problem has al- ready been tackled by Ph. Bougerol in 3] where were established local limit theorems on some solvable groups with exponential growth in particular, for a class R of probability measures on the ane group of the real line (that is d= 1 and N = 0) he showed that the sequence
n n3=2 ( )n n
o
n0
converges weakly to a non-degenerate measure. In 7] we extend this result to a quite large class of probability measures the new ingredient in our proof was the fact that there exists closed connections between this problem and the theory of the uctuations of a random walk on the real line. In the present paper, we extend this result to the case N 1 we rst obtain uniform upperbounds in the Local limit theorem for a random walk on Rd and, secondly, we use a generalisation of the Wiener-Hopf's factorisation due to Ch. Sunyach 9].
This study is also related with the work by N. T. Varopoulos 10], 11] where upperbounds and lowerbounds for the asymptotic behaviour
of the convolution powers n of a large class of probability measures are given.
From now on, we will suppose that N 1 and we set G =Gd N. We introduce the following conditions on :
Hypothesis G1.
There exists >0 such thatZ
G(ejlogaj+kuk+kbk2) (dadudb)<+1:
Hypothesis G2.
ZGLoga (dadudb) = 0 andZ
Gb (dadudb) = 0:
Hypothesis G3.
The support of is included in R + (R+)d RN, the image of by the mapping (aub) 7;! (Logab) is aperiodic inRN+1 (see Denition 2:1) and there exists >0 such that
Z
Gkuk; (dadudb)<+1:
Hypothesis G'3.
The measure is absolutely continuous with respect to the Haar measure mD on G and its density satisesZ
]0 1]RN
q s
Z
R
q(aub)du dadba <+1: for some and q in ]1+1:
We have the
Theorem 1.1.
Let be a probability measure on G satisfying hypothe- ses G1, G2 and G3 (or G'3). Then, the sequence of nite measuresfn(N+3)=2 ngn0 converges weakly to a non-degenerate Radon mea- sure on G.
Note that the asymptotic behavior of the sequencef ngn1 does not depend on d.
When is not centered, that is
Z
GLoga (dadudb)6= 0
or Z
Gb (dadudb)6= 0
we bring back the study to the centered case as in 7]. We introduce the following conditions on :
Hypothesis G 1.
There exists > 0 such thatZ
G(at+kuk+ exp(tkbk)) (dadudb)<+1 for any t2R.
Hypothesis G 2.
One hasZ
GLoga (dadudb)6= 0
with fg2G: a(g)<1g>0 and fg2G: a(g)>1g>0.
When satis es these two conditions, there exists a unique (s0t0)
2R RN such that
Z
Gas0eht0 bi (dadudb) = (s t)inf
2RR N
Z
Gaseht bi (dadudb): Furthermore,
( ) =Z
Gas0eht0 bi (dadudb) belongs to ]01]. Note that the probability measure
0(dg) = 1( )a(g)s0eht0b(g)i (dg)
satis es hypotheses G1 and G2. The following condition is the equiva- lent of Hypothesis G'3 in the non centered case:
Hypothesis G 3.
The measure is absolutely continuous with respect to the Haar measure mD on G and its density satisesZ
]0 1]RN
q s
Z
R
q(aub)du dadba <+1
for some q2]1+1 and 2]1;s0+1.
Theorem 1.2.
Let be a probability measure on G satisfying condi- tions G 1,G 2 and G3 (or G 3) and let( ) =(s t)inf
2RR N
Z
Gaseht bi (dadudb): Then, the sequence of nite measures
nn(N+3)=2 ( )n n
o
n1
weakly converges to a non-degenerate Radon measure on G.
The demonstration of Theorem 2.1 is closely related to the study of the uctuations of a random walk (X1nY1n)n0 onRN+1. In Section 2, we rst state the classical local limit theorem on RN+1 but we add in its statement uniform upperbounds relatively to the starting point of the random walk (X1nY1n)n0. This result is thus very usefull to obtain a precise equivalent in Theorem 2.5 of the joint law of the random walk (X1nY1n)n0 with its rst entrance time T+ in the half spaceR+RN a local limit theorem for the process
(X1nmaxf0X11:::X1ngY1n)n0
is thus obtained (Theorem 2.6). In Section 3 we give the main steps of the proof of Theorem 1.1.
2. Fluctuations of a random walk on
RN+1.
Fix an integer N 1 and let (X1Y1)(X2Y2)::: be indepen- dent R RN-valued random variables with distribution p de ned on a probability space (FP). Let (X1nY1n)n0 be the associated random walk onRRN starting from (00) and de ned byX01 = 0Y01 = 0 and X1n =X1++XnY1n =Y1++Ynforn 1 the distribution of the couple (X1nY1n) is the nth power of convolution p n of the measure p. Denote byFnthe -algebra generated by (X1Y1):::(XnYn)n 1.
Let us rst recall the
Denition 2.1.
Let p be a probability measure on Rk, k 1. The measure p is aperiodic on Rk if there is no closed and proper subgroup H of Rk and no 2Rk such that p(+H) = 1.Denote by ^p the characteristic function of p de ned by ^p(uv) =
EeiuX1+ihv Y1i] for any (uv) 2 R RN. Recall that the probability measure p is aperiodic if and only if jp^(uv)j<1 for (uv)6= (00).
For any A R RN let fTA(k)gk0 be the the successive times of visit of the random walk (X1nY1n)n1 to the set A one has TA(0) = 0TA(1) = inffn 1 : (X1nY1n)2 Ag and TA(k+1) = inffn TA(k)+ 1 : (X1nY1n)2Ag. Note that theTA(k) are stopping times with respect to the ltrationfFngn1. We will associate to (pA) the transition kernel PA de ned by
PA((xy)B) =
Z
RR N
1
Ac\B(x+x0y+y0)p(dx0dy0)for any Borel set B in R RN note that for any k 1 one has PAk((00)B) = ETA > k](X1kY1k)2B]. In order to simplify the no- tations we will set T; = TR;RNP; = PR;RN and T;(k) = TR(k;)RN similar notations will hold, with obvious modi cations, when A =
R
;
RNR+ RN and R + RN.
Troughout this paragraph, for any k 1, we denote by k the Lebesgue measure on Rk. Furthermore, for any > 0, H(Rk) is the space of C-valued functions 'on Rk such that
xsup2Rk(1 +kxk)kj'(x)j<+1:
2.1. Preliminaries.
The local limit theorem gives the asymptotic behaviour of the se- quence fp n(')gn1 for any continuous function ' with compact sup- port onRN+1 we state it here and we precise some uniform upperbound for the sequencefp n(')gn1 when'belongs toH(RN+1) with > 4.
Theorem 2.2.
Assume that:i) the common distribution p of the variables (XnYn), n 1, is aperiodic on RN+1,
ii) EjX1j2+kY1k2]<+1 and EX1] = 0, EY1] = 0. Then:
i) for any continuous function ' with compact support on RN+1 one has
n!lim+1n(N+1)=2E'(X1nY1n)]
= 1
(2)(N+1)=2pjCj
Z
R N+1
'(xy)1(dx)N(dy) wherejCjdenotes the determinant of the positive denite quadratic form
C(uv) =E(uX1 +hvY1i)2]:
ii) For any function ' in H(RN+1) with > 4, the sequence
fn(N+1)=2E'(x+X1ny+Y1n)]gn1 is bounded uniformly in (xy) 2
R RN.
Proof. The rst assumption is the classical local limit theorem. To obtain the second claim, x a non negative function whose Fourier transform has a compact support K(^). Recall that
p^(uv) = 1; 1
2C(uv)(1 +"(uv))
with lim(u v)!(0 0)"(uv) = 0 so there exists > 0 such that for juj+
kvk< one has
jp^(uv)j1; 1
4 C(uv)e;C(u v)=4:
On the other hand, by the aperiodicity ofp there exists=(pK(^)) such thatjp^(uv)j as soon as (uv) belongs toK(^) andjuj+kvk . It follows that
(2 n)(N+1)=2E(X1nY1n)]
n(N+1)=2Z
juj+kvk<j^(uv)jjp^(uv)jn1(du)N(dv) +n(N+1)=2Z
juj+kvkj^(uv)jjp^(uv)jn1(du)N(dv)
n(N+1)=2Z
juj+kvk<n(N+1)=2 ^ u
pn v
pn
e;(n=4)C(u=pn v=pn)
1(du)N(dv) +n(N+1)=2nk^k1
k^k1Z
RR N
e;C(u v)=41(du)N(dv) +n(N+1)=2nk^k1 : Now setx y(x0y0) =(x+x0y+y0) for any (xy)2RRN and note that ^x y(uv) = eiux+ihv yi^(uv) the functions ^x y and ^ thus have the same compact support and satis es the equalities k^x yk1 = k^k1
and k^x yk1 =k^k1. For any (xy)2R RN one thus has
j(2 n)(N+1)=2Ex y(X1nY1n)]j
k^k1Z
RR N
e;C(u v)=41(du)N(dv) +n(N+1)=2nk^k1 : The assertion ii) thus holds for any function whose Fourier transform has a compact support. To achieve the proof of ii) it suces to show that for any function'in H(RN+1) with >4 there exists a function whose Fourier transform has a compact support and j'j . It is an immediate consequence of the following result we thank here J. P.
Conze for helpfull discussions about this fact.
Lemma 2.3.
Seth"(x) = 1 1 +jxj4+"
for any x2 R. If " >0 there exists a function h" greater than h" and whose Fourier transform has a compact support in R.
Proof. Set
h"(x) =Csin2x
x2 + sin2x x2
for some and C in R + which will depend on ". Assume =2 Q, the function h" is strictly positive on R it thus suces to show that there exists =2Q such that
x!lim+1x2+"(sin2x+ sin2(x)) = +1:
If such a real did not exist, then for any =2 Q there should exist a sequencefxngn1 which tends to +1and a constantC" >0 such that for all n 1,
sin2xn+ sin2(xn) C x2+n " :
So there should exist two strictly increasing sequences of integers
fkngn1 and flngn1 such that
jxn;knj C0
x1+n "=2 jxn;lnj C0 x1+n "=2
which implies
; ln
kn C00 k2+n "=2
for some positive constants C0 and C00. This leads to a contradiction because for almost all2R (with respect with the Lebesgue measure), this last inequality has at most a nite number of solutions in N2 2].
The lemma is proved.
2.2. A local limit theorem for a killed random walk on a half space.
In 7], we proved the following
Theorem 2.4.
Let the hypotheses of Theorem 2:2 hold. Then for any continuous function with compact support ' on R; we haven!lim+1n3=2ET+ > n]'(X1n)] = 1 (X1)p2
Z 0
;1
'(x);1 U ;(dx) where ;1 denotes the restriction of the Lebesgue measure on R; and U ; is the -nite measure on R; dened by
U ;(B) =X+1
k=1
E
1
B(X1T;(k ))]for any Borel set B. In the same way, one has
n!lim+1n3=2ET + > n]'(X1n)] = 1 (X1)p2
Z 0
;1
'(x);1 U;(dx)
where U; is the -nite measure on R; dened by U;(B) =+X1
k=1
E
1
B(X1T;(k ))]for any Borel set B.
Recall that the random walksfX1T;(k )gk1andfX1T;(k )gk1are tran- sient on R; it follows that the seriesP+k=01ET+ > k]'(x+X1k)] and
P+1
k=0ET + > k]'(x+X1k)] do converge. Furthermore one has
+1
X
k=0
ET+ > k]'(x+X1k)] =
Z 0
;1
'(x)U ;(dx)
and +
1
X
k=0
ET + > k]'(x+X1k)] =Z 0
;1
'(x)U;(dx): Let us now state the following
Theorem 2.5.
Let the hypotheses of Theorem 2:2 hold. Then:i)For any continuous function'with compact support onR;RN one has
n!lim+1n(N+3)=2ET+ > n]'(X1nY1n))]
= 1
(2)(N+1)=2pjCj
Z
R
;
R N
'(xy);1 U ;(dx)N(dy): ii) For any continuous function f with compact support on R and any g in H(RN) with >4, the sequence
fn(N+3)=2ET+ > n]f(X1n)g(y+Y1n)]gn1
is bounded, uniformly in y2RN. In the same way, one has
n!lim+1n(N+3)=2ET + > n]'(X1nY1n)]
= 1
(2)(N+1)=2pjCj
Z
R
;
R N
'(xy);1 U;(dx)N(dy)
and the sequence
fn(N+3)=2ET+ > n]f(X1n)g(y+Y1n)]gn1
is bounded, uniformly in y2RN.
Proof. We prove this theorem by induction overN. Theorem 2.2 deals with the caseN = 0 we will suppose that this result hold for someN 0 and we consider a sequence (XnYnZn)n1 of independent identically distributed random variables on R RN R. By a classical argument in probability theory, it suces to show the above convergence hold for '(xyz) = eax
1
R;(x)(y)(z) where a 2 R + and are C- valued functions whose Fourier transform are continuous with compact supports. By the inverse Fourier transform one hasIn=ET+ > n]eaX1n(Y1n)(Z1n)]
= 1
(2)(N+1)=2
Z
R N
R
^(v) ^(w)n(avw)N(dv)1(dw) with n(avw) = ET+ > n]eaX1n+ihv Y1ni+iwZ1n].
The Spitzer's factorisation for random walks on R gives for all a >0, for alls 201
+1
X
n=0snET+ > n]eaX1n] = expX+1
n=1
sn
n EX1n <0]eaX1n]: Using the fact that R+ RN+1 and R ; RN+1 are semi-groups in
RN+2, Ch. Sunyach extended this factorisation to the multidimension- nal case (9, Corollary 3, p. 553 and Theorem 5, p. 556]) for anya > 0, v2RN, w2R and s201 one thus has
+1
X
n=0snET+ > n]eaX1n+ihv Y1ni+iwZ1n]
= exp+X1
n=1
sn
n EX1n <0]eaX1n+ihv Y1ni+iwZ1n] that is
(n+ 1)n+1(avw) =Xn
k=0n+1;k(avw)k(avw)
with n(avw) = EX1n <0]eaX1n+ihv Y1ni+iwZ1n]. Finally In= 1n+ 1
n
X
k=0In k
with
In k = 1 (2)(N+1)=2
Z
R N
R
n+1;k(avw)k(avw)
^(v) ^(w)N(dv)1(dw): Set
I = 1
(2)(N+2)=2pjCj
Z
R N
R
+1
X
k=0
E
hT+ > k]eaX1k a
i
(y)(z)N(dy)1(dz) since
I =;1 U ;(ea)N()1()
it suces to show that fn(N+4)=2Ingn1 converges to I, that is 1) for allk >0, limn
!+1n(N+2)=2In k =I k, 2) +X1
k=0
jI kj<+1 and X+1
k=0I k=I, 3) limsupl
!+1 limsupn
!+1 n(N+2)=2Xn
k=l
jIn kj= 0.
To prove the assertion 1), note that
In k =ET+ > k]\Xkn+1+1 >0]eaX1n+1(Y1n+1)(Z1n+1)]
by the local limit theorem onRN+2 the assertion 1) follows with I k = 1
2(N+2)=2pjCj
ET+ > k]eaX1k] a
Z
N
(y)N(dy)
Z
(z)1(dz):
The fact that the series +k=0jI kj converges is a direct consequence of Theorem 2.4. To prove the assertion 3), note that
jIn kjET+ > k]\Xkn+1+1 <0]eaX1n+1j(Y1n+1)jj(Z1n+1)j]
E
hT+ > k]eaX1kZ
R
;
R N
R
eaxj(y+Y1k)j
j(z+Z1k)jp (n+1;k)(dxdy dz)i
C(a)
(n+ 1;k)(N+2)=2 ET+ > k]eaX1k] by Theorem 2.2.ii)
C1
(n+ 1;k)(N+2)=2k3=2 by Theorem 2.4.
On the other hand
jIn kjkk1Z
R
;
R N
R
ET+ > k]eaXk1j(y+Y1k)j
eaxp (n+1;k)(dxdy dz)]
kk1C(a) k(N+3)=2
EXkn+1+1 <0]eaXk +1n+1] by hypothesis of induction
C2
k(N+3)=2pn+ 1;k :
The assertion 3) follows since for any " >0 one has n(N+2)=2Xn
k=l
jIn kjC1
n(1X;")]
k=l
n(N+2)=2
k3=2(n+ 1;k)(N+2)=2 +C2 Xn
n(1;")+1]
n(N+2)=2 k(N+3)=2pn+ 1;k
C1
"(N+2)=2
n(1X;")]
k=l
k31=2
+ C2
pn(1;")(N+3)=2
n
X
n(1;")+1]
1
pn+ 1;k
C"(N+2)1=2pl +
p"
(1;")(N+3)=2
: Since " is arbitrarily small, the assertion 3) follows.
The proof of ii) is also made by induction overN. Ifg2H(RN+1) there exist2H(RN) and 2H(R1) such thatjgj. We set
In(yz) =ET+ > n]eaX1n(y+Y1n)(z+Z1n)]
and we have
In(yz) = 1n+ 1
n
X
k=0In k(yz) with
In k(yz) =ET+ > k]\Xkn+1+1 <0]eaX1n+1(y+Y1n+1)(z+Z1n+1)]: As above, one has
jIn k(yz)jinfn C1
(n+ 1;k)(N+2)=2k3=2 C2
k(N+3)=2pn+ 1;k
o
which proves that the sequence
nn(N+2)=2Xn
k=0
jIn k(yz)jon
1
is uniformly bounded in yz. This achieves the proof of ii).
The convergence of the sequence
fn(N+3)=2ET + > n]'(X1nY1n)]gn1
is obtained with similar arguments.
2.3. Behaviour of the process
((X1nmaxf0X11:::X1ngY1n))n0.
For any n 0 set X1n = maxf0X11:::X1ng and let Tn be the random variable de ned on (FP) by Tn = inff0 k n : X1n =
X1kg for any continuous function 'with compact support onRN+1 we have
E'(X1nX1n;X1nY1n)]
= Xn
k=0
ETn =k]'(X1k;Xnk+1Y1n)]
= Xn
k=0
E0 < X1kX11 < X1k:::X1k;1 < X1k
X1k+1 X1k:::X1nX1k]'(X1k;Xnk+1Y1n)]
= Xn
k=0
EX11 >0:::X1k >0]\Xkk+1+1 0:::Xnk+1 0]
'(X1k;Xnk+1Y1n)]: One obtains the following factorisation
E'(X1nX1n;X1nY1n)] =Xn
k=0Jn k(') with
Jn k(')
=
Z
R N+1
'(x;x0y+y0)P;k((00)dxdy)Pn+;k((00)dx0dy0): The behaviour of the process (X1nX1n;X1nY1n) is thus closely related to the one of the iterates of the transition kernels P; and P +. Using this factorisation one proves the
Theorem 2.6.
Suppose that the hypotheses of Theorem 2:2 hold.Then, for any continuous function with compact support on R+
R+ RN the sequence
fn(N+3)=2E'(X1nX1n;X1nY1n)]gn1
converges to (2)(N+1)1=2pjCj
Z
R +
R +
R N
'(s;ty)U +(ds);1 U;(dt)N(dy)
+ 1
(2)(N+1)=2pjCj
Z
R +
R +
R N
'(s;ty)+U +(ds)U;(dt)N(dy): Furthermore, for any continuous function f with compact support on
R+ R+ and any g in H(RN), the sequence
fn(N+3)=2Ef(X1nX1n;X1n)g(y+Y1n)]gn1
is bounded, uniformly in y2RN.
Proof. We only proof the rst assertion the second one may ob- tained with obvious modi cations as in Theorem 2.5. Set '(xty) = '1(x)'2(t)'3(y) where '1'2 and '3 are continuous with compact support. Fixk 0 by Theorem 2.5, the sequence
nn(N+3)=2Z
R
;
R N
'2(x0)'3(y+y0)Pn+;k((00)dx0dy0)on
1
is bounded uniformly in y2RN and converges to (2)(N+1)1=2pjCj
Z 0
;1
'2(;t);1 U;(dt)N('3):
By the dominated convergence theorem, one thus obtains, for any xed i 1
n!lim+1n(N+3)=2Xi
k=0Jn k(')
= 1
(2)(N+1)=2pjCj
i
X
k=0
ET; > k]'1(X1k)]
Z 0
'2(;t);1 U;(dt)N('3):