2 Dévissage, Convergence, Equivariance and Regularity Conditions
Tải bản đầy đủ - 0trang
Dévissage of a Poisson Boundary Under Equivariance and Regularity Conditions
203
2.3.1 On the Dévissage Condition
The starting point of the dévissage method is that the state space E of the original
diffusion can be written as X G (resp. X G=K) in an appropriate coordinate
system, where the corresponding first projection .xt /t 0 is a sub-diffusion of
.xt ; gt /t 0 (resp. .xt ; yt /t 0 ). This “splitting property” occurs in a large number of
situations, in particular when considering diffusion processes on manifolds that
show some symmetries.
For example, any left invariant diffusion .zt /t 0 with values in a semi-simple Lie
group H can be decomposed in Iwasawa coordinates as zt D nt at kt where nt 2 N,
at 2 A, kt 2 K take values in Lie subgroups and .kt /t 0 and .at ; kt /t 0 are subdiffusions. In other words, the state space can be decomposed as the product of
X D A K and G D N. Under some regularity conditions (see e.g. [11]), it can
be shown that the Poisson boundary of the sub-diffusion .at ; kt / is trivial and that nt
converges almost-surely to a random variable n1 2 N when t goes to infinity. Thus,
our results ensure that the Poisson boundary of the full diffusion .zt /t 0 is generated
by the single random variable n1 .
Another typical situation where the dévissage condition is fulfilled is the case
of standard Brownian motion on a Riemannian manifold with a warped product
structure, a very representative example being the classical hyperbolic space Hd seen
in polar coordinates .r; Â/ 2 RC Sd 1 , i.e. X D RC and G=K D SO.d/=SO.d
1/. In that case, the radial component .rt /t 0 is a one-dimensional transient subdiffusion whose Poisson boundary is trivial and the angular component .Ât /t 0 is a
time-changed spherical Brownian motion on Sd 1 that converges almost surely to a
random variable Â1 2 Sd 1 . Again, the dévissage method ensures that the Poisson
boundary of the full diffusion is generated by the single random variable Â1 . This
example generalizes to the case of a standard Brownian motion on a rotationally
symmetric manifold, see Sect. 4.1.
The hypothesis that the first projection .xt /t 0 is a sub-diffusion of the full diffusion .xt ; gt /t 0 (resp. .xt ; yt /t 0 ) is convenient and easy to check when considering
examples. Nevertheless it is not necessary in the sense that there are cases where
the couple .xt ; gt /t 0 does not a priori satisfy the dévissage condition, but where a
simple change of coordinates allows to implement the method, see Remark 1 below
for such an example.
Finally, remark that the absolute continuity condition required when Inv..xt /t 0 /
is non-trivial, is ensured for example if the infinitesimal generator of the diffusion
process .xt /t 0 is hypoelliptic. Moreover, without loss of generality, we can suppose
in that case that the measure on .S; G / is a probability measure, see [10].
2.3.2 On the Equivariance Condition
The main hypothesis that allows to implement the dévissage scheme is the third
one i.e. the equivariance condition. To emphasize its role, let us first consider the
following example where the diffusion process .xt ; gt /t 0 with values X G D R R
204
J. Angst and C. Tardif
is solution of the stochastic differential equations system:
dxt D dt C e
x2t
dBt ;
dgt D e
xt
dt;
.x0 ; g0 / 2 R
R;
(1)
where .Bt /t 0 is a standard real Brownian motion. The infinitesimal generator L of
the diffusion is hypoelliptic, so that Hypothesis 4 is fulfilled. Naturally, the process
.xt /t 0 is a one dimensional sub-diffusion of .xt ; gt /t 0 and from the Lemma 1 below,
Hypothesis 1 is also fulfilled.
Lemma 1 There exists a process .ut /t 0 that converges P.x;g/ -almost surely to a
random variable u1 in R when t goes to infinity such that for all t 0
xt D x0 C t C u t :
Moreover, the invariant sigma field Inv..xt /t 0 / is trivial.
Rt
2
Proof For all t
0, we have xt D x0 C t C ut ; where ut WD 0 e xs dBs :
R t 2x2
The martingale ut satisfies huit D 0 e s ds Ä t so that from the law of iterated
logarithm, we have almost surely xt
t=2 for t sufficiently large. In particular,
hui1 < C1 almost surely and ut is convergent. Since xt goes almost surely
to infinity with t, standard shift-coupling arguments apply and we deduce that
Inv..xt /t 0 / is trivial. Note however that the tail sigma field of .xt /t 0 i.e. the
invariant sigma field of the space-time process Inv..t; xt /t 0 / is not trivial. Indeed,
x0 C u1 D limt!C1 .xt t/ is a non-trivial shift invariant random variable.
Rt
From Lemma 1 again, the second projection gt D g0 C 0 e xs ds converges P.x;g/ almost surely when t goes to infinity to a random variable g1 in R and Hypothesis 2
is satisfied. Finally, considering the action of G D .R; C/ on itself by translation,
Hypothesis 3 is also satisfied since, for f 2 C.R R; R/ and .x; g; h/ 2 R3 we have
1
L .h f /.x; g/ D .@x f /.x; g C h/ C e
2
x2
.@2x f /.x; g C h/ C e x .@g f /.x; g C h/
D h .L f /.x; g/:
Hence, from Theorem A, the invariant sigma field Inv..xt ; gt /t 0 / coincide with
Inv..xt /t 0 /_ .g1 / D .g1 / up to P.x;g/ -negligeable sets i.e. the dévissage scheme
applies. Let us now consider a very similar process, namely the diffusion process
.xt ; gt /t 0 with values X G D R R which is solution of the new following
stochastic differential equations system:
dxt D dt C e
x2t
dBt ;
dgt D gt dt;
.x0 ; y0 / 2 R
R;
(2)
where .Bt /t 0 is again a standard real Brownian motion. With a view to apply the
dévissage method, the context seems favorable because the infinitesimal generator
L of the diffusion is hypoelliptic, .xt /t 0 is a one dimensional sub-diffusion of
Dévissage of a Poisson Boundary Under Equivariance and Regularity Conditions
205
.xt ; gt /t 0 , and gt D g0 e t converges (deterministically) to g1 D 0 when t goes to
infinity. In particular, the sigma field Inv..xt /t 0 / _ .g1 / is trivial. Nevertheless,
we have the following proposition:
Proposition 1 Let .x; g/ 2 R R with g Ô 0, then the sigma field Inv..xt ; gt /t 0 /
differs from Inv..xt /t 0 / _ .g1 / by a P.x;g/ -non-negligeable set.
Proof If g Ô 0, the sigma field Inv..xt ; gt /t 0 / is not trivial under P.x;g/ because the
process xt C log.jgt j/ converges P.x;g/ -almost surely to x0 C log.jg0 j/ C u1 which,
from the proof of Lemma 1, is a non-trivial invariant random variable.
The reason for which the dévissage method does not apply in this last example
is that Hypothesis 3 i.e. the equivariance condition is not fulfilled. Indeed, the
generator of the full diffusion writes
1
L D @x C e
2
and in general, for f 2 C.R
x2 2
@x
g@g ;
R; R/ and .x; g; h/ 2 R3 we have
1
L .h f /.x; g/ D .@x f /.x; g C h/ C e
2
Ô
1
h .L f /.x; g/ D .@x f /.x; g C h/ C e
2
x2
.@2x f /.x; g C h/
g.@g f /.x; g C h/
x2
.@2x f /.x; g C h/
.g C h/.@g f /.x; g C h/:
Remark 1 The equivariance condition is relatively strong and forces .xt /t 0 to be a
sub-diffusion (which is already supposed in Hypothesis 1). Indeed, since a function
f W X G ! R does not depend on its second variable if and only if g f D f for all
g 2 G, the equivariance condition implies that L maps C1 .X/ onto C1 .X/ (and
thus .xt / is a sub-diffusion). Nevertheless, some cases where this assumption is not
fulfilled can be solved by the dévissage method. For example, consider the diffusion
process .xt ; gt /t 0 solution of following system of stochastic differential equations
8
ˆ
ˆ
< dxt D
ˆ
ˆ
: dgt D
xt g2t
x2t Cg2t
Á
C gt dt C gt dBt ;
(3)
g3t
dt;
x2t Cg2t
where, clearly, there is no equivariance. It is yet possible to show that, almost
surely, xt escapes to infinity with t, gt converges to a random variable g1 and that
Inv..xt ; gt /t 0 / D .g1 / almost surely. Indeed the invariant sigma-field of .xt ; gt /
coincides with the one of .ut ; vt / WD .xt =gt ; log.gt // (since the map is bijective). But
206
J. Angst and C. Tardif
now .ut ; vt / solves the system
8
ˆ
< dut D dt C dBt ;
ˆ
: dvt D
(4)
1
dt;
1Cu2t
and one can easily check that, for this new diffusion, all the hypotheses of the
dévissage method are now fulfilled. Therefore, applying Theorem A, we can
conclude that Inv..xt ; gt /t 0 / D Inv..ut ; vt /t 0 / D .v1 / D .g1 /.
2.3.3 On the Regularity Condition
As already noticed in the examples of the last section, the regularity condition is
automatically satisfied for a large class of diffusion processes, namely when the
infinitesimal generator L is elliptic or hypoelliptic. The role of this assumption will
be clear at the end of the proof of Theorems 1 and 2, since it allows to go to the limit
in the regularization procedure. In a more heuristical way, the regularity condition
can be seen as a mixing hypothesis which prevents pathologies that may occur when
considering foliated dynamics.
To be more precise on the kind of pathologies we have in mind, consider
the following discrete and deterministic example that was suggested to us by S.
Gouëzel. The underlying space is the product space X Y D S1 S1 where S1 is
identified to R=Z. Fix ˛ … Q, and define the transformation T W X Y ! X Y
such that T.x; y/ WD .x C ˛; y/. Now let X.x; y/ WD x and Y.x; y/ WD y be the first and
second projections and for n 0 define Xn WD X ı T n i.e. Xn .x; y/ D .x C n˛; y/ and
Yn WD Y ı T n Á Y. In this discrete time context, the resulting sequence .Xn ; Yn /n 0
plays the role of .xt ; yt /t 0 in the framework described in Sect. 2.1. The dynamics of
.Xn / does not depend on .Yn /, which is constant, and thus converges when n goes to
infinity. It is thus natural to ask if the devissage method applies or not in this context.
The answer is negative in general. To see this, for y 2 S1 , consider the probability
measure
y
WD C
X
n2Z
1
ıyCn˛ ;
1 C n2
where C is a normalizing constant and define a measure P on X
ÄZ
Z
Z
h.x; y/P.dx; dy/ WD
X Y
y2S1
Â
x2S1
h.x; y/
1
1
y .dx/ C
2
2
Y such that
Ã
yC1=2 .dx/
dy:
R
Note that the first marginal PX . / D Y P. ; dy/ of P is the Lebesgue measure hence
the invariant sigma field Inv..Xn /n 0 / is trivial under P. Since Y is T-invariant, the
invariant sigma field Inv..Yn /n 0 / is composed of events that do not depend on the
Dévissage of a Poisson Boundary Under Equivariance and Regularity Conditions
207
first coordinate x. Thus, under P, the sigma field Inv..Xn /n 0 / _ Inv..Yn /n 0 / is only
composed of events that do not depend on the first coordinate. Now consider the
sets A WD f.y C n˛; y/; y 2 S1 ; n 2 Zg and B WD f.y C 1=2 C n˛; y/; y 2 S1 ; n 2 Zg.
Both sets are invariant by the dynamics but they do depend on the first coordinate.
Hence, the global invariant sigma field Inv..Xn ; Yn /n 0 / differs from Inv..Yn /n 0 /
by some P-non negligeable events and the dévissage method does not apply here:
Inv..Xn ; Yn /n 0 / Ô Inv..Xn /n 0 / _ Inv..Yn /n 0 /:
3 Proof of the Main Result
We now give the proof of our results. To highlight the main ideas behind the proof,
we first consider the simplest case when the invariant sigma field of .xt /t 0 is trivial
and when Y D G is a finite dimensional Lie group. Then, we extend the result in the
case where the invariant sigma field of .xt /t 0 is non-trivial and finally, we consider
the homogeneous case.
3.1 Starting from a Trivial Poisson Boundary
Let us first prove the following result:
Theorem 1 Suppose that the full diffusion .xt ; gt /t 0 satisfies Hypotheses 1–4.
Suppose moreover that for all .x; g/ 2 X G, the invariant sigma field Inv..xt /t 0 /
is trivial for the measure P.x;g/ . Then the two sigma fields
Inv..xt ; gt /t 0 / and
.g1 /
coincide up to P.x;g/ -negligeable sets. Equivalently, if H is a bounded L -harmonic
function, then there exists a bounded measurable function on G such that H can
be written as H.x; g/ D E.x;g/ Œ .g1 /, for all .x; g/ 2 X G.
Proof (Proof of Theorem 1) The first step of the proof is the following lemma,
which is valid under Hypotheses 1–4 (the triviality of Inv..xt /t 0 / is not required
here). From Hypothesis 2, for all .x; g/ 2 X G, the process .gt /t 0 converges
P.x;g/ -almost surely to a random variable g1 D g1 .!/ in G.
Lemma 2 Under Hypotheses 1–4, and for all starting points .x; g/ 2 X G and
h 2 G, the law of the process h:.xt ; gt /t 0 D .xt ; h:gt /t 0 under P.x;g/ coincides with
the law of .xt ; gt /t 0 under P.x;h:g/ . In particular,
1. the law of the limit g1 under P.x;h:g/ is the law of h:g1 under P.x;g/ ;
2. for all .g; g0 / 2 G2 , the push-forward measures of both P.x;g/ and P.x;g0 / under
the measurable map ! D .! X ; ! G / 7! hg11 :! D .! X ; hg11 .!/:! G / coincide.
208
J. Angst and C. Tardif
Proof (Proof of Lemma 2) The result is an direct consequence of the equivariance
Hypothesis 3. Indeed, if f 2 C1 .X G; R/ is compactly supported, from Itô’s
formula, under P.x;g/ we have for all h 2 G:
Z
f .xt ; h:gt / D .h f /.xt ; gt / D .h f /.x; g/ C
Z
D f .x; h:g/ C
0
L .h f / .xs ; gs /ds C Mt
t
h .L f / .xs ; gs /ds C Mt
0
Z
D f .x; h:g/ C
t
t
0
.L f / .xs ; h:gs /ds C Mt ;
where Mt is a martingale vanishing at zero. Otherwise, under P.x;h:g/ we have:
Z
f .xt ; gt / D f .x; h:g/ C
t
0
.L f /.xs ; gs /ds C Nt ;
where Nt is again a martingale vanishing at zero. In other words, under P.x;g/
and P.x;h:g/ respectively, both processes h:.xt ; gt /t 0 and .xt ; gt /t 0 solve the same
martingale problem, hence their laws coincide.
Let us go back to the proof of Theorem 1. From Hypothesis 2, for all starting
points .x; g/ 2 X G, the process .gt /t 0 converges P.x;g/ -almost surely to a random
variable g1 D g1 .!/ in G. We define
.x;g/
˝0
WD f! 2 ˝; lim gt .!/ existsg;
t!C1
.x;g/
.x;g/
and consider gQ 1 such that gQ 1 WD g1 on ˝0 and gQ 1 WD IdG on ˝n˝0 . Let
H be a bounded L -harmonic function. By the standard duality between bounded
invariant random variables and bounded harmonic functions, see e.g. Proposition 3.4
p. 423 of [14], there exists a bounded variable Z W ˝ ! R which is measurable with
respect to Inv..xt ; gt /t 0 /, i.e. Z is F 1 -measurable and satisfies Z.Âs !/ D Z.!/ for
all ! 2 ˝, such that for all .x; g/ 2 X G:
H.x; g/ D E.x;g/ ŒZ:
Moreover, .x; g/ 2 X
G being fixed, for P.x;g/ -almost all paths !, we have:
Z.!/ D lim H.xt .!/; gt .!//:
t!C1
The first idea here is to use the Lie group structure to condition the diffusion to
escape at a prescribed point in G. Remark that standard conditioning methods such
as Doob h-transform can not be implemented here since the law of the limit g1 is
Dévissage of a Poisson Boundary Under Equivariance and Regularity Conditions
209
not known a priori. For h 2 G, consider the random variable
Z h .!/ WD Z.h:Qg11 :!/ D Z.! X ; hQg1 .!/ 1 ! G /:
This new variable Z h can be seen as modification of the initial variable Z so that
the value of Z h .!/ is the value of Z but conditioned by the event that the G-valued
component ! G of the sample path ! does not exit at the random point g1 .!/ but
at the fixed point h. This variable is again Inv..xt ; gt /t 0 /-measurable. Indeed, since
the constant function equal to h and the random variable Z are shift-invariant, we
have
Z.h:Qg11 .Âs !/:Âs !/ D Z.Âs .h:Qg11 :!// D Z.h:Qg11 :!/:
Since Z h is bounded and is measurable with respect to Inv..xt ; gt /t 0 /, the function
.x; g/ 7! E.x;g/ ŒZ h is also a bounded L -harmonic function. But from the second
point of Lemma 2, for all starting points .x; g; g0 / 2 X G2 , we have
E.x;g/ ŒZ h D E.x;g0 / ŒZ h :
In other words, the harmonic function .x; g/ 7! E.x;g/ ŒZ h is constant in g and its
restriction to X is L X -harmonic, where L X denotes the infinitesimal generator of
the sub-diffusion .xt /t 0 . Since Inv..xt /t 0 / is supposed to be trivial, we deduce that
the function .x; g/ 7! E.x;g/ ŒZ h is constant. In the sequel, we will denote by .h/
the value of this constant. Note that h 7! .h/ is a bounded measurable function
since h 7! Z h is. The resulting function
is precisely the one appearing in the
statement of Theorem 1. By construction, .h/ is the common value, for all starting
points .x; g/ 2 X G, of E.x;g/ ŒZ h , i.e. the expectation of Z “conditioned” by the
event that ! G exit in h instead of g1 .!/. The second step of the proof consists in
considering a “smooth version” of the map h 7! Z h , that will allow us to deal with
non-countable union of negligeable sets, which is necessary if we want to mimic
the above approach replacing h by g1 .!/. So let us introduce an approximate unity
. n /n 0 on G, fix g 2 G, n 2 N and consider the “conditioned and regularized”
version Z, namely:
Z
Z h .!/ n .gh 1 / .dh/:
Z g;n .!/ WD
G
The exact same reasoning as above shows that Z g;n is bounded and measurable
with respect to Inv..xt ; gt /t 0 / so that the function .x; g/ 7! E.x;g/ ŒZ g;n is constant.
Hence, for all g 2 G, n 2 N and .x; g/ 2 X G, there exists a set ˝ g;n;.x;g/
˝
such that P.x;g/ .˝ g;n;.x;g/ / D 1 and such that for all paths ! in ˝ g;n;.x;g/ , we have:
Z g;n .!/ D lim E.xt .!/;gt .!// ŒZ g;n D E.x0 .!/;g0 .!// ŒZ g;n D E.x;g/ ŒZ g;n :
t!1
210
J. Angst and C. Tardif
Let D be a countable dense set in G and consider the intersection
\
˝ .x;g/ WD
˝ g;n;.x;g/ :
g2D;n2N
We have naturally P.x;g/ .˝ .x;g/ / D 1 and for ! 2 ˝ .x;g/ :
8g 2 D; n 2 N;
Z g;n .!/ D E.x;g/ ŒZ g;n :
Since the above expressions are continuous in g, we deduce that the last inequality
is true for all g 2 G. In other words, we have shown that for all g 2 G and for all !
in ˝ .x;g/ :
Z
Z g;n .!/ D E.x;g/ ŒZ g;n D
.h/ n .gh 1 / .dh/:
G
In particular, taking g D gQ 1 .!/, we obtain that for all ! 2 ˝ .x;g/ and for all n 2 N:
Z gQ1 .!/;n .!/ D
Z
.h/ n .Qg1 .!/h 1 / .dh/:
(5)
G
Recall that the Haar measure
Z
gQ 1 .!/;n
is right invariant so that
Z
.!/ D
Z
1
Z.h:!// n .h 1 / .dh/;
Z .!/ n .g1 .!/h / .dh/ D
h
G
G
and
Z
Z
.h/ n .Qg1 .!/h 1 / .dh/ D
G
.hQg1 .!// n .h 1 / .dh/:
G
Thus, Eq. (5) is equivalent to
Z
Z
1
.hQg1 .!// n .h 1 / .dh/:
Z.h:!/ n .h / .dh/ D
G
G
Taking the integral in ! with respect to P.x;g/ on ˝ .x;g/ , we deduce that for all n 2 N:
Z
Z
1
E.x;g/ ŒZ.h:!/ n .h / .dh/ D
G
E.x;g/ Œ .hQg1 / n .h 1 / .dh/;
G
which, from Lemma 2 yields
Z
H.x; hg/ n .h 1 / .dh/ D
G
Z
E.x;hg/ Œ .Qg1 / n .h 1 / .dh/:
G
Dévissage of a Poisson Boundary Under Equivariance and Regularity Conditions
211
From Hypothesis 4, bounded L -harmonic functions are continuous, hence we can
let n go to infinity in the above expressions to get the desired result, namely:
H.x; g/ D E.x;g/ Œ .g1 /:
t
u
3.2 Starting from a Non-trivial Poisson Boundary
Let us now consider the general case when Inv..xt /t 0 / is not trivial but generated
by a random variable `1 with values in a separable measure space .S; G /. We will
prove the following result:
Theorem 2 Suppose that the full diffusion .xt ; gt /t 0 satisfies Hypotheses 1–4.
Then, for all starting points .x; g/ 2 X G, the two sigma fields
Inv..xt ; gt /t 0 / and
.`1 ; g1 /
coincide up to P.x;g/ -negligeable sets. Equivalently, if H is a bounded L -harmonic
function, there exists a bounded measurable function on S G such that H can
be written as H.x; g/ D E.x;g/ Œ .`1 ; g1 / for all .x; g/ 2 X G.
Proof The proof is very similar to the one of Theorem 1, but it requires an extra
argument to ensure the measurability of the function . So let H be a bounded L harmonic function and Z W ˝ ! R the associated bounded random variable which
is measurable with respect to Inv..xt ; gt /t 0 /. For g; h 2 G and n 2 N, we consider
the random variables
Z
Z h .!/ WD Z.h:Qg11 :!/;
Z g;n .!/ WD
Z h .!/ n .gh 1 / .dh/:
G
As in the proof of Theorem 1, the element h being fixed, the variable Z h is
bounded and Inv..xt ; gt /t 0 /-measurable, so that the function .x; g/ 7! E.x;g/ ŒZ h
is bounded and L -harmonic. From Lemma 2, this function is constant in g and its
restriction to X is thus L X -harmonic. Hence, there exists a bounded measurable
function h W S ! R such that
8.x; g/ 2 X
G; E.x;g/ ŒZ h D E.x;g/ Œ
h .`1 /:
(6)
By Hypothesis 1, the random variable `1 admits a density k with respect to the
reference probability measure on .S; G /, so that the last equation can be written
Z
8.x; g/ 2 X
G; E.x;g/ ŒZ h D
h .`/k.x; `/
.d`/:
212
J. Angst and C. Tardif
The difficulty here is that, a priori, the function .h; `/ 7! h .`/ is not measurable in
both variables. To deal with this difficulty, note that for any A 2 G , we have also:
Z
Ex;g Œ1`1 2A Z h D Ex;g Œ1`1 2A
h .`1 /
1A .`/
D
h .`/k.x; `/
.d`/:
(7)
Let us fix x0 2 X and define
Qh .d`/ WD
h .`/
E.x0 ;g/ ŒZ h
k.x0 ; `/ .d`/:
For each h, the measure Qh is absolutely continuous with respect to k.x0 ; `/ .d`/
and, by Eq. (7), the one parameter family .Qh /h is a measurable family of probability
measures. Recall that by Hypothesis 1, the measurable space .S; G / is separable,
thus Theorem 58 p. 57 of [6] applies and there exists a measurable map X W S G !
R such that X.:; h/ is a density of Qh with respect to k.x0 ; `/ .d`/, i.e. for all h 2 G
X.`; h/ D
h .`/
E.x0 ;g/ ŒZ h
almost all `:
for
The map h 7! E.x0 ;g/ ŒZ h being measurable, the function
Q .`; h/ WD X.`; h/E.x0 ;g/ ŒZ h
is also measurable and for all .x; g/ 2 X G, we have E.x;g/ ŒZ h D E.x;g/ Œ Q .`1 ; h/:
For all g 2 G, n 2 N and .x; g/ 2 X G, we thus have:
Z
E.x;g/ ŒZ h n .gh 1 / .dh/ D
E.x;g/ ŒZ g;n D
G
Z
E.x;g/ Œ Q .`1 ; h/ n .gh 1 / .dh/
G
ÄZ
Q .`1 ; h/ n .gh 1 / .dh/ :
D E.x;g/
G
Hence, .x; g/ 2 X
G being fixed, we obtain that P.x;g/ -almost surely
Z g;n D lim E.xt ;gt / ŒZ g;n
t!C1
ÄZ
Q .`1 ; h/ n .gh 1 / .dh/ D
D lim E.xt ;gt /
t!C1
G
Z
Q .`1 ; h/ n .gh 1 / .dh/:
G
In other words, .g; n; .x; g// being fixed, there exists a set ˝ g;n;.x;g/
measure i.e. P.x;g/ .˝ g;n;.x;g/ / D 1 such that for all ! 2 ˝ g;n;.x;g/
Z
Z g;n .!/ D
G
Q .`1 .!/; h/ n .gh 1 / .dh/:
˝ of full
Dévissage of a Poisson Boundary Under Equivariance and Regularity Conditions
If D a countable dense set in G, we get that for all ! 2 ˝ .x;g/ WD
Z
8g 2 D; n 2 N;
213
T
g2D;n2N
˝ g;n;.x;g/ :
Q .`1 .!/; h/ n .gh 1 / .dh/:
Z g;n .!/ D
G
The above expressions being continuous in g, we can take g D gQ 1 .!/ to get
8! 2 ˝ .x;g/ ; 8n 2 N;
Z gQ1 .!/;n .!/ D
Z
Z
Q .`1 .!/; h/ n .Qg1 .!/h 1 / .dh/
G
Q .`1 .!/; hQg1 .!// n .h 1 / .dh/:
D
G
Taking the expectation under P.x;g/ , the left hand side gives :
E.x;g/ ŒZ
gQ 1 ;n
Z ÄZ
D
˝
Z h .!/ n .Qg1 .!/h 1 / .dh/ P.x;g/ .d!/
G
Z ÄZ
Z hQg1 .!/ .!/ n .h 1 / .dh/ P.x;g/ .d!/
D
˝
G
Z ÄZ
Z.h:!/ n .h 1 / .dh/ P.x;g/ .d!/
D
˝
G
Z ÄZ
Z.!/ n .h 1 / .dh/ P.x;hg/ .d!/
D
˝
Z
G
E.x;hg/ ŒZ n .h 1 / .dh/
D
G
and the right hand side
ÄZ
Q .`1 ; hQg1 / n .h 1 / .dh/ D
E.x;g/
Z
G
E.x;g/ Q .`1 ; hQg1 /
n .h
1
E.x;hg/ Q .`1 ; gQ 1 /
n .h
1
/ .dh/
G
Z
D
/ .dh/:
G
Since L -harmonic functions are continuous, letting n go to infinity, we deduce
E.x;g/ ŒZ D E.x;g/ Œ Q .`1 ; gQ 1 /: