VOL.110 2008 NO.1
ERGODIC AUTOMORPHISMS WHOSE WEAK CLOSURE OF OFF-DIAGONAL MEASURES CONSISTS OF ERGODIC
SELF-JOININGS BY
Y.DERRIENNIC(Brest),K.FRCZEK(Toru«andWarszawa), M.LEMACZYK(Toru«andWarszawa)andF.PARREAU(Paris)
Abstra t. Basi ergodi properties of the ELF lass of automorphisms, i.e.of the lassofergodi automorphismswhoseweak losureofmeasuressupportedonthegraphsof iteratesof
T
onsistsofergodi self-joiningsareinvestigated.DisjointnessoftheELF lass with:2-foldsimpleautomorphisms,intervalex hangetransformationsgivenbyaspe ial typepermutationsandtime-onemapsofmeasurableowsisdis ussed.Allergodi Poisson suspensionautomorphismsaswellasdynami alsystemsdeterminedbystationaryergodi symmetriα
-stablepro essesareshowntobelongtotheELF lass.Introdu tion. The notion of disjointness between measure-preserving automorphisms of standard probability Borel spa es was introdu ed by Furstenberg [9 ℄ in 1967. Sin e then many results showing disjointness of some lasses have been proved (see e.g. [9 ℄, [12 ℄, [14 ℄, [19 ℄, [21 ℄, [26 ℄, [28 ℄, [29 ℄,[46 ℄, [47 ℄).
In[6 ℄ these ond and the third named authors of this paper introdu ed the notion of ELF (
1
) ow. An ELF ow is, by denition, an ergodi ow su hthatwhenwepasstotheweak losureofitstime-
t
maps onsideredas Markov operators of the underlyingL
2
-spa e, then all the weak limits are inde omposable Markov operators. The ELFpropertyis interesting only in thenon-mixing ase,andindeedin ontrastwiththisproperty,some lassi al weaklymixingbutnon-mixingspe ialowsoverirrationalrotationsor,more generally, over interval ex hange transformations turn out to have in the weak losure of Markov operators given by their time-
t
maps su iently de omposableMarkovoperators.Su howsareoftenspe ialrepresentations of some smooth ows on surfa es and a motivation to introdu e the ELF propertywastoprovedisjointness(inthesenseofFurstenberg)ofsu hows2000Mathemati sSubje tClassi ation:37A05,37A50. Keywordsandphrases:joinings,ELFproperty,disjointness.
Resear hpartiallysupportedbyKBNgrant1P03A03826andMarieCurieTransfer ofKnowledge program,proje tMTKD-CT-2005-030042(TODEQ).
(
1
) Thea ronymELF omesfromtheFren habbreviationofergodi itédeslimites faibles.
[81℄
from the ELF lass (see [6 ℄, [7 ℄, [8 ℄). In parti ular, some lassi al smooth weakly mixing ows on surfa es (e.g. onsidered in [25 ℄) turn out to be disjoint fromtheELF lass.
Ontheotherhand,theELFpropertywasalso introdu edinthehopeof expressingthefa tthatagivenowisofprobabilisti origin.Indeed,arst attempt to dene a system to be of probabilisti origin might be via the Kolmogorovgroup propertyof the spe trum. However, ea h weakly mixing systemhas anergodi extension whi h has theKolmogorovgroup property, simply by taking the innite dire t produ t of the system. Therefore this spe tralproperty istoo weak to singleout systemsof probabilisti origin. As noti ed in [6 ℄ Gaussian ows enjoy the ELF property (this result also follows from some earlier results of [29 ℄). The present paper and, indepen-dently,the PhD thesis of E.Roy[36 ℄ are a further onrmation of the fa t that dynami al systems whose origin are well-known lasses of stationary pro esses(see below)are insidetheELF lass. We also mentionthat inthe general ase,in luding mixing,another joiningproperty(satised for exam-plebyowswithRatner'sproperty[35 ℄)hasbeenintrodu edin[43 ℄toshow disjointness fromGaussian systems.
In this paper, instead of ows, we onsider the ELF property for au-tomorphisms. One of the main results of the paper states that all ergodi Poisson suspension automorphismsenjoytheELFproperty.Thisresultisa onsequen e of Theorem 1 belowsaying thatPoissonian joiningsof ergodi Poisson automorphisms remain ergodi ; the same result is also proved in there ent,independent paper[36 ℄.Moreover, we onsiderso alled
α
-stable automorphisms, i.e. ergodi automorphisms a ting on a spa e whose mea-surable stru ture is determined by an invariant real subspa e in whi h all variablesaresymmetriα
-stable (0 < α < 2
,forα = 2
we omeba kto the Gaussian ase). We prove (Theorem 3 below) thatα
-stable self-joinings of su hautomorphismsmustne essarily beergodi , fromwhi htheELF prop-ertydire tlyfollows.Intheaforementionedthesis[36℄,afurtherstepforward isevenmade:itisprovedthatgivenanergodi stationaryinnitelydivisible pro ess, ea h innitely divisible self-joining of the orresponding measure-preserving automorphismremains ergodi , andin parti ular we also obtain theELFpropertyinthismost general ase.Furthermore, we show (Proposition 12 below) that weakly mixing but non-mixing2-foldsimpleautomorphismsaredisjointfromtheELF lass.Itis alsoshownthatthetime-onemapsofows onsideredin[8℄aredisjointfrom any ELF automorphism, and therefore the time-one maps of Ko hergin's smoothows from[25 ℄ aredisjoint fromanyELFautomorphism.
Re ently, some attention has been devoted to joining properties of in-terval ex hange transformations(see e.g. [4℄,[5℄).Here we areableto prove
ingathree-intervalex hange transformationweobtaindisjointnessfromthe ELF lass. In fa t, this resultis a onsequen e of a more general statement provedinthepaper.Namelygiven
k ≥ 3
we onsiderspe ialpermutationsof{1, . . . , k}
andweprovethatfora.a. hoi esoflengthsofpartitionintervalsof[0, 1)
theresultingautomorphismsaredisjointfromallELFautomorphisms. Some results in this paper have been obtained during the visit of the fourth-named author at Ni olaus Coperni us University inSeptember 2003 and during the visit of the third-named author at Université de Bretagne O identale intheSpring 2004.1.Preliminaries
1.1. Fa tors, joinings and Markov operators. Assume that
T
is an er-godi automorphism of a standard probability Borel spa e(X, B, µ)
. The asso iated unitarya tion ofT
onL
2
(X, B, µ)
isgivenby
U
T
(f ) = f ◦ T
(but wewilloftenwriteT (f )
insteadoff ◦ T
).WedenotebyC(T )
the entralizer ofT
, thatis, the set of all automorphisms of(X, B, µ)
ommuting withT
. EndowedwiththestrongoperatortopologyofU (L
2
(X, B, µ))
the entralizer be omes a Polish group. Any
T
-invariant sub-σ
-algebraA ⊂ B
is alled a fa tor ofT
. The quotient a tion ofT
on the quotient spa e(X/A, A, µ|
A
)
will be denoted byT |
A
or even byA
if no onfusion arises. We say thatT
is rigid if the set{T
n
: n ∈ Z}
has an a umulation point in
C(T )
. It follows thatinthe rigidity ase the entralizer is un ountable and for some in reasing sequen e(q
n
)
,T
q
n
→ Id
.Automorphismswhi h have no rigidity at all are alledmildly mixing (see [11 ℄). More pre isely,
T
is alled mildly mixing if its only rigidfa tor isthe one-point fa tor.Assumenowthat
S
isanotherergodi automorphismofastandard prob-ability Borel spa e(Y, C, ν)
.By a joining ofT
andS
we mean anyT × S
-invariant measure̺
on(X × Y, B ⊗ C)
whose marginals̺
X
and̺
Y
satisfy̺
X
= ̺|
X
= µ
and̺
Y
= ̺|
Y
= ν
respe tively.ThesetofjoiningsbetweenT
andS
isdenoted byJ(T, S)
.WhenevertheautomorphismT × S
a ting on(X ×Y, B ⊗C, ̺)
(forshortwewillalsowrite(T ×S, ̺)
)isergodi ,thejoining̺
is alledergodi andthesetofergodi joiningsisdenotedbyJ
e
(T, S)
.The formula \X×Y
f ⊗ g d̺ =
\Y
Φ
̺
(f ) · g dν
establishesaone-to-one orresponden ebetweentheset
J(T, S)
and thesetJ (T, S)
ofallMarkovoperatorsfromL
2
(X, B, µ)
to
L
2
(Y, C, ν)
intertwining
U
T
andU
S
(seee.g. [42 ℄,[29℄for moredetails). Re all thatapositive linear operatorΦ : L
2
(X, B, µ) → L
2
(Y, C, ν)
is alled Markov if
Φ(1
X
) = 1
Y
andΦ
∗
(1
Y
) = 1
X
,and thenΦ = Φ
̺
where̺(A × B) =
TB
Φ(1
A
) dν
for measur-able setsA ∈ B
andB ∈ C
. The set of Markov operators is losed in the weak operator topology ofB(L
2
(X, B, µ), L
2
(Y, C, ν))
and
J(T, S)
are ompa t (on the latter set we transport the topology ofJ (T, S)
). Ergodi joinings orrespond to so alled inde omposable Markov operators,i.e.to theextremalpointsinthesetJ (T, S)
,whi hhas anatural stru tureofaChoquetsimplex.NotethattheMarkovoperator orrespond-ingtothe produ tmeasureµ ⊗ ν
equalsΠ
X,Y
(f ) =
T
X
f dµ
.Ifonemore er-godi automorphismR
on(Z, D, η)
isgivenandΦ
̺
∈ J (T, S)
,Φ
κ
∈ J (S, R)
thenΦ
κ
◦ Φ
̺
∈ J (T, R)
and the orresponding joining ofT
andR
will be denotedbyκ ◦ ̺
.Whenever
S = T
we will writeJ
2
(T )
andJ
e
2
(T )
instead ofJ(T, T )
andJ
e
(T, T )
respe tively.Note thatifW ∈ C(T )
thenthe formulaµ
W
(A × B)
= µ(A ∩ W
−1
B)
determines a self-joining, alled a graph joining, of
T
, and moreoverµ
W
∈ J
e
2
(T )
(forW = T
n
we speak about o-diagonal self-joinings). We saythat
T
is 2-fold simple (see [49 ℄, [21 ℄)if the only ergodi self-joinings ofT
are graph joinings or the produ t measureµ ⊗ µ
. The measureµ
Id
will alsobe denotedby∆
X
or∆
µ
.Wesaythat
T
isrelativelyweaklymixingwithrespe ttoafa torA ⊂ B
if theself-joiningλ
( alledtherelativelyindependent extensionof thediagonal measure onA
) given byλ(A × B) =
\
X/A
E(1
A
| A) · E(1
B
| A) d(µ|
A
)
is ergodi . If
A
1
⊂ A
is another fa tor andT |
A
is relatively weakly mixing overA
1
thenT
is still relatively weakly mixing overA
1
(forthis hainrule seee.g. [20 ℄).Following[9℄wesaythattwoergodi automorphisms
T
andS
aredisjoint ifJ(T, S) = {µ ⊗ ν}
. Re all thatJ
e
(T, S) = {µ ⊗ ν}
implies disjointness of
T
andS
. Given a lassR
of ergodi automorphisms, we denote byR
⊥
the lassofallergodi automorphismsdisjointfromanymemberof
R
.Then bya multiplier (see [12 ℄) ofR
⊥
we mean an ergodi automorphism ea h of whose ergodi joinings with an automorphism belonging to
R
⊥
gives rise to another member of
R
⊥
. The lass of multipliers of
R
⊥
is then denoted by
M(R
⊥
)
.
Inwhat follows, we willneed thefollowing.
Proposition1([1 ℄). Let
T
be anergodi automorphismof(X, B, µ)
.If̺ ∈ J
e
2
(T )
andalso̺ ◦ ̺ ∈ J
e
2
(T )
then(T × T, ̺)
isrelatively weakly mixing over the two marginal fa torsB ⊗ {∅, X}
and{∅, X} ⊗ B
.Assumethat
T
isweaklymixing and̺ ∈ J
e
2
(T )
.Then dire tlyfrom the hainrulefor therelative weakmixing propertywe obtain thefollowing. (1) If(T × T, ̺)
isrelativelyweaklymixingoverthemarginalfa tors,thenWe will alsoneed the following simple lemma.
Lemma 2. Assume that
T
is a weakly mixing automorphism of a stan-dard probability Borelspa e(X, B, µ)
. Assume thatN
0
⊂ N
andthe density ofN \ N
0
equals zero. Assume moreover that for ea hf, g ∈ L
2
(X, B, µ)
,
hf ◦ T
n
, gi → hf, 1ih1, gi
as
n → ∞
,n ∈ N
0
.Assumethat̺ ∈ J(T )
.Thenforallf, g, h ∈ L
∞
(X, B, µ)
, \X×X
f (T
n
x)g(x)h(T
n
y) d̺(x, y) →
\X
f (x)h(y) d̺(x, y)
\X
g(x) dµ(x).
Proof. Wehave \X×X
f (T
n
x)g(x)h(T
n
y) d̺(x, y) =
\X
f (T
n
x)g(x)Φ
∗
̺
(h ◦ T
n
)(x) dµ(x)
=
\X
(f · Φ
∗
̺
(h)) ◦ T
n
· g dµ
→
\X
f · Φ
∗
̺
(h) dµ
\X
g dµ =
\X×X
f ⊗ h d̺
\X
g dµ.
Formoreinformationonjoiningswereferthereadertothemonographby E.Glasner[13 ℄.Forthespe traltheoryofdynami alsystemsseee.g.[3℄,[33 ℄. 1.2. Sub-joiningsand sub-Markovoperators in innite measure-preserv-ing ase. Given two automorphisms
T
andS
a ting onσ
-nite standard Borelspa es(X, B, µ)
and(Y, C, ν)
respe tively,byasub-joining ofT
andS
we meanea hpositiveσ
-niteT × S
-invariant measure̺
on(X × Y, B ⊗ C)
whose marginals̺
X
and̺
Y
satisfy̺
X
≤ µ
and̺
Y
≤ ν
.By theformula\
X×Y
f (x)g(y) d̺(x, y) =
\Y
V (f ) · g dν,
there is a one-to-one orresponden e between the set of sub-joinings and theset of sub-Markov operators
V : L
2
(X, B, µ) → L
2
(Y, C, ν)
intertwining
U
T
andU
S
, where by a sub-Markov operator we mean a positive operatorV : L
2
(X, B, µ) → L
2
(Y, C, ν)
su h thatV f ≤ 1
for allf ∈ L
2
(X, B, µ)
satisfying0 ≤ f ≤ 1
,andV
∗
g ≤ 1
forallg ∈ L
2
(Y, C, ν)
satisfying0 ≤ g ≤ 1
. Remark1. Notethateveninthe aseT = S
,althoughtheo-diagonal measuresµ
T
n
havethepropertythattheirmarginalsareequaltoµ
(equiva-lently,T
X
T
n
(1
A
) dµ = µ(A)
forea hA ⊂ X
ofnitemeasure),thefa tthat the onstant fun tion1
X
is not integrable may ause thatthe marginalsof aweaklimit̺
ofasequen eofo-diagonal measuresneed notbe equaltoµ
(nevertheless,wewill have̺
X
≤ µ
).1.3.Co y lesand ompa tgroupextensions. Assumethat
T
isanergodi automorphism of a standard probability Borel spa e(X, B, µ)
. LetG
be a ompa t metri group with theσ
-algebraB(G)
of Borel sets and Haar measurem
G
. Letϕ : X → G
be a measurable map. It generates a o y leϕ
( · )
( · ) : Z × X → G
bythe formulaϕ
(n)
(x) =
ϕ(T
n−1
x) · ϕ(T
n−2
x) · . . . · ϕ(x)
ifn > 0
,1
ifn = 0
,(ϕ(T
−1
x) · . . . · ϕ(T
n
x))
−1
ifn < 0
.We denote by
T
ϕ
the skew produ t automorphism dened on(X × G,
B ⊗ B(G), µ ⊗ m
G
)
bythe formulaT
ϕ
(x, g) = (T x, ϕ(x) · g).
We all
T
ϕ
a ompa t group extension ofT
.Denote by
τ
g
the map onX × G
given byτ
g
(x, g
1
) = (x, g
1
g
−1
)
. Note that
τ
g
∈ C(T
ϕ
)
for ea hg ∈ G
.Compa t group extensions have the so alled relative unique ergodi ity (RUE) property: whenever the produ t measure
µ ⊗ m
G
is ergodi , it is theonlyT
ϕ
-invariant measureof(X × G, B ⊗ B(G))
whose proje tion onX
equalsµ
(seee.g. [10℄).We say that a o y le
ϕ : X → G
is ergodi ifT
ϕ
onsidered withµ ⊗ m
G
isergodi . Inthis aseergodi self-joinings ofT
ϕ
whose proje tions onX × X
are∆
X
arene essarily graphjoinings orrespondingtoτ
g
,g ∈ G
(see [21 ℄).1.4. Gaussian automorphisms. An ergodi automorphism
T
of a stan-dard probability Borel spa e(X, B, µ)
is alled Gaussian if there exists aU
T
-invariant subspa eH ⊂ L
2
(X, B, µ)
of real-valued fun tions generating
B
andsu h thatea h non-zerovariablefromH
has aGaussiandistribution. For a joining theory of Gaussian automorphisms we refer the reader to [29 ℄ (seealso [3 ℄forageneral theory ofGaussian automorphisms).Inparti ular, it is proved in [29 ℄ that there is a spe ial subsetJ
g
2
(T ) ⊂ J
2
e
(T )
alled the set ofGaussian self-joinings (for̺ ∈ J
g
2
(T )
,(T × T, ̺)
remains a Gaussian automorphism).Roughlyspeaking,thisset orrespondstoall ontra tionsof therst haosH
intertwiningthe unitarya tion ofT
onH
(all o-diagonal self-joiningsµ
T
n
are inJ
g
2
(T )
). It follows thatJ
g
2
(T )
is losed in the weak topology ofjoinings.AGaussianautomorphism
T
isentirely determinedbythespe tral mea-sureσ
ofU
T
onH
(c)
= H + iH
.Moreover,
T
isergodi iσ
is ontinuous. Themaximal spe tral type ofT
isthesumof onse utive onvolutionsσ
(n)
automor-Ea hvariable
f ∈ H
,viewed as a mapf : X → R
,is alleda Gaussian o y le.Itis alleda Gaussian oboundary iff = g − g ◦ T
for someg ∈ H
. The subspa eH
onsists entirely of Gaussian oboundaries i 1 is not in the topologi al support ofσ
([27℄). We refer the reader to [27 ℄ for more information about ergodi ity of ir le group extensions of the formT
e
2πif
, wheref
is aGaussian o y le.1.5. Integral automorphisms andspe ial ows. Let
T
be an ergodi au-tomorphism of a standard probability Borel spa e(X, B, µ)
. Assume thatf : X → N
is a measurable fun tion with nite integral. LetX
f
⊂ X × N
be given byS
n∈N
X
n
× {n}
, whereX
n
= {x ∈ X : f (x) ≤ n}
. LetB
f
denote the restri tion of the produ tσ
-algebra ofB
and theσ
-algebra of all subsets ofN
to the setX
f
. Letµ
f
denote the restri tion of the prod-u t measureµ ⊗
P
n∈N
δ
{n}
toX
f
.Bytheintegral transformation built over theautomorphismT
andunderthefun tionf
we meanthe transformationT
f
: (X
f
, B
f
, µ
f
) → (X
f
, B
f
, µ
f
)
dened byT
f
(x, k) =
(x, k + 1)
if
f (x) < k
,(T x, 1)
iff (x) = k
.Supposethat
A ∈ B
haspositivemeasure.Itiseasyto he kthat(T
A
)
τ
A
andT
aremetri allyisomorphi ,whereT
A
: A → A
istheindu edautomorphism andτ : A → N
stands fortherst returntimefun tion (see[3, Chapter1℄). Denote bym
R
the Lebesgue measure onR
. Assume thatf : X → R
is a measurable positive fun tion su h thatT
X
f dµ = 1
. The spe ial owT
f
= {(T
f
)
t
}
t∈R
built fromT
andf
is dened on the spa eX
f
=
{(x, t) ∈ X × R : 0 ≤ t < f (x)}
( onsidered withB
f
, the restri tion of the produ t
σ
-algebra, andµ
f
, the restri tion of the produ t measure
µ ⊗ m
R
ofX × R
). Under the a tion of the spe ial ow ea h point inX
f
moves verti allyatunitspeed,andweidentifythepoint
(x, f (x))
with(T x, 0)
(see e.g. [3 , Chapter 11℄). In the spe ial ase wheref ≡ 1
the spe ial owT
f
a tson
X × [0, 1)
andis alledthe suspensionow forthe automorphismT
. Then we writeT
b
instead ofT
f
and
( b
X, b
B, b
µ)
instead of(X
f
, B
f
, µ
f
)
. Let
π : b
X = X × [0, 1) → X
denote thenatural proje tion. Then theσ
-algebraπ
−1
(B) ⊂ b
B
is
( b
T )
1
-invariant andπ : ( b
X, π
−1
(B), b
µ) → (X, B, µ)
establishes an isomorphism between automorphisms
( b
T )
1
of( b
X, π
−1
(B), b
µ)
and
T
of(X, B, µ)
.Finally,noti e thattheowsT
b
f
andT
f
areisomorphi whenever
f : X → N
.Lemma 3. Let
T
be an ergodi automorphism of(X, B, µ)
and letf :
X → N
be a measurable fun tion with nite integral. Suppose that(a
n
)
is a sequen e of integers su h that(T
f
)
a
n
→ p((T
f
)
1
)
weakly, wherep
is a trigonometri polynomial. ThenT
a
n
Proof. Sin e the operators
(T
f
)
1
a ting onL
2
(X
f
, B
f
, µ
f
)
and( b
T
f
)
1
a ting onL
2
( b
X
f
, b
B
f
, b
µ
f
)
areunitarily isomorphi ,( b
T
f
)
a
n
→ p(( b
T
f
)
1
)
intheweakoperatortopologyon
L
2
( b
X
f
, b
B
f
, b
µ
f
)
.Letπ : b
X
f
= X
f
×[0, 1) →
X
f
be the natural proje tion. Sin eL
2
( b
X
f
, π
−1
(B
f
), b
µ
f
) ⊂ L
2
( b
X
f
, b
B
f
, b
µ
f
)
is an invariant subspa e with respe t to the operators
( b
T
f
)
a
n
(n ∈ N
),( b
T
f
)
a
n
→ p(( b
T
f
)
1
)
in the weak operator topology onL
2
( b
X
f
, π
−1
(B
f
), b
µ
f
)
. Sin ethe operatorsT
f
onL
2
(X
f
, B
f
, µ
f
)
and( b
T
f
)
1
onL
2
( b
X
f
, π
−1
(B
f
), b
µ
f
)
are unitarily isomorphi ,
T
a
n
f
→ p(T
f
)
in the weak operator topology onL
2
(X
f
, B
f
, µ
f
)
.2. Basi properties of ELF automorphisms. An ergodi automor-phism
T
of a standard Borel spa e(X, B, µ)
is said to have theELF prop-erty if{µ
T
n
: n ∈ Z} ⊂ J
e
2
(T )
,or equivalently,the weak losureofthesetof Markov operators{T
n
: n ∈ Z}
onsists of inde omposable Markov opera-tors.For short,wewill speakabout ELF automorphisms.
It is lear that ergodi dis rete spe trum automorphisms and mixing automorphisms are examplesof ELF automorphisms. By what was said in Se tion 1.4, Gaussian automorphisms also enjoy the ELF property (see [6℄ for adire tproof of thatfa t).
The following two onsequen es of Proposition 1 have already been no-ti ed in[6 ℄.
Proposition 4 ([6 ℄). If
T
is an ELF automorphism and if̺ ∈
{µ
T
n
: n ∈ Z}
then(T × T, ̺)
is relatively weakly mixing with respe t tothe two natural marginalσ
-algebras.Proposition 5 ([6 ℄). Assume that
T
is an ELF automorphism andlet̺ ∈ {µ
T
n
: n ∈ Z}
. LetS
be an ergodi automorphism on(Y, C, ν)
.Assume that̺
1
isan ergodi joining ofT
andS
. Then̺
1
◦ ̺
isstill ergodi .2.1. Disjointness of ELF automorphisms from time-one maps of some measurableows. Proposition 5,similarlyto [6 ℄,allowsus toprove disjoint-nessofthe lassofELFautomorphismsfromautomorphismshavingapie e of integral Markov operator in the weak losure of its powers. Indeed, as-sumethat
S
isanautomorphismof(Y, C, ν)
.LetP
beaprobabilitymeasure dened on the Borelσ
-algebra ofC(S)
. We dene a Markov operatorM
P
onL
2
(Y, C, ν)
byputtingM
P
(f ) =
\C(S)
f ◦ R dP (R).
Theintegral ontherighthand side ismeant intheweaksense,i.e.for ea h
f, g ∈ L
2
(Y, C, ν)
,D
\C(S)
f ◦ R dP (R), g
E
=
\C(S)
hf ◦ R, gi dP (R).
Inorder to seethatthis denitionis orre twe dene
hhf, gii =
\
C(S)
hf ◦ R, gi dP (R)
and he kthatwehaveobtainedabilinearformon
L
2
(Y, C, ν)
whi h,bythe S hwarzinequality,is bounded. Clearly,
M
P
∈ J
2
(S)
.Proposition 6. Let
S : (Y, C, ν) → (Y, C, ν)
be an ergodi automor-phism. Assume that there exist an in reasing sequen e(t
n
)
of natural num-bersand a probability BorelmeasureP
onC(S)
su h thatS
t
n
→ a
\
C(S)
R dP (R) + (1 − a)Φ
intheweakoperatortopologyon
B(L
2
(Y, C, ν))
,where
a > 0
andΦ ∈ J
2
(S)
. Assume thatP ({R ∈ C(S) : R
isweakly mixing}) > 0
. ThenS
is weakly mixing. If moreoverP
isnot Dira andeither(i)
P
is on entrated on{S
i
: i ∈ Z}
,or
(ii)
P
is on entrated on{S
t
: t ∈ R}
, whereS
1
= S
(i.e. we assume in parti ular thatS
isembeddable in a measurable ow),then
S
isdisjoint fromall ELF automorphisms.Proof. First, let us show that
S
is weakly mixing. Indeed, iff
is its eigenfun tion thenkf k
2
L
2
= |hS
t
n
f, f i| →
a
\C(S)
hf ◦ R, f i dP (R) + (1 − a)hΦ(f ), f i
.
Sin e|hf ◦ R, f i| ≤ kf k
2
and|hΦ(f ), f i| ≤ kf k
2
, a onvexity argument shows that we must havehf ◦ R, f i = kf k
2
for
P
-a.e.R ∈ C(S)
(and alsohΦ(f ), f i = kf k
2
provided
a < 1
).So for su hanR
,wehavef ◦ R = c(R)f
(c(R) ∈ C
), andsin eR
maybe taken weakly mixing,f
is onstant.Let
T
be an ELF automorphism on(X, B, µ)
. LetΨ : L
2
(Y, C, ν) →
L
2
(X, B, µ)
be an inde omposable Markov operator intertwining
S
andT
. ThenΨ ◦ S
t
n
= T
t
n
◦ Ψ
andbypassing toasubsequen e of
(t
n
)
ifne essary, we ndΨ ◦ (aM
P
+ (1 − a)Φ) = Φ
̺
◦ Ψ,
inde om-posable. Ontheother hand,
Ψ ◦ (aM
P
+ (1 − a)Φ) = a
\
C(S)
Ψ ◦ R dP (R) + (1 − a)Ψ ◦ Φ,
and hen e we must have
Ψ ◦ R = Φ
̺
◦ Ψ
forP
-a.e.R ∈ C(S)
. This means that for a set of fullP ⊗ P
-measure of(R
1
, R
2
) ∈ C(S) × C(S)
, we haveR
2
◦ R
−1
1
◦ Ψ
∗
= Ψ
∗
.Noti ehoweverthat both assumptions (i)and (ii)and the fa t thatP
is not Dira imply that for some weakly mixing elementR ∈ C(S)
we haveR ◦ Ψ
∗
= Ψ
∗
and therefore
Ψ = Π
Y,X
.Supposenowthat
(S
t
)
t∈R
isa measurable,weaklymixing owa tingon(Y, C, ν)
.Supposethatfora sequen e(r
n
)
ofreal numberswithr
n
→ ∞
we have (2)S
r
n
→ a
\R
S
t
dQ(t) + (1 − a)Φ,
where
Q
isnotDira .Bypassingtoasubsequen eifne essarywe anassume that the sequen e({r
n
})
of fra tional parts ofr
n
onverges to0 ≤ b ≤ 1
. Sin e the ow is measurable,S
{r
n
}
→ S
b
in the strong operator topology. It follows that the sequen e(S
1
)
[r
n
]
= S
r
n
◦ S
−{r
n
}
onverges in the weak operatortopology and we have(S
1
)
[r
n
]
→ a
\
R
S
t−b
dQ(t) + (1 − a)Φ ◦ S
−b
.
We have provedthe following.
Corollary7. Assumethat
(S
t
)
t∈R
isameasurable,weaklymixingow for whi h(2)
holds withQ
whi h is not Dira . Then the time-one mapS
1
isdisjoint fromallELF automorphisms.Remark 2. The assumptions of Corollary 7 are satised for time-one maps of some lassi al examples of spe ial ows over irrational rotations and overinterval ex hange transformations(see [6 ℄[8 ℄)and inparti ular it issatised for some smooth owson surfa es(see [8 ℄).
2.2. Fa tors and dire t produ ts of ELF automorphisms. The following propositionshowsthatthe lassofELFautomorphismsis losedundersome basi operations.
Proposition 8. The lass of ELF automorphisms is losed under fa -tors and inverse limits. The dire t produ t of weakly mixing ELF automor-phisms remains an ELFautomorphism.
Proof. Closedness undertakingfa tors and inverse limitsisobvious. Assume that
T
i
is a weakly mixing ELF automorphism of(X
i
, B
i
, µ
i
)
,Suppose that
T
n
i
→ Φ
̺
for some̺ ∈ J
2
(T )
.By applying thediagonalizing pro edure if ne essary, we an assume that for ea hj ≥ 1
,T
n
i
j
→ Φ
̺
j
for some̺
j
∈ J
e
2
(T
j
)
.Iteasilyfollowsthat̺ = ̺
1
⊗ ̺
2
⊗ · · ·
andbe auseof(1),̺
isergodi , whi h ompletestheproof.Remark 3. Notehoweverthat an ergodi self-joining of an ELF auto-morphism need not be an ELF automorphism. Indeed, even if
T
is mixing thenanergodi self-joiningneednot give risetoanELFautomorphism.For example,bySmorodinskyThouvenot's resultfrom[45 ℄itfollowsthatgiven an ergodi zero entropy automorphismS
and a Bernoulli automorphismT
we an nd an ergodi self-joining̺
ofT
su h that(T × T, ̺)
hasS
as its fa tor.2.3.Lifting theELF propertyto ompa t groupextensions. Wewill now dis uss the possibility of lifting theELF property by a ompa t group ex-tension.Soassumethat
T
isanELFautomorphismandletϕ : X → G
bea o y le, whereG
isa ompa tmetri group.Re all rst thatifT
is mixing and the extensionT
ϕ
is weakly mixing thenT
ϕ
is infa t mixing (see [37 ℄). A lookat ashort joining proof (dueto A. delJun o) of thatfa t gives rise to a riterion oflifting the ELFproperty.Proposition9. Assumethat
T
hastheELFpropertyandϕ : X → G
is ergodi . Assumemoreoverthat for ea h̺ ∈ {µ
T
n
: n ∈ Z}
the o y leϕ × ϕ
over(T × T, ̺)
isergodi .ThenT
ϕ
has the ELF property.Proof. Assume that
(T
ϕ
)
m
i
→ Φ
e
̺
. We must show that̺
e
is ergodi . We an assume thatm
i
→ ∞
, otherwise the result is lear. We then haveT
m
i
→ Φ
̺
, where̺
is theproje tion of̺
e
onX × X
.Now,̺
e
is aT
ϕ
× T
ϕ
-invariant measure whose proje tion is̺
.However, byour standing assump-tionthemeasure̺ ⊗ m
G
⊗ m
G
has thesame propertyanditisergodi .The resultnowfollows fromthe relative uniqueergodi ityproperty for ompa t groupextensions.The above proof suggests that in general we have no han e to lift the ELF property and in fa t we will loose this property when the base has dis rete spe trum.
Proposition 10. An ergodi isometri extension
T
b
of a dis rete spe -trumautomorphismT
hastheELFpropertyitheextensionalsohasdis rete spe trum.Proof. We an assumethat
T
isan ergodi rotation(T x = x + x
0
) of a ompa t metri monotheti groupX
.Moreoverassume thatϕ : X → G
is anergodi o y leforwhi hT
istheKrone kerfa torandT
b
isthequotient a tion ofT
ϕ
onX × G/H
. All we need to show is that under all these assumptionsT
b
doesnot have theELFproperty.To thisend rst hoosea sequen e
(n
i
)
of density1 su hthat (3)U
n
i
b
T
→ 0
weaklyonL
2
(X × G/H, m
X
⊗ m
G/H
) ⊖ L
2
(X, m
X
),
whi h is possible be ause
T
is the Krone ker fa tor ofT
b
and therefore the spe traltypeofU
b
T
onL
2
(X ×G/H, m
X
⊗m
G/H
)⊖L
2
(X, m
X
)
is ontinuous. Sin e the density of(n
i
)
is 1, there exists a subsequen e(m
i
)
of(n
i
)
su h thatT
m
i
→ Id
.Indeed,givenaneighbourhood
W ∋ 0
inX
,bythepointwise ergodi theoremfor stri tly ergodi systems the average timeof visitingW
bytheorbit of an arbitrary point ofX
is equal tom
X
(W )
,hen e positive. Therefore we an ndn
j
= n
j
(W )
so thatn
j
x
0
∈ W
. LettingW → {0}
provesthe laim.It follows from (3) that
T
b
m
i
onverges weakly to the operator
E( · | X)
whi h orresponds to the joining∆
X
⊗ m
G/H
⊗ m
G/H
. However, this last joiningisnotergodi :thefun tionF (x, gH, x, g
′
H) = g
−1
g
′
H
isnot onstant but it is
T × b
b
T
-invariant∆
X
⊗ m
G/H
⊗ m
G/H
-a.e. Therefore,T
b
does not have theELFpropertyand theresultfollows.Thefollowing orollaryfollows dire tlyfromProposition 10.
Corollary 11. If an extension of a rotation
T
has the ELF property, then the extension isrelatively weaklymixing overT
.Remark 4. In [52 ℄ there are expli it onstru tions of ELF automor-phisms whi hare relatively weakly mixing extensionsof some irrational ro-tations.
Let us now show however that the riterion of Proposition 9 may work insome ases ofmildlymixing ELFautomorphismswhi harenot mixing.
We onsidersymmetri probabilitymeasures
σ
onT
su hthat (4) all weak losurepointsofthesequen e{z
n
: n ∈ Z}
in
L
2
(T, σ)
arein the set
{az
n
: |a| < 1, n ∈ Z}
.
Sin e
σ
isasymmetri measure,thenumbersa
in(4)havetobereal.Under the above assumption, the Gaussian automorphism asso iated toσ
has to bemildlymixing.Re allthat lassi alRieszprodu tsyieldexamplesofsu h measures,in ludingexampleswhi harenotRaj hmanmeasures sothatthe setof weak losurepointsisnot trivial (see[16 , Ch. II, Se t. 7℄).Proposition 12. Assumethat
T
isamildly mixingGaussian automor-phism determined bya measure satisfying(4)
forwhi ha ertaina 6= 0
isin theweak losureof hara ters(su h aT
isnotmixing).Takef
fromtherst real haos. Assume thatf
is not a Gaussian oboundary. ThenT := T
Proof. Assume that
(5)
(T
e
2πif
)
n
t
→ Φ
̺
e
forsomesequen e
(n
t
)
withn
t
→ ∞
.ThenT
n
t
→ Φ
̺
,where̺
isthe proje -tionof̺
e
onX ×X
.Withoutlossofgeneralitywe anassumethatz
n
t
→ a
in theweaktopologyof
L
2
(T, σ)
forsomereal
a
with|a| < 1
.Wehavetoprove that̺ ∈ J
e
2
(T
e
2πif
)
isergodi . If now̺
isthe produ t measurethen so is̺
e
, sin eT
e
2πif
isweaklymixing([27℄)andwemayapplytherelativeunique er-godi itypropertyfor ompa tgroupextensionsto on lude.Notethat̺
an-not be agraphmeasure, sin eT
isassumed tobemildlymixing. Moreover, sin ez
n
t
→ a
intheweaktopologyof
L
2
(T, σ)
,
T
n
t
restri tedtotherstreal haostendstomultipli ationby
a
,andhen eΦ
̺
ismultipli ationbya
onthe rst real haos. ByProposition 9 allwe need to showisthatT
e
2πif
× T
e
2πif
isergodi as a
T × T
-extensionof(T × T, ̺)
.Following Proposition6in[27 ℄ itis su ient to show thatthe o y lelf (x) + mf (y)
(with(l, m) 6= (0, 0)
) isnot aGaussian oboundary (fortheGaussian automorphism(T × T, ̺)
). If for ea hr ∈ N
we putf
(r)
(x) = f (x) + f (T x) + · · · + f (T
r−1
x)
then we haveklf
(r)
(x) + mf
(r)
(y)k
2
L
2
(̺)
= (l
2
+ m
2
)kf
(r)
k
2
+ 2lmhf
(r)
(x), f
(r)
(y)i
L
2
(̺)
= (l
2
+ m
2
)kf
(r)
k
2
+ 2lm
\X
(Φ
̺
f
(r)
)(y)f
(r)
(y) dµ(y)
= (l
2
+ m
2
+ 2ml · a)kf
(r)
k
2
.
Now,
f
is not aGaussian oboundary,sokf
(r
t
)
k → ∞
alonga subsequen e
(r
t
)
(see [27 ℄)andsin e|a| < 1
,(l
2
+ m
2
+ 2ml · a)kf
(r
t
)
k → ∞
or equivalently
klf
(r
t
)
(x) + mf
(r
t
)
(y)k
L
2
(̺)
→ ∞,
whi hmeans(see[27 ℄)thatindeed
lf (x)+mf (y)
isnotaGaussian obound-ary.3.PoissonautomorphismshavetheELFproperty. Inthisse tion we will dene and study a spe ial lass of self-joinings for the lass of au-tomorphisms obtained byPoisson suspension of innite measure-preserving maps.
3.1.Poisson suspension automorphisms. Assumethat
T
isan automor-phismofastandard Borelspa e(X, B, µ)
,whereµ
isσ
-nite.Wedenotebye
T
thePoisson suspension automorphism a ting on( e
X, e
B, e
µ)
.The points ofe
nite measurewe dene
N
A
: e
X → N
byputtingN
A
(e
x) = #{n ∈ N : x
n
∈ A}.
Then we dene
B
e
as the smallestσ
-algebra of subsets ofX
e
for whi h all variablesN
A
,µ(A) < ∞
, are measurable. The measureµ
e
is dened by the requirement that the variablesN
A
satisfy the Poisson law with parameterµ(A)
and moreover that for ea h family of pairwise disjoint subsets ofX
of nite measure the orresponding variables are independent (see [24 ℄ for details).Finally,weletT
e
a tbytheformulaT ({x
e
n
}) = ({T x
n
})
toobtainan automorphism of( e
X, e
B, e
µ)
.The spa eL
2
( e
X, e
B, e
µ)
admits a de omposition into invariant haos
L
n≥0
H
(n)
, whereH
(0)
is the subspa e of onstants,
H = H
X
= H
(1)
is the subspa e generated by the entred variablesN
0
A
=
N
A
− µ(A)
andH
(n)
is the ortho omplement of the sum of haos
H
(i)
,
0 ≤ i ≤ n − 1
,in the subspa e generated by the produ ts ofn
variables of theformN
A
(see[32 ℄).Themap1
A
7→ N
0
A
anbeextendedtoanisometryI
ofL
2
(X, B, µ)
onto
H
andit onjugatesU
T
withU
e
T
|
H
.Moreoverweobtain a natural isometry betweenH
(n)
and the
n
th symmetri tensor produ tH
⊙n
ofH
underwhi h⊙
n
i=1
N
A
0
i
orresponds to theproje tion of
Q
n
i=1
N
A
i
inH
(n)
. TheoperatorU
e
T
preserves the haosand,for ea hn ≥ 0
,its restri tion toH
(n)
orresponds to
(U
e
T
|
H
)
⊙n
by this natural isometry. In su h a ase, we will saythatan operator a ts well on the haos.If
0 6= f ∈ L
2
(X, B, µ)
is an eigenfun tion of
U
T
orresponding toc
(with|c| = 1
), thenf
is an eigenfun tion ofU
T
orresponding toc
. ThenI(f ) ⊙ I(f ) ∈ H
(2)
and it is aU
e
T
-invariant fun tion. Furthermore, ifσ
denotesthemaximalspe traltypeofU
e
T
onH
(whi hisequaltothemaximal spe tral type ofU
T
onL
2
(X, B, µ)
) then the maximal spe tral type of
U
e
T
on the
n
th haos is equal to then
th onvolutionσ
(n)
= σ ∗ · · · ∗ σ
. Re all that
σ
(n)
is ontinuous i
σ
is ontinuous.ThereforethePoissonsuspensionautomorphism
T
e
on( e
X, e
B, e
µ)
isergodi ifandonlyifthespe traltypeofT
onL
2
(X, B, µ)
is ontinuous;equivalently, ithere areno
T
-invariant subsetsofX
ofnite positive measureor elseiL
2
(X, B, µ)
doesnot ontainnon-zeroT
-invariant fun tions. In this aseT
e
is weakly mixing. Finally, note that, in parti ular, if a Poisson suspension automorphism isergodi then ne essarily themeasureµ
is innite.3.2. Fa tors and Poisson joinings. If
X
1
is aT
-invariant subset ofX
, thenT
e
isthedire tprodu toftwoPoissonsuspensionsofT
a tingonX
1
and onX \ X
1
,inparti ular,]
T |
X
1
isafa torofT
e
.AssumenowthatS
a tingon anotherσ
-nite standardBorelspa e(Y, C, ν)
isafa tor of(X
1
, µ|
X
1
, T |
X
1
)
inthesense thatthereis ameasurable map
F : X
1
→ Y
su hF
∗
(µ|
X
and
F ◦ T = S ◦ F
onX
1
.ThenS
e
a ting on( e
Y , e
C, eν)
isafa tor of]
T |
X
1
via themapF : e
e
X
1
→ e
Y
given byF ({x
e
n
}) = {F (x
n
)}
.Thenthe asso iated operator
V
e
F
: L
2
( e
Y , e
ν) → L
2
( e
X
1
, g
µ|
X
1
)
,g 7→ g ◦ e
F
, a ts wellon the haos.By Poisson fa tors of
T
e
we will mean fa torsS
e
obtained as above. We willalso saythatthe mapF
isapartialmapofX
toY
semi- onjugatingT
andS
.Note thatifF : X
1
→ Y
establishes a semi- onjugation ofT
andS
thenthe asso iated isometryV
F
isa sub-Markovoperator fromL
2
(Y, ν)
to
L
2
(X
1
, µ|
X
1
)
.Assume that
T
andS
are automorphisms ofσ
-nite spa es(X, B, µ)
and(Y, C, ν)
respe tively. A joiningη
ofT
e
andS
e
is alled a Poisson join-ing if the asso iated Markov operatorV = Φ
e
η
a ts well on the haos and, via the natural isomorphisms of the rst haosH
X
of( e
X, e
B, e
µ)
andH
Y
of( e
Y , e
C, eν)
withL
2
(X, B, µ)
andL
2
(Y, C, ν)
, the operatorV |
e
H
X
or-responds to the sub-Markov operatorV
asso iated to a sub-joining ofT
andS
.Proposition 13. The lass of Poisson joinings between
T
e
andS
e
is losed in the weak topology of joinings, in parti ular, the lass of Poisson self-joiningsofT
e
ontains the weak losure of{ e
T
n
: n ∈ Z}
.
Moreover, the relative produ t of
T
e
over a Poisson fa tor is a Poisson self-joining.Proof. The rst part follows dire tly from the fa t that the set of sub-Markovoperatorsis losed inthe weakoperator topology.
Toprove these ondparttake aPoissonfa tor whi hisdetermined bya partialfun tion
F : X
1
⊂ X → Y
.ThentheMarkovoperator orresponding totherelativeprodu toverthisfa torisgivenbyV
e
F
◦V
F
e
∗
◦P
L
2
( e
X
1
)
.Sin eV
e
F
,V
∗
e
F
,P
L
2
( e
X
1
)
a twellonthe haosandtheirrestri tionsto therst haos an naturally be identied withV
F
,V
∗
F
,L
2
(X) ∋ f 7→ f |
X
1
∈ L
2
(X
1
)
resp., so aresub-Markovoperators,the ompositionV
e
F
◦V
F
e
∗
◦P
L
2
( e
X
1
)
istheasso iated Markovoperator ofa Poisson self-joining.
3.3. Ergodi ity of Poisson joinings. Assume that
̺
is a sub-joining ofT
andS
and denote byV
the orresponding sub-Markov operator fromL
2
(X, B, µ)
to
L
2
(Y, C, ν)
.We will nowpass to a onstru tion of a Poisson joining
η
ofT
e
andS
e
orrespondingto̺
,i.e.ifweputV = Φ
e
η
thenV |
e
H
≡ V
. This Poisson joining turns out to be unique, so the stru ture of Poisson joiningswill be understood.Set
µ
′
= µ − ̺
X
andν
′
= ν − ̺
Y
.Letusdene aσ
-nite standardspa e(Z, ̺
′
)
as a formaldisjoint union of
(X, µ
′
)
,
(Y, ν
′
)
we dene
R
on(Z, ̺
′
)
byputting
R|
X
= T
,R|
Y
= S
andR|
X×Y
= T × S
. Sin e̺
X
and̺
Y
areT
- andS
-invariant respe tively,̺
′
is
R
-invariant. We nowhave the partial mappingF : X ∪ (X × Y ) ⊂ Z → X
whi h tox ∈ X
orto(x, y) ∈ X ×Y
asso iatesx
,andforea hA ⊂ X
ofnitemeasure we have̺
′
(F
−1
(A)) = µ
′
(A) + ̺
X
(A) = µ(A)
.Clearly,F ◦ R = T ◦ F
,soF
establishesasemi- onjugationofR
andT
.ThemappingG : Y ∪ (X × Y ) ⊂
Z → Y
whi h toy ∈ Y
or to(x, y) ∈ X × Y
asso iatesy
has similar properties. Hen e, the two mapsF : e
e
Z → e
X
andG : e
e
Z → e
Y
are fa tor mappingsofR
e
toT
e
andS
e
respe tively.It followsthat( e
F , e
G) : e
Z → e
X × e
Y
denesajoiningη = ( e
F , e
G)
∗
( e
̺
′
)
of
T
e
andS
e
,thatis,forea hf ∈ L
2
( e
X, e
B, e
µ)
andg ∈ L
2
( e
Y , e
C, e
ν)
we have \f (e
x)g(e
y) dη(e
x, e
y) =
\f ◦ e
F · g ◦ e
G d e
̺
′
.
It follows that theMarkov operator asso iated to
η
is equal toV = V
e
∗
e
G
V
F
e
. Hen e,V
e
a ts well on the haos and learly its restri tion toH
X
an be naturally identied withV
∗
G
V
F
. Let us now show thatV
∗
G
V
F
= V
. Indeed, takeA ⊂ X
andB ⊂ Y
of nite measure. Noti e that1
A
◦ F · 1
B
◦ G
equalszerooutside ofX × Y
,andonX × Y
thisfun tion isequalto1
A×B
. Therefore \V
G
∗
V
F
1
A
· 1
B
dν =
\1
A
◦ F · 1
B
◦ G d̺
′
= ̺(A × B) =
\V 1
A
· 1
B
dν,
when eV
∗
G
V
F
= V
.Theorem 14. Ea h Poisson joining of two ergodi Poisson suspension automorphisms remainsergodi .In parti ular,ea hergodi Poisson suspen-sion automorphism has the ELF property.
Proof. Noti e that the se ond assertion follows from the rst one and Proposition 13.
Assume that
T
e
andS
e
are ergodi . It follows thatX
andY
have no invariant sets of nitepositive measure. Letus showthatinZ
there arenoR
-invariantsetsofnitepositive̺
′
-measure.Indeed,supposethat
h = 1
C
∈
L
2
(Z, ̺
′
)
is
R
-invariant. ThenV
∗
F
h
isT
-invariant, so equal to zeroµ
-a.e.In parti ular for ea h subsetA ⊂ X
ofnite measurewehave0 = hV
F
∗
h, 1
A
i = hh, V
F
(1
A
)i =
\h · (1
A
◦ F ) d̺
′
that is,̺
′
(C ∩ F
−1
(A)) = 0
. Sin eF
−1
(A)
is a formal disjoint union of
A
andA × Y
,̺
′
(C ∩ X) = 0
and by a similar argument
̺
′
(C ∩ Y ) = 0
togetherwith
̺
′
(C ∩ (A × B)) = 0
for ea h
A ⊂ X
,B ⊂ Y
ofnitemeasure. Therefore̺
′
(C) = 0
.ItfollowsthatthePoisson suspension
R
e
ofR
isergodi and thereforeits fa tor( e
T × e
S, η)
remains ergodi .Remark 5. Independently, using dierent arguments, the resulton er-godi ityofPoissonian joiningshas also been proved byE.Royin[36 ℄.
4. Self-joinings of symmetri
α
-stable automorphisms. In this se tionwewill dene and studyα
-stableself-joinings forα
-stable automor-phisms, i.e. automorphismsgiven by stationaryergodi symmetriα
-stable pro esses (see [17 ℄, [30 ℄, [44 ℄). We will show that ea h ergodi symmetriα
-stableautomorphism has the ELFproperty.4.1. Auxiliary lemmas. The proofs of the following two elementary in-equalities areslight adaptations ofthe proofs from [44 ,pp.9192℄.
Lemma 15. If
0 < α < 1
then for allx, y ∈ R
we have(6)
|x|
α
+ |y|
α
− |x + y|
α
≥ (2 − 2
α
) min(|x|
α
, |y|
α
).
If
1 ≤ α < 2
then for allx, y ∈ R
we have (7)2(|x|
α
+ |y|
α
) − (|x + y|
α
+ |x − y|
α
) ≥ 2(2 − 2
α/2
) min(|x|
α
, |y|
α
).
Inparti ular, (6)implies (8)
| |x + y|
α
− |y|
α
| ≤ |x|
α
for
x, y ∈ R
and0 < α ≤ 1,
and bytheHölder inequality(9)
|x + y|
α
≤ max(1, 2
α−1
)(|x|
α
+ |y|
α
)
for
x, y ∈ R
and0 < α ≤ 2.
Thefollowing resultis a onsequen e of (8)and the Hölderinequality.Lemma 16. Assume that
0 < α ≤ 2
. Let(Ω, F , m)
be a nite measure spa e. Let(A
n
)
n≥1
⊂ F
. Assume that(f
n
), (g
n
) ⊂ L
α
(Ω, m)
satisfy \A
n
|f
n
|
α
dm → 0
and \Ω
|g
n
|
α
dm =
O(1)
asn → ∞.
Then \A
n
(|f
n
+ g
n
|
α
− |g
n
|
α
) dm → 0
asn → ∞.
4.2. Symmetri
α
-stable pro esses. Re all that a real random variableX
has a stable distribution if for anya, b > 0
we an ndc > 0
and a real numberd
su h that the distributions ofaX
1
+ bX
2
and ofcX + d
are the same,whereX
1
, X
2
areindependent opies ofX
(oneprovesthenthatthere existsα = α(X)
,0 < α ≤ 2
,su hthatc = (a
α
+ b
α
)
1/α
)
.Inwhatfollowswe will onsideronly the symmetri ase(i.e. thedistribution of
X
andof−X
arethesame, f.theGaussian ase).Inthis ase, the hara teristi fun tion ofX 6= 0
isof theformE(e
itX
) = e
−|t|
α
σ
for some positive
σ
(t ∈ R
). Let0 < α ≤ 2
. LetS
be an arbitrary ountable set. LetX = (X
s
)
s∈S
be a pro ess dened on a probability spa e(Ω, F , P )
. We say thatX
is (symmetri )α
-stable if ea h nite linear ombinationY =
P
m
i=1
a
i
X
s
i
(
a
i
∈ R
,i = 1, . . . , m
) is a symmetriα
-stable variable, i.e. there existsσ ≥ 0
su h thatEe
itY
= e
−|t|
α
σ
for all
t ∈ R
(andσ > 0
wheneverY 6= 0
). We thenwrite((Y ))
α
= σ
1/α
.
Remark 6. For
1 ≤ α < 2
(α = 2
leads to the Gaussian ase)((Y ))
α
turns outto bea normintheBana hspa eT
0<r<α
L
r
(Ω, P )
,andfor ea h0 < r < α
thereexistsc = c
α,r
su hthat((Y ))
α
= c
r,α
kY k
r
forea hY
whi h isα
-stable. For0 < α < 1
ina similarmanner we obtaina Fré het spa e.Thefollowing theoremhas beenprovedin[30 , pp.127128℄.
Theorem 17. Assume that
0 < α ≤ 2
. Assume moreover thatX =
(X
s
)
s∈S
isanα
-stablepro ess.Then there existsanite positive Borel mea-sure ( alled a spe tralmeasure ofX
)m
onR
S
su h thatE exp
i
n
X
j=1
a
j
X
s
j
= exp
−
1
2
\R
S
n
X
j=1
a
j
x
s
j
α
dm(x)
for arbitrary
a
1
, . . . , a
n
∈ R
ands
1
, . . . , s
n
∈ S
, wherex = (x
s
)
s∈S
. Remark7. Itfollowsthat((
P
n
j=1
a
j
X
s
j
))
α
α
=
1
2
TR
S
|
P
n
j=1
a
j
x
s
j
|
α
dm(x)
. Remark 8. When0 < α < 2
,the measurem
is notunique.4.3.
α
-stableautomorphisms. WesaythatanautomorphismT
ofa stan-dard probability Borelspa e isα
-stable if there exists a linear spa eB
0
of real fun tions onX
su hthat1.
B(B
0
) = B
,2. for ea h
0 6= f ∈ B
0
,f
is anα
-stablevariable, 3.B
0
isT
-invariant.The following riterion as well as the method of proof arevery lose to the ergodi ity riteria in[17 ℄and [15 ℄.
Proposition 18.
T
is ergodi i for ea hf, g ∈ B
,((f ◦ T
n
− g))
α
α
→ ((f ))
α
α
+ ((g))
α
α
along a subsequen e of
n
'swhose omplementhas densityzero. Moreover,ifT
isergodi thenT
isweaklymixing.4.4. Self-joinings of ergodi
α
-stable automorphisms. From now on we assumethatT
isanergodiα
-stableautomorphismof(X, B, µ)
andB
isitsα
-stablesubspa e.Assume that
̺ ∈ J(T )
. We say that this self-joining isα
-stable if the variableF (x, y) = f (x) + g(y)
as a variable on(X × X, ̺)
isα
-stable for ea hf, g ∈ B
.Remark 9. A ording to the denition of
̺
the automorphismT × T
a ting on(X × X, B ⊗ B, ̺)
isα
-stable with itsα
-stable spa e being the losureofB
0
(̺) = {f (x) + g(y) : f, g ∈ B}
.In this se tion we will prove that ea h ergodi
α
-stable automorphism has theELFproperty.Proposition 19. Let
T
be an ergodiα
-stable automorphism. Assume thatΦ = lim
t→∞
U
T
nt
,Φ = Φ
̺
.Then̺
isα
-stable.Proof. Take
f, g ∈ B
ands ∈ R
.We have \X×X
e
is(f (x)+g(y))
d̺(x, y) =
\X×X
e
isf (x)
e
isg(y)
d̺(x, y)
=
\X
Φ(e
isf
)(y)e
isg(y)
dµ(y) = lim
t→∞
\X
e
isf ◦T
nt
· e
isg
dµ
= lim
t→∞
\X
e
is(f ◦T
nt
+g)
= lim
t→∞
e
−|s|
α
((f ◦T
nt
+g))
α
α
.
Hen efor some
σ ≥ 0
and anys ∈ R
,lim
t→∞
e
−|s|
α
((f ◦T
nt
+g))
α
α
= e
−|s|
α
σ
.
Itfollowseasilythatif
σ = 0
thenf (x)+g(y) = 0
for̺
-a.e.(x, y) ∈ X ×X
. From now on we xN
0
⊂ N
su h thatN \ N
0
has density zero and for ea hf, g ∈ B
,((f ◦ T
n
− g))
α
α
→ ((f ))
α
α
+ ((g))
α
α
as
n → ∞
,n ∈ N
0
(whi h usesthefa tthatT
isweakly mixing).Using Proposition 18 and the denition of an
α
-stable self-joining we obtain the following.Lemma20. Assumethat
T
isan ergodiα
-stableautomorphism withB
itsα
-stable subspa e andN
0
as above. Assume that̺ ∈ J(T )
isα
-stable. Then̺ ∈ J
e
(T )
i for ea h
f, g, h, j ∈ B
,((f (T
n
x) + g(T
n
y) − h(x) − j(y)))
α
α,̺
→ ((f (x) + g(y)))
α
α,̺
+ ((h(x) + j(y)))
α
α,̺
as
n → ∞
,n ∈ N
0
.We an now apply Lemma 2 to
e
if
, e
ig
, e
ih
and to
e
if
, e
ig
, e
ij
to obtain thefollowing.
Lemma21. Assumethat