Internat. J. Math. & Math. Scl. VOL. 16 NO. 2 (1993) 277-282
277
A METRIC SPACE ASSOCIATED WITH A PROBABILITY SPACE
KEITHF. TAYLORandXIKUI WANG
Department
ofMathematics Universityof SaskatchewanSaskatoon,Sask. CanadaSTN 0W0
(Received January 7, 1992)
ABSTRACT.
For
acomplete probabilityspace(9t, I,P),
theset of allcomplete sub-(z-algebrasof )2,
S(Y:.),
is given a natural metric and studied. The questions of when5’(Z)
is compact or connected are awswered and the important subset consisting of all continuous sub-a-algebras is shown tobe closed. Con,aections with Christensen’s metricon theyonNeumann
subalgebrasofaType
II]-factorarebriefly discussed.KEY
WORDSAND
PHRASES.
a-algebras,conditional expectations,metricspace,yonNeu-mannalgebra.
1992 AMS SUBJECT CLASSIFICATION
CODES.
Primary"28A05,
28A20. Secondary:46L10.
1. INTRODUCTION.
Let
(f/, E,
P)
beacomplete probabilityspaceandletS(I)
denotethe set ofallcomplete sub-(z-algebrasofI]. Wewill defineametric don5’(Z)
and investigatesometopological propertiesoftheresultingmetricspace
(S(Z),d).
In[1],
Boylanintroduceda metricd’
on3’(Z)
for thepurpose of studying convergence propertiesoftheconditional expectationoperatorsdefinedby varying sub-(z-algebrasofZ. Ourmetricturns out to beequivalent toBoylan’sand seemstobemore convenient forstudyusing fitnctionalanalytic methods.Weprove, in section3,that
(S(Z),d)
iscompact if and only ifZ
ispurelyatomic. Wealso show that(S(Y_,),d)
isconnected if andonlyifZ
has at mostoneatom.In
section4,weconsider the continuous sub-(z-algebrasofZ
and show that they formaclosed nowheredense subsetof(S(Z),
d).
Thereis acloseanalogy between probabilityspaces and von
Neumann
algebras witha faithful finitenormal trace.In
fact,ourdefinitionof dismodeledon ametricdefinedbyChristensen[2]
on the set of all yon Neumann subalgebrasofaType II-factor.
Christensen’s metrichasbeen usefulin the studyofthe index in
II-factors
(see
[3]
and[5]).
Wegiveashort discussionofacommongeneralizationofourmetricand Christensen’s in section5.
2.
THE
METRICS.In
thissection,wedefinethemetric dandBoylan’smetricd’
and show thattheyareequivalent. IfZo
S(Z),
then letL(Zo)
L(f, Zo, P)
be consideredas aclosed subspaceofL(f, Z,
P)
L
(o),{f
GL’’(0)
ilflloo
_<
1}.
ForEl,
E
ES(E),letd(Vl, 2)
7.J-{
sll|) infIlf
gll2,
sup infII.f
gll2}.
Itiseasytocheckthat, disa metricon
S().
Forthe metric
d’,
we use a definitionduetoRogge
[6]
which isaslight variationon Boylan[1].
For
EI,E
ES(Z), letd’(Ea,_E)
sup infP(AAB)V
sup infP(AAB).
AZBEE BEEA
d
(Boylan’s metric d" has Vreplacedby
+
andclearly1/2d*
< d’ _<
d’,so and d* areessentially thesame).
Thearguments in[1]
showthat(S(E),d’)is
acompletemetricspace.Forany
Eo
S(E),
let0
denote theorthogonal projectionofL(,Z,P)
ontoL(fl,
E0,P),
considered as asubspace ofLz(E).
LetE
denote the rtriction of e toL(E).
As is wellk.own, e and
E
are restrictions of the conditional expectation mapping ofL(fl, E,P)
ontoL(,
E0, P). We will use any of the well known special properties ofconditional expectationwithoutgiving references.
Wenownsider the relationship between themetricdand
d’.
THEOREM 1. Forany
E,
E
S(E),
d’(E,,
E)
d(E,,)
22d’(,,
Z)(1
d’(E,,
E))
Thus
d’
and dareequivalent metrics.PROOF. To prove the left. hand inequality, fix
A
E.
It is shown in 2.1 of[4]
thatif
P(AAB)
isachieved at thesetC
{EE(XA)
>
},
where XA isthe indicator function of theset A.Some
elementary manipulations show thatP(AC)
-
I-
Er(XA)
dPNow
II-
XAII,
,
oP(AC)
I1
x.ll,
-I
E(XA)II,
IIX*
E(*)ll,
IIX*
E(XA)II*
Thus
inf
P(AAB)
d(, ),
for allA
.
Symmetric arguments apply for2,which givesd’(,,
)
5
d(,,
,).
E
,
For
the right inequality, note that, for anyf
EL(),
(f)
f
andE*(f)
is the(L
-norm)
closest element ofL(E)
tof.
Thus,
sup inf
Ilf-gll
supI[f-E(f)ll,
g
supIIW’(f)-E(f)ll,
lL(hai() IL() fL(h
Similarly for theother term in d(E,),sowehave the inequality
(,,2:)
pIIE’(f)-S(f)ll
METRIC SPACE ASSOCIATED WITH A PROBABILITY SPACE 279
and theorem 3in
[6]
shows thatThus,
lIE
-’
(f’)
E2(.f’)[[2
<_
v/2d’(E,,
E2)(1
he(E,,
Z;))
lIE"’
(.f)
E(/)II
_<
Thiscompletes theproofofthetheorem.
COROLLARY. Theidentity map isahomeomorphismof
(S(E),d’)
with(S(E),d).
REMARK
1.Not
onlyarcd andd’
equivalentinthesensethat theydefine thesameopenscts inS(E), but thesamesequences areCauchy ineachof(._q(E),d)
and(S(E),d’).
Since(S(),d’)
iscomplete,wehave that(._q(S).d) isacomplete metricspace.
REMARK
2. The right hand inequality intheorem cannot be improved tooneoftheform d_<
kd’
forsomepositiveconstant I,’,asisshownbythefollowing example.EXAMPLE.
Let12[0,
1],
withP
denotingLebesguemeasureonZ,the Lebesgue measurable stbsets of[0,
1].
For each a>
0, letE
denote the a-algebra generated by an atom[0, a]
andEl"l(a,
1]
{B
-
S C (a,1]}.
ForanyA
,
letB
Af’l(a,
1]
andB
[O,
a]
U(
A
l"l(
a,1]).
Then
B,B
fiE,
and eitherP(AAB)
orP(AAB)
islessthan orequalto.
IfA
[0,
],
thenP(AAB) >
,
forallB
E,
Thusd’(E,,
E)=
To
computed(E,Z),
note that, for anyf
L(E),
thefunctionf..(,,,]
is inL(Y:.o)
and]]f
f.Y(.]]l
-<
.
On theotherhand,iff
.([0,]- .X(,,],thenEo(])
0almosteverywhere.Thisimplies that
]].f-
gll
->
ll.f[]2
v/.
forany g(L(E,).
Thusd(Zo, E)
yrs.
With
E
E,
andE
,
the inequalities in theorem become_<
_<
2-()
(_<
’).
Thusnok existswithd(Z,, E)
_<
kd’(,E)
for alla(0,1).
3.
THE MAIN
RESULTS.Weconsiderthemetric space
(S(), d)
andestablishthefollowing two theorems. Since(S(),
d’)
ishomeomorphicto
(S(F_,),d),
these characterizationsholdfor eithermetric.THEOREM
2.(S(F_,),d)
iscompact ifand only if:E
ispurely atomic.THEOREM 3.
(S(F.,),d)
isconnectedifand only if has atmost oneatom.Before proving these theorems,weneed thefollowinglemma whichisafolkloreresult ofmeasure theory andcan beproven withaneasy applicationof
Zorn’s
lemma.LEMMA
If(X,.T’,u)
is a finite measurespace with no atoms, then there existsa mapF"
[0,1]
.T"
such thatF(s)
C_F(t)
and,(F(t))
tI(X)
forall s,t.
[0,1]
withs_<
t.Wewill usethe notation, forA
E
andAo
{AB"
B
c= 0}
PROOF
OFTHEOREM
2.Assume
that(12, E,P)
is notcompletely atomic.
Let fie
and12
be the continuous and atomic parts of fl respectively.We
are assumingP(fl) >
0.Let
F[0,1]
E
be such thatF(I)
C_12,P(F(t))
280 K.F. TAYLOR AND X. NANG
A,,
U{
F[(k
)/2"-’,
{2k
I)/2"]
k 1, :2 ,2"-foreachn 1,2,... Then P(A,,)
P(f/c)/2,
for eachn andP(A,AA,,)
P(f/c)/2,
forn-
m. LetE,
debotethe a-algebragenerated
byA,,
f]E
and theatomf\A,,
for eachn.Let
f
denote the indicatorfunction ofF/\A,.
Thenf
E L(E,,) and forany9 5L(Em),
ifm n,then9differs fromf
byatleast 7oneither(F/\A,,)
CI
A,or(f\A,)f’l(fl\A,,).
Eachofthesesetshas probabilityP(flc)/4.
ThusIll
91122
_>
p(flc)/16,
whichimpliesd(E,
E,,,)
_>
/P(fc)/4,
ifn-fl
m.Hence,
the sequence(E,,),,__
has nocluster points andS(E)
isnotcompact.Conversely, suppose
(f,E,P)
is completelyatomic. Since(S(.,),d)
iscomplete, compactness will follow ifwe show thatS(E)
istotally bounded. IfE
hasonly afinitenumber of atoms, thenS(E)
is afinite set and obviously compact. So suppose{A,
A2
aredisjointatomsgeneratingE. For
>
0, chooseNsuch that:__N+
P(A,,)
<
.
There
areonlyafinitenumberofalgebras
ofsetscontained in the
algebra generated
by{A
A2v
}.
Let
.A,...,
,4 denote thesealgebras. For
<
j<
k,let denotethea-algebra generated by,4j
and{AN+I,
AN+,.
..}.
LetA
For
any’
S(E) A
Iq’
.Ai
forsomej,<
j_<
k. Then’
C_E
andit is easy tocheck thatd(E;,
Zj)
<
c. Therefore,S(Z)
iscoveredbythe finite setof-balls centered at,
<
j<
k. Sincec
>
0wasarbitrary,S(E)
istotally bounded.PROOF OF THEOREM 3. Suppose Aand
B
aretwo distinct atoms in.
LetS’
denote the set of allE0
inS(E)
which haveanatomcontainingbothA
andB,
letS
denotethe set ofall0
inS(Z)
whichhave two disjoint atoms containingA
andB,
respectively. Itiseasy toseethatS(E)
S
[JS.
LetS
andE
S.
LetA’
andB’
be disjoint atoms ofE
containingA
andB,
respectively. Let Cbe an atom ofEl
which containsAI.JB.
Let
f
XA’
X’
-
L()
Foranyg
L(Z),g
a,almost everywhereonC,
forsomeconstant a,lal
<
1. ThusIIf
gll
>
fA
If
gl
de+/
If
gl
de/(
-a)P(a)
+
(
+
a)P(B)
>
/2min{P(a),P(B)}
Therefore,
d(,E) >
/min{P(A),P(B)},
for allr
S
andZ
ES.
Noticealso that the trivialo-algebraE
{f,
1}
S
andE
S.
SoS
andS
arenonempty disjointopensubsets ofS(E)
with unionS(E). Hence,
S(E)
is notconnected if therearetwoor moreatomsin ).?,.Conversely,suppose
(f/, E, P)
has at mostoneatom. If thereis anatom, callitA
and if there isnoatom, letA
.
IfP(A)
1, thenS(E)
isaonepoint space, whichisconnected.So
assumeP(Ft\A)
>
0. Wewillshow that anyE0
S(E)
canbe connected toE; byanarcinS(E).
Firstassume
E0
S(E)
andE0
iscompletelyatomic,generated bythe disjoint atomsA0,A,...
(a
finiteor infinite collection). We may assumeA
C_Ao.
Foreach n_>
1, by the lemma, there existsF,,
[0,1]
A,,f’IE such thatP(F,(t))
tP(A)
andF(s)
C_F,(t)if
s,t[0,1]
with s<
t. Similarly, letFo
[0,1]
A01DE
be such thatF0
is increasing andP(Fo(t))
P(A)
+
tP(Ao\a),
for all[0, 1].
Foreach[0,1],
defineZ(t)
S’()
as ther-algebra generated by{F,(t)CIE, A,,\F,(t)’n
0,1,2}.
It
isclear thatE(0)
Eo,
E(1)
E
andE(s)
C_Z(t)if
METRIC SPACE ASSOCIATED WITH A PROBABILITY SPACE 281
For
any.[
6_.L((t))l
and for each,
0,1,2,...,f
isconstant,saya,,,almost everywhereon the atom.4,,\F,,(t)i,,
E(/). Define g EL’’(E()),,
by making 9 a,, onA,,\F,,(s)
and9f
onFn(s),
forn 0, 1,2 Thenf
andy differonlyonU(F,,(t)\F,,(s))
and therebyatmost2.Now
P{(F,,(t)\F,(s)))
(ts)[P(Ao\A)
+
P(A,)
+
P(A2)
+...]
Th
I1- 11
<
This implies that
d(V(s),V(t))
<
2//-’Z-
s. Therefore,any atomic element:E0
ofS()
iscon-nected byan arcto v
If
:E0
is not completely atomic, let ftic
(J fl, whereflc["l
0
is continuous andfl
:E0
is atomic.Let
:E
be theatomic a-algebra generated by an atomequal tofie
andfl
]0.By
methodsimilar tothe above, but justworking with fie,an arcfrom
:El
to:E0
canbefound. Also,beingatomicis connected byan arc to
.
Thusanyelement:E0
ofS(:E)
can be connectedby an arcto :E. ThereforeS()
isconnected, in fact,arcwise connected if(fl, :E,
P)
has atmostone atom.The two theorems ofthis sectionshow that thereis somemeaningflfl connection between the
topologyof
S(:E)
and thestructure of theoriginal probabilityspace. Since thistopology
isreallyanembodimentof the equiconvergence property for conditional expectations,wefeel that
S(:E)
with this metricd(or
ifoneprefersd’)
will provideagoodlocale forstudyingconditionalexpectations. 4.CONTIIIUOUS SUB-a-ALGEBRAS.
Let
S(:E)=
{0
6S(:E)’
(fl:E0,
P)has
noatoms},
the space ofcontinuous sub-a-algebrasof E. Ofcourse, if is not acontinuous a-algebra, then
S()
.
On
the otherhand,
if:E
is continuous, thenS(:E)
isavery rich set tolookat. For example,ifG
isagroup of :E-measurable transformations of12, let denote thea-algebraof G-invariant sets in.
IfGis afinitegroup, thene S(:E).
This providesaway in whichtoconstructmanyinteresting examplesofelements ofS(:E).
In
spite ofthis,Sc(:E)
isa"small" subset ofS(:E).
THEOREM
4.S(E)
isaclosed nowhere dense subset ofS(]).
PROOF. We
firstobserve that if:E0
ES(:E)
andA
E,
then{P(AB)"
B
e :Eo}
[0,
P(A)].
Nowfix
e S()\S(:E)
and letAbeanatomof.
By
the aboveobservation,forany:Eo
e
there exists
B
e
:Eo
such thatP(A["IB)
1/2P(A).
Let
f
X -Xu\L(0).
Forany gL(F,),g
is constantonA
andoneeasily checks thatIll-
gll2
>-(P(A)) /:.
Thus
d(:E0, :E) _> (P(A))
/forany0
e
S(:E).
HenceS()\S(:E)
isopen andS(:E)is
closed.To
see thatS(])
is nowhere dense, let:E0
ESc()
and e>
0. ChooseA
forthis particular
E0
and1,d(
x",-,o,)
(P(A))
/ Thusthe openballof radiusecenteredat
E0
is not contained inS.(E)
forany>
0.Since5’c(E)
isclosed,
it isnowhere denseinS().
Thus
(So(E),
d)
is itselfaconplete metric space whose topological propertiesshould beinter-esting to study.
We
donot consider suchastudyinthis paper.5. CONNECTIONS WITH VON
NEUMANN ALGEBRAS.
There is aone-.to-onecorrespondence between the sub-a-algebrasof
E
and the yonNeumann subalgebrasofL(2, E,
P).In
this section,weassumethe readeris familar with the basictheoryofvon
Neumann
algebrasasfound ina reference such asSa"kai[7].
Let
.M
beafixedfiniteyonNeumann
algebrawithadistinguished faithful, finite, normaltracer. For example,
.M
could
beL(,E,P)
or aType
lI-factor.
Let
S(.JM)
denote thesetofallyon Neumann subalgebrasof21 whichhave thesameidentityas.A4.In
the case that.M
is aType
ll-factor, Christensen defined a metric onS(.M)
in[2].
He showed thatS(AJ)
then becameacompletemetricspace.He
used hismetrictostudy perturbationproperties of subfactors and his metrichas also been used in thestudyof the index ofasubfactor in a
Type
II-factor
(see
[3]
and[5]).
Christensen’s definition works well in thesituation we are considering. Forx .A4,letIIllz
9r(’)
ForA/,
S(.A4),
letA/’
and denote theunit balls inA/’and,
respectively.Then define
d(A/’,
)
max{
sup inf xYlI,
sup infII
xYlI}
It
is routinetocheck thatd isametriconS(.A4)
and Christensen’sproof thatS(.A4)
iscompletecarriesovertothis moregeneralsituation.
Someoftheproofsofsection3canbeadaptedtotheyon
Neumann
algebra situationbut not all.We
list below what we can prove and leave the details of the adaption of the proofs totheinterestedreader.
THEOREM 5. Let .A4 be a finite yon Neumann algebra with a fixed faithful, finite trace
r. With the Christensen’smetric,
S(.M)
iscompact if and onlyif.A4i.s
generated byitsminimal projections.PROPOSITION6. If
S(./t4)
isconnected,then.A4 hasat mostoneminimalprojection.OPEN PROBLEM Istheconverseof proposition 6
true?
ACKNOWLEDGEMENT.Thisresearchwassupported byan
NSERC
Canada operating grant.!EFERENCES
1. Boylan, E. S. EquiconvergenceofMartingales,
Ann_.
Math._.__.Stst.
42
(1971),
552-559. 2. Christensen, E. SubalgebrasofaFiniteAlgebra, Math____:.Ann___..
243 (1979),
17-29. 3.Jones,
V. F.
1.Index forSubfactors, Invent.
Math.(1983),
1-25.4. Kudo, H.
A Note
ontheStrong Convergence
ofr-algebras, Ann____. _2(1974),
"/6-83. 5.Mashhood, B.
and Taylor,K. F. On Continuityof theIndexofSubfactorsofa FiniteFactor,
J.
Func.
Anal. 12(195S),
56-66.6.
Rogge,
L.
Uniform Inequalities forConditional Expectations,Ann;
Prob.___..2(1974),
486-489.