You are on page 1of 9

On the Construction of Almost Uniformly Convergent Random Variables with Given Weakly

Convergent Image Laws


Author(s): Michael J. Wichura
Source: The Annals of Mathematical Statistics, Vol. 41, No. 1 (Feb., 1970), pp. 284-291
Published by: Institute of Mathematical Statistics
Stable URL: http://www.jstor.org/stable/2239738
Accessed: 05/07/2010 14:12
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=ims.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and extend access to The
Annals of Mathematical Statistics.

http://www.jstor.org

Statistics
TheAnnalsofMathematical
1970,Vol. 41, No. 1,284-291

ON THE CONSTRUCTION OF ALMOST UNIFORMLY


CONVERGENT RANDOM VARIABLES WITH GIVEN WEAKLY
CONVERGENT IMAGE LAWS'
BY MICHAEL J. WICHURA

of Chicago
University
d, and
metricspace,withdistancefunction
Let S be an arbitrary
1. Introduction.
distributions
letY' be itsBorela-algebra.DenotebyY(S) theclass ofall probability
on (S, 9). A net(Py)y of probabilitiesP e (S) is said to convergeweaklyto a
probabilityP eG (S) if P(f) = lim Pf(f) foreach real-valuedboundedcontinuous
functionf on S; here P(f) = ffdP, PY(f) = ffdPY. Let gs(S) denote the subclass of CA(S)consistingof thoseprobabilitiesP forwhichthereexistsa separable
one. gS(S) includestheso-calledtightprobabilities
subsetofS in f ofP-probability
i.e. probabilitiesP such thatsup {P(K): Kcompact} = 1 ([5] page 29). The chief
resultofthispaperis statedin thefollowing.
be a netofprobabilities
THEOREM1. Let (S, d) be a metricspace and let (Py)yEr
P E gs(S). Thenthereexistsa probability
weaklytoa probability
PYG i7(S)converging
on Q
X andX/(yE F) defined
S-valuedfunctions
space (Q, ., ji) andA-9- measurable,
- 1 ofXyare respectively
uX- ' ofX and luXy
P andPy(yE F)
suchthatthedistributions
toXalmostuniformly.
andsuchthatXyconverges
One sometimes([1], [8]) has occasion to considerthe weak convergenceof
distributions
probability
Pywhichare definedonlyon certainsub-c-algebrasof9,
in Theorem1 thatthe
of interestto knowthatthe requirement
and it is therefore
PY belongto 3d(S) can be weakened.To make thisprecise,let us say thata net
ofprobabilities
(Py)yEr
Pydefinedon sub-c-algebras-4yof " convergesweaklyto a
P E ^(S) iflimyPY(f) = Pff) = limyPY(f) foreach real-valuedbounded
probability
theupperand lower
continuousfunctionfon S; here Pyand Pydenoterespectively
associatedwithPY:
probabilities
PY(f) = inf{Py(g):f < g, Py(g)defined}
PY(f)

sup { Py(g):f?> g, P,(g) defined}

of thisdefinition
see Theorem1 of [8]). It is clearthat
(forequivalentformulations
thisdefinition
of weak convergence
reducesto theusual one ifall the -4yequal Y.
Let Y' denotethesub-c-algebraof S9 generatedby the open balls of S. We then
extensionofTheorem1:
havethefollowing
THEOREM2. Let S, 9, and Yo be definedas above and let (Py)yerbe a net of
in 9, which
probabilities
g'o and contained
Py,definedon a-algebras 4y containing
ReceivedOctober23,1968;revisedAugust7, 1969.
l Thisresearch
was supported
in partbyResearchGrantNo. NSF-GP-8026fromtheDivision
of Mathematical,
Physicaland Engineering
Sciencesof theNationalScienceFoundation,
and,in
partfromtheStatistics
Branch,Office
ofNavalResearch.
284

ALMOST UNIFORMLY CONVERGENT RANDOM VARIABLES

285

Pe Ys(S). Thenthereexistsa probability


convergesweaklyto a probability
space
onQ suchthat:
(Q, X, ji) andS-valuedvariablesX andXy(yeF) defined
(1)

X is . -J9 measurable, Xy is M -.y

(2)

=x-

(3)

- P,

Xy-X

measurable (yeF)

uX,-1 = PY(yeFr)

almostuniformly.

withall theusual axiomsof settheoryto assume


We remarkthatit is consistent
in Theorems1
thatYS(S) = Y(S) (see [2] page 252). In thissense,therequirement
and 2 thatPG Y,(S) can be replacedbythetrivialone thatP Y
g(S).
In the construction
used to validateTheorems1 and 2, Q0is theproductspace
Sx HlyerSy,
whereeach Sy is a copy of S, M is a a-algebra whichcontainsthe
1uis theprolongationto (Q, X) of a mixture
sl = 9 x Hlyerdy,
product F-algebra
of productprobabilitieson (Q, d), and X and Xyare thecanonicalprojectionsof
LI onto S and Sy(ye F). When F is countableand S separable,one has -=a.
havebeenusedto validatespecialcases ofTheorem1. Working
Otherconstructions
is equivalentto almostsure
withsequences(forwhichalmostuniform
convergence
convergenceby Egoroff'stheorem)instead of nets, Skorokhod ([7], Theorem
3.1.1) has provedTheorem1 for S separableand complete;in his construction
Q = [0, 1], theunitinterval,4 is thea-algebraof its Borel sets,and p is Lebesgue
measure.Again workingwith sequences,Dudley ([3], Theorem 3) has proved
Q is a countableproductof copies
Theorem1 forS separable;in his construction,
of S x [0, 1], a is the producta-algebra on Q2,and p is a mixtureof product
of almostsurely
probabilitieson (Q, X). For applicationsand otherconstructions
see
convergent
processeswhichare of interestin thetheoryof weak convergence,
thesurveypaperbyPyke[6].
2. Proof for F countableand S finite.The simplicityof our constructionis
obscuredin thegeneralcase by severaltechnicalconsiderations;in orderto illustratethe generalidea we willin this sectionproveTheorem1 underthe assumptionsthat F is countableand S is finite.To this end, let (Sy,fy) be a copy of
(S, S9) foreach y, and let (Q, A)=(S xHlyr Sy,Y x flyrJS"y)be the productof
the measurablespaces (S, f) and (SY, Vy)(yeF). Let the canonical coordinate
mappingsX and Xy(yE F) be definedon Q by
(4)

X( (s,(so)oer)) = S,

Xy((s, (SO)oer))

The requiredmeasurability
properties
clearlyhold.
Let k: y-- k(y)be anyfunction
fromF to {0, 1, 2, ..,
(5)

sy(y F).

o} suchthat

limye r k(y)= oo

(k(y) should be thoughtof as a measureof the largenessof y and later will be


further
specified).For I < k < oo, set
(6)

= X}.
Uk = nyy:k(y)>k{XY

286

MICHAEL J. WICHURA

Observethat each UkE X since F is countableand that Xy-+ X uniformly


over
each Ukin viewof(5).
Let Qy(yeF) be any familyof probabilitieson (S, f). It laterwill be further
givingmass one to thepoints E S, let
specified.Letting3, denotetheprobability
Pi=s =

(7)

bs x HlyerMj,s,y

(1 ? j < oo,s eS) denotethe productprobability([4] page 166) on (Q, X) whose


componentsare respectively:
bs definedon (S, f), and
= QY
Pj,s,y
= bs

if

0 ? k(y)<j,

if

j < k(y)< oo,

definedon (Sy. Y.). Clearly Pui,X = bs Pi'sXY-I = pjsy; moreover,since F


is countable,Mj s(Uk) = 1 forj1? k. Next, defineprobabilities,j(1 < j < oo) on

(Q, X) by

Hj= sEs P{S}',s.

(8)
Clearly 1X-1 = p

tjXy- = Qy
=P

if

0 ? k(y)<j

if

j<k(y)<

oo,

= I ifj? k.
and puj(Uk)
Finally,let (wk)1
(9)

Wk _

(10)

k<0,

be any sequenceof numbersWksatisfying


Z
Ej<k

Z Wk = 1
Ek
Clk

Wk<

1(1 k < oo)

andput

El<j_ k Wi(O< k _ o).

Note w0 = 0, w,, = 1. Definetheprobabilityp on (Ql,X) by


p=

(I1)

Ejwjpj.

Clearly iX- 1 = p
( 12)
(13)

1elXY -

)k(y) P + (I -t)k(y))Qy

/(Uk) >?

and

< k < oo).

k(l

withrespect
Since limk-o Ck = 1, (13) impliesthat Xy-+ X almost uniformly
in
of (12), to
in
it
view
this
suffices,
setting
to p. To comp1-tethe proof
special
of Pyto P impliestheexistenceof k(y)'ssatisfying
showthattheweak convergence
(5) and probabilitiesQysatisfying
(14)

Py=

0k(y) P +

-(l

Wk(y))QY

(14) ifand onlyif


forall yCF. Now ifk(y) = oo,thereexistsa Qysatisfying
(1 5)

P1. = P,

287

ALMOST UNIFORMLY CONVERGENT RANDOM VARIABLES

and thenany Q, willdo. On theotherhand,if0 < k(y) < oo,we see,aftersetting


(16)

qk=s,y= Py{s} +(w)k!(l -ck))(Py{s}

(17)

Mk.y=

min ES

P{S})

qk,s,y

that thereexistsa probabilityQysatisfying


(14) if any only if
Es Es qk(y),s,y = 1, and thenone musttake
(18)

Ink(y),y?

and

Qy = EseSqk(y),s,y6s.

> 0. Thus itsuffices


We notethatEs Esqk,s,y = 1 forall k(O < k < oo) and thatmo0y
to showthat(5) is satisfied
and (15) holdsfork(y)= oo ifwe put

k(y)=sup {j _ 0: mj, > 0}.

(19)

Now sincePy-+ P, we have


(20)

Py{s}

-+

P{s}

foreach se S(Ifs1beinga continuousboundedfunctionon the discretespace S).


Hence limyerqksy= P{s} foreach se S, 1 ? k < oo; this,togetherwiththe fact
thatqk,s,y > 0 if P{s} = 0, impliesthatqk,,,y is ultimately
foreach
nonnegative
k (1 ? k < oo). Thus since S is finite,thereexistsforeach k an indexykeIT such
that mk,y_ 0 for all y _ y(k). Since mk,y ? 0 impliesk(y)? k, (5) is satisfied.
Next,ifk(y)= oo, we have (recallZs s qk,s,y = 1)

(21)

0?

Py{s}?(+

k/(1-

k) )(Py{s}-P{S})

< 1

for each s e S and arbitrarilylarge k; since Wk/(l - Wk) -+ oc, it follows that
Py{s} = P{s} for each seS, i.e., that (15) holds. This completesthe proof of
Theorem1 forF countableand S finite.
2 in the general case. Let P, Py(yeF), ?, Y'0, and
3. Proof of Theorenm
e
be as in Theorem2. Let W(P) = {C e : P(boundaryof C) = 0} be the
y(y-F)
class of P-continuity
sets.We recall ([5] page 50) thatW(P) is an algebraand that
foreach s e S, theopen ball
(22)

{t: d(t,s) < r} E W(P)

forall but at mostcountablymanyvalues of r. The followinglemmashowsthat


theanalogueof(20) holdsforsetsCe-(P) (conferT1.1 of [8]):
LEMMA

1. In thepresentcontext,C e(P)

implies

limye rPy(C) = P(C) = limye r PY(C).


PROOF. Let F be a closed subsetof S. Since the continuousboundedfunctions
fn:s - max((1 - nd(s,F)), 0) decrease to the indicatorfunctionof F, the weak
of Py to P impliesthatlimsupyPY(F) _ limsupyPY(fn)= P(fn)I P(F).
convergence
The dual relationforopen sets is seen to hold by takingcomplements;thus for
anyC S/'we have

(23)

P(C) ? liminfy
PY(C) ? limsupyPY(C) < P(C),

288

MICHAEL J. WICHURA

where C (resp. C) denotesthe interior(resp. closure) of C. When Ce%'(P), the


extrememembersof(23) are equal. 0
We shall need a sequenceof "finiteapproximations"to S. For this,choose and
fixanytwo numericalsequences(Ak)l I k< and (8k)1 <k< . suchthat
(24)

Ak >

(25)

Ak = O

limk,

Ek Sk <

Ek > ?,

00-

Lettingd(C) = sup{d(y, z): y, z e C} denotethe diameterof a subsetC of S, we


thenhave
LEMMA2. In thepresentcontext,thereexistpositiveintegersnk(l _ k < oo) and
disjointsubsetsCm. . ., mk(O < mj nj, 1 j ? k) ofS suchthat

(26)
(27)

Cm1,... mk-1

CmI,

EO<mkk<nk

1<j<kmax1

maxO<j<nj,

(28)

-,,nk

_mk_nkd(Cml,

ZO<mj_nj,l 1j<kP(Cml

_1,1k

,mk)

Ak

-k

nmk- l,O)

n7Y?(O _ mj _ nj, 1 _ j < k).


PROOF.Let E be a separablesubsetof S suchthatP(E) = 1 and let {sn, n _ 1} be
a countabledensesubsetofE. In viewof(22), thereexistsforeach n ? 1 an open ball
in S, call itEn,centeredat Sn withradiusgreaterthan'A1 butlessthanA1,suchthat
En6W(P). Since the union of these balls covers E and hence has P-probability
one, there exists a positive integern1 such that P(Un n, En) ? 1-81. Setting

(29)

CmI = EmI -E<m<m

CMl,

c-

* * *, mk

(P)

? m1 ? n1),

Cm(1

CO =

S-Un<nlEn

max,?<m?nid(Cml)< A1,P(CO)

S-E

CO,C
81C,

1m?nm?nCm1Iweget

** ,Cn,1e(P)nft0
The proofis completedbyinductionon k. 0
Let Hk(1 ? k < oo) be the finitepartition of S whose members are the Cml,...mk,
and put Ho = {S}. Choose and fixnumbersWksatisfying
(9) and defineokkby(10).
<
For 0 k < oo, CeHik,and y eF, set(confer(16), (17), and (19) )
S=EOZmi?niCmi,

qk,C,y =

(30)

Py(C) + (Py(C)

P(C) )(k/(l

mk,y = minEenk qk,E,y

k(y)=sup {j

0: m., _ 0}.

of Pyto P implies(see theargument


In viewof(29) and Lemma 1,theconvergence
following
(20)) that
limyErk(y)= oo

(31)

For y suchthat0 < k(y)< oo,put(confer(18) )


Qy = EC E nk(y)qk(y),C,yPy(

(32)

C),

on (S, sly) obtainedfromPybyconditioning


wherePy( C) denotestheprobability
on theoccurrenceof theeventC. It is easy to see thatQyis itselfa probabilityon
(Sya,sy)and that(confer(14))
(33)

(-)k(y)

(ZC e

rk(y)

Py(|

C)P(C))

+(1

Ok(y))Qy

Py

289

ALMOST UNIFORMLY CONVERGENT RANDOM VARIABLES

Now for each yeF, let Sy be a copy of S, and let (Q,)-(Sx


fl, rSY,
the productof themeasurablespaces (S, 9) and (SY,.y)(y E F).
s E S, and let(confer(7))
Let Ck,S denotetheelementofHk containing
y X fly e FS1y)be

Vj'S= bs X flyvj,s,y
be the productprobabilityon (Q, d) whose componentsare respectively:b
definedon (S, 9), and
(34)

= PYO Ck(y)s)

if

0 ? k(y)<i,
j < k(y)< oo,

= bs

if

k(y)= oo,

= QY
VJ,S,Y

if

definedon (SY,sy). For eachj, the mappings -+ vj,s(A) is a randomvariableon


base (sinceforeach
A E a? is a cylindersetwitha finite-dimensional
(S, S9) whenever
manyC in Hk(y) belongsto ), and hence([4] page 74)
yeF, each of thefinitely
thismappingis a randomvariableforeach A E d/. Thus ([4] page 76) we maydefine
a probabilityVj on (Q,4) by the formula(confer(8)) vj = fsvjsP(ds). Finally
v on (Q, d) by
(confer(11) ), definetheprobability

v=

(35)

<-< < wjvj.


Once again,letthecoordinatemappingsX and Xy(ye F) be definedon Q by (4).
We have
LEMMA 3.

In thepresentcontext,

(37)

(y eF)
X is sl'-9' measurable,Xyis '-.4y measurable
vX-' = P

(38)

vX,- I = PY(yeF).

(36)

For (38),
PROOF. Relations(36) and (37) followdirectlyfromthe definitions.
observethat
vX

' (=

k(y)(ECenk(,)

Py(

IC)P(C))+(1

-Ok(y))Qy

if
if

to a?y
restricted

0 < k(y)<
k(y)= ci.

00,

In viewof (33), (38) holds when0 < k(y)< oo. It remainsto show that(38) holds
hereis similarto, butmorecomplicatedthan,thatat
whenk(y)= oo; theargument
(21). Put
(39)

Dk =

EO0mj?nj,1?j<kEl?mk<nk

CmI,

,mk(k _

1);

D = liminfkDk

and observethat(28) and (25) imply


(40)

P(D) = 1.

Let Wk be the sub-algebraof Y? made up of sums of membersof Hk and put


tWk; since(in viewof(26) )

W = Uk

290

MICHAEL J. WICHURA

Witselfis a sub-algebraof SO. Let u<K>nD (resp.synD,fnD) be thetraceon


D ([4] page 19) of C<@> (resp..?y, f), and let YD denotetheBorelc-algebraofD.
In viewof (24), (27), and (39), each open subsetof D is a union,necessarily
countable, ofsetsoftheformCnD withC Ec ; itfollowsthat YD C cr<KrD> = C<KC>nD.
Since Dk belongsto t7k, we have D E c<W>,and since([5] page 5) YD = fnD, we
have
(42)

De

ryrDcrD=

/DCCU<W>.

In viewof (41) and theadditivity


of P and Py,theconditionk(y)= oo implies(see
(30) and (21)) that for each Ce-W the inequalities0 < Py(C) + (Py(C) - P(C))
(wk/(l - Wk)) ? 1 hold forarbitrarily
largevaluesofk; sincelimk,
O k/(l - wk) = 00
it followsthatPy and P coincideover W,henceovero<W>, and hence,in viewof
(42), over . r)D. But then,in viewof (40), we have PY(J) = P(D) = 1, so thatP
and Pycoincideover.y Thiscompletestheproofofthelemma. ]
Now putA,, = 0 and set(confer(6) and (24))
(43)

Uk=

nY:k(y)>k{d(Xy, X)

?k(y)}

The Ukneednotbelongto ./ in general,althoughtheywillifF is countableand S


is separable(so thatd(Xy,X) is d1-measurable
(confer[5] page 6)). For any subset
QOof Q, let v*(Qf)= inf{v(A): Q. a A e .} denote the outer probabilityof QO
underv.
LEMMA4. In thepresentcontext,

overeach
XY-*X uniformly

(44)

Uk

Iimk, V*(Uk)= I

(45)

PROOF. We get (44) from(24), (31), and (43). For (45) put Ek = infm,kDm(1 <
k < oo),whereDm is definedby (39). Suppose that UkcA Ed/. Then thereexists
([4] page 81) a countablesubset1A of f suchthatA dependsonlyon X and theX,
withy [A; it followsthat
er

ny

rA;

k(y)> k {d(Xy,

X) < Ak(y)}C A

(theseton theleftneed notbelongto d/). Thus forj < k we have (confer(34) and
(27))
vj(A) = fsvj,s(An {X = s})P(ds)
>

_V__S(nY6rA;k(y)?k{d(XY,S) ? Ak}(y)P(ds)

= P(Ek).

By (35), v(A) > Zj<kWjVj(A) >? OkP(Ek); it followsthatv*(Uk)> O)kP(Ek).But (28)


impliesP(Ek) _ 1 -LmEkEm; (45) nowfollowsfrom(9) and (25).
We notethatthe Uk increasewithk. In viewof Lemmas3 and 4, to completethe
to establish
proofofTheorem2 itsuffices

ALMOST UNIFORMLY CONVERGENT RANDOM VARIABLES

291

LEMMA5. Let (Q, .4,v) be anyprobability


space and let (Uk)k>1 be an increasing
sequenceof subsetsof Q of outerprobabilitiesv*(Uk). Let X be the a-algebra
generatedbysl and theUk(l < k < oo). Thenv maybeprolongedto a probability
It
on(Q, .) suchthat
(46)

JU(Uk)

foreachk.
PROOF. Put
Bk*E sl

Bk

Uk-

Uk-

such that v(Bk *)

(1

V*(Uk)

? k < oo), put Boo= (supkUk)c, and choose

v*(Bk)(1

<

k ? oo). According to [4] page 43, %

coincideswiththe class of sets of the formEkAkBk, whereAkeC-;


formula
(47)

I(Yk

Ak

Bk)

Ek fAkfk

moreoverthe

dv

definesa probabilityju on (Q, 9), whose restriction


to (Q, ) is v, providedthat
each fk is a nonnegative,d-measurable random variable vanishingoff of
Bk*(l < k? oo) and thatEkfk = 1.
Let fk be the indicator function of Bk*

(47). Then
(48)

,u(Uk)

hi(Zj_kBj)

= YjkJffjdv

<kBj *(

<k ? oo) and define[t by

v(UjikBj*).

Since U j kBj* is an dW-measurable


setcontainingUk,we have

(49)

v(Uj <k Bj*) > v*(Uk).

On the otherhand, suppose UkcA E X, so thatBjc A forj < k. Then each Bj*,
and hencealso U .<kB*, is containedin A up to a v-equivalence.It followsthat
v(A) ? v(Ui<kBi*) andthat
(50)

V*(Uk) > V(Uj_kBj*)

Together(48), (49), and (50) imply(46). E]


REFERENCES
on non-separable
metricspaces
of probabilities
[II DUDLEY, R. M. (1966a).Weakconvergence
on Euclideanspaces.Ill.J.Math.10 109-126.
andempirical
measures
ofBairemeasures.
StudiaMath.27 251-268.
[2] DUDLEY,R. M. (1966b).Convergence
measuresand randomvariables.Ann.Math.
[3] DUDLEY, R. M. (1967).Distancesofprobability

Statist.39 1963-1572.
[4] NEVEU,J. (1965). MathematicalFoundationsof the Calculus of Probability.Holden-Day, Inc.,
San Francisco.
D. R. (1967). ProbabilityMeasures on Metric Spaces. Academic Press,
[51 PARATHASARATHY,
New York.
ofweaklyconvergent
ofalmostsurelyconvergent
constructions
[6] PYKE,R. (1968).Applications
Berlin.
processes.Proc. Internat.Symp.Prob. Inform.Theor.Springer-Verlag,
[7] SKOROKHOD,A. V. (1956). Limit theoremsforstochasticprocesses (translatedby SIAM). Th.
ProbabilityAppl. 1 261-290.
ofnon-Borel
on a metric
[8] WICHURA,MICHAELJ.(1968).On theweakconvergence
probabilities

ColumbiaUniv.
dissertation,
space.Unpublished

You might also like