Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

 

The Development of Rigor in Mathematical Probability (1900-1950)


Author(s): Joseph L. Doob
Source: The American Mathematical Monthly, Vol. 103, No. 7 (Aug. - Sep., 1996), pp. 586-595
Published by: Mathematical Association of America
Stable URL: http://www.jstor.org/stable/2974673
Accessed: 07-11-2015 11:14 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/ 
info/about/policies/terms.jsp 

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.

Mathematical Association of America  is collaborating with JSTOR to digitize, preserve and extend access to  The American
Mathematical Monthly.

http://www.jstor.org

This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
All use subject to JSTOR Terms and Conditions
 

THE EVOLUTIONOF...
Edited by:Abe Shenitzer
Mathematics, orkUniversity,orthYork,OntarioM3J1P3, Canada

The Developmentof Rigor


in MathematicalProbability(1900-1950)
Joseph L. Doob

1 Introduction.This paper is a brief informal outline of the history of the


introduction f rigour nto mathematical robabilityn the first half of this century.
Specificresultsare mentionedonly in so far as they are important n the historyof
the logical developmentof mathematicalprobability.
The developmentof science is not a simpleprogression romone advance o the
next. Judged by hindsight,the development s slow, proceeds in a zigzag course,
with many wrong turns and blind alleys, and frequently moves in directions
condemnedby leading scientists.In the 1930'sBanach spaces were sneered at as
absurdlyabstract, ater it was the turn of locally convex spaces, and now it is the
turn of nonstandardanalysis. Mathematiciansare no more eager than other
humans o embracenew ideas, and full acceptanceof mathematical robabilit was
not realized until the second half of the century.In particular,many statisticians
and probabilistsresented the mathematization f probabilityby measure theory,
and some still place mathematicalprobabilityoutside analysis. The following
quotations in translation)are relevant.

Planck: A new scientific ruthdoes not triumphby convincingts opponents nd


making hemsee the light,but rather ecause ts opponents ventuaGyie,
and a newgeneration rowsup with t.
Poincare: Formerly,when one inventeda new function, it was to furthersome
practicalpurpose; odayone invents hem in order o make incorrecthe
reasoning f ourfathers,and nothingmore will everbe accomplished y
these nventions.
Hermite: (in a letter to Stieltjes)I recoilwithdismayand horror t this amentable
plagueof functionswhichdo not hauederivatives.

Probabilitytheory began, and remained for a long time, an idealization and


analysisof certain real life phenomenaoutside mathematics,but gradually, n the
first half of this centuiy, mathematicalprobabilitybecame a normal part of
mathematics.The mathematizationof probabilityrequired new ideas, and in
particularrequireda new approach o the idea of acceptabilityof a function. In

Reprinted with kind permission of Birkhauser Verlag AG, from Deuelopment f Mathematics
 

view of the above quotations t is not surprising hat acceptanceof this mathemati-
zation wasdited
1900-1950, slowbyand facedPier,
Jean-Paul resistance.In fact even
Basel, 1994, ISBN now some probabilistsear that
0-8176-2821-5
mathematization as removed he intrinsiccharm rom their subject.And they are
rightin the sense that the charmof the old, vague probability-mathematics,ased
586 HE EVOLUTION OF . . . [Aug.-Sept.
on nonmathematical efinitions,has split into two quite differentcharms: hose of
real world probability nd of mathematicalprecision.But it must be stressed that
manyof the most essentialresultsof mathematical robability avebeen suggested
This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
by the nonmathematical subjectof
ontext
All use real world
to JSTOR Terms and Conditions which has never even
probability,
had a universallyacceptable definition. In fact the relation between real world
probability nd mathematicalprobabilityhas been simultaneously he bane of and
inspiration or the developmentof mathematicalprobability.
2 VVhats the real world (nonmathematical)problem? What is usually called
(real world) probabilityarises in many contexts. Besides the obvious contexts of
gamblinggames, of insurance,of statisticalphysics, here are such simplecontexts
as the following.Suppose an individual ides his bicycle to work.The riderwould
be surprised f, when the bicycleis parked,the valve on the front tire appeared n
the upper half of the tire circle 10 successive days, just as surprised as if 10
successive osses of a coin all gave heads. However, t is clear that (tire context) f
the ride is very short, or (coin context) if the coin starts close to the coin landing
place and the initial rotationalvelocity of the coin is low, the surprise would
decrease and the probability ontextwould become suspect. The moralis that the
specific context must be examinedclosely before any probabilisticstatement is
made. If philosophy s relevant,an arguablequestion, t must be augmentedby an
examinationof the physicalcontext.
3 The law of large numbers. In a repetitivescheme of independent rials,such as
coin tossing,what strikesone at once is what has been christened he law of large
numbers. n the simple context of coin tossing it states that in some sense the
numberof heads in n tosses dividedby n has limit 1/2 as the numberof tosses
increases.The key words here are in some sense. If the law of large numbers s a
mathematical heorem,that is, if there is a mathematicalmodel for coin tossing, n
which the law of large numbers s formulatedas a mathematical heorem,either
the theorem is true in one of the variousmathematical imit concepts or it is not.
On the other hand, if the law of large numbersis to be stated in a real world
nonmathematicalcontext, it is not at all clear that the limit concept can be
formulated n a reasonableway. The most obvious difficulty s that in the real
world only finitelymanyexperiments an be performed n finite time. Anyonewho
tries to explain o studentswhathappenswhen a coin is tossed mumbleswords ike
in the long run, tends, seems to clusternear, and so on, in a desperateattempt to
give form to a cloudy concept. Yet the fact is that anyone tossing a coin observes
that for a modestnumberof coin tosses the numberof heads in n tosses dividedby
n seems to be getting closer to 1/2 as n increases.The simplestsolution, adopted
by a prominent Bayesian statistician,is the vacuous one: never discuss what
happenswhen a coin is tossed. A more commonequallysatisfactory olution is to
leave fuzzy the question of whether the context under discussion is or is not
mathematics.Perhapsthe fact that the assertion s called a law is an example of
this fuzziness. The following statements have been made about this law (my
emphasis):

Laplace: (1814) This theorem, mpliedby commonsense, was difficult o prove by


 

Ville: (1939) One sees no reason or this proposition o be true; but as it is


analysis.
impossible o prove experimentallyhat it is false, one can at leastsafely
state t.
Bauer: (translated romthe context of dice to that of coins) It is an experimen-
1996] HE EVOLUTION OF . . . 587
tally establishedact that the quotient. . exhibitsa deviation rom 1/2
whichapproaches for largen.
This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
These statements illustrate the to
All use subject enduring charm
JSTOR Terms of discussions of real world
and Conditions
probability.Mathematicians,unfortunately,have felt forced to think about the
followingquestion,or at least to write aboutit.

4 What is probability? Here are some attemptsto answerthis questionand to


discussthe teachingof the subject.

Poincare: (1912) Onecan scarcely ivea satisfactoryefinition f probability.


Mazurkiewicz:(1915) The theoryof probabilitys not an independent lementof
mathematicalnstruction; everthelesst is verydesirablehata math-
ematicianknows ts generalprinciples. tsfundamental onceptsare
incompletelyetermined. heycontainmanyunsolveddifficulties.
v. Mises: (1919) In fact, one can scarcelycharac-terize hepresent tateother
thanthatprobabilitys not a mathematicaliscipline.He proceeded
to make it into a mathematicaldisciplineby basing mathematical
probability n a sequenceof observations <<Beobachtungen>>) ith
properties hatcannotbe satisfiedby a mathematically ell defined
sequence.In a lighter he is said to have defined probability
as a numberbetween0 and 1 aboutwhich nothingelse is known.)
E. Pearson: (1935) (Oral communication)Probabilitys so linkedwithstatistics
that,.although t ispossible o teachthe twoseparately,ucha project
wouldbejust a tourdeforce.
Uspensky: (1937) In a useful textbook,he gave the followingonce common
textbook definition. If, consistentwith conditionS, there are n
mutually xclusive,and equally ikelycases, and arefavorable o
the eventA, then the mathematical robability f A is definedas
m/n.

The foregoing should make obvious the advisabilityof separatingmathematical


probability heory from its real world applications.Note however that no one
doubtsthe real world applicability f mathematicalprobability.Gambling,genet-
ics, insuranceand statisticalphysicsare here to stay.
Only mathematical robabilitywill be discussedbelow, except for the following
remarkon coin tossing. Newtonian mechanicsprovides a partial mathematical
model for coin tossing. In coin tossing, a solid body falls under the influence of
gravity. ts motion s determined n Newton'smodelby his laws,and anydiscussion
of whatthe coin does cannotbe completeunlessthese lawsare applied.Onlythese
laws,ratherthan philosophical emarks, an explain he quantitativenfluenceand
importanceof the initialand final conditionsof the coin motion in order to justify
allusionsto equal likelihood of heads and tails. Of course these laws can at best
reducethe analysis o considerations f the initialand final conditionsof the toss,
but these conditions can show what the <<equalikelihoods>>epends on and
therebygive it a plausible nterpretation.
 

5 Mathematicalprobabilitybefore the era of precise defilnitions.There were


many mportantadvances n mathematicalprobability efore 1900, but the subject
was not yet mathematics.Although nonmathematicalprobabilisticcontexts sug-
gested problems n combinatorics, ifferenceequationsand differentialequations,
588
there was a minimumof attentionpaid HE EVOLUTION OF . . .
to the mathematical asis of the[Aug. Sept.
contexts a
maximumof attention to the pure mathematicsproblems they suggested. This
unequaltreatmentwas inevitable,because measuretheory,needed for mathemati-
This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
cal modelingof real world probabilistic
All use subject to JSTORontexts,had not yet been invented.
Terms and Conditions
It was alwaysclear that, howeverclassicalmathematicalprobabilitywas to be
developed, the concept of additivityof probability s applied to incompatiblereal
worldeventswas fundamental.Additive unctionsof sets were of coursefamiliar o
mathematiciansromconceptsof volume,mass and so on, long before 1900. It was
realizedthat contextsinvolvingaverages ed to probability. t was frequentlyclear
how to use the contextsto suggestproblems n analysis,but it was not clear how to
formulatean overall mathematical ontext, that is, how to define a mathematical
structure n which the variouscontextscould be placed.
A weaker condition than additivitywas less familiar but turned out to be
essential later. The standard oose languagewill be used here. If x1, x2, . . . are
numbersobtainedby chance,and if A is a set of numbers,consider he probability
that at least one of the membersof this sequence lies in A, or, in more colorful
language,considerthe probability hat an orbit of this motion throughpoints of a
line hits A. The usual calculation(ignoringhere all notions of rigour),defines a

inequality

(t(A) + (h(B)-+(A U B) 2 +(A n B), (5.1)

whereasadditivity f + would implyequality n (5.1). The point is that the left side
of (5.1) is the probability hat the sequence x hits both A and B, a probability t
least equal to, and in general greater than, +(A n B), the probability hat the
sequence hits A n B. The inequality(5.1), the strongsubadditivitynequality, s
satisfiedalso by the electrostaticcapacityof a body in R3, and this fact hints at the
close connection between potential theory and probability,developed in great
detail in the second half of the century with the help of Choquet'stheory of
mathematical apacity.

6 The development f measuretheory. Recall that a Borelfield(-(r al,kebra) f


subsetsof a space is a collectionof subsetswhich is closed under the operationsof
complementationand the formationof countable unions and intersections.The
class of Borelsets of a metric space is the smallest set (r algebracontaining he
open sets of the space. A measurablepace is a pair, (S, S), where S is a space and
s is a (r algebraof subsetsof S. The sets of S are the measurableets of the space.
In the following, f S is metric,the coupled ff algebramaking t into a measurable
space will alwaysbe the cr algebraof its Borel sets. In particular RN,RN) denotes
N dimensionalEuclideanspace coupledwith its Borel sets. The superscriptwill be
omitted when N = 1. A measurablefunction from a measurablespace (S1,S1)
into a measurable pace tS2, S2) iS a function from S1 into S2 with the property
that the inverseimage of a set in 52 is a set in S1.
Measure theory started with Lebesgue's thesis (1902), which extended the
definitionof volumein RN to the Borel sets. Radon (1913) made the furtherstep
 

to generalmeasuresof Borel sets of RN (finite on compactsets). These measures


are usually extended to slightly larger classes than the class of Borel sets, by
completionFinally Frechet (1915), 13 years after Lebesgue'sthesis, pointed out
that all that the usual definitionsand operationsof measure require s a Cr
1996] HE EVOLUTION OF . . . 589
algebra of subsets of an abstractspace on which a measure, that is, a positive
countablyadditiveset function,is defined. At each step of this progressionnot
necessarilypositive countablyadditive set functions-signed measures were in-
This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
corporated nto the theory.As noted
All use subject below,
to JSTOR the
Terms andRadon-Nikodymheorem
Conditions (1930),
whichgives conditionsnecessaryand sufficientthat a countablyadditivefunction
of sets can be expressed as an integralover the sets, turned out to be the final
essentialresultneeded to formulate he basicmathematical robabilitydefinitions.
Thus it was 28 years before Lebesgue'stheory was extended far enough to be
adequatefor the mathematicalbasis of probability.This extensionwas not devel-
oped in order to provide a basis for probability,however. Measure theory was
developedas a part of classicalanalysis,and applications n analysiswere immedi-
ate, for example to the (Lebesgue measure)almost eveiywherederivabilityof a
monotonefunction.
There has been criticismof the fact that mathematicalprobability s usually
prescribednot only to be additivebut even to be countablyadditive.The question
whether real world probability s countablyadditive, if the question is to be
meaningful,asks whether a mathematicalmodel of real world probabilisticphe-
nomena necessarily lways nvolvescountablyadditiveset functions.In fact there
maywell be real worldcontextsfor which the appropriatemathematicalmodel is
based on finitelybut not countablyadditiveset functions.But there havebeen very
few applicationsof such set functions n either mathematicalor nonmathematical
contexts,and such set functionswill not be discussed urtherhere.
7 Earlyapplicationsof explicit theory o probability.Some probabilistic
slang will be needed, enduringrelics of the historicalbackgroundof probability
theory.A probability pace is a triple(S, X,P), where (S, S) is a measurable pace
and P is a measureon s with P(S) = 1. A measurewith this normalizations a
probabilitymeasureA randomvariables a measurable unctionfroma probability
space (S, X, P), into a measurable pace (S', '). The space S', or, when one writes
carefully, S', '), is the statespace of the randomvariable.Mutualindependence
of randomvariables s defined in the classicalway. The distribution f a random
variablex is the measurePxon ' definedby setting

PX(X)=P{SES X(S)Xt}@

The joint distributionof finitely many randomvariables defined on the same


probability pace is obtainedby making x into a vector and specifying and S'
correspondingly. stochasticprocess s a familyof randomvariables x(t,), t E >}
fromsome probability pace (S, X,P), into a state space (S', '). The set o is the
index set of the process. Thus a stochastic process defines a function of two

from S into S' is the tth randomvariableof the process; he functionx(, s) from
o into S' is the sth samplefunction,or samplepath, or samplesequence if o is a
sequence.
Borel (1909) pointed out that in the dyadicrepresentationx = xtx2 . . . of a
numberx between0 and 1, in whicheach digit Xj is either 0 or 1, these digitsare
functions of x, and if the interval[0,1] is providedwith Lebesgue measure, a
 

variables which have exactly the distributionsused in calculatingcoin tossing


probability measureon
probabilities. That is, 2-nthisis interval, hese functionsmiraculously
the probability ssignedto the event that, ecome
in arandom
tossing
experiment, he first n tosses yield a specifiedsequenceof headsand tails, and 2-n
is also the total length (= Lebesguemeasure)
590 of the finite set of intervalswhose
HE EVOLUTION OF . . . [Aug.-Sept.
points x have dyadicrepresentationswith a specifiedsequenceof 0's and l's in the
first n places. Thus a mathematical ersionof the law of large numbers n the coin
tossingcontext s thedownloaded
This content existence from some sense of
n 213.233.188.210 a limit of the sequenceof function
on Sat, 07 Nov 2015 11:14:07 UTC
averages{(x1 + *@+xn)/n, All use n2
subject1}. Classicalelementaryprobabilitycalculations
to JSTOR Terms and Conditions
implythat this sequence of averagesconverges n measureto 1/2, but a stronger
mathematicalversion of the law of large numbers was the fact deduced by
Borel-in an unmendablyaultyproof-that this sequenceof averagesconverges
to 1/2 for (Lebesguemeasure)almost everyvalue of x. A correctproofwas given
a year later by Faber, and much simpler proofs have been given since. [Frechet
remarked actfully:<<Borel'sroof is excessively hort.It omits several ntermediate
argumentsand assumes certain results without proof.>>] his theorem was an
important tep, an exampleof a new kind of convergence heorem in probability.
Observe hat (fortunately)pure mathematicians eed not interpret his theorem n
the realworldof real people tossingreal coins. Someof the quotationsgiven above
indicatethat they not only need not but should not.
Daniell (1918)used a deep approach o measuretheory in which integralsare
definedbefore measures o get a (ratherclumsy)approach o infinite sequencesof
randomvariablesby way of measures n infinitedimensionalEuclideanspace.
The Brownianmotion stochastic process in R3 is the mathematicalmodel of
Brownianmotion, the motionof a microscopicparticle n a fluid as the particleis
the molecules of the fluid. The process is normalizedby supposing t starts
at the originof a cartesiancoordinatesystemin R3, and a (normalized)Brownian
motionprocess n R is the processof a coordinate unctionof a normalizedprocess
in R3, vanishing.initially.A (normalized)Brownianmotion process in RN is a
process defined by N mutually ndependentBrownianmotion processes in R. It
was well knownwhat the joint distributions f the randomvariablesof a Brownian
motion process should be, and it had been taken for granted that in a proper
mathematicalmodel the class of continuous paths would have probability1. By
1900, Bachelier had even derivedvarious importantdistributionsrelated to the
Brownianmotionprocessin R, such as that of the maximum hangeduringa time
interval,by findingcorresponding istributionsor a certaindiscreterandomwalk
and then going to the limit as the walk steps tended to 0. More precisely,what
Bachelierderivedwere distributions alid for a Brownianmotionprocess if in fact
there was such a thing as a Brownianmotion process,and if it was approximable
by his randomwalks. Observethat there was no question about the existenceof
Brownianmotion;Brownianmotion is observableunder a microscope.But there
was as yet no proof of the existence of a stochastic process, a mathematical
construct,with the desired properties. Wiener (1923) constructedthe desired
Brownianmotion process, now sometimescalled the Wiener process, by applying
the Daniell approachto measure theory to obtain a measure with the desired
propertieson a space S of continuousfunctions: f x(t,) is the randomvariable
defined by the value at time t of a function in S, the stochasticprocess of these
randomvariables s a stochasticprocesswith samplefunctionsthe membersof S,
and with the joint randomvariabledistributionshose prescribed or the Brownian
motion process.
Bachelier's esultsremainedunnoticedfor years, and in fact were rediscovered
 

little immediateinfluence because it was published in a journalwhich was not


widelydistributed.
severaltimes. t was an aspect
Wiener'swork, like hisof fundamentalwork
his geniusthat he carriedout
in potential theory,had
his Brownian
motion researchthen and later withoutknowledgeof the slang and some of the
useful elementarymathematical echniquesof probabilityheory.
1996] HE EVOLUTION OF . . . 591
Steinhaus (1930) demonstratedthat classical argumentsto derive standard
probability heorems could be placed in a rigorouscontext by taking Lebesgue
measureon a linearintervalof length 1 as the basicprobabilitymeasure, nterpret-
This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
ing random variables All as use
Lebesgue measurablefunctions
subject to JSTOR Terms and Conditions on this interval, and
expectationsof randomvariablesas their integrals.No new proofswere required;
all that was requiredwas a propertranslationof the classicalterminology nto his
context.If this were all mathematization f probabilityby measuretheoryhad to
offer, the scorn of rigorousmathematicsexpressed by some nonmathematicians
would be justified.

8 Kolmogorov's 933 monograph.Kolmogorov 1933) constructed he following


mathematicalbasisfor probability heory.

(a) The contextof mathematical robabilitys probability pace(S, X, P). The


sets in S are the mathematical ounterparts f real worldevents;the points
of S are counterpartsof elementaryevents, that is of individual possible)
real worldobservations.
(b) Random variableson (S,S,P), are the counterpartsof functions of real
world observations.Suppose {x(t,),t e>} is a stochastic process on a
probability pace (S, X, P), with state space S'. A set of n of the process
randomvariableshas a probabilitydistributionon stn. Such finite dimen-
sional distributions re mutuallycompatible n the sense that if 1 < m < n,
the joint distributionof x(t1,),. . ., x(tm,) on S'm is the m-dimensional
distribution induced by the n-dimensional distribution of x(t1,),....
x(tn ) onS
(c) Conversely,Kolmogorovprovedthat given an arbitraryndex set X, and a
suitablyrestrictedmeasurablespace (S , ) (for example,the measurable
space can be a completeseparablemetricspacetogetherwith the (r algebra
of its Borel sets) and a mutuallycompatibleset of distributions n stn, for
integers n 21, indexed by the finite subsets of X, there is a probability
space and a stochasticprocess{x(t,), t E >} defined on it, with state space
S', with the assignedjoint random variable distributions.To prove this
result he constructeda probabilitymeasureon a (r algebraof subsetsof the
productspace S'>, the space of all functionsfrom > into S', and obtained
the requiredrandomvariablesas the coordinate unctionsof S'>.
(e) The expectationof a numericallyvalued integrablerandomvariable is its
integralwith respectto the givenprobabilitymeasure.
(f) The classicaldefinitionof the conditionalprobability f an event (measura-
ble set) A, given an event B of strictlypositiveprobability, s P(A r) B)/
P(B). In this way, for fixed B, new probabilitiesare obtained,and expecta-
tions of randomvariables or given B are computed n terms of these new
conditional robabilities.More generally,given an arbitrarycollection of
random variables, conditional probabilitiesand expectations relative to
givenvalues of those randomvariablesare needed, functionsof the values
assigned to the conditioningrandomvariables.If (S, , P) is a probability
space, and if a collection of randomvariables s given,let IFbe the smallest
 

measurable.This S algebra s the (r algebraeneratedyconditionsmposed


sub (r algebra
on thegiven of S relative
andom ariables. to which all thenterpretation
reasonable given randornvariables are
of a measurable
real valued function of the given collection of random variables is a
592 measurable unction rom (S, IF)nto R. The Kolmogorov onditional
HE EVOLUTION OF . . . xpecta-
[Aug.-Sept.
tionof a real valued integrablerandomvariablex on (S, X, P), relativeto a
(r algebra(Cof measurablesets, is a randomvariablewhich is measurable
relativeto (Cand
This content has the
downloaded fromsame integralas
213.233.188.210 x on
on Sat, everyset
07 Nov in (C.UTC
2015 11:14:07 The existence
of such a randomvariable,and its uniquenessup to P-nullsets, is assuredby
All use subject to JSTOR Terms and Conditions
the Radon-Nikodymheorem.The conditionalexpectationof x relativeto a
collectionof randomvariables s defined as the conditionalexpectationof x
relativeto the (r algebrageneratedby conditionson the randomvariables.
A conditionalprobabilityof a measurableset A is defined as the condi-
tional expectationof the randomvariablewhich is 1 on A and Oelsewhere.

Kolmogorov's1933 exposition paints a discouragingpicture of mathematical


progress In the first pages of his monographhe states explicitly hat real valued
randomvariables are measurable unctions and expectationsare their integrals.
Even as late as 1933,however,he must have thought hat mathematicianswere not
familiarwith measuretheory.In fact in the body of his monograph,when he comes
to the definitionof a real valued randomvariable,he does not simplyrefer back to
the first pages of the monographand say that a randomvariable s a measurable
function.Instead he actuallydefines measurability f a real valued function, and
similarlywhen he defines the expected value of a randomvariable he does not
simplystate that it is the integralof the randomvariablewith respect to the given
probabilitymeasure,but he actuallydefines the integral.Later in the monograph,
when he needs Lebesgue's heorem allowingtaking limits of convergent unction
sequences under the sign of integration,he does not simplyrefer to Lebesguebut
gives a detailed proof of what he needs. As confirmationof Kolmogorov's aution
in invokingmeasuretheory,the authorrecallshis student experience n 1932when
there were professorial disapprovingremarks on the extreme generality of a
seminar ecture givenby Saks on what is now called the Vitali-Hahn-Saksheorem,
a theorem which has since become an importanttool in probability heory. [He
also recalls that he did not understand he point of Kolmogorov'smeasure on a
functionspace until long after he had read the monograph.]
It was some time before Kolmogorov's asis was accepted by probabilists.The
idea that a (mathematical) andomvariable s simplya function,with no romantic
connotation, eemed ratherhumiliating o some probabilists.A prominent tatisti-
cian in 1935 wonderedwhether two orthogonalreal valued randomvariableswith
zero means (integrals)are necessarily ndependent, as they are under the added
hypothesis that they have a bivariateGaussian distribution.He was rather sur-
prisedby the exampleof the sine and cosine functionson the interval O, qr],with
probabilitymeasure defined as Lebesgue measure divided by 2qr. These two
functions,orthogonaland with zero means but not independent,are not the kind
of random variablesprobabilistswere used to. Some analysts may be gratified,
some humiliated, o learn that in discussingFourierseries they can be accused of
discussingprobabilitiesand expectations.

9 Expansionbackwardsof the Kolmogorov asis. Kolmogorov's asis for mathe-


maticalprobability an be expanded,and shouldbe expanded n the view of some
probabilists,who want to start with some not necessarilynumericalmathematical
 

proceedpostulationally o numericalevaluationsof this confidence,and finallyto


version of Sthe
additivity. uchconfidence
an analysismay of observers that certain
be enlightened events will
n discussing occur, and to
he appropriateness
of mathematicalprobabilityas a model for real world phenomena, but any
approach o the subjectwhichends with a justificationof the classicalcalculations
HE EVOLUTION OF . . . 593
1996]
and is mathematically sable,will end with Kolmogorov's asis, howeverphrased,
because all the measure basis to probability to give a formal precise
mathematical ramework or the classical calculationsand
This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 their
11:14:07present
UTC refine-
ments.This framework All
haduse made
subject to
itJSTOR Terms
possible toand Conditions
applymathematical probabilityn
many other mathematicalfields, for example to potential theory and partial
differentialequations.Althoughsuch applicationswere made in the past before
the acceptance of measure theory as the basis of probability, he probabilistic
context served only to suggest mathematicsand was not an integralpart of the
mathematics.The meaningof solutionsas probabilitiesand expectations ould not
be formulatedand exploited.
Uncountable ndex sets. If the index-set > of a stochasticprocess {x(t,-),
t E >} is an intervalof the line, and if the state space of the randomvariables s R,
the class of continuoussample functionsmay not be measurable.This difficulty
arises in the processesderivedby the Kolmogorov onstructionof a measureon a
function space, for example, whatever the choice of joint distributionsof the
process randomvariables.To understand he difficulty,observethat if the index
set > of a stochasticprocesswith state space 1Rs an interval,and if Z is a subset

not be measurable f Z is uncountable. f boundednessand continuityof sample


functionsare to be discussed,some modificationof the probability elationsof the
randomvariablesof a stochasticprocessshould be devised to make such suprema
measurablefunctions.A clumsy approachwas proposed by Doob (1937) but a
more usableone was not deviseduntil after 1950.
11 Reluctance o acceptmeasuretheoly by probabilists.Therewas considerable
resistanceto the acceptanceand exploitationof measure theory by probabilists,
both in Kolmogorov's ay and later. The followingquotation s an exampleof the
reluctanceof some mathematicians o separatethe mathematics rom the context
that inspired t.
Kac (1959) Howmuch iwss vermeasure heozy s necessazyorprobabilityheoty s
a matterof taste. Personally preferas little iwssas possiblebecauseI
firmlyelieve hatprobabilityheozy s morecloselyrelated o analysis,
physicsandstatistics han to measure heotyas such.
12 New relations betweenfunctions made possible by the mathematizationof
probability.Probability heory suggested new relations between functions. For
exampleconsiderthe sequence x1,x2, . . . of real valued integrablerandomvari-
ables on a probability pace (S, Si,P) and supposethat the conditionalexpectation
of xn given x1, . ., xn_1 vanishes(P) almost everywhere, or n > 1, that is, the
integralof xn over any set determinedby conditions on the precedingrandom
variablesvanishes.If these randomvariablesare squareintegrable, his condition
is equivalent o the condition,muchstronger hanmutual hat xn is
orthogonal to every square integrable function of x1,. .., xn. Bernstein (1927)
seems to have been the first to treat such sequencessystematically.This condition
on a sequence of functions means that in a reasonablesense the sequence of
partialsums of the given sequence is the counterpartof a fair game. In fact, the
fair game. In fact, the
partialsumsY1,Y2, . . are characterized y the property hatthe expectationof Yn
 

594 HE EVOLUTIONOF . . . [Aug.-Sept.

This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
relative to Y1 XYn-ls subject to
All useequal Yn-lTerms
to JSTOR almost everywhereon the probability
and Conditions
space. Processeswith this property,called martingales, irstused explicitlyby Ville
(1939),have had manyapplications, or exampleto partialdifferentialequations,
to derivation,and to potential theory. Another importantclass of sequences of
random variables is the class of sequences with the Markov property.These
sequencesare characterized y the fact thatwhen n 2 1 the conditionalprobabili-
ties for xn relative to xl, . . ., xn_1 are equal almost everywhere o those for xn
relativeto xn_1. Roughlyspeaking,the influence of the present, given the past,
depends only on the immediatepast. The Markovproperty, ntroduced n a very
special case by Markov n 1906, (namedin his honor by others) has provedvery
fruitful,for example,leading in the second half of the centuryto a probabilistic
potentialtheory,generalizingand includingclassicalpotential theory.
13 What is the place of probability heoly in measuretheoly, and moregenerally
in analysis? It is considered by some mathematicians hat if one deals with
analyticpropertiesof probabilitiesand expectationsthen the subject is part of
analysis,but thatif one dealswithsamplesequencesand samplefunctions hen the
subject s probability ut not analysis.These authorsare in the interesting ituation

stochasticprocesses they call it analysisif the family of functions x(t,) as t


variesis studied,but call it probabilityand definitelynot analysis f the familyof
functions x(, s) as s varies is studied. More precisely,they regard discussions
of distributions nd associatedquestionsas analysis,but not discussions n termsof
samplefunctions.This point of view is expressed n the followingquotation.
ntegraln 1944with tochasticrocessess integrands,to
Protter By developinghis
ech-
iffusions ithpurely robabilistic
wasable o studymultidimensional
niques, n improvement methodsf Feller.
ver heanalytic
The following remark on the convergence of a sum of orthogonal functions
illustratesthe difficulty n separating mathematical)probability rom the rest of
analysis.The measure space is a probability pace, but with trivialchanges the
discussion s valid for any finite measurespace.
If x is an orthogonalsequence of functions,on a probabilitymeasurespace,
and if xn2has integral(rn2,hen (Riesz-Fischer)Exn converges n the mean if
E 2 < + (13.1)

The orthogonal series converges almost everywhereif either (Mensov-Rade-


macher) 13.1)is strengthened o
E (rn2 log2 n < + oo, ( 13.2)
or (Levy, 1937) the condition (13.1) is kept but the orthogonalitycondition is
o the condition n Section 12.
The readershouldjudgewhichof these resultsis measuretheoreticand which
probability
is probabilistic,whether there is any point in evictingmathematicalprobability
from analysis,and if so whethermeasuretheoryshould also be evicted.

101 WestWindsor oad,Apt.1104


Urbana,llinois 1801-6663
doob@symcom.ath. iuc.du

THE EVOLUTION OF . . . s9s


1996]

This content downloaded from 213.233.188.210 on Sat, 07 Nov 2015 11:14:07 UTC
All use subject to JSTOR Terms and Conditions

You might also like