Hsu's Work in Probability Author(s): K. L. Chung Source: The Annals of Statistics, Vol. 7, No. 3 (May, 1979), pp.

479-483 Published by: Institute of Mathematical Statistics Stable URL: http://www.jstor.org/stable/2958733 . Accessed: 14/08/2011 19:24
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.

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

http://www.jstor.org

The Annals of Statistics 1979, Vol. 7, No. 3, 479-483

HSU'S WORKIN PROBABILITY BYK. L. CHUNG StanfordUniversity Hsu retumedfrom Englandto Kunmingin the middle of the Sino-Japanese war, apparentlyin 1940. He gave a long course beginningwith the theory of measure and integration,throughprobabilitytheory,and leading to mathematical statistics. For the first part he based his lectureson Caratheodory's Vorlesungen uberReelle Funktionen, for the second on Cramer's RandomVariables and Probability Distributions. He was a polished and vivacious lecturerwith complete notes written out beforehandin a notebook he carried.He enjoyedmakinga fine point in class. For instance,when he was doing the inversionformulafor characteristic functions,he took delight in the fact that one could integrate over a single point by the Lebesgue-Stieltjes integral. He had a tatteredcopy of Cramer'sbook with many marginalnotes and used to say that it was written in a more difficult way than necessary but contained all that was essential for probability. He was a true virtuoso in the method of characteristic functions. His papers[17], [23], [25], [26], [31]and [35] all showedhis fascinationfor, as well as masteryof, this precioustool. Mathematicalliteraturewas hard to come by in Kunmingduringthose years, and some of the old volumes shipped from Peking (then called Peiping)were kept in caves to preserve them from air raids. (We did not actually live in caves, but frequentlyhad to run to them for hours at a stretchunder raids or alerts.)Among the books so kept was, for example, Kolmogorov'sGrundbegriffe der Wahrscheinlichkeitsrechnung. At my requestHsu had this book extractedfrom the caves, but I rememberhis saying of it: "Thisis anotherkind of mathematics". He was not, of course, a probabilistper se at a time when such an appellationhardly existed anywhere in the world; by education and inclination he was more fond of than generalities.Howeverhe never tried to dissuadeothers from problem-solving following their own bent and getting interestedin those other kinds of things. Indeed, if he could be drawninto a new topic, he would work at it in earnestand quickly produce somethingof value. For instance,when Borel'swork on what he called "denumerableprobabilities"attracted my attention, Hsu took an interest also, which led to the paper [21]. This was an early approach to what became known as "recurrent events"underFeller'sdevelopment.Papers[23] and [30] were probablywrittenunder similarcircumstances. But, of course, Hsu's major contributionto probability and its applications functions.Let me characteristic flowed from his consummateskill in manipulating describeand commenton these papersin their chronologicalorder.
ReceivedJune 1978. AMS 1970subjectclassifications. Primary OIA70;secondary60E05. andphrases. Obituary, Key words probability theory.

479

480

K. L. CHUNG

Paper [17] is, without doubt, one of Hsu's most importantworks. He wrote it shortly after A. C. Berry obtained the correct order of magnitudeof the error estimatein the classicalcentrallimit theorem.Namely, if {(, n > 1} are independent and identically distributedrandom variableshaving distributionP(x) with mean 0 and variance 1, and finite third absolute moment /3, then for all x the distribution
F(x)
= Pr{
2Er= r

<

differs from the standardnormaldistribution by less than Af33/n 2, whereA is an absolute constant. (This resultwas also provedby Esseen, but Esseen'spaperwas publishedlater and not known to Hsu at the time.) Hsu extendedBerry'smethod to give a simplerproof of Cramer'stheoremon the asymptoticexpansionof F(x) when P(x) is nonsingular.But much deeper and fartherreachingare the correspondingresultswhen the sample mean
=

n zr=

ltr

is replacedby the samplevariance

In Hsu's own words: "so much for the known results for the approximate distribution of {. By a purelyformaloperationalmethodCornishand Fisherobtain terms of successive approximationto the distributionof any random variableX with the help of semiinvariants. It is hardlynecessaryto emphasizethe importance of turning Cornish and Fisher's formal result (asymptotic expansion without appraisalof the remainder) into a mathematicaltheoremof asymptoticexpansion which gives the orderof magnitudeof the remainder. In this paperwe achieve this for the simplestfunction next to C,viz. q."The basic idea is as follows. Let
X-1
En
r2

n=

n2

(a4-

1)2

n2

wherea4 is the fourthmomentof P(x). These are relatedto q by

(
4

n
- l

(11)

= X-

y2

((a4

n)

and the crux of the methodis to approximate the distribution of the randomvector (X, Y) through its characteristicfunction. Even a cursory perusal of the proof would make it obvious that this is a task of a higherdegreeof complexityand not just a routine analogue of the results of Berry and Cramer,since X and Y are highly correlated.To achieve the goal Hsu literallyadded a new dimensionto the method of approximationused before him. He was still teaching the course

HSU'S WORK IN PROBABILITY

481

mentionedabove when he obtainedthe result,and was obviouslyelated. I recallhis sayingthat he made severalpreviousattemptsat the problemand the breakthrough came when he realizedthat he must attack the bivariateapproximation ratherthan evading the difficulties by shortcuts. Hsu's method is applicableto many other functionscommonlyused in statisticssuch as samplemomentsof higherorderand Student'sstatistic.Some of these were carriedout by his students.AlthoughHsu's paper was cited by Bikjalisin his work on normal approximations, a recent book on the subjectfailed to mention it. In the course of writingthis article I asked the authorsabout this glaringomission and receivedthe regretfulreply that they were not awareof Hsu's paper!This is a seriousomissionand suggeststhat the analytic power generatedby Hsu's method has yet to be fully explored. In paper [23] the main result is that for a sequence of independent and identically distributedrandom variables with zero mean and finite variance, we
have for every E > 0:

nP {I+XI+

+Xnl > E}<

X.

This is an interestingstrengthening of the classical stronglaw of large numbersin the directionof the Borel-Cantelli lemma.The idea of such a resultis probablydue to Robbins,but the method of proof is vintage Hsu. The problemis reduced,after a convoluted Fourier inversion,to the estimationof several pieces of an integral involving the characteristicfunction-a remarkabletour de force. Erdos (1949) later sharpenedthe result and showed that the conditions on the moments are necessaryas well as sufficient. (According to Robbins, their original manuscript also contained a proof of the necessity of the conditionswhich was not published as being too complicated.)Erdos used a combinatorialtype of argument.Baum and Katz extendedthe result to momentsof fractionaland higherorder. Paper [35] forms Appendix III in the translationof Gnedenko-Kolmogorov (1968)and its storymay be worthtellinghere. Hsu had alwaysbeen impressedwith the necessaryand sufficient conditions for the general form of the central limit theoremdue to Feller,if only for the adroituse of characteristic functionsthere.In a letter dated May 12, 1947,from Chapel Hill, he told me that he had just solved the same problem for all symmetricstable laws. He wrote: "I believe that in a domain of the weak laws this is the best theoremyet. The methodin this work also points to solutionof the most generalproblemof weak law, namely[whenthe limit is] an infinitelydivisiblelaw. But furtherdifficultyis great.... The problemin my hand is so natural that I am in constant worry of clashing with people. So, wheneveryou communicatewith probabilitypeople, please spread the news." A latter dated May 26, 1947,announcedthe final result to me. It gives the necessary and sufficient conditions under which the row sums of a triangulararray of infinitesimalrandomvariables,independentin each row, convergein distribution to a given infinitely divisible distribution.His conditions differ somewhat from those in Gnedenko(1944) in that the "two tails"are combined.The proof is direct

482

K. L. CHUNG

and makes use of the kernel (sin t/t)3. The complete manuscriptcontainingthis result and its self-containedproof is paper [35]. It was given to me shortly before Hsu's departurefor China in the summerof 1947. (I have erroneouslyplaced the year to be 1946 in the preface to the translation cited above.) Despite his premonition,Hsu did not learn of Gnedenko's 1944 paper until later. In the only letter he wrote me after his return to China, dated March 19, 1950, he acknowledged Gnedenko'spriorityand urgedme in most emphatictermsto keep the only copy of his paper for him. He had clearly staked his own strengthon solving this challengingproblem,the culminationof a seriesof papersby such names as Levy, Khintchine,Kolmogorov,Marcinkiewicz (whom Hsu met), Feller and Gnedenko. He toiled hard and triumphedquickly. Does it matter that another striver has climbed the same peak before?Hsu's method is direct and "startsfrom nothing". The last threewords give a free renderingof a more elusive Chinesesaying which Hsu used on another occasion, meaning minimalreliance on precedents.He had always considered this nonreliance as a virtue of mathematicalwork, and it is apparentin all his papers where use of previousresultsfor intermediate steps are shunned so far as possible. The new edition of the translation of GnedenkoKolmogorov(1968) appearedin the middle of the culturalrevolutionin China; I do not know if Hsu ever saw this paperin print. Papers [25] and [26] were reviewed in some detail by me in the Mathematical Reviews,Vol. 17 (1956), page 274. Since this is easily available,I need not repeat the contentshere. Sufficeit to say that they are again productsof his expertisewith characteristic functions. Paper [31] falls in the same categorybut apparentlywas never reviewed.In it it is proved that every distribution in the class L is absolutely continuous.So far as I know this was a new result.It was rediscovered by Fisz and
Varadarajan in 1963 (Z. Wahrscheinlichkeitstheorie und Verw. Gebiete, Vol. 1).

Paper[30] standsout in Hsu's work in probabilityas it has nothingto do with a characteristicfunction. Like [31], it is published in Chinese with an English abstract, appeared in the same issue of the journal, and apparentlywas never reviewed.It deals with the differentiability propertiesof the transitionprobability function of a homogeneousMarkov process whose state space is Euclideanand whose sample paths are purely jumping. Such processes were studied by D. G. Kendall in 1955, who analysed the differentiationof the transitionprobability functionP(t, x, E) at t = 0. In the case of Markovchains,wherethe state space is discrete,the term-by-term differentiability of the Chapman-Kolmogorov equations:
t > 0, was proved by Austin analytically,and by myself probabilistically (for references to D. G. Kendall and Austin see Chung (1967).) Hsu extendedthese equationsto the Eucidean case, "usingmore elementarymethods and obtainingmore precise results".For the first equationabove he actuallyderivedan integralrepresentation of P(t, x, E) in terms of the post-exit process,thus generalizing my approachbut P'(s + t) = P'(s)P(t), P'(s + t) = P(s)P'(t), s > 0,

HSU'S WORK IN PROBABILITY

483

by purelyanalyticmeans.The intimaterelationbetweenthe analyticand stochastic structures is implicitin his work,and can easily be made explicit. I did not learn of Hsu's death until my arrivalin Pekingin June 1972.Hsiao-fu Tuan told me that Hsu spent his last years re-doinga lot of classicalcombinatorial analysis in his own way and left copious notes. So far the contents of these notes remainunknown.
REFERENCES
CHUNG, K. L. (1967). Markov Chains With Stationary Transition Probabilities, 2nd ed. Springer-Verlag.

ERDos, P. (1949).On a theoremof Hsu and Robbins.Ann.Math.Statist.20 286-291. Mat. Nauk. B. V. (1944).Limittheorems randomvariables.Uspehi GNEDENKO, for sumsof independent Translation No. 45, Amer.Math.Soc., New York, 1951. 10 115-165. Englishtranslation: A. N. (1968).Limit Distribution of Sums of IndependentRandom GNEDENKO, B. V. and KOLMOGOROV, by K. L. Chung,with Appendicesby J. L. Doob and P. L. Variables, rev. ed. (translated Hsu). Addison-Wesley.
DEPARTMENT OF MATHEMATICS STANFORDUNIVERSrrY

94305 STANFORD, CALIFORNiA