a conversation with eugenio regazzini - arxiv · eugenio regazzini was born on august 12, 1946 in...

arX

iv1

205

4807

v1 [

stat

OT

] 2

2 M

ay 2

012

Statistical Science

2011 Vol 26 No 4 647ndash672DOI 10121411-STS362ccopy Institute of Mathematical Statistics 2011

A Conversation with Eugenio RegazziniAntonio Lijoi1 and Igor Prunster1

Abstract Eugenio Regazzini was born on August 12 1946 in Cremona(Italy) and took his degree in 1969 at the University ldquoL Bocconirdquo ofMilano He has held positions at the universities of Torino Bologna andMilano and at the University ldquoL Bocconirdquo as assistant professor and lec-turer from 1974 to 1980 and then professor since 1980 He is currentlyprofessor in probability and mathematical statistics at the University ofPavia In the periods 1989ndash2001 and 2006ndash2009 he was head of the Insti-tute for Applications of Mathematics and Computer Science of the ItalianNational Research Council (CNR) in Milano and head of the Depart-ment of Mathematics at the University of Pavia respectively For twelveyears between 1989 and 2006 he served as a member of the Scientific Boardof the Italian Mathematical Union (UMI) In 2007 he was elected Fellowof the IMS and in 2001 Fellow of the ldquoIstituto LombardomdashAccademia diScienze e Lettererdquo His research activity in probability and statistics hascovered a wide spectrum of topics including finitely additive probabili-ties foundations of the Bayesian paradigm exchangeability and partialexchangeability distribution of functionals of random probability mea-sures stochastic integration history of probability and statistics Overallhe has been one of the most authoritative developers of de Finettirsquos legacyIn the last five years he has extended his scientific interests to probabilisticmethods in mathematical physics in particular he has studied the asymp-totic behavior of the solutions of equations which are of interest for thekinetic theory of gases The present interview was taken in occasion of his65th birthday

Key words and phrases Bayesian inference Dirichlet process exchange-ability de Finetti finitely additive probabilities History of Statistics andProbability in Italy subjective probability

Antonio Lijoi is Associate Professor of Statistics at theDepartment of Economics and Quantitative MethodsUniversity of Pavia via San Felice 5 27100 PaviaItaly e-mail lijoiunipvit Igor Prunster is AssociateProfessor of Statistics at the Department of AppliedMathematics and Statistics ldquoDiego De CastrordquoUniversity of Torino cso Unione Sovietica 218bis10134 Torino Italy e-mail igoreconunitoit

1Also affiliated to Collegio Carlo Alberto Moncalieri Italy

This is an electronic reprint of the original articlepublished by the Institute of Mathematical Statistics inStatistical Science 2011 Vol 26 No 4 647ndash672 Thisreprint differs from the original in pagination andtypographic detail

1 PROBABILITY AND STATISTICS AT

BOCCONI UNIVERSITY

Antonio You received your degree in economicsfrom ldquoL Bocconirdquo University in Milano Why didyou decide to study economicsEugenio I enrolled in an economics degree essen-

tially because it was the only choice I had Hav-ing attended a Technical High School for Accoun-tants at the time the law did not allow studentsfrom this kind of secondary school to study mathe-matics at university which would have been my fa-vorite option You needed to attend college prepara-tory schools to enroll in subjects like mathematicsMy family could not afford my university studiesand I was expected to get a job right after com-

1

2 A LIJOI AND I PRUNSTER

pleting high school The choice I made at the ageof 14 was coherent with this scenario By the timeI got my diploma from high school the situationhad improved a little bit and I could afford goingto university This was also due to a financial aidprogram adopted by the farseeing Italian centermdashleft governments of the time for students belongingto economically disadvantaged families The moneyfrom this program (it was around 200 euros peryear in 1965) and the earnings deriving from privatelessons I used to teach to other students allowed meto obtain a degree at Bocconi UniversityIgor Why did you choose Bocconi University and

not another universityE In the 1960s Bocconi was considered a presti-

gious university a degree from it represented the keyfor obtaining a good and rewarding job on the mar-ket These elements exerted an influence on me andmy family However there was also a practical rea-son Milano is just one hour by train from my home-town Cremona and I could go back home everyweekend Yet another reason is the influence of myfriend Lorenzo Peccati a mathematical economistfrom Cremona who was a student at Bocconi whileI was still at high school He was well aware of mybent for mathematics and suggested a few advancedtextbooks where I started reading the mathemati-cal tools used in economic modeling In particularI was excited at reading the Italian translation of themonograph Allen (1956) on mathematical analysisfor economists and this convinced me that Bocconiwould still have allowed me to study MathsI This is a funny coincidence since I was con-

vinced by Lorenzo Peccatirsquos son Giovanni by nowa well-known probabilist to enroll at Bocconi andfor precisely the same reasons Where did your pas-sion for mathematics come fromE I was very lucky at high school because I had

a brilliant maths teacher Sidomo Vailati He hada variety of scientific and cultural interests and alsodid consulting for a few private companies thanksto his unusual at least in that period knowledgeof statistics probability and operations research Hecertainly was a self-taught man in the area of Stochas-tics At that time probability and statistics unlikeanalysis geometry and algebra were not perceivedas relevant topics within mathematics degrees theywere only present in a few optional courses To myknowledge the only exception was the Universityof Roma due to the presence of Bruno de FinettiIn fact this unfortunate situation lasted until the

1970s when the first full professors in probabilityapart from de Finetti were recruited after nationalcompetitions Turning back to Vailati it is worthrecalling that among some courses for high schoolteachers organized by the Italian Ministry of Educa-tion he also took a course in probability which wasdelivered by de Finetti As a consequence at theage of 16 I was introduced to the realm of subjec-tivism and learned the first elements of probabilityand its applications These first years of exposure tode Finettirsquos approach have stimulated an intellectualand scientific interest that has certainly influencedmy later researchA How was the environment at Bocconi Univer-

sity in the years you have been thereE Bocconi had very few professors among its own

faculty and heavily relied upon adjunct faculty hold-ing positions in other universities These few pro-fessors were all influential personalities of the timeplaying significant roles in the Italian social polit-ical and economic life of the 1960s For instanceGiovanni Demaria was a Paretian economist whoacted as economic consultant for the constituent as-sembly that created the Constitution that lies atthe foundation of the modern Italian Republic afterWorld War II There was also a special feeling be-tween Bocconi and Milano a city that had been ableto overcome the disasters of World War II and wasexperiencing dramatic economic growth led by themanufacturing sector Bocconi looked to me andmany others as a vital part of Milano and con-tributed to consolidate this process Then duringthe last couple of years the student protests of 1968started Despite being a private university Bocconiexperienced serious clashes and some of its studentsplayed an active role in the movementI Did you like studying economicsE I was very fond of economics The professors

I was interacting with were quite enthusiastic aboutmy inclination toward developing mathematical toolsuseful for economic modeling There is an episodethat occurred during my third year that I like torecall I was attending a course in Political economywhich included a series of seminars and one of themconcerned the relationship between the Italian Cen-tral Bank and the Department of Treasury whichat that time was the subject of a lively debate Foran economic interpretation of the relationship be-tween the two institutions we were suggested to re-fer to an article by Giorgio La Malfa and FrancoModigliani the latter was later awarded the Nobel

A CONVERSATION WITH EUGENIO REGAZZINI 3

prize in economics in 1985 The main contribution ofthe paper was the proposal of a static model Playinga bit around with that model I was able to derivea dynamic version of it which seemed in line withthe real situation in Italy This was appreciated bythe other students and the teaching assistants Alsoin connection with the course on Public FinanceI devised a model describing the evolution of cer-tain taxing decisions Overall I think I had quitea good economic intuitionA What did lead you to study probability and

statisticsE I was both impressed and fascinated by the first

year course in mathematics that was taught by Gio-vanni Ricci It was more advanced than a traditionalcalculus course The second year maths course de-livered by Giuseppe Avondo-Bodino included alsoa part devoted to probability which actually cov-ered essentially the same material nowadays taughtin first probability courses in maths degrees Duringthe second year of my degree I also attended a courseon statistical Inference held by Francesco Brambillawhich was important for my education Finally mythird maths course by Eugenio Levi contained someprobabilistic applications This experience revivedmy curiosity dating back to high school for founda-tional aspects of probability Moreover I perceivedprobability as a tough subject and therefore morechallenging and stimulating than others I was study-ingI What was the topic of your degree thesisE I asked Avondo-Bodino to be my thesis super-

visor He was a passionate Fisherian and hostile to-ward the Bayesian paradigm it might thus seemcurious that the title of my thesis was ldquoThe Bayesianapproach to hypothesis testingrdquo Indeed he chosethat topic with the aim of proving the fallacy of theBayesian paradigm this is revealed by the fact thatthe potential of the Bayesian approach was going tobe assessed with respect to hypothesis testing prob-lems that had already received well-established an-swers within the frequentist framework To be hon-est he did not even like the NeymanndashPearson ap-proach according to him it introduced subjectiveelements since it relied on decision theory My taskwas essentially to (i) collect as much material aspossible on hypothesis testing (ii) evaluate the pos-sible impact of the Bayesian approach and (iii) es-tablish whether it could be a sensibile alternative tothe frequentist approach And my supervisor obvi-ously expected a negative answer to the last ques-tion

A And how did it work out What were your firstimpressions on the Bayesian approachE While working on the thesis I developed some

skepticism about the automatic implementation ofBayesrsquo theorem which was a legacy from Laplaceand his followers However my viewpoint was lim-ited In fact writing the thesis was not an easy jobespecially because I could not rely on many sys-tematic and exhaustive treatments There were ofcourse de Finettirsquos papers but given the unortho-dox way they were written I was not able to un-derstand the connection between his theory and theAnglo-American neo-Bayesian approach typicallyadopted in papers appearing in statistics journalsat that time De Finettirsquos work did not follow thestandard BayesndashLaplace paradigm in contrast here-constructed it and recast it in a way to be coher-ent with his approach to prediction The books I re-ferred to were Lindley (1965) Raiffa and Schlaifer(1968) and mostly Ferguson (1967) which containeda beautiful part on the Bayesian approach from theviewpoint of Waldrsquos decision theory The 1959 lec-ture notes of de Finettirsquos course at a Summer Schoolin Varenna [later translated in de Finetti (1972)] andSavage (1954) were also helpful I obtained a few mi-nor results in terms of interpretation and compari-son and also derived a ldquorulerdquo for the choice of type-I error probability α As soon as I completed thethesis the monograph DeGroot (1970) appearedI found it very interesting and it proved to be veryuseful for my statistical educationA One of your best friends and main coauthors

is certainly Donato Michele Cifarelli Did you meethim while studying at BocconiE Yes he is actually 10 years older than me and

was my teaching assistant while I was attendingthe statistics course He delivered insightful lectureswhere it was apparent that he had remarkable math-ematical skills and also a deep knowledge of the bookby E Lehmann Therefore it was quite natural toseek his help when I started working on the thesiswhich required me to study also the frequentist ap-proaches His advice was very important althoughat the time he looked at least to me not interestedinto the frequentist versus Bayes debate I was im-pressed by his vast knowledge of frequentist meth-ods both parametric and nonparametric as well asof probability theory and stochastic processes I verymuch liked the fact that he preferred fundamentalthough maybe difficult at least for us books tomuch more immediate cookbooks For instance heknew in great detail the celebrated monograph by


J L Doob on stochastic processes We have beenbound by a deep friendship and reciprocal esteemever sinceI Who were other important scholars you met at

the time and who influenced your early approach tomathematicsE Overall Bocconi was an intriguing place at

least in Italy for probability and statistics giventhese subjects were as I said almost absent frommost maths degrees At the time Italian statisticsand academia in general did not have systematiccontacts with the international community and thehead of the Institute of Statistics Brambilla had themerit of introducing and spreading the great devel-opments in statistics and operations research whichtook place in the UK and the US He had contactswith foreign scholars for example he is also citedby Leonard J Savage in his 1954 book At the heartof these scientific activities was the Centre for Op-erations Research which among its sponsors couldcount on Adriano Olivetti an enlightened and revo-lutionary entrepreneur of the time who ran the com-pany producing the celebrated typewriter ldquoLettera22rdquo now displayed at the Museum of Modern Artsin New York City His company is also well-knownfor its pioneering contributions to the developmentof personal computers Among the various culturaland scientific activities created and supported byAdriano Olivetti it is worth mentioning the jour-nal ldquoTecnica e Organizzazionerdquo Brambilla was a co-editor of the journal when de Finetti published animportant paper on the essentials of computationaltechniques based on Monte Carlo methods ldquoMac-chine che pensano e che fanno pensarerdquo (ldquoMachinesthat think and that make you thinkrdquo) Brambillawas a remarkable figure he had been the assistantof Ferruccio Parri who besides being one of the firstItalian Prime Ministers after the war was also im-prisoned by the Germans during World War II sincehe had been one of the antifascist opposition leadersA What happened after you graduatedE I really enjoyed working on my thesis and I was

eager to continue at least for some time with re-search PhD programs did not exist in Italy sincethey were only introduced in the mid-1980s So I wasdoomed to the military service which was compul-sory and would have lasted for 15 months I triedto postpone my entry for as long as possible sinceI wanted to compete for a scholarship from the Ital-ian Ministry of University Had I obtained it I couldhave freezed it until the end of the military ser-vice Thankfully my strategy was successful and

when I was discharged in January 1972 I was ableto go back to university After graduation and be-fore starting the military service I shared the officeat Bocconi with Cifarelli and together we attendedvarious maths courses at the State University in Mi-lano I then sat the exams during my military ser-vice but in the end I did not complete a maths de-gree since I was already involved in developing myown research and I was willing to publish Nonethe-less those studies turned out to be very useful forme

2 FROM TORINO TO BOLOGNA MILANO

AND PAVIA

I Unlike many Italian academics and more in linewith what happens abroad you have been workingin many different universities Was this importantfor your professional developmentE Definitely In addition to working in various

universities I also experienced very different envi-ronments It has been very helpful from both a sci-entific and personal point of view I met many statis-ticians and mathematicians with very different back-grounds My first experience outside Bocconi was inTorino it was a small Department most colleagueswere of my age and so it was pretty easy to settlein Afterward I moved to Bologna in a much largerDepartment more in line with the Italian statisticstradition Then I got back to Milano first at theMathematics Department of the State Universityand then to Bocconi Finally Pavia which is one ofthe oldest universities founded in 1361 and in a veryprestigious Mathematics Department of which I ama proud memberA Tell us a bit about your first steps in the Italian

academia in TorinoE My supervisor Avondo-Bodino was full pro-

fessor in Torino and a lecturer at his department re-signed and decided to leave the academia Since theyneeded a replacement in December 1973 I movedthere with the concrete opportunity of obtaininglater a permanent position Due to absurd bureau-cratic reasons I finally obtained an Assistant Pro-fessorship only in 1978I You obtained a full professorship position in

a national competition at a young age in 1980 andtherefore moved to Bologna which hosted one ofthe few faculties of statistics in Italy Then back toMilanoE That was a time Italy was investing in univer-

sities unlike now Therefore the career perspectiveswere quite good also in the academia if you worked


hard Incidentally the head of the selection commit-tee was de Finetti I must say a recurrent figure inmy life In 1980 the only autonomous statistics fac-ulties were Roma and Padova whereas Bologna andPalermo were offering statistics degrees but withineconomics faculties Statistics became a faculty inBologna only toward the end of the 1980s In factand in contrast to what happens outside Italy facul-ties are pivotal players in Italian academia mainlybecause they manage the recruitment Now it seemsthat things will change but as we say in Italy every-thing changes so that nothing changes Anyhow inBologna I mostly taught probability courses but myties to Milano were still strong especially becauseof my collaboration with Cifarelli Therefore I ac-cepted the offer from the Mathematics Departmentof the University of Milano in 1984 where they didnot have a faculty member doing research in prob-ability and mathematical statistics until my arrivalA key role for my transfer was played by an analystMarco Cugiani whom I also replaced as director ofa research institute of the CNR nowadays the Mi-lano branch of the Institute of Applied Mathematicsand Information TechnologyA In 1989 you moved back to your beloved Boc-

coni University Did you find any substantial changessince the last time you had been thereE At Bocconi things had changed a lot the most

apparent being that it had turned from an elite toa larger and more open university with somethinglike 10000 students Hence I think that some changeswere necessary As for myself I was in a somehowprivileged position since I was mostly teaching ad-vanced and not compulsory courses which were muchmore challenging than most other courses There-fore starting from a yearly basin of more than 2000students of very good quality by self-selection I hadsmall numbers of students who were highly moti-vated and of the highest quality I guess I have tomention you two will not I But let me also mentionChiara Sabatti and Giovanni Peccati among manyothersI What convinced you to move to another univer-

sity in 1998E During the 1990s while I was there an even

more radical reorganization was occurring coursesof the type I was teaching were perceived as tooldquoaristocraticrdquo and had too few students so that theywere doomed to be shut down And the same des-tiny was foreseen for the most challenging degreesIn fact when I moved back in 1989 I did it withthe aim of setting up a statistics degree I was very

disappointed when the project was officially turneddown in 1997 Therefore I decided that my experi-ence at Bocconi was concluded and that I wantedto move back to a maths department NonethelessI still have a special affective relation with Boc-coniI Then you moved to Pavia And I followed you

given I had just graduated from Bocconi and startedmy PhD in Pavia It was a new challenge for youat a mature age was it notE The only science faculty members and future

colleagues of mine in Pavia I knew in person prior tomoving were three brilliant mathematicians Mau-rizio Cornalba (we were both members of the scien-tific committee of the Italian Mathematical Union)Franco Brezzi (we got to know each other at themeetings of the Italian National Research Council)and Enrico Magenes Enrico Magenes who passedaway last November shaped the department in itscurrent form and significantly contributed to its in-ternational standing When I moved I was one ofonly two probabilists and since then we have beenable to hire two additional Assistant Professors Oncearrived with great enthusiasm I immediately got in-volved in the PhD program and the outcome hasbeen rewarding as also witnessed by the achieve-ments of some former PhD studentsA Throughout your career you have been working

at faculties of science and economics at public andprivate universities What environment do you thinkbetter fits the needs of a researcher in probabilityand statistical scienceE I think that the ideal environment relatively

to the Italian experience which is the only one I amaware of is a maths department which is open to-ward the modern trends and research directions ofmathematics and therefore not too much bound totraditional subjects like analysis geometry and math-ematical physics And thankfully there are variousdepartments that comply with these criteria How-ever I must admit I am a bit pessimistic about thefuture the Italian university system in general anddepartments involved in basic research in particu-lar are struggling and suffering due to the indis-criminate financial cuts in recent years These havebeen implemented somehow light-heartedly by thegovernment since cuts in basic research funding areunlikely at least in Italy to cause immediate socialupheaval In fact it would be important to abandonthe habit of uniformly distributed cuts and aim atcreating or consolidating niches of excellence AndI have seen many such niches during my career


3 BAYESIAN INFERENCE

I At the beginning of your academic career youstarted working on inferential problems according tothe frequentist approachEMy interest in frequentist inference started soon

after completing my degree thesis and heavily bene-fited from the collaboration with Cifarelli And oneof the first topics we started working on was hypoth-esis testing Corrado Gini and other Italian statisti-cians had introduced a considerable number of sum-mary statistics that were originally used only forexploratory data analysis to measure for instanceconcentration variability dependence and similar-ity between sets of data and so on Our idea was touse such summary measures for inferential purposesand specifically as test statistics for studying de-pendence in nonparametric problems An early con-tribution in this direction was achieved by Cifarelli(1975) who studied the asymptotic distribution ofa statistic arising in a test of homogeneity for two-sample problems The paper contained a remarkableresult on the distribution of the integral of the abso-lute value of the Brownian bridge Our initial effortsled to a paper (Cifarelli and Regazzini 1974) that weare very proud of there we determine the limitingdistribution of a measure of monotone dependenceintroduced by Gini The program we set was veryappealing and consisted in checking whether thesestatistics when used for hypothesis testing yieldedtests that were more efficient than those commonlyused at the time For example the index of mono-tone dependence I was mentioning was comparedwith Spearmanrsquos ρ and with Kendallrsquos τ and in somecases it featured better performancesAAround the mid-1970s you turned back to Baye-

sianism What about your skepticismE Yes and my experience at University of Torino

was fundamental in this respect The department li-brary held the collection of all de Finettirsquos paperswell kept and easily accessible I started looking atcontributions cited by Savage as decisive for thefoundations of the Bayesian paradigm My curiositywas fueled by the fact that as I said de Finettirsquoswork appeared to me completely disconnected fromthe Bayesianism I had studied on books and journalarticles It was a challenging task since de Finettirsquoswriting style which was actually one of the mainaspects he was criticized for was unorthodox andsometimes seemingly cryptic Nonetheless hard workand stubbornness finally allowed me to understandwhy de Finetti introduced exchangeability and therole such a form of symmetry plays in the recon-

struction he gave of the BayesndashLaplace paradigmThis really opened my eyes on a new world provid-ing a coherent and unified view of statistical infer-ence where subjective probabilities play an impor-tant role In fact suddenly the subjective interpreta-tion of probability was the only one that made senseto me from both a philosophical and a mathematicalpoint of viewI You were then able to convince Cifarelli to enter

the realm of Bayesian statisticsE I have to say that Cifarelli shared my same

doubts on the foundations of the Bayesian approachto statistical inference However after completingmy study program in Torino I pointed him to thereferences where de Finetti was answering our ques-tions and solving our doubts Besides de Finettirsquoswell-known papers I had discovered many other ldquomi-norrdquo contributions that were important for under-standing the unified framework he had in mind Andafter struggling to understand I started to love hisstyle entering his world had been very demandingbut once I succeeded the reward was incomparableIn his work one could find ideas hints and conceptswhose expressive force was much more powerful thana standard presentation of definitions theorems andcool mathematical technicalities Spurred by the en-thusiasm I had been able to convince Cifarelli andwe started working together in this directionI Is this when you started your research on Bayes-

ian nonparametricsE In some sense yes On the one hand we were

hoping to be able to tackle in a Bayesian setting thesame issues we had addressed within classical non-parametric inference On the other hand we guessedthat our starting point should have been de Finettirsquosrepresentation theorem as stated in de Finetti (1937a)which we could consider as being nonparametricIn this fundamental paper the law of an exchange-able sequence is described as a mixture on a spaceof probability measures and the prior is the almostsure limit in a weak sense of the empirical measuregenerated by the data This motivated the investi-gation of random probability measures (rpmrsquos) forstatistical inference and might have led to extendthe BayesndashLaplace paradigm We planned to con-sider estimation of functionals of rpmrsquos such as themean the variance or other characteristic parame-ters of the unknown distribution The necessary pre-liminary step was to determine the posterior distri-bution of these functionals A helpful reference wasa short paper (de Finetti 1935) where de Finettiprovides a reformulation in Bayesian terms of meth-


ods used in exploratory data analysis for smooth-ing the empirical distribution Moving from this hebasically addressed in a nonparametric frameworkboth the issue of prediction and of evaluation of theposterior distribution on a set of probability mea-sures Unfortunately we had no clue on how to de-fine a probability distribution on a space of proba-bility measures that would be analytically tractableOf course we were not aware of T S Fergusonrsquospaper (Ferguson 1973) We were stuck and all theattempts we made led us nowhereA Was there any decisive event that helped you

overcoming these difficultiesE In 1976 I met Andrew L Rukhin who had left

the Soviet Union and was in Italy just before mi-grating to the US We discussed our research activ-ities and I described to him the technical problemsCifarelli and I were dealing with He suggested wego through Fergusonrsquos paper in order to find an an-swer to our questions And indeed that was thecase that paper allowed us to resume our projectSo we started considering linear functionals of theDirichlet process with the aim of determining theirprobability distributions analyticallyA and I Let us also recall that the study of the

Dirichlet process suited your passion for classicalmusic very wellE Gustav Dirichlet is associated with the distri-

bution because he evaluated the integral on the sim-plex The musical connection is that he married Re-becka Henriette Mendelssohn younger sister of FelixMendelssohn the famous German composerI Were there other Bayesians in Italy at the timeE A few years after its re-flourishing at an inter-

national level due to the work of Leonard J Sav-age the Bayesian approach was sort of rediscoveredin Italy as well This may sound surprising given deFinetti is Italian however one has to consider thatde Finetti only entered academia in 1946 at theage of 40 when his research was already focusedon different topics Interestingly he had obtainedthe position already in 1939 but could only starthis job in 1946 after the fall of the fascist regimedue to a law forbidding the appointment of unmar-ried professors as was de Finettirsquos case Anyhow inthose years there was a large group led by GiuseppePompilj in Roma and some scholars started to workon Bayesian statistics like Ludovico Piccinato InRoma there were also some of de Finettirsquos studentslike for instance Fabio Spizzichino I should alsomention a group based in Trieste and coordinatedby Luciano Daboni who started working under de

Finettirsquos supervision soon after gaining his univer-sity degree Besides actuarial mathematics they fo-cused mainly on exchangeable processes and foun-dational issues of Bayesian inference and during theyears I had many fruitful interactions with themA Even if more interested in the Bayesian para-

digm you did not avoid doing research based ona frequentist approach It seems you did not andstill do not see any ideological contraposition be-tween Bayesianism and frequentismE I have never seen this as an ideological con-

traposition I think that ideological positions makesense only outside the realm of mathematics Any-how even when I was working on statistical prob-lems according to the frequentist approach I alwayshad the feeling that the Bayesian framework was farmore complete and logically sound I was not en-thusiastic about the automatic use of priors on un-observable parameters the subjective views I hadon probability were in conflict with such a treat-ment of the BayesndashLaplace paradigm as I believethat inference must concern quantities that can beempirically observed But on the other hand theFisherian attitude appeared to me as too drasticbecause prior beliefs should play a role in statisticalinference Once able to fully understand the conse-quences of de Finettirsquos results I became convincedthat Bayesianism was the only acceptable way ofinductive reasoningA Current developments in Bayesian inference in-

volve a heavy use of simulation algorithms Do youstill think there is a need for putting a strong effortin determining exact forms of Bayesian inferences(or at least error evaluation when approximationsare used) even when these are difficult to use inpracticeE Computational techniques have been decisive

in making Bayesian models applicable to real worldproblems and some recent applications I saw aresimply amazing I definitely think that the advan-tages they yield largely surpass some drawbacks as-sociated with their uses That said I would still liketo make a point which I think is important since ithas to do with how statistical modeling is conceivedIndeed models should be devised as simple as pos-sible while still preserving the capability of captur-ing the essential features of the phenomenon understudy Such a simplification could be achieved byfirst detecting inessential elements and then drop-ping them when it comes to the point of specifyingthe model This attitude is natural when one aims atachieving exact estimates of the quantities of inter-


est However if the need for pushing analytic resultsas far as possible disappears it is likely that themodels become more loose and unnecessarily com-plex Both parsimony and extreme care in the for-malization of models are still important guidelinesfor research the only difference is that they nowneed to be spelled out clearly while they were im-plicitly followed in the past Another related andimportant point concerns approximation When ex-act inferences are not possible one should put someeffort in providing an upper bound to the error ofapproximation yielded by the numerical techniquesthat are used I have tried myself to work in thisdirection for instance in relation to approximatingthe probability distribution of the mean of a Dirich-let process I know this is a challenging task but itcannot be avoided

4 DE FINETTI AND THE INFLUENCE OF

DE FINETTIrsquoS WORK

I There is no doubt your research has been deeplyinfluenced by de Finettirsquos work Which was the firstpaper of de Finetti you read throughE While I was completing my thesis at Bocconi

I came across his joint paper with Savage (de Finettiand Savage 1962) It contained a discussion on thechoice of the prior distribution and was mainly illus-trative with no deep mathematics involved but stillevocative for a noviceA His most renowned piece of work certainly is

the two-volume book on probability theory de Finet-ti (1970) What else would you suggest to a studentwho is willing to study and understand de Finettirsquosstance in probability and statisticsE I would certainly suggest de Finetti (2006) two

volumes containing selected papers by de Finettiwhich have been published by the Italian Mathemat-ical Union in 2006 in occasion of the centenary of hisbirth The first volume is on probability and statis-tics whereas the second is on applied maths and onthe teaching of maths As for his subjective views onprobability one should refer to de Finetti (1931)One should also read de Finetti (1937a) Anotherimportant piece of work is de Finetti (1972) Fi-nally de Finetti (1992) contains a selection of someof de Finettirsquos papers with English translation Un-fortunately some significant contributions at leastto my knowledge have been only published in Ital-ian such as those related to independent incrementsprocesses and some others on the subjectivistic def-inition and interpretation of probability

I As you just mentioned the fact that he wasnot writing in English hindered the circulation ofhis ideas and results in the scientific communityE This is definitely true For example it is prob-

ably unknown to many that de Finetti introducedthe celebrated τ index a few years before Kendall(de Finetti 1937b) In 1939 he obtained some impor-tant results on optional stopping indeed de Finetti(1939) deals with the gamblerrsquos ruin problem whereone can also find an embryonic version of the Gir-sanov theorem Another important contribution wasthe continuity theorem for characteristic functionshe proved it in the appendix of de Finetti (1930a)Besides these it is worth listing a few other con-tributions for which a priority to de Finetti shouldbe acknowledged he completed what is now knownas the GlivenkondashCantelli theorem before FrancescoP Cantelli in de Finetti (1933) in de Finetti (1940)he devised a model that anticipated the portfoliotheory for which Markowitz was awarded the Nobelprize he proved the theorem on almost everywherenondifferentiability of the trajectories of the Brow-nian motion in de Finetti (1929)A With reference to the de Finetti (1929) pa-

per which is actually our favorite should we not asItalians propose Levy processes be called de FinettindashLevy processes insteadE As I mentioned before the answer is affirma-

tive Indeed de Finetti started from a more gen-eral problem of providing the random counterpartsof a Volterra classification for the ordinary laws ofphysics In this context he identified processes withindependent and homogeneous increments as thosewhose characteristic function satisfies the first of theequations in Volterrarsquos classification namely X prime =f(λ) As a by-product he also introduced implic-itly the notion of infinite divisibility In a subse-quent paper de Finetti (1930b) he further char-acterized the class of infinitely divisible laws as theclass of distribution limits of compound Poisson pro-cesses thus providing a representation theorem forinfinitely divisible distributions Levy was not awareof de Finetti (1929) and resorted to a different ap-proach to obtain more general and deep results Thecontribution by Khintchine to the well-known LevyndashKhintchine representation originates from a paperpublished in 1937 (see Khintchine 1937) Khint-chinersquos paper builds upon Kolmogorov (1932) whereKolmogorov explicitly mentioned (even in the titleof the article) that he was resorting to the approachset forth by de Finetti So yes it should definitelybe de FinettindashLevy processes


I And what were his connections with the broaderinternational scientific communityE His first international contacts before gradu-

ating in mathematics at the University of Milanoare related to a paper de Finetti (1926) he wroteon Mendelian inheritance which had quite an im-pact in biology It was his first paper and appearedon the Italian journal Metron His results also at-tracted the attention of Alfred J Lotka and JacquesS Hadamard The latter was so impressed by de Fi-nettirsquos achievements that he suggested Georges Dar-mois to study the paper as witnessed by one of theletters that de Finetti wrote to his mother in 1929and that have recently been published by his daugh-ter Fulvia This research also originated the so-calledde Finetti diagrams that are extensively used in pop-ulation geneticsA An important event at which de Finetti drew

attention on his research in probability was the In-ternational Congress of Mathematicians (ICM)which was held in Bologna in 1928E That conference was definitely important for

the development of de Finettirsquos interactions withforeign scholars On that occasion he presented hisfirst results on exchangeability and made contactwith Maurice R Frechet who later invited him tothe Institut Henri Poincare in 1935 and to the Col-loque de Geneve in 1937 where he then also metJerzy Neyman and others He had frequent interac-tions with Paul Levy and Aleksandr Khintchine re-spectively on independent increment processes andon the proof of the representation theorem for ex-changeable sequences He was also in contact withAndrey N Kolmogorov as witnessed by the Kol-mogorov (1932) paper on infinite divisibility whosetitle contains an explicit reference to ldquoA problem ofde Finettirdquo Both Kolmogorov and de Finetti alsoworked at the same time on the derivation of a repre-sentation theorem for associative means now knownas the de FinettindashKolmogorovndashNagumo TheoremHe also got in contact with many eminent math-ematicians via mail In fact he used to have a note-book in which he recorded to whom he had sentwhich of his papers de Finettirsquos daughter Fulviaonce showed it to me and the names are impres-sive After World War II he had significant scientificcollaborations with Leonard J Savage and LesterDubins and he interacted also with William Fellerand Abraham WaldI Were his views on the subjective approach to

probability theory held in high regard

E In mathematics his work has been largelyignored and not only because of the subjective in-terpretation Indeed the mathematical approachyielded by such interpretation does not require σ-additivity In fact finitely additive laws also becomeadmissible and the traditional measurendashtheoretic ap-proach to probability theory represents obviouslya particular case Countably additive probabilitiesare coherent in de Finettirsquos sense but are just a sub-class of coherent laws And de Finetti himself waswell aware that many results could have been neaterby assuming countable additivity We may reason-ably conjecture that his position in favor of includingfinitely additive probabilities somehow put him offfrom focusing on the particular countably additivesetup This could explain for example why he didnot further investigate processes with independentincrements It is to be noted that the framework forhis subjective approach had been settled by 1931and as evident from his published mail exchangewith M Frechet he fought for it for a whileI And what about the impact on statistical prac-

ticeE In Bayesian statistics references to subjectivism

are quite frequent but I actually see little of de Finet-ti behind them First in the subjective approachalso finitely additive laws are allowed and thereforea proper subjectivist should try to analyze statisti-cal problems in this setup This point is very impor-tant in the case where ldquotranscendentrdquo conditionsmdashsuch as convergence of sequences of random ele-ments forms of the corresponding limits etcmdashareinvolved one should then establish the extent towhich the conclusions depend on the specific σ-addi-tive extension (usually unique) of the original finite-dimensional distributions Second from an interpre-tation point of view subjectivism and objectivismare often mixed up and Bayes theorem is appliedin an automatic way whereas subjectivism wouldrequire probabilistic statements to be made on ver-ifiable events Subjectivism seems more a kind ofcatch-phrase than a real commitment In my opin-ion the papers of L J Savage L Dubins J Pit-man P Diaconis and D Freedman are the ones thatadhere most closely to de Finettirsquos viewsA Did your convinced support of subjective prob-

ability affect the way you teach probability coursesE This represented a sort of dilemma through-

out my career Focusing solely on de Finettirsquos math-ematical theory of probability would have impliedproviding students with an unorthodox background


in probability it could have been an enrichment forsome of them but also a drawback for some othersespecially for those who needed to use probabilityas a mere tool in other disciplines Therefore mostof the courses I taught were within the σ-additivityframework Nonetheless I have always tried to il-lustrate extensively some distinctive features of thesubjective viewpoint in one of my first lectures Thiswas useful since it provided students with a morecomplete picture of the subject and allowed themto understand that the results I was going to stateand prove were valid on a special class of proba-bilities sharing the property of countable additiv-ity Students were then aware that it was somehowlike teaching them a course in analysis that was justabout analytic functions The connection with con-ditional properties was far more difficult to pointout As for the subjective interpretation it is stillpossible to preserve it even when confining to σ-additive probabilitiesA Can you provide some further insight on this

last issueE The difficulty I am referring to arises due to the

fact that Kolmogorovrsquos definition cannot be seen asa special case of coherent conditional probabilitiesIn fact the Kolmogorov approach lacks an appropri-ate axiomatization and interpretation of conditionalprobability the definition is by means of a limit-ing procedure The perspective is then completelydifferent For example de Finettirsquos approach nec-essarily leads to conditional probabilities that areregular and proper whereas it is well known thatKolmogorovrsquos definition does not In order to graspthese mathematical and conceptual differences onconditional expectations and probabilities one canrefer to the works by L Dubins David BlackwellCzeslaw Ryll-Nardzewski William D Sudderth Ro-ger A Purves Pietro Rigo Patrizia Berti and alsomyselfI Tell us about your meetings with de FinettiE I first met him in 1969 at a summer course on

mathematical economics in Urbino Since I was work-ing on my thesis I took the opportunity to ask hima few questions about his paper with Savage (de Fi-netti and Savage 1962) I had read He was not verytalkative and probably thought I was not understan-ding anything He was right but I still went awaywith the impression that it was not simple at all to in-teract with him Afterward I met him at some confe-rences during the 1970s but at that time he was notworking on statistics and probability with the same

intensity and creativity of the early days he wasmore inclined to elaborate on general philosophicaland foundational aspects The only thing I can sayabout our meetings is that I had the impression hewas interested in nontrivial and original approachesor attitudes that to some other people might haveappeared as singularities For example in Bolognahe once told me he had been fascinated by the math-ematical physics lectures held at the Polytechnic inMilano by a lecturer Bruno Finzi whose assign-ments were notoriously challenging and containedexercises that Finzi himself could not solve He re-called the solutions he had been able to give werevery original and much appreciated by Finzi He alsotold me he had appreciated lectures on economicsof insurance companies delivered by Ulisse Gobbi atthe Polytechnic in Milano because they had been thesource of inspiration for the mathematical modelingof many aspects of economics he had later investi-gated I am surprised by this since in Gobbirsquos workI did not find any mathematical formalismA He was also engaged in public life and gained

some popularity because of his political experienceE His political experiences can be well under-

stood if one refers to the environment where he grewup De Finettirsquos family was wealthy and highly ed-ucated They were part of the Italian community interritories of the Habsburg Empire and his fatherwas an engineer working for the Austro-Hungarianrailway During his childhood he had learned aboutthe irredentist ideas of the Italian minority that wasaiming at unification with Italy Such aspirationsquite naturally developed into strong nationalist feel-ings once the area became part of Italy Moreoverhaving been part of a minority he developed a strongsensitivity toward injustice in all respects and there-fore also a strong criticism toward some social im-plications of capitalism of the time This blend ofideas somehow naturally led him to support the ris-ing fascist party its initial political and social pro-gram included a series of reforms whose goal was thecomplete State control of the economy As de Finettihimself wrote a few years before dying the direc-tion of the whole economy once freed from the ter-rible tangle of individual and interest group selfish-ness should lean toward the collective achievementof a Paretian ldquooptimumrdquo and should be further in-spired by ldquofairnessrdquo criteriaA Hence his support to fascism was mainly the

result of ideal feelings that were fueled by strongsocial and economic views


E This is further witnessed by the fact that af-ter the fall of fascism he sympathized with left-wing movements without adhering to a large politi-cal party Finally during the 1970s he started beinginvolved in important campaigns for civil rights andfor social justice The Italian party that better fit-ted his political thoughts of the time was the RadicalpartyI Can you tell us something about it It seems

that while being involved in political activities setforth by the Radical Party he spent one night in jailE In fact he did not end up in jail because the or-

der to release him arrived before being imprisonedTo make a long story short he was editor of a news-paper of the Radical Party which was publishingletters of conscientious objectors who refused to per-form the compulsory military service This was ille-gal at the time The day he learned he was goingto be arrested he asked the police whether it waspossible to arrest him at the Accademia dei Lin-cei the most prestigious Italian science academywhere he was going to have an official meeting theday after He motivated such a seemingly bizarre re-quest with the fact that the police could have savedsome money by not picking him up by car at homethe Accademia dei Lincei building was indeed justa few steps away from the prison he was supposedto go to However the order to release him arrivedas soon as he got to jail This episode had a hugeecho in the pressAWe also recall a story you told us about Kolmo-

gorov visiting Roma and wanting to meet de FinettiE In 1962 Kolmogorov was awarded the Balzan

prize for Mathematics the other awardees beingPope Giovanni XXIII for Peace Paul Hindemith forArts Samuel E Morison for Humanities and Karlvon Frisch for Biology Two well-known mathemati-cians Gaetano Fichera and Olga A Oleinik went tocollect him at the Roma airport and asked him whatthey could do for him And as Fichera reported hisanswer was ldquoIf you know him then you should or-ganize a meeting with de FinettirdquoA and I De Finettirsquos papers are scattered with

brilliant ideas sometimes only sketched What arethe aspects of de Finettirsquos work which still need tobe developedE As for some specific topics such as exchange-

ability and processes with independent incrementsin my opinion most of his ideas have already beenextensively developed and not much is left to in-vestigate in the precise direction he had originallythought of On the other hand I believe that much

is still left to investigate on the general foundationsof probability theory that emerge from his work andthat he strongly supported These studies might havea relevant impact in statistics in physics and inother research areas The advances I am thinkingof concern both the interpretation of probability andthe enlargementmdashalong with its mathematical impli-cationsmdashof the class of admissible probability lawsto include also the finitely additive ones

5 PROBABILITY AND STATISTICS IN ITALY

A You investigated quite extensively the devel-opment of statistics and probability in Italy in thefirst half of the 20th century (eg Regazzini 2005)Can you tell us about itE In contrast to what happened in the Anglondash

American world or in Russia in Italy probabilityand statistics developed along almost separate pathsProbability started growing in mathematical envi-ronments As far as I know the first to deal withthe topic in a comprehensive way was Guido Castel-nuovo a famous mathematician who was mainlydoing research in algebra and geometry His 1919book on probability (Castelnuovo 1919) was used asa textbook for quite some time in those few math-ematics degrees where probability was taught Theinterpretation of probability was frequentist in linewith a view that would have been later shared alsoby Frechet Levy and Kolmogorov and covered re-sults of the Russian school up to Andrey Markovand Aleksandr M Lyapunov Already back in 1915he had the idea of setting up a school of statisticsand actuarial sciences at the University of Romawhich was then created in 1927 It had consider-able success with many enrolled foreign students andthen became a proper faculty in 1936 with Gini Inthe preparation of his book Castelnuovo was helpedby Cantelli who is considered also at an interna-tional level one of the first modern probabilists Hederived among other contributions versions of thelaws of large numbers the BorelndashCantelli lemmaa mathematical theory of risk that was named afterhim and developed an autonomous abstract meas-urendashtheoretic theory of probability which appearedbefore Kolmogorovrsquos It is interesting to recall thatin this last development a crucial point was theproof of the existence of measurable maps definedon [01] endowed with the uniform distribution insuch a way they have prescribed probability lawssuch an approach also reflects the idea of adher-ing to the classical definition of probability due toLaplace Anyhow this problem led him and his stu-


dents to anticipate at least part of what is nowa-days known as the Skorokhod representation A dis-tinguished scholar who obtained important resultsalong the lines of research undertaken by Cantelliwas Giuseppe Ottaviani who is also known for hisinequalities that are related to Cantellirsquos theory ofrisk Francesco G Tricomi eminent analyst andfriend of Cantelli also gave some contributions toprobability as did Carlo E Bonferroni who is wellknown for his inequalitiesA Given such a glorious tradition it is quite sur-

prising as you said earlier that the first full profes-sors in probability were appointed by Italian univer-sities only in the 1970s with the notable exceptionof de FinettiE Actually at the beginning of the 1970s only

two professors in probability were recruited namelyGiorgio DallrsquoAglio and Giorgio Letta DallrsquoAglio wasat the Faculty of Statistics in Rome and was a mem-ber of the before mentioned group led by PompiljLetta is from Pisa and spent several research periodsin Germany and France The latter experience stim-ulated collaborations between Italian probabilistsmdashsome of whom were Lettarsquos studentsmdashand Frenchprobabilists in Paris and Strasbourg a fruitful trendwhich is still ongoing Then a larger group of peoplewere appointed at the end of the 1970s in variousItalian universitiesI And what about statisticsE In the last three decades of the 19th century

topics that are today ascribed to Mathematical Sta-tistics were taught in geodesy or astronomy coursesLectures by a not well-known Italian mathemati-cian Paolo Pizzetti were very interesting and con-tained some innovative ideas on significance testsMore conventional at least according to the Ital-ian framework statistics courses were in law facul-ties many academic statisticians actually had a de-gree in law Most of them were involved in Offi-cial statistics and it was therefore natural that theinteractions between statisticians and probabilistswere rather limited The first modern Italian statis-tician was Rodolfo Benini who had a law degreefrom the University of Pavia and developed statisti-cal methods for demographic sociological and eco-nomic problems around the end of the 19th and thebeginning of the 20th century I recall once I cameacross historical documents presented in notewor-thy conferences of the American and British Eco-nomic Societies where Benini is referred to as one ofthe founders of econometrics I think this due to hisanalysis of income and wealth distributions and to

the pioneering use of multiple regression methods toestimate for example demand curves He also hadthe idea of studying contingency tables with fixedmarginals Among his successors the main figure iscertainly Gini also a graduate in law His method-ological contributions to statistics were praiseworthyand were later studied not only in relation to meredata analysis Gini dominated Italian statistics un-til his death in 1965 and created a school of faithfulfollowers A prominent group of scholars was led byPompilj at the Faculty of Statistics in Roma AsI recalled earlier DallrsquoAglio was one of its membersand he obtained noteworthy mathematical resultsthat can be traced back to the Ginian analysis ofstatistical relationships His results however havea remarkable independent interest for example heprovided a relevant contribution to the definitionand to the properties of what is today known as theWasserstein distance See DallrsquoAglio (1956)A You mentioned Paolo Pizzetti who seems to be

a neglected figure within the Italian statistics com-munity was he not We have never heard of him inour statistics coursesE Yes he was unfairly neglected His contribu-

tions which appeared in the 1880rsquos were very inno-vative and relied on an original approach that some-how anticipated a few distinguishing ideas lying atthe foundations of statistics as set forth by KarlPearson and by Ronald A Fisher As an examplehe proposed procedures that were very similar to thesignificance tests Fisher would have later adopted asa distinctive feature of his methods Pizzetti also hadremarkable mathematical skills that allowed him todetermine the exact distribution of certain statis-tics used for data analysis And he was well awareof the results achieved in this direction by a Ger-man geodesist Friedrich R Helmert He reprovedHelmertrsquos results with the aim of extending themand relied on innovative methods and techniquesthat Fisher himself would have later proposed in-dependently This is very well documented in a re-cent historical monograph by Anders Hald As youcan easily guess Pizzettirsquos ideas were totally dif-ferent from those that Gini would have later ex-pressed apropos of the Fisherian tests Indeed Giniwas very critical about the use of significance testsand his criticisms were shared by de Finetti Thismay partly explain why Pizzetti is not known bymany statisticians His work was somehow consid-ered as heterodox for quite some time as demon-strated by the 1960s reprint of Pizzettirsquos 1892 book(Pizzetti 1963) In the preface written by V Castel-


lano P Fortunati and G Pompilj it is claimed thatparts of Pizzettirsquos work were ldquomisleading and con-tained errors that had been masterfully pointed outby Gini in Gini (1939)rdquo And the ldquomisleading partsrdquothey were referring to are exactly those where Pizzet-ti uses his results for devising statistical testsA and I The excerpt you read can partly explain

the isolation of the Italian statistics community inthose yearsE It partially does Indeed I think that Ginirsquos

critical remarks make sense The point is that theywere not complemented by alternative proposals thatcould take his concerns into account Hence it wasalmost inevitable that Ginirsquos position would havebecome marginal and isolated within the broaderinternational community It should be recalled thatisolation fitted very well with the political climatefavoring autarkic tendencies during the fascist regimeand it unfortunately further consolidated over theyears in the Italian statistics community at leastin academia This obviously had a long-lasting neg-ative impact from which Italian statistics startedrecovering only in the 1970sA Gini was appreciated both for his scientific

achievements and for his praiseworthy services asa scientific expert within various important ItalianinstitutionsE Definitely He was founding President of the

Italian Central Institute of Statistics in 1926 and setup first the School and then the faculty of statis-tics demographic and actuarial sciences in Romain 1936 He was in constant contact also meetinghim in person with Mussolini who used to pay at-tention to statistical analyses for taking decisionson policy issues For example he acted as a tech-nical advisor within the programs of demographicand eugenics policies pursued by the fascist regimeLater he also founded the Italian Statistical Societyof which he has been President for 20 years Besidesthe scientific and institutional authoritativeness hegained in Italy it should be recalled that he ob-tained countless recognitions abroad as well Amongthem I could mention that he became Honorary Fel-low of the Royal Statistical Society Vice Presidentof the International Sociological Institute and Hon-orary Member of the International Statistical Insti-tute In 1920 he was the founding Editor of the jour-nal Metron which published papers by many emi-nent statisticians of the time such as R A FisherA A Chuprov A J Lotka S S Wilks E E Slut-sky S Kullback H Wold and A L Bowley

I We have also heard of some funny stories aboutGini bearing ill-luck Can you tell us something moreE Yes this is somehow true but it is to be con-

sidered within the typical Italian attitude of mak-ing fun of powerful people as Gini certainly wasThere are various minor anecdotes and a dramaticepisode that would allow to conjecture a ldquocorrela-tionrdquo of the type you are referring to As for the lat-ter something incredible happened in 1927 he wason the steamboat ldquoPrincess Mafaldardquo which ship-wrecked off the Brazilian coast between Salvador deBahia and Rio de Janeiro and he was among thefew survivors the ldquolegendrdquo says thanks to his row-ing skills a sport he had practiced in youth A lessdramatic and funnier story I have heard of concernsan episode where chatting with a colleague of his hepaid a compliment to a young female studentrsquos legswhom they met on the stairs after a few steps shefell down and broke her leg I remember that Otta-viani did not mention his name he referred to himas the unnamed since mentioning his name couldhave led to something bad happening All kiddingaside after the shipwreck in Brazil he criticized theItalian authorities for the poor assistance from theItalian Navy and more in general from the Ital-ian government These complaints caused him a lotof troubles with the fascist regime in Italy He hadindeed a strong and straight attitude that helpedhim to protect scientific matters and appointmentsfrom political influence Of course this position at-tracted the aversion of many Fascist party officialswho strove for Mussolini to remove him as presi-dent of the Italian Central Institute of StatisticsAnd his criticisms on the occasion of the shipwreckwere added to the list of Ginirsquos ldquooffencesrdquo to theregime that led to his resignation in 1932 How-ever as I said before he kept collaborating with theregime as a scientific expert in demography statis-tics and eugenicsA Cantelli de Finetti and Gini were the tower-

ing figures in probability and statistics before WorldWar II in Italy They were also completely differentcharacters How did they get alongE Gini published de Finettirsquos work on Mendelian

inheritance in Metron and offered de Finetti a jobat the Italian Central Office of Statistics before hegraduated While at the Italian Central Office deFinetti was involved in a project for predicting theevolution of the Italian population and crucially de-signed all modeling aspects of the project He thenwanted this to be credited as his contribution but


Gini was reluctant to do so This episode is welldocumented in one of the letters de Finetti wrote tohis mother and contained in the collection publishedby his daughter that I have already mentioned Inany case at the end of his four year contract in1931 de Finetti moved back to Trieste and startedto work for the insurance company Generali The re-lationship between de Finetti and Cantelli was quitea difficult one since they were in strong disagree-ment on the interpretation of probability Cantellidid not want to hear anything about finite additiv-ity and he also tried to prove that σ-additivity wasa necessary propertyI In addition to Metron there was also the Gior-

nale dellrsquoIstituto Italiano degli Attuari (GIIA) whichwas a top journal in statistics and probability dur-ing the 1930s Why have they both lost their inter-national reputation since thenE The GIIA was established in 1930 the same

year The Annals of Mathematical Statistics pub-lished their first issue It was edited by Cantelliand the most distinguished scholars of the timesuch as Cramer Frechet Kolmogorov KhintchineLevy Neyman and von Mises published fundamen-tal contributions on it World War II ruined every-thing since its publication was suspended and theGIIA lost its elite status among the top probabilityand statistics journals which during and soon af-ter the war included The Annals of MathematicalStatistics along with Biometrika and the Journalof the Royal Statistical Society The other Italianprestigious journal Metron which was establishedin 1920 paid a high price for the line of developmentof Italian methodological statistics and actually al-ready declined before the warA In 1978 you actually published a very inter-

esting paper characterizing the Dirichlet process interms of linear predictive distributions (Regazzini1978) on the GIIA Why did you decide this wasa suitable outlet for your paperE After my discussions with Rukhin the Dirich-

let process became a main ingredient of my researchagenda In fact I was dealing with risk premiummodels for insurance companies which were linearcombinations of an empirical part and an expectedvalue related to some prior guessmdashthey identify theso-called credibility premium I thought to revisitthe problem coherently with the predictive distri-butions generated by an exchangeable sequence andasked myself what the underlying de Finetti mea-sure was it turned out to be the law of a Dirichlet

process I wrote this paper while I was working withCifarelli on the distribution of linear functionals ofthe Dirichlet process I then presented it at a con-ference where Luciano Daboni an editorial boardmember of GIIA was present he liked the papera lot invited me to give a seminar in Trieste andproposed for me to publish it in GIIA Some yearslater the same result was independently obtained byAlbert Y Lo (Lo 1991)I During the 1980s you were probably one of the

few statisticians in Italy who published their papersin international journals Do you have any idea whythis happened at the timeE Well first of all most people both in statistics

and probability did not even try to submit theirwork abroad It was simply not necessary for theprogress in academic careers Even many mathe-maticians only published in Italian journals Overallthe need for trying to spread onersquos own work at aninternational level was not felt yet Actually it wasprobably not even felt in the Anglo-American worldit just happened that their journals then becamethe ldquointernationalrdquo ones By the way papers thatwere published in Italian journals with a very lim-ited spreading were not all necessarily of bad qual-ity On the contrary some of them are very wellknown even abroad and contain innovative ideasAnyhow in recent years things have changed sub-stantially and young researchers submit their workto the best international journalsA We as students have nice memories of sum-

mer schools organized by the Italian scientific com-munity to support the spreading of probability andstatistics You have been an active part of this ini-tiativesE After attending some of them as a student

I have been involved several times in organizing andteaching at summer schools that took place in var-ious beautiful locations in Italy such as CortonaPerugia Livigno and Rheme-Notre Dame In addi-tion to being an opportunity to meet talented stu-dents summer schools also allowed me to get in con-tact and actually build up friendships with someauthoritative scholars such as Alan Agresti PatrickBillingsley Albert Y Lo Slava Sazonov Henry Te-icher and Jon Wellner Unfortunately the cuts op-erated through the years by the Italian governmentshave made it more difficult to sustain the organiza-tion of such praiseworthy initiativesA You then also started to collaborate with Sazo-

nov In fact one of your articles we really enjoyed


Fig 1 Patrick Billingsley with Eugenio in Cortona Sum-mer School 1989

reading for the wealth of results and techniques itoffers is Regazzini and Sazonov (2000) How did youconvince him to do research on Bayesian statisticsE Slava was a wonderful person I really miss He

was a loyal friend to me and an extraordinary math-ematician I first met him at a conference at thebeginning of the 1990s and then invited him to de-liver a course jointly with Albert Lo at a summerschool organized by Bocconi University in 1992 Hethen taught also in the 1993 and 1994 editions Westarted collaborating in 1996 while he was teachinga course on ldquoProbability Theory in Hilbert spacesrdquoat the Italian National Research Council in MilanoOur first joint work concerned central limit theo-rems for partially exchangeable arrays of random el-ements taking values in a Hilbert space At the timeI was also preparing my lectures for a PhD courseon Bayesian nonparametrics to be taught in Romaand I was dealing with the problem of estimatinga statistical model by means of a mixture of Dirich-let processes Such a problem was suggested by Di-aconis and Ylvisaker (1985) with Slava we showedthat it is possible to construct a mixture of laws ofDirichlet processes that approximates the distribu-tion of any random probability measure with re-

spect to the topology of weak convergence And wehave been able to obtain under suitable assump-tions the corresponding approximation bounds forthe posterior measures These results were presentedat the 1st Workshop on Bayesian nonparametricsthat took place in Belgirate (Italy) in 1997 Whilewe were working on this paper my mother becameseriously ill and Slava has been very important insupporting me in such a difficult periodI You have always had good relationships with

probabilists and statisticians from Russia For exam-ple Ildar Ibragimov is another good friend of yourswho has been several times in Pavia contributing tothe PhD program I had the pleasure to attend hislectures and really enjoyed themE It was actually Slava who suggested I contact

Ildar Ibragimov In fact I had asked Slava indica-tions for possible instructors for PhD courses AndSlava mentioned about Ildar and told me that inaddition to being a great scientist he was an excel-lent teacher Of course I knew Ildar by fame andI feared he would have not accepted my invitationbut he did I got the chance to meet in person notonly a brilliant mathematician but also a wonderfulperson His courses in Pavia were greatly appreci-ated and I liked the fact that he and also Slavawas trying to adapt his lectures to the studentsrsquobackground Our PhD classes are quite compositewith most students having either mathematics oreconomics degrees The former typically have goodbackgrounds in pure maths but not in statistics andprobability whereas for the latter it is the oppositeI remember Ildar asking me ldquoHow come I have stu-dents in my class who know about RadonndashNikodymderivatives of stochastic processes but struggle withFourier transform coefficientsrdquo

6 RESEARCH

61 Bayesian Nonparametrics

A Your papers with Cifarelli on functionals ofthe Dirichlet process (Cifarelli and Regazzini 19791990) are probably your most well-known contribu-tions to Bayesian nonparametrics And it is amazinghow many connections your results have with a va-riety of research areas such as combinatorics math-ematical physics theory of stochastic processes themoments problem and so on Were you aware oftheseE As I said our original problem was merely

of a statistical nature From an analytical point of


Fig 2 Patrick Billingsley standing in the far right and Eugenio with some students in Cortona Summer School 1989

view the task we were facing was very challengingbut we were not aware of the connections with seem-ingly unrelated areas of mathematics We learnedabout some of these relations thanks to the paperby Persi Diaconis and Johannes Kemperman thatwas presented at the Valencia meeting in 1994 seeDiaconis and Kemperman (1996) In addition to em-bedding the whole problem in a wider mathemati-cal context it is also very well written and sketchesa few open problems I strongly recommend readingit It is also thanks to this very same paper that mywork with Cifarelli gained some popularityI The basic trick you resorted to was the inver-

sion of a CauchyndashStieltjes transform for the meanof the Dirichlet process How did you arrive to thisintuitionE The procedure actually relied on the deter-

mination of recursive relations for the moments ofthe linear functional Such a strategy was inspiredby the work of M Kac who used it to obtain thewell-known FeynmanndashKac formula see for exam-ple Kac (1949) This closeness is further revealed bythe adoption in our paper Cifarelli and Regazzini(1979) of the same symbols used by Kac Cifarellihad successfully used it to establish a closed formexpression for the probability distribution of the in-tegral of the absolute value of the Brownian bridgein Cifarelli (1975) These recursive relations we ob-tained allowed us to determine the Laplace trans-form whose iteration yields the CauchyndashStieltjestransform We then resorted to the inversion formu-

lae of the CauchyndashStieltjes transform to deduce anexact form for the probability distribution of a lin-ear functional of the Dirichlet process Most of theseideas were already contained in Cifarelli and Regazz-ini (1979) In Cifarelli and Regazzini (1990) we basi-cally completed that paper and provided some fur-ther insightA More recently you developed an alternative

method based on an inversion formula for the char-acteristic functionE The approach you are referring to was inspired

by the representation of the Dirichlet process as thenormalization of a gamma process that was firstpointed out by Ferguson himself in his 1973 paperThis representation combined with a suitable inver-sion formula led to new forms for the probability dis-tribution of the mean of a Dirichlet process whichare recorded in a paper with Alessandra Guglielmiand Giulia Di Nunno I have then extended withthe two of you the approach to deal with means ofrandom probability measures induced by the nor-malization of a generic process with independent in-crementsI At the moment Bayesian nonparametric regres-

sion is a hot topic In this respect a paper of Ci-farelli and yourself has been recently ldquorediscoveredrdquo(Cifarelli and Regazzini 1978) Can you talk to usabout its origin and contentsE The original goal of our research was to de-

termine a probability distribution for partially ex-changeable arrays of random elements In particular


Fig 3 Workshop on ldquoRecent developments in exchangeabilityrdquo Cortona October 1991 Among others Luigi Accardi Do-nato M Cifarelli Guido Consonni Persi Diaconis Joe Eaton Colin Mallows Jan von Plato Maurizio Pratelli WolfgangRunggaldier Marco Scarsini Brian Skyrms Fabio Spizzichino Piero Veronese Wolfgang Woess and Eugenio

Fig 4 From the left Alan Agresti Eugenio and SlavaSazonov in Livigno Summer School July 1993

we were looking for a solution that could be treatedanalytically while avoiding the independence assump-tion among rows These were the two reasons whichled us to the idea of resorting to the mixture of prod-ucts of Dirichlet processes We have been able to de-termine the associated system of predictive laws andthe distribution of vectors of functionals In a para-metric setting partial exchangeability had been in-corporated in a paper by Lindley and Smith (Lind-ley and Smith 1972) I then used our model to studycredibility formulae with collateral data Cifarelli

had also developed the model for applications toANOVA and linear models the latter in collabora-tion with Marco Scarsini and Pietro Muliere We didnot even submit the paper to a journal since asI said at the time a technical report or a journalpublication counted the same for us Nowadays I amreally pleased to see the recent explosion of propo-sals on dependent nonparametric models somehowin the spirit of our 1978 paper developed by S Mac-Eachern P Muller D Dunson and many othersI In our opinion the work of two probabilists

John F C Kingman and Jim Pitman has to belisted among the main and most far reaching con-tributions to Bayesian nonparametrics even if notdirectly focused on it Do you share this viewE I am strongly in favor of a Bayesian approach

that solely relies on the specification of distributionsfor observable random elements Therefore in gen-eral I like all those contributions and tools that aimat providing systems of predictive distributions re-lated to modeling and applications These do not re-sort to conditional distributions given parameters(either finite or infinite-dimensional) that in someapplications would be devoid of any empirical mean-ing And the works by Kingman and Pitman al-though originated in different research areas havean important impact on Bayesian statistics Even


Fig 5 Slava Sazonov and Patrick Billingsley (on the left) and Eugenio and Andrew Rukhin (on the right) with some studentsin Rhemes-Notre-Dame Summer School July 1994

though I read their papers only recently I have ap-preciated them very much since they open up thepossibility of implementing the Bayesian paradigmin the direction I lean toward

62 Exchangeability

A The contributions of Kingman and Pitman youjust mentioned are closely related to exchangeabil-ity a topic you extensively worked on both froma statistical and probabilistic point of viewE My interest in exchangeability was stimulated

by reading de Finettirsquos papers The first place whereI came across the statement of de Finettirsquos represen-

Fig 6 From the left Giorgio DallrsquoAglio Henry Teicher andEugenio in Perugia Summer School August 1995

tation theorem was the monograph by Loeve ButI could not understand its statistical implicationsI could appreciate its relevance for inductive rea-soning only through a careful study of de Finetti

Fig 7 Eugenio in Belgirate 1st Bayesian NonparametricsWorkshop June 1997


Fig 8 From the left Donato M Cifarelli Persi Diaconisand Eugenio at Stanford University July 2002

(1930a 1937a) in my opinion these papers reallystand out in terms of conceptual and mathemati-cal rigor and effectiveness in highlighting the role ofexchangeability for induction and remain unbeatento date Of course the modern uses of exchangeabil-ity and the key role it plays in modeling a varietyof phenomena are probably beyond what de Finetticould have expectedI You have also been working on characterization

theorems in this contextE You are probably referring to results I have ob-

tained with Sandra Fortini and Lucia Ladelli and

that characterize systems of predictive distributionsassociated with exchangeable sequences of randomelements I have also noted that these kinds of re-sults have recently attracted more and more interestin Bayesian nonparametrics practice Another in-teresting characterization was obtained in a paperI coauthored with Giovanni Petris where we dealtwith exchangeability in the presence of finitely ad-ditive probabilities we stated and proved a weakversion of the representation theorem that reducesto the celebrated de Finetti theorem (strong version)if one specializes to the case of σ-additive probabil-ities In this situation we were also able to use therepresentation theorem to show existence of a ran-dom probability measure defined by means of a sys-tem of finite-dimensional distributions agreeing withFergusonrsquos frameworkI You have also provided nice contributions to the

investigation of properties of partially exchangeabil-ityE Indeed I have been and I still am interested

in forms of dependence more general than exchange-ability as witnessed by some contributions I havealready mentioned before such as the paper on mix-tures of products of Dirichlet processes or the for-mulation of a central limit theorem for partially ex-changeable arrays Besides these I wish to men-tion a nice characterization of partially exchange-

Fig 9 Jon Wellner and Eugenio with some students in Cortona Summer School August 2004


Fig 10 Eugenio at a conference on ldquoNonndashlinear PDEs homogeneization and kinetic equationsrdquo Wien June 2006 In thepicture among others Peter Markowich Pierre Degoud Eric Carlen Maria C Carvalho Ester Gabetta Giuseppe ToscaniCristian Ringhofer Anton Arnold and George Zubelli

able arrays that has been established in a paperI wrote with Fortini Ladelli and Petris Indeed weproved a conjecture formulated in de Finetti (1959)according to which a suitable random matrix re-lated to the transitions of a recurrent process ispartially exchangeable if and only if the law of theprocess can be represented as a mixture of laws ofMarkov chains Moreover we have been able to showthat de Finettirsquos definition of partial exchangeabil-ity is equivalent to the one provided by Diaconis andFreedman in a couple of papers they wrote in 1980

63 Subjective Probability

I In some of your work you have also providedsome insight into an approach to Bayesian statisti-cal inference based on finitely additive conditionalprobabilitiesE I started getting involved into research on fini-

tely additive conditional probabilities after readingsome papers by R Scozzafava in the first half ofthe 1980s In fact I grew convinced that count-able additivity was not justifiablemdashas a necessaryconditionmdashunlike finite additivity which is neces-

sary for the validity of de Finettirsquos coherence prin-ciple Therefore finitely additive probabilities haveto be considered as admissible and I became inter-ested in revisiting known results in probability asparticular cases of the finitely additive frameworkIn particular I found the interpretation of the def-inition of conditional probability as given by Kol-mogorov unsatisfactory Conditioning is based onclasses of events that partition the whole samplespace and that become finer and finer conditionalprobability is then obtained through a limiting pro-cess in terms of a RadonndashNikodym derivative anddepends on the class of events one conditions onIn de Finettirsquos approach a conditional probabilitygiven an event is defined through a natural and un-avoidable strengthening of the coherence principleDe Finetti himself had hinted at such a possibilitywithout developing his idea in general mathemati-cal terms I tried to make this more explicit in somepapers I wrote during the 1980s in Regazzini (19851987) These topics have been object of further in-vestigation by my friends P Berti and P Rigo Animportant point is that many situations that ap-


Fig 11 IMS President Jim Pitman with Eugenio at theIMS Fellows Ceremony 70th IMS Annual Meeting Salt LakeCity July 2007

pear as paradoxical if one refers to Kolmogorovrsquosconditional probabilities can be justified within thefinitely additive frameworkA Can you provide us with an exampleE The most well-known is probably Borelrsquos para-

dox Indeed if a uniform distribution on the sur-face of a sphere is defined with respect to a specificchoice of geographic coordinates (namely latitudeand longitude) one would expect that the condi-tional distribution for latitude given a fixed longi-tude is uniform However this does not happen inKolmogorovrsquos framework In de Finettirsquos approachinstead one can adopt the more intuitive probabil-ity assessment even if it would be nondisintegrableThe reason for such a behavior can be traced backto the specific notion of conditional probability ac-cording to Kolmogorovrsquos approach since it does notadmit the evaluation of the probability of an ldquoiso-lated eventrdquo with probability zero On the contraryde Finettirsquos setup is open to different solutions in-deed disintegrability turns out to be not necessaryfor coherence

I Another amusing aspect of finitely additive con-ditional probabilities emerges from your work onwell-calibration of systems of predictive distributionsE Loosely speaking well-calibration corresponds

to situations where the distance between weightedaverages of forecast probabilites and empirical ob-servations converges to zero as the number of obser-vations and forecasts increases Kolmogorovrsquos the-ory always yields well-calibrated predictions or fore-casts This corresponds to a somehow unrealistic sit-uation in practice since one would also expect casesof not well-calibration With P Berti and P Rigo wewere interested in checking whether the same wastrue within de Finettirsquos theory as well Our curios-ity to this problem was stimulated by a paper ofPhil Dawid (Dawid 1982) The answer we got wasnaturally affirmative for strategic conditional prob-abilities The term strategic was coined by Dubinsand Savage in their well-known monograph wherethey resorted to de Finettirsquos theory to solve quitecomplicated measurability problems Strategic con-ditional probabilities do indeed preserve in a finitelyadditive setting the disintegrability property thatcharacterizes Kolmogorovrsquos definition As for well-calibration we were able to show that beyond strate-gic evaluations there exist not well-calibrated co-herent Bayesian predictors with positive probabil-ityAMany critics of de Finettirsquos subjectivistic stand-

point in probability theory use as an argument forsupporting their position de Finettirsquos sentence ldquoprob-ability does not existrdquo What can be replied to suchobjectionsE First of all one should consider the provoca-

tive nature of de Finettirsquos sentence Moreover itsmeaning should not be decontextualized Accordingto de Finetti if one wants to give probability anobjective meaning one should prove its existenceIn other words there should be an existence the-orem a clear proof of the existence of an objecttermed ldquoprobabilityrdquo For example the interpreta-tion of probability as a limiting frequency cannot beconsidered as a proof even if just empirical of its ex-istence Hence he used the expression ldquoprobabilitydoes not existrdquo just to make the point that proba-bility has simply a subjective meaning It is also tobe said that most of the criticism raised against sub-jectivism basically refers to the contents of his two-volume monograph de Finetti (1970) which is ac-cording to its subtitle ldquoa critical introductory treat-mentrdquo In my opinion de Finettirsquos position can bebetter discussed by relying on his early works which


are more concise go straight to the point and dis-play more mathematical and formal detailsI A noteworthy scholar who contributed to the

theory of finitely additive probabilities was LesterDubins You were also a good friend of his and hadthe chance to host him in MilanoE Dubins had been in Italy several times and he

delivered courses at summer schools He was veryfond of Italy and in the second half of the 1980sI invited him once to stay for a month in MilanoWe had discussions on various research topics Hewas the source of many ideas that I later devel-oped in my research In those years I was mainlyworking on technical aspects of nonparametric in-ference whereas he could provide me with many in-sights into theoretical issues related to finite addi-tivity that turned out to be of great importance tome

64 Probabilistic Methods for Mathematical

Physics

A and I You have lately become interested insome problems in mathematical physics How didit happenE Moving to Pavia in 1998 I joined a Mathe-

matics Department with a few internationally well-known mathematical physics scholars I started in-teracting with them and at some point a colleague ofmine Ester Gabetta showed me some papers wherethe Central Limit Theorem was used to describe theconvergence to equilibrium of the solution of certainkinetic equations In particular I read two papersMcKean (1966 1967) that spurred my enthusiasmfor the topic I tried to understand and extend theconnections with probability and could count on thecollaboration of my colleagues to help me under-stand the problem from the perspective of physicsFurthermore the encouragement from Eric Carlenand Maria Carvalho has been important for pursu-ing my research in this direction In fact they likedour first results and suggested us to publish them(see Gabetta and Regazzini 2006) there we ob-tained some identities that came in handy for laterdevelopments of the work in this areaA and I Was this line of research as rewarding as

others you have pursued in your careerE I would say I am happy about what I have

achieved so far with my coauthors Starting from theKac model which is generally considered as a toymodel we obtained some interesting results concern-ing the characterization of the initial data in order togain convergence to equilibrium We have also con-

sidered situations where the energy interpreted asthe variance of the initial datum is infinite and weperformed an analysis of the speed of convergenceIn these studies I have also collaborated with LuciaLadelli and Federico Bassetti Later I have super-vised the thesis of Emanuele Dolera a PhD studentin Pavia This work has required a strong effort thatwas rewarded by the achievement of a noteworthyresult proving the validity of a conjecture formulatedin the 1966 McKean paper In the last 40 years manyscholars have worked hard with the aim of provingit

7 THOUGHTS ON FOUNDATIONAL ISSUES

AND RESEARCH IN STATISTICS

I In some of the previous questions we have lin-gered on the subjectivistic interpretation of proba-bility What is the most relevant impact this has onstatisticsE A crucial point to understand is whether it is

worth preserving an axiomatization based on count-able additivity Of course I think it does not gener-ally have a statistical justification that makes its usenecessary If finitely additive probabilities are alsoadmissible then a considerable number of resultsin the literature should be revisited I have alreadymentioned that one should reconsider the definitionof conditional expectation Moreover a number oflimiting theorems should be reformulated in orderto account for this more general framework Theseissues are also of great relevance in statistics regard-less of the approach either frequentist or Bayesianone adoptsA Does this lead among others to a rethinking

of Bayesian proceduresE Indeed Bayesian procedures are typically im-

plemented by assuming complete additivity and thisleads to assume some of its implications as neces-sary Let us consider as an example the Dirichletprocess A well-known result is that the Dirichletprocess selects almost surely discrete probabilitymeasures However such a property holds true forthe countably additive extension of the collectionsof finite-dimensional probability distributions of theprocess There are other non σ-additive extensionsfor which the Dirichlet process selects nondiscretedistributions with positive probability This pointsto the fact that in statistical practice one shouldavoid assessing a probability for objects devoid ofempirical evidence For example take the proposi-tion stating that de Finettirsquos measure is the law of


the (almost sure) weak limit of the empirical distri-bution thus it depends on infinitely many observa-tions and concerns ldquotranscendentrdquomdashin de Finettirsquoswordsmdashconditions not directly verifiable The con-clusion of such a proposition could be obviously falsewith non σ-additive extensions On the other handthe fact that de Finettirsquos measure is the weak limitof the low of the empirical distribution as the sam-ple size increases is in any case true in my opinionthis suffices with respect to sound statistical goalsI think this is an important foundational aspectwhich is often neglected and should be further in-vestigatedI Are you saying that one should have clear in

mind the different levels at which mathematics andstatistical applications operateE More or less that is what I mean Indeed it

is true that mathematics makes parameters inter-pretable as limits of (or of functionals of) empiricalprocesses but it does not automatically grant thatinference on them are legitimateA Does this position contrast with the usual way

of presenting a Bayesian model as the combinationof a likelihood and a priorE Let me start by making an important point that

reflects my view on statistics if inference is seen asa decision problem to be solved under uncertaintyand if one agrees that probability is a tool to resortto then there is no other choice but the Bayesianapproach Nonetheless I agree with what Diaconisand Ylvisaker say at the beginning of their paperDiaconis and Ylvisaker (1985) Bayesian statisticscannot be reduced to the elicitation of a prior andthe automatic application of Bayesrsquo theorem HenceI would give an affirmative answer to your question ifone conditions on unobservable quantities But thisis not limiting the scope of Bayesian inference atall Indeed one can think of inferential proceduresthat can still be implemented in this more generalframework even when unobservable parameters areinvolved The previously mentioned ldquoweakrdquo inter-pretation of the de Finetti measure says that a priordistribution can be viewed in any case and withno distinction between observable and not observ-able parameters as an approximation of the law ofa frequency distribution or of some functional of itMoreover prediction can be carried out without re-lying on the BayesndashLaplace paradigm it is enoughto specify the system of predictive distributions con-nected to the exchangeable sequence And I have ap-preciated very much the work by J Pitman which

in the spirit of de Finettirsquos stance relies on the pro-posal of systems of predictive distributions that arethen proved to be associated to an exchangeable se-quenceA I also guess that a subjectivist would not agree

on the notion of posterior consistency as a frequen-tist validation criterion of Bayesian nonparametricmethodsE I have to admit that besides the Bayesian con-

text I am skeptical on the use of consistency ina frequentist setting as well On the one hand theselimiting results are very neat and beautiful froma mathematical point of view But on the otherthey lack a sensible statistical interpretation Thisis very well discussed in de Finetti (1970) Volume 2in the section devoted to the laws of large num-bers where he motivates why results such as con-sistency do not represent justifications of statisticalprocedures under the assumption of stochastic inde-pendence The same can be said for the GlivenkondashCantelli theorem Of course my position encom-passes commonly used frequentist validation crite-ria adopted in a Bayesian framework A differentrole must be attributed to approximation resultslike Central Limit Theorems or also the ldquoweakrdquointerpretation of the de Finetti measure for whichthese concerns do not applyI So what are the kind of asymptotic problems

that you think are interesting for BayesiansE The kind of results I like are those in the spirit

of Blackwell and Dubins (1962) where they inves-tigate the phenomenon of the merging of opinionsThis is a very nice finding both from a mathemati-cal and from a statistical point of view On the onehand it is a general result valid even beyond ex-changeability On the other hand it has a nice sta-tistical interpretation for Bayesians since it hints atthe predominance of empirical findings over differ-ent subjective prior opinions as the sample size or inother terms the amount of information increasesOther results of great interest are those that cur-rently are designated as Bernsteinndashvon Mises typetheorems for the posterior distribution It is worthnoticing that among the first contributions to thistopic there is also an important paper Romanovsky(1931) published on GIIAA From what you said up to now on the inter-

play between Bayesian inference and de Finettirsquos in-terpretation of probability a valuable research topicwould focus on the analysis of the asymptotic be-havior of the predictive distributions


E You are right In my opinion an importantissue to address is the analysis of the distance be-tween the predictive and the empirical distributionsInstead of looking at the limiting behavior it wouldbe more interesting to analyze how such a discrep-ancy changes for any sample size n and a fortiorias n increases Since the predictive can also be ob-tained as a functional of the posterior distributionone can also gain some insight if one relies on con-vergence theorems which say that the posterior con-verges in some sense to a distribution concentratedon the limit of the empirical process In this respectBernsteinndashvon Mises type results are usefulI You have had a large number of students and

by now also descendants working in many differentuniversities in Italy and abroad In your opinionwhat is the background a statistics student needs toperform well in nowadays research and what are thetopics you would suggest to pursueE As I said earlier I see statistics as inductive

reasoning under the supervision of probability the-ory Therefore it is natural that I firmly believethat statisticians should have a solid background inprobability the more the better However a statis-tician must also be able to think through the logi-cal and philosophical aspects of what shehe is do-ing This concerns modeling the understanding ofpractical implications yielded by the mathematicalformulation that is used and the interpretation ofresults Mathematical skills are not enough logicaland conceptual rigor being a necessary complementOne needs to be able to handle statistics since it isa powerful instrument which allows one to makesubstantial steps forward compared to traditionaldeterministic procedures Statistics can get you closeto the best solutions avoiding overwhelming tech-nical and mathematical difficulties that often arisewithin deductive deterministic reasoning The latterapproach lacks the flexibility of a learning mecha-nism whereas in the probabilistic framework every-thing is kept under control you have a law whichgoverns everything and unless you change the learn-ing mechanism it allows one to learn from experi-ence in a way that is transparent and controlled byBayesrsquo theoremA In modern science the specialization of re-

searchers is constantly increasing Even probabilityand statistics which have grown in close relation-ship to each other seem to be drifting apartE You can observe the fragmentation of fields

all over the place This phenomenon also originatesfrom an excessive specialization that characterizes

most undergraduate studies The situation was inthe past quite different and there were many schol-ars with a wide spectrum knowledge and diversi-fied cultural and scientific interests De Finetti andGini are excellent examples in this respect Thatsaid fragmentation in research is unavoidable andit would be unrealistic to try reversing it It is justa pity to see that it tends to create duplications andrepetitions whereas a more cohesive scientific com-munity could produce better results in a collectiveeffort In statistics Bayesian statisticians have keptto themselves for some time in reaction to the thenmainstream statistics which was certainly not in fa-vor of Bayesian methods Now with Bayes statisticswell-established I note that younger generations aremore open to interactions with non-Bayesian whichin my opinion is certainly beneficial A different issueis the specialization in education which should becontrasted to some extent because it precludes pos-sible paths to future researchers As I have alreadysaid every statistician should have a solid back-ground in probability and every probabilist shouldknow the basics of statistics which is a noble andfascinating at least to me field of application ofprobabilityI How should in your opinion a good statistics

paper be structuredE Well first of all the definition ldquogoodrdquo is to be

considered with reference to the historical periodUntil some years ago theoretical papers were veryappreciated whereas nowadays applied work playsan increasing role thanks to the computational toolsHowever I think that in general different formsof motivation are equally valid an enrichment ofthe available tools an improvement over other ex-isting contributions or a useful application are allfine However in all cases it is crucial that the paperis logically sound and coherent with its motivationThis is essential since we write for the scientific com-munity and not for the general public which is an-other job For instance I do not like methodologicalpapers to which an illustration has been evidentlyadded only for editorial needs A methodologicalcontribution can stand on its own if its motivation issound While I was young I experienced some of thelast manifestations of Ginirsquos school which as a ruleof thumb required publications to include data a ta-ble and a plot To me this does not make senseI also do not like applied papers in which one setsforth a model analyzes a couple of data sets andconcludes that the model works well Any so-calledempirical validation does not show anything and is


not enough to assess the suitability of a model In-deed there should also be a sensible methodologi-cal motivation in the sense that one should explainwhich features of a certain model make it more ap-propriate for the problem at handA and I Moving away from statistics and proba-

bility we already mentioned your passion for musicHow did you get fond of music and what else areyou interested in when you do not do researchE Being born in Cremona my passion for music

is quite natural it is 20 km away from the placesGiuseppe Verdi grew up in melodrama is popularand there is a great tradition It is also the home-town of Claudio Monteverdi and of Amilcare Pon-chielli two famous composers Last but not leastit is the town of lute makers the most renownedbeing Antonio Stradivari Even the general publicknows opera very well Then starting from operawhen I was young my interest extended to sym-phonic music I have also been fond of visual artssince I was a kid I loved paintings architecture andsculpture since I related them to Italian history I re-member having a great teacher at school who usedto emphasize links between history arts and litera-ture A peculiar feature of Italy is that if you areinterested in any historical aspect you necessarilyend up considering also painting sculpture and ar-chitecture since they are all intimately connectedWe obviously benefited from Christian culture thatplayed a fundamental role after the fall of the Ro-man Empire in preserving the wonders inheritedfrom classical Greek and Roman traditions and inpromoting arts in forms we can today admire whilevisiting churches historical buildings squares andmuseums Moreover during the Renaissance therewere a large number of small states and many ofthem had patrons who liked and could afford beingsurrounded by artists Hence many towns developedtheir peculiar artistic heritageA and I Eugenio thanks a lot for patiently an-

swering all our questionsE Thanks to you for listening to all this

ACKNOWLEDGMENTS

Many thanks to Elena Di Biase for a careful read-ing of the manuscript and several useful suggestionswhich improved the presentation This work is par-tially supported by MIUR Grant 2008MK3AFZ

REFERENCES

Allen R G D (1956) Mathematical Economics MacMil-lan London

Blackwell D and Dubins L (1962) Merging of opinionswith increasing information Ann Math Statist 33 882ndash886 MR0149577

Castelnuovo G (1919) Calcolo delle Probabilita 1st edSoc Ed Dante Alighieri Milano

Cifarelli D M (1975) Contributi intorno ad un testper lrsquoomogeneita tra due campioni Giorn Econom AnnEconom (NS) 34 233ndash249 MR0433704

Cifarelli D M and Regazzini E (1974) Ancorasullrsquoindice di cograduazione del Gini Technical ReportSerie III No 5 Istituto di Matematica FinanziariadellrsquoUniversita Torino

Cifarelli D M and Regazzini E (1978) Problemistatistici non parametrici in condizioni di scambiabibilitaparziale Impiego di medie associative Quaderni Istitutodi Matematica Finanziaria DellrsquoUniversita di Torino SerieIII N 12 1ndash36 English translation available at http

wwwunibocconiitwpsallegatiCTPCR-Scamb-parz[1]

20080528135739pdfCifarelli D M and Regazzini E (1979) A general ap-

proach to Bayesian analysis of nonparametric problemsThe associative mean values within the framework of theDirichlet process II Riv Mat Sci Econom Social 2 95ndash111 MR0574040

Cifarelli D M and Regazzini E (1990) Distributionfunctions of means of a Dirichlet process Ann Statist 18429ndash442 [Correction in Ann Statist (1994) 22 1633ndash1634]MR1041402

DallrsquoAglio G (1956) Sugli estremi dei momenti delle fun-zioni di ripartizione doppia Ann Scuola Norm Sup Pisa(3) 10 35ndash74 MR0081577

Dawid A P (1982) The well-calibrated Bayesian J AmerStatist Assoc 77 605ndash613 MR0675887

de Finetti B (1926) Considerazioni matematichesullrsquoereditarieta mendeliana Metron 6 3ndash41

de Finetti B (1929) Sulle funzioni a incremento aleatorioAtti Reale Accademia Nazionale dei Lincei Serie VI Rend10 163ndash168

de Finetti B (1930a) Funzioni caratteristiche di legge is-tantanea Atti Reale Accademia Nazionale dei Lincei SerieVI Rend 12 278ndash282

de Finetti B (1930b) Funzione caratteristica di unfenomeno aleatorio Atti Reale Accademia Nazionale deiLincei Mem 4 86ndash133

de Finetti B (1931) Sul significato soggettivo della prob-abilita Fund Math 17 298ndash329 [English translation in deFinetti (1992) 291ndash322]

de Finetti B (1933) Sullrsquoapprosimazione empirica di unalegge di probabilita Giornale dellrsquoIstituto Italiano degli At-tuari 4 415ndash420

de Finetti B (1935) Il problema della perequazione InAtti della Societa Italiana per Il Progresso delle Scienze 2

227ndash228de Finetti B (1937a) La prevision Ses lois logiques

ses sources subjectives Ann Inst H Poincare 7 1ndash68MR1508036

de Finetti B (1937b) A proposito di ldquocorrelazionerdquo Sup-plemento Statistico Ai Nuovi Problemi di Politica Storia edEconomia 3 41ndash57 English translation available at httpwwwbrunodefinettiitOpereAboutCorrelationspdf


de Finetti B (1939) La teoria del rischio e il problemadella ldquorovina dei giocatorirdquo Giornale dellrsquoIstituto Italianodegli Attuari 10 41ndash51

de Finetti B (1940) Il problema dei ldquoPienirdquo GiornaledellrsquoIstituto Italiano degli Attuari 11 1ndash88 MR0019248

de Finetti B (1959) La probabilita e la statistica neirapporti con lrsquoinduzione secondo i diversi punti di vistaIn Centro Internazionale Matematico Estivo (CIME) In-duzione e Statistica 1ndash115 Cremonese Roma [Englishtranslation in de Finetti (1972) 147ndash227]

de Finetti B (1970) Teoria delle Probabilita Sintesi Intro-duttiva Con Appendice Critica Vol 1 2 Einaudi Torino[English translation De Finetti B (1974 and 1975) The-ory of Probability A Critical Introductory Treatment Vol1 2 Wiley London]

de Finetti B (1972) Probability Induction and StatisticsThe Art of Guessing Wiley London MR0440638

de Finetti B (1992) Induction and Probability A supple-ment to the journal Statistica 52 (P Monari and D Cocchieds) CLUEB Bologna

de Finetti B (2006) Opere Scelte Voll I e II UnioneMatematica Italiana

de Finetti B and Savage L J (1962) Sul modo discegliere le probabilita iniziali In Biblioteca del MetronSerie C Note e Commenti 82ndash15 Istituto di StatisticaDellrsquoUniversita Roma [English summary in de Finetti(1972) 143ndash146]

DeGroot M H (1970) Optimal Statistical DecisionsMcGraw-Hill New York MR0356303

Diaconis P and Kemperman J (1996) Some newtools for Dirichlet priors In Bayesian Statistics 5

(J M Bernardo J O Berger A P Dawid and

A F M Smith eds) 97ndash106 Oxford Univ Press NewYork MR1425401

Diaconis P and Ylvisaker D (1985) Quantifying prioropinion In Bayesian Statistics 2 (A K BernardoM H DeGroot D V Lindley and A F M Smitheds) 133ndash156 North-Holland Amsterdam MR0862488

Ferguson T S (1967) Mathematical Statistics A DecisionTheoretic Approach Probability and Mathematical Statis-tics Vol 1 Academic Press New York MR0215390

Ferguson T S (1973) A Bayesian analysis of some non-parametric problems Ann Statist 1 209ndash230 MR0350949

Gabetta E and Regazzini E (2006) Some new resultsfor McKeanrsquos graphs with applications to Kacrsquos equationJ Stat Phys 125 947ndash974 MR2283786

Gini C (1939) I pericoli della Statistica SupplementoStatistico ai Nuovi problemi di Politica Storia ed Econo-mia 5 1ndash44

Kac M (1949) On distributions of certain Wiener function-als Trans Amer Math Soc 65 1ndash13 MR0027960

Khintchine A Y (1937) A new derivation of a formula

by P Levy Bulletin of the Moscow State University 1 1ndash5 (In Russian) [English translation in the Appendix of

Mainardi F and Rogosin S (2006) An origin of infinitelydivisible distribution From de Finettirsquos problem to Levyndash

Khintchine formula Math Meth Econ Fin 1 37ndash56]Kolmogorov A N (1932) Sulla forma generale di un

processo stocastico omogeneo Un problema di Bruno de

Finetti Atti Reale Accademia Nazionale dei Lincei Ser

VI Rend 15 805ndash808Lindley D V (1965) Introduction to Probability and Statis-

tics from a Bayesian Viewpoint Part I Probability Cam-bridge Univ Press New York MR0168083

Lindley D V and Smith A F M (1972) Bayes estimatesfor the linear model J Roy Statist Soc Ser B 34 1ndash41

MR0415861Lo A Y (1991) A characterization of the Dirichlet process

Statist Probab Lett 12 185ndash187 MR1130354

McKean H P Jr (1966) Speed of approach to equilibriumfor Kacrsquos caricature of a Maxwellian gas Arch Rational

Mech Anal 21 343ndash367 MR0214112McKean H P Jr (1967) An exponential formula for solv-

ing Boltmannrsquos equation for a Maxwellian gas J Combi-natorial Theory 2 358ndash382 MR0224348

Pizzetti P (1963) I Fondamenti Matematici per la Criticadei Risultati Sperimentali (reprint of 1892 edition) Cap-

pelli BolognaRaiffa H and Schlaifer R (1968) Applied Statistical

Decision Theory MIT Press Cambridge MA MR0226757Regazzini E and Sazonov V V (2000) Approximation

of distributions of random probabilities by mixtures of

Dirichlet distributions with applications to nonparamet-ric Bayesian statistical inferences Theory Probab Appl 45

93ndash110Regazzini E (1978) Intorno ad alcune questioni relative

alla definizione del premio secondo la teoria della credi-bilita Giornale dellrsquoIstituto Italiano degli Attuari 41 77ndash89

Regazzini E (1985) Finitely additive conditional probabil-ities Rend Sem Mat Fis Milano 55 69ndash89 MR0933711

Regazzini E (1987) de Finettirsquos coherence and statisticalinference Ann Statist 15 845ndash864 MR0888444

Regazzini E (2005) Probability and statistics in Italyduring the First World War I Cantelli and the laws of

large numbers J Electron Hist Probab Stat 1 1ndash12MR2208347

Romanovsky V (1931) Sulle probabilita ldquoa posteriorirdquo

Giornale dellrsquoIstituto Italiano degli Attuari 4 493ndash511Savage L J (1954) The Foundations of Statistics Wiley

New York MR0063582

1 Probability and Statistics at Bocconi University

2 From Torino to Bologna Milano and Pavia

3 Bayesian Inference

4 de Finetti and the Influence of de Finettis Work

5 Probability and Statistics in Italy

6 Research


62 Exchangeability


64 Probabilistic Methods for Mathematical Physics

7 Thoughts on Foundational Issues and Research in Statistics

Acknowledgments

References


pleting high school The choice I made at the ageof 14 was coherent with this scenario By the timeI got my diploma from high school the situationhad improved a little bit and I could afford goingto university This was also due to a financial aidprogram adopted by the farseeing Italian centermdashleft governments of the time for students belongingto economically disadvantaged families The moneyfrom this program (it was around 200 euros peryear in 1965) and the earnings deriving from privatelessons I used to teach to other students allowed meto obtain a degree at Bocconi UniversityIgor Why did you choose Bocconi University and

not another universityE In the 1960s Bocconi was considered a presti-

gious university a degree from it represented the keyfor obtaining a good and rewarding job on the mar-ket These elements exerted an influence on me andmy family However there was also a practical rea-son Milano is just one hour by train from my home-town Cremona and I could go back home everyweekend Yet another reason is the influence of myfriend Lorenzo Peccati a mathematical economistfrom Cremona who was a student at Bocconi whileI was still at high school He was well aware of mybent for mathematics and suggested a few advancedtextbooks where I started reading the mathemati-cal tools used in economic modeling In particularI was excited at reading the Italian translation of themonograph Allen (1956) on mathematical analysisfor economists and this convinced me that Bocconiwould still have allowed me to study MathsI This is a funny coincidence since I was con-

vinced by Lorenzo Peccatirsquos son Giovanni by nowa well-known probabilist to enroll at Bocconi andfor precisely the same reasons Where did your pas-sion for mathematics come fromE I was very lucky at high school because I had

a brilliant maths teacher Sidomo Vailati He hada variety of scientific and cultural interests and alsodid consulting for a few private companies thanksto his unusual at least in that period knowledgeof statistics probability and operations research Hecertainly was a self-taught man in the area of Stochas-tics At that time probability and statistics unlikeanalysis geometry and algebra were not perceivedas relevant topics within mathematics degrees theywere only present in a few optional courses To myknowledge the only exception was the Universityof Roma due to the presence of Bruno de FinettiIn fact this unfortunate situation lasted until the

1970s when the first full professors in probabilityapart from de Finetti were recruited after nationalcompetitions Turning back to Vailati it is worthrecalling that among some courses for high schoolteachers organized by the Italian Ministry of Educa-tion he also took a course in probability which wasdelivered by de Finetti As a consequence at theage of 16 I was introduced to the realm of subjec-tivism and learned the first elements of probabilityand its applications These first years of exposure tode Finettirsquos approach have stimulated an intellectualand scientific interest that has certainly influencedmy later researchA How was the environment at Bocconi Univer-

sity in the years you have been thereE Bocconi had very few professors among its own

faculty and heavily relied upon adjunct faculty hold-ing positions in other universities These few pro-fessors were all influential personalities of the timeplaying significant roles in the Italian social polit-ical and economic life of the 1960s For instanceGiovanni Demaria was a Paretian economist whoacted as economic consultant for the constituent as-sembly that created the Constitution that lies atthe foundation of the modern Italian Republic afterWorld War II There was also a special feeling be-tween Bocconi and Milano a city that had been ableto overcome the disasters of World War II and wasexperiencing dramatic economic growth led by themanufacturing sector Bocconi looked to me andmany others as a vital part of Milano and con-tributed to consolidate this process Then duringthe last couple of years the student protests of 1968started Despite being a private university Bocconiexperienced serious clashes and some of its studentsplayed an active role in the movementI Did you like studying economicsE I was very fond of economics The professors

I was interacting with were quite enthusiastic aboutmy inclination toward developing mathematical toolsuseful for economic modeling There is an episodethat occurred during my third year that I like torecall I was attending a course in Political economywhich included a series of seminars and one of themconcerned the relationship between the Italian Cen-tral Bank and the Department of Treasury whichat that time was the subject of a lively debate Foran economic interpretation of the relationship be-tween the two institutions we were suggested to re-fer to an article by Giorgio La Malfa and FrancoModigliani the latter was later awarded the Nobel

















AND PAVIA


























































































































6 RESEARCH
























62 Exchangeability



































Physics






































ACKNOWLEDGMENTS


REFERENCES





































































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References

















AND PAVIA


























































































































6 RESEARCH
























62 Exchangeability



































Physics






































ACKNOWLEDGMENTS


REFERENCES





































































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References




















































































































6 RESEARCH
























62 Exchangeability



































Physics






































ACKNOWLEDGMENTS


REFERENCES





































































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References





















62 Exchangeability



































Physics






































ACKNOWLEDGMENTS


REFERENCES





































































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References






























Physics






































ACKNOWLEDGMENTS


REFERENCES





































































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References



























ACKNOWLEDGMENTS


REFERENCES





































































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References















































New York MR0063582






6 Research


62 Exchangeability




Acknowledgments

References

a conversation with eugenio regazzini - arxiv · eugenio regazzini was born on august 12, 1946 in...

Documents