958 AD-A246 AFS-89-0261 : THEORY AND APPLICATION OF RANDOM FIELDS ROBERT J. ADLER Faculty of Industrial Engineering and Management Technion - Israel Institute of Technology Technion City, Haifa, ISRAEL Arrrr .1- A ( ' ) Final Technical Report 1 January 1989 - 31 December 1991 DTIC S ELECTE MARO0 5 1992D Approved for public release. . Distribution unlimited. Prepared for: Technion Research & Development Foundation; European Office of Aerospace Research and Development, London, England; Directorate of Mathematical Sciences, AFOSR, Washington. 1 92-05652 ! 8 03 120 mf1h1111 14t1cCC S42 SECURITY CLASSIFICATION OF THIS PAGE RD E O j REPORT DOCUMENTATION PAGE I&. REPORT SECURITY CLASSIFICATION lb. RESTRICTIVE MARKINGS Unclassif ied 2a. SECURITY CLASSIFICATION AUTHORITY 3. DISTRIBUTION IAVAILABILITY OF REPORT 2b. DECLASSIFICATION/DOWNGRADING SCHEDULE Appoved for public release; distribution unlimited 4. PERFORMING ORGANIZATION REPORT NUMBER(S) S. MONITORING ORGANIZATION REPORT NUMBER(S) 6a. NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATION Technion-Israel Inst of Technoogy( European Office of Aerospace f Research and Develomepnt 6c. ADDRESS (City, State, and ZIP Code) 7b. ADDRESS(Cty. State, and ZIP Code) 5WOtLsf Industrial Engineering & Managem-it Box 14 ISRAEL _FPO New York 09510-0200 8a. NAME OF FUNDING/SPONSORING j8b. OFFICE SYMBOL 9. PROCUREMENT INSTRUMENT IDENTIFICATION NUMBER ORGANIZATION Air Force Office (If applicable) of Scientific Research NM AFOSR 89-0261 8. ADDRESS (City, State, and ZIP Code) 10. SOURCE OF FUNDING NUMBERS PROGRAM I PROJECT TASK WORK UNIT Bolling AFB, Washington DC 20332-6448 INO. T NO. NO ACCESSION NO. 2301 D1 ON, 11. TITLE (Include Security Classification) Theory and Application of Random Fields 12. PERSONAL AUTHOR(S) R.J. Adler 13a. TYPE OF REPORT 13b. TIME COVERED 14. DATE OF REPORT (Year, Month, Day) I1S. PAGE COUNT Final IFROM 1,7419 - 3.9,&~ 92-1-9 22 16. SUPPLEMENTARY NOTATION 17. COSATI CODES 18. SUBJECT TERMS (Continue on reverse if necessary and identify by block number) FIELD GROUP SUB-GROUP Random fields. 19. ABSTRACT (Continue on reveise if necessay and identify by block number) This report covers research performed under the three years of the grant. Main progress has been in ihe areas of modelling problems related to measure valued diffusions, in the completion of a monograph on the properties and structure of Gaussian processes on general parameter spaces, and level crossing problems. '0. DISTRIBUTION /AVAILABILITY OF ABSTRACT 121. ABSTRACT SECURITY CLASSIFICATION PUNCLASSIFIED/UNLIMITED OSAME AS RPT. OOTIC USE.S Unclassifed 2A, NAME OfRESPONSIBEINDIVIL 22b. TELEPHONE Incu Area Code) 22c.OFFICE SYMBOL D FORM 1473, 84 MAR 83 APR edition may be used until exhausted. SECURITY CLASSIFICATION OF .HIS PAGE All other editions are obsolete. UNlCLASS IF IED CONTENTS 1. Introduction 1 2. On random fields 2 3. Specific projects: 4 (a) Monograph on Gaussian processes 4 (b) Super processes and density processes 7 (c) Visualisation of super processes 13 (d) Level crossings 15 (e) Miscellaneous 18 4. Publications prepared under the grant 20 5. Conferences attended and professional visits 22 Accesiori For COP NTIS CRA&I 1INPSE CT E DTI iC iA 6 6 U. ".;:oI.ed L By .............. .... ......I.... Availabiiy CodeS, Avail aiid Ior Dist Special 1. INTRODUCTION This report covers work carried out under the support of the AFOSR, under contract 89-0261, during the three year period January 1, 1989 until December 31, 1991. This three year project was a natural continuation and extension of similar work supported under contracts 83-0068, 84-0104, 85-0384 and 87-0298. The research program is cen- tered on the study of various properties of random fields (stochastic processes whose "time" parameter is multi-dimensional) and includes the development of the requisite theoretical foundation to enable the application of these properties to specific modelling problems. The last three years have been particularly successful, having seen the completion of a number of projects that have been going on for a while (in particular, the completion of a monograph on Gaussian processes) and the commencement of some major new directions of research. This makes the (hopefully temporary) lack of further sup- port even more disappointing than it would normally be. As one would expect, there will be a considerable overlap between this final report and the previous two annual reports. i 2. ON RANDOM FIELDS Random fields are stochastic processes, X(t), whose parameter, t,v aries over some general space rather than over the real line, in which case t is usually interpreted as time. The simplest of these occur when the parameter space is some multi-dimensional Euclidean space, and it is these fields that were at the centre of my research for a number of years. Of these, the most basic arise when the parameter space is the two-dimensional plane, so that we are dealing with some kind of random surface. When the parameter space is three-dimensional then we have a field (such as ore concentration in a geological site) that varies over space, while when the dimension increases to four we are generally dealing with space-time problems. In space-time problems it is clear that one of the parameters is qualitatively different from the others, and so there is a natural tendency to denote this explicitly by writing the random field as X(t, x), where t is now one-dimensional time and x an N-dimensional space parameter. Of course, one could also write this as X =(xX) (tx), in which case one could consider X either as a random field on (N + 1)-dimensional space, or, as a process in t, (one-dimensional) taking values X(-) in a space of N- dimensional random fields. In the latter case, since "time" has once again become one- dimensional, one can begin talking about Markov properties, diffusion processes (albeit infinite-dimensional from the point of view of their state space), etc. Both of the above views of random fields will appear in this report, with differing methodologies and types of result appearing in each case. As well as this, random fields over very general parameter spaces that have no Euclidean structure at all will appear, with results of yet another nature. Despite the different types of random fields that will appear in this report, there are two common threads running through it. One is of a theoretical nature, and is a result of my belief that despite the vast differences between many of these random fields, there is much to be gained by placing all of them within as common conceptual and mathematical frameworks as possible, and that this gain shows itself both in terms of developing intuition and in terms of developing streamlined proofs. The second thread is of a more practical nature, and comes from the fact that I believe that problems that have no possible application are unlikely to be even of theoretical interest. Thus, whereas most of the results in this report are not what one would normally call "applied", all 2 the processes that I consider are related to real life modelling problems, and it is this, I believe, that ultimately makes them interesting and, often, difficult. Wherever it is possible to do so succinctly, the connection between theory and model will be described. 3 3 3. SPECIFIC PROJECTS Much of my time over the past three years was dedicated towards the completion of two major projects. The first of these was the completion of a monograph on Gaussian processes described in the proposal. The second was the commencement of a major project in the study of measure and distribution valued stochastic processes, which were described there under the heading "random field valued processes". These two projects are described in some detail in the subsections (b) and (c) following. Over the past twelve months a good deal of effort has also gone into the more classical "level crossing problems", which are of more immediate applicability and closer in spirit to the work done under AFOSR support in previous years. One subsection is devoted to these. A number of other problems that have also been looked are described in the final, "miscellaneous", subsection. Yet others are not mentioned in the report at all, as can be gleaned by noting that the list of publications at the end of the report contains a goodly number of topics not discussed at all. (A) MONOGRAPH ON GAUSSIAN PROCESSES: A project that demanded a considerable amount of my time over the first two years of the grant period was the preparation of a monograph on the sample path continuity, boundedness, and extremal behaviour of Gaussian processes on general parameter spaces. This monograph appeared as Volume 12 of the prestigious Lecture Notes-Monograph Series of the Institute of Mathematical Statistics. The main theme running through this work is that Gaussian processes should be studied in a general, unifying, framework, without particular emphasis being given to the specific geometric structure of their parameter spaces. I presented series of lectures based on these notes in Sweden (February 1988, 12 lectures) Technion (second semester 1988/89, 26 lectures), and gave shorter sets of 2-3 lectures on a number of occasions. There were also groups at a number of institutions, in- cluding the University of California, Berkeley, and at the University of British Columbia, Vancouver, that held workshops based on my manuscript. A good idea of what the manuscript is about can be gleaned from the following excerpt from the Preface. "...o n what these notes are meant to be, and what they are not meant to be. They are meant to be an introduction to what I call the "modern" theory of sample path properties of Gaussian processes, where by "modern" I mean a theory based on 4 I concepts such as entropy and majorising measures. They are directed at an audience that has a reasonable probability background, at the level of any of the standard texts (Billingsley, Breiman, Chung, etc.). It also helps if the reader already knows something about Gaussian processes, since the modem treatment is very general and thus rather abstract, and it is a substantial help to one's understanding to have some concrete examples to hang the theory on. To help the novice get a feel for what we are talking about, Chapter 1 has a goodly collection of examples. The main point of the modem theory is that the geometric structure of the pa- rameter space of a Gaussian process has very little to do with its basic sample path properties. Thus, rather than having one literature treating Gaussian processes on the real line, another for multiparameter processes, yet another for function indexed pro- cesses, etc., there should be a way of treating all these processes at once. That this is in fact the case was noted by Dudley in the late sixties, and his development of the notion of entropy was meant to provide the right tool to handle the general theory. While the concept of entropy turned out to be very useful, and in the hands of Dudley and Fernique lead to the development of necessary and sufficient conditions for the sample path continuity of stationary Gaussian processes, the general, non-stationary case remained beyond its reach. This case was finally solved when, in 1987, Talagrand showed how to use the notion of majorising measures to fully characterise the continuity problem for general Gaussian processes. All of this would have been a topic of interest only for specialists, had it not been for the fact that on his way to solving the continuity problem Talagrand also showed us how to use many of our old tools in more efficient ways than we had been doing in the past. It was in response to my desire to understand Talagrand's message clearly that these notes started to take form... I rather hope that what is now before you will provide not only a generally accessible introduction to majorising measures and their ilk, but also to the general theory of continuity, boundedness, and suprema distributions for Gaussian processes. Nevertheless, what these notes are not meant to provide is an encyclopeadic and overly scholarly treatment of Gaussian sample path properties. I have chosen material on the basis of what interests me, in the hope that this will make it easier to pass on my interest to the reader. The choice of subject matter and of type of proof is therefore highly subjective." The subjectiveness comes primarily from the fact that the examples and motivation of the monograph come from the area of set indexed empirical processes. The contents page of the monograph is as follows: 5 AN INTRODUCTION TO CONTINUITY, EXTREMA, AND RELATED TOPICS FOR GENERAL GAUSSIAN PROCESSES CONTENTS I INTRODUCTION 1.1 The basic ideas 1 1.2 The Brownian family of processes 6 1.3 A collection of examples 13 1.4 Exercises 39 II TWO BASIC RESULTS 2.1 Borell's inequality 42 2.2 Slepian's inequality 48 2.3 Applications in Banach spaces 59 2.4 Exercises 61 III PRELUDE TO CONTINUITY 3.1 Boundedness and continuity 62 3.2 Zero-one laws and continuity 66 3.3 The Karhunen-Lo~ve expansion 75 3.3 Exercises 77 IV BOUNDEDNESS AND CONTINUITY 4.1 Majorising measures 80 4.2 Upper bound proof 88 4.3 Lower bound proof 95 4.4 Entropy 104 4.5 Ultrametricity and discrete najorising measures 109 4.6 Discontinuous processes 112 4.7 Exercises 114 V SUPREMA DISTRIBUTIONS 5.1 Introduction 117 5.2 Some easy bounds 119 5.3 Processes with a unique point of maximal variance 121 5.4 General bounds 127 5.5 The Brownian sheet on the unit square 131 5.6 Exercises 139 VI AFTERTHOUGHTS 6.1 Two topics that were left out 142 6.2 Directions for research 145 VII REFERENCES 149 VIII INDICES 157 6 77-~ (B) SUPER PROCESSES AND DENSITY PROCESSES: The general area of infinite dimension diffusions is one of the most important and vital areas in Probability and Stochastic Processes today. For the first time since the introduction of interacting particle systems over a decade and a half ago, the theory of Markov and related processes has found a totally new family of processes that not only are of intrinsic interest, but require the kind of major extensions of an existing body of knowledge that rejuvinate mathematics and mathematicians. That this is the case can be seen from the calibre of the people now developing this area in the U.S. and Canada, and the flurry of conferences and sessions being dedicated to this topic. There is no doubt that this subject is about to blossom into the "bandwagon" of the early 1990's in Probability and Stochastic Processes, and the results and techniques that have been developed to date promise a very fruitful and rewarding ride. One of the reasons that these processes were initially of particular interest to me is the fact, noted in Section 1, that they can be considered either as stochastic processes in one-dimensional time whose values are N-dimensional random fields, or as real valued (N + 1)-dimensional random fields. Thus there was the natural challenge to see what could be done with them from the latter point of view. This challenge become even more interesting when, as time went by, it became clear that even the usual formulation was interesting from a purely random fields point of view. To describe what these processes are, we start with a particle picture based on N > 0 initial particles which, at time zero, are independently distributed in R', d > 1, according to some finite measure m. We define a branching rate p > 0. Each of these N particles follows the path of independent copies of a Markov process Y, until time t = p/N. At time p/N each particle, independently of the others, either dies or splits into two, with probability I for each event. The individual particles in the new population then 2 follow independent copies of Y in the interval [p/N, 2p/N], and the pattern of alternating critical branching and spatial spreading continues until, with probability one, there are no particles left alive. The process of interest for us is the measure valued Markov process () X (A) Number of particles in A at time t} N where A E Bd Borel sets in R,. Note that, for fixed t and N, Xf' is an atomic measure. - Note also that if p = oo there is no branching occuring. It is now well known that under very mild conditions on Y the sequence {XN)N>I converges weakly, on an appropriate Skorohod space, to a measure valued process which, when p = 1, is called the superprocess for Y. We shall be interested primarily in the 7