STOCHASTIC FINITE ELEMENTS A Spectral Approach Revised Edition RoGER G, GHANEM Johns Hopkins University Po D. Spanos LB. Ryon Chair in Engineering Rice DOVER PUBLICATIONS, IN c > Mineola, New YorkCopyright Copyright © 1991 by Springer-Verlag New York, Inc All rights reserved, Bibliographical Note This Dover edition, first publish 3 is a newly revised edition of the work originally published by Sf Now York, Ino, i 1991. A new Preface has been prepared especially for this edition, Publication Data Library of Congress Catalogin Ghanem, Roger Stochastic finite Spanos Rev ed tral approach / Roger G, Ghanem, Pol D. 2818-4 (pbk) method, 2. Stochastic proce 2s 1. Spanos, PD. (Pol D) 2003046062 Manufactured in the United States of America Publications, Inc, 31 East 2nd Street, Mineola, NY To my Mother and my Father To my parents Demetri and Aicat thinking; to my wife Olympia, my permanent catalyst in substantive living; and to my children Demetri and Bvie delightful randomness. P.D.S.wat the entize system of theory (of probability) analogy—are based on proba ge is connected with th human knowle Pierre Simon de Laplace 1816 ‘Nature permits us to calculate only probabilities, yet science has not collapsed. Richard P. Feynman, QED: The Strange Theory of Light and Matter, 1985, Preface to the Dover Edi n Since the publication of the first edition of this book in 1991, the field of stochastic finite elements has greatly benefited from the concepts of spec tral representation of stochastic processes in terms of thematic diversity and mathematical foundation, This, in fact, was anticipated in the epilogue of the first printing of the book. In this contest, it has been tempting to ed with a second edition towards incorporating certain of these devel- opments, However, it has been felt that the simplicity and tutorial effec: tiveness of the original version could be compromised by some of the lo tic and conceptual details which would have to be incorporated. Therefore, the authors have decided to proceed with this Dover publication with the hope that the original, and in many respects seminal, concepts would become widely available to broad audience for a longer period of time. It is hoped that in this manner, the many requests that we had for reprinting. worldwide, will be answered ‘Thanks are expressed to Dover Publications for accommodating this need and helping disseminate the original contribution towards a rational, foundation for uncertainty quantification. Nevertheless, a somewhat exhaustive list of references to the utilization, refinement, and extension of the concepts of this monograph is included as an addendum at the end of the book. It is hoped that this addition will help as a catalyst for even more accelerated use of spectral approaches to new applications and ‘mathematical formulation, The Authors September 2002Preface This monograph considers engine vandom parameters. Its wnd timing are correlated with the intention of accelerating context, format. the evolution of the challenging field of Stochastic Finite Elements, The random system parameters are modeled as second order stochastic defined by their mean and covariance functions. Relying on the spectral properties of the covariance function, the Karinmnen-Loeve expansion is used to represent these processes in terms of a countable set of uncorrelated ‘andom variables, Thus, the problem is cast in a finite dimen: otting, Then, various spectral approximations for the stoc m are obtained based on different criteria. Implementing the concept of Generalized Inverse as defined by the Neumann Expansion, leads to an explicit expression for the response process as a multivariate polynomial functional of a set of uncorrelated random variables. Alte ly, the solution process is treated as an element in the Hilbert space of random anctions, in which a spectral representation in terms of the Polynomial Chooses is identified. In this context, the solution process is approximated by its projection onto a finite subspace spanned by these polynomials. 7 of the jo response of the oncepts presented in this monograph can be construed as extensions ectral formulation of the deterministic finite element: method to of random functions. These concepts are further elucidated by nies, The nt with those obtained by a the spa applying them to problems from the field of structu correspondix Monte-Carlo simulation solution of the problems. a : to thank Rica University, the Houston Advanced Re- inter (HARC), and the National Center for Supercomputing Appli- ations (NCSA) for the extensive use of their computational facilities during the course of the studies which have led to # on, development, and integration of the material of this monograph. The financiel supporperiod of years, from the National Science Foundation for Earthquake Engineei Buffalo, the is gratefully acknowledged. Further, the plethora of students and colleagues are greatly appreciated, R.G. Ghanem P.D. Spanos October 1990, J 1 2 : Sap plata quelques withade, deo duvdeyuncals cles fovshins odeslnte the National Center Research at the State University of New York Air Force Office of Scientific Research, and Rice University imulating discussions with a Contents 1 INTRODUCTION 1.1 Motivation 1.2 Review of Available Techniques. 1.3. The Mathematical Model 14 Outline Preliminary Remarks lel PRESENTATION OF STOCHASTIC PROCESSES a 2 Review of the Theory Karlunen-Loeve Expansion Derivation | 2.3.1 2.3.2 Properties 2.3.3 Solution of the Inte | | 3 4 ‘al Equation 24 Homogeneous Chaos Preliminary Remar‘ Definit as and Propertie suction of the Polynomial Chaos SFEM: Response Representation 1 Preliminary Remarks 2 Deterministic Finite Elements Problem Definition ‘Variational Approach tral Methods and Hierarchi- cal Finite Blement Bases 3.3 Stochastic Finite Elements 63 63 693.3.1 Preliminary Remarks 3.3.2 Monte Carlo Simulation ( 3.3.3 Perturbation Method Neumann Expansion Method Improved Neumann Expansion Projection on the Homogeneous Chaos Geometrical and Variational Exten 4 SFEM: Response Statistics 4.1 Reliability Theory Backs 4.2 Statistical Moments 42.1 Mon id Curmmulants Equi Second Order Statistics 13 Approximation to the Probability Distribution 4.4 Reliability Index and Response Surface Simulation nd ns 5 NUMERICAL EXAMPLES 5. Preliminary Remarks 5.2 One Dimensional Static Pro! Formulatio: 5.22 Results 5.8 Two Dimensional Static Problem 5.3.1 Formulation 5.3.2 Results 5.4 One Dimensional Dynamic Problem Description of the Problem Implementation Results 6 SUMM. el SU RY AND CONCLUDING REMARKS IMARY AND CONCLUDING REMARKS BIBLIOGRAPHY ADDITIONAL REFER INDEX CONTENTS 189 189 195, STOCHASTIC FINITE ELEMENTS[denmrtediye| { Chapter 1 INTRODUCTION 1.1 Motivation 1d as a(lack)offpattern) or regularity. This fe are andomness are generally recog- egulerity in th nd the impossibility of an exhaustive det for instance, with the Uncertain gas. The other can be related to a generalized lack of{kno Rendomnessican be def can be observed in physical realizations of most objects tha in a space-time context. Two sed (Matheron, 1989), The first one is an inherent i phenomenon being observed, ministic description. Such is the of quan about the processes involved. The level of uncertainty associated with tl 1m mechanies and the kinetic theory ced by recording more observations class of problems can usually be of the process at hand and by improving the measuring devices through is being observed. In this category fall econometric series whereby the behavior of the financial matket is modeled ag a stochastic process. This process, however, can be entirely specified, in the deterministic sense, from a complete knowledge of the flow of goods at the smallest levels. Another example that is of more direct concern in the present study pertains to the properties of a soil medium. These properties fro uniquely dofined at @ quite impractical, however, to me wen spatial laeation within the medium. Tt i fe them at all points, or even at a relatively large number of points. From a finite number of observations .e properties may be modeled as random variables or, with « higher leve of sophistication, as randozn processes with the actual medium properties 12 CHAPTER 1, INTRODUCTION 1s a particular realization of these processes. From the precedi spparent that deterministic models ean be considered as approx corresponding physical problems, in a similar sense that linear models approximations to the actual nonlinear behavior of these problem tions The response of a system ean be induced either by the input or the system operator. From a causality perspective, the output sence of an excitation or some initial conditions on the operator. A parallel probabilistic argument can be made to the extent that the output cannot be random without either the input or the operator boing random, The case of a random input and a deterministic operator has been amply studied and numerous results have been obtained in the field of engineering mechanics. The second case, that of a random operator, with or without random inputs, is mmuch more complicated and the underlying mathematica tools are still in the developing stage. This operator induced randomness is also termed par complexity depend: model. A major factor in this respe lex ne requisite accuracy an tric, Within this class of problen jain on the e the degroe of ent of sophistication required from the is whether the random aspect of the process. Obviously hence sophistication of the model depends on he relative importance of the system being investigated. This importance fact ;pts as the generalized cost to be incu a potential failure. w the development of accurate and realistic 1m ideled as a random variable or as a r is related to such ne startling technological growth witn ed during recent decades, of engineering has become feasible. ‘The functionality of many modem struct t on their ability to perform adequately and lity under not absolutely controllable conditions. Such s, nuclear power plants, offshore platforms space v The randomness in these snomena as soil variability, earthquak indueed ground motion, random ocean waves, thermal and acoustic loadings, and uncertain fatigue effects. ‘These are exam that should be designed with a negligible p to being invaluable for the s of important structures obability of failure. In addition fe design of such structures, the probabilistic rovides crucial concepts for the task of ende development 1.2. REVIEW OF AVAILABLE 7} \CHNIQUES. 3 n for more sophisticated models, svay in relation to the analysis, red by the need for Paralleling this technology induced opti basic theoretical developments have been und of random processes. These developments mote accurate models in theoretical statistics and el thapter II concen portion of this monograph, namely the representation ore will be said in theory for random processes. Random aspects aside, modern engineer to analyze. Such complexity results from t different parts of the structure, as with the surrounding medinmn. Addressing such complexity require to accurate and efficient numerical algorithms. Therefore, itis clear that for y probabilistic modeling of a physical system to be useful, it has to be apatible with the available numerical schemes. ‘The finite element method has proven to be well euited for a large class of engineerin hms have a sound and well developed theoretical basis and has long been proven and tested on a variety of problems. ‘The formulation presented herein is, indoed, a natural extension of the ba ideas of the deterministic finite element method to accommod functions. ectly related to the theme structures are quite complex intricate interaction among the n of the structure fas the interact 1.2 Review of Available Techniques. ‘om en engineering mechanies perspective, the most common stochastic differential equation with random coef the system under gystem problem involves a line. icients, These coefficients represent the pr vestigation. They ca f curately and andom variab complexity, as random processes ‘the problem can be ith an ix ied probability structure. Mathematica formulated as Au=f (at jal operator, u is the random response, ation. Note that equation (1.1) with where A is a linear stochastic differe and f is the possibly random o extensively (YK. Lin, 1967; Yang and and Kapania, 1984; YK. Li 1986). ‘The case where the operat. considerably more di mate solution: have boon reported alt and only app:4 CHAPTE 1. INTRODUCTION in the literature. A rigorous mathematical theory has been developed for the solution process when equation (I) is of the Ito type (It In this case, the solution is a Markov process the probability distribution of which satisfies a Fokker-Planck, partial differential, equation whic t amenable to analytical solutions. Implicit in the It6 equation 1s the assumption that the random coefficients associated with the operator are white noise processes, The relationship between ordinary equations and equations of the It6 type was considered by Wong and Zakai (1965). The stability of linear stochastic equations with Poisson process coefficients vias also investigated (Li and Blankenship, 1986). Along similar lines, anc through @ judicious use of the relationship between diffusion and random walk, the Navier-Stokes equation for flows with high Reynolds numbers was numerically analyzed by Chorin (1978). ‘The approach which was used featured a grid-free numerical solution of the governing equations, which would otherwise require & prohibitively fine discretization. ‘This solution strategy hinges on an analogy between the diffusive part of the Navies equation and rendom-walk, thus permitting the simulstion of the icity generation and dispersal by using computer-generated peeudo-random mumbers, Various convergence studies of this method have since been carvied out (Roberts, 1989), long with « number of applications, specially in connection with turbulence related problema (Ghoniem and Oppenlieim, 1984), A more realistic problom is that of a stochastic operator the random coefficients of which poss class of problems the It6 calcul ss smooth sample paths. For this breaks down, and no general mathematical theory is available for obtaining exact solutions. Kozin (1969) presented a comprehensive review of the st problem of these equations as well as equations of the Ité type. Bharrucha-Reid (1959) and Hans (1961) were among the first to investigate the problem from a mathematical paint of view. They studied the existence and uniqueness of the solution t operator equations, The problem in dealing with stochastic equations is two-fold. Firstly, the random properties of the system must be modeled adequately as random variables or processes, with a realistic probability distribution. A good treatment of this modeling phase is presented by Benjamin and Comell (1970}, and by Vanmarcke (1983). Also, compact probabilistic repre- sentation has been suggested by Masei and Miller (1982), and by Tr et.al (1986). Secondly. ina, e resulting differential equation must be solved, 1.2, REVIEW OF AVAILABLE TECHNIQUES, 5 and response quant t are obtained, usually as determined cond order statistics. It is this second st cent study. Namely, the solution of equation (1.1) when the coefficients of the operator A possess a prescribed probabilistic structure, Interest in this class of differential equations mechanics, wave propagation, turbulence theory, random eigenval functional integration. Good accounts of these applications are presented by Bharrucha-Rei(1968) and by Sobczyk (1985). The methods of solution range from solving the averaged equations for an equi perturbing the random fields and keeping one or two terms in the expansion, or using a Born approximation for the random fields. Some attempts have also been directed at obtaining exact solutions to the problem by functional integration (Hopf, 1952), He suggested the construction of a differential equation for the characteristic functional of the random solution, from which a complete probability description can be obtained. The method was applied bby Hopf to a turbulence problem. Lee (1974) applied functios to the problem of wave propagation in random media, and obtained a Folsker- Planck equation for the characteristic functional of the solution. Later, Ko- tulski and Sobeayk (1984) developed a scheme to construct the characteristic functional of the stochastic evolution equation and obtained expressions that concern in the pr has its origins in quantum ss, and lent solution, to functional approaches tend, in ge . nature. yee from the above discussion it becomes clear that the c rent state of) knowledge ochastic differential equations” precludes) exac lutions except in some rare instances. ‘The problem becomes ever intractable when applying the results deseribed above to engineeting g iams with intricate geometries and boundary conditions and various types of excitations. Comprehensive reviews of the current trends for analyzing tas- in engineering are given by Iyengor and Dash (1976)ulbrahiaa™ a (1987B)yand Benaroya and Reha dom syst (1987), Shinozuk The most widely used technique for analyzing random systems in engi- necting io the perturbation method. This fact is mainly cine fo the math= ematical simplicity of the method, which does not, however, guarantee its validity as a method of solution. The perturbation scheme consists of ex- panding all the random quantities around their respective mean values via ‘@ Taylor series. Secular terms that cause instability of the approxin6 CHAPTER 1. INTRODUCTION solution appear in higher order terms. Also, computations beyond the frst second terms are not practical. These facts restrict the applicability the method to problems involving small randomness, es witnessed in the available literature. Boyce and Goodwin (1964) used the perturbation approach for the solution of the eigenvalue problem of random strings and beams. They used the integral equation approach to invert the governing ferential operator, and considered randomness properties and boundary conditions. Soong and Bogdanoff (1963) used transfer metrix techniques to investigate the behavior of disordered linen wins. They expressed the frou ‘aused both by material \cy response of the chain ag a perturbation type expansion in terms of the small deviations of the random variables involved. They concluded that even for small levels of disorder in the system, the higher frequencies showed considerable spread around their mean values Later, Collins and Thompson (1969) used the perturbation approach to treat the general problem of computing eigenvalue and eigenvector statistics were random variables described by their covariance matrix. A second order Taylor series about the mean value was o represent the eigenvalues, eigenvectors and the random parameters the system. They also investigated the stability problem of a column and concluded that stability problems were more sensitive to cross-sectional area uncertainty than frequency problems. Hart and Collins (1970) and He and Hart (1972) extended the work to develop a perturbation scheme compatible with the finite element method, Dendrou and Houstis sis of a soil medium (1978a) Wve a soil-structure interaction problem as +r perturbation scheme which again, restricted their approach to rather small levels of random fluctuations, Using the same perturbation type expansion, Nakagiri and Hisada (1982) and Hisada and ed again n eigenvalue problem, allowing for random boundary conditions. concluded that the second order perturbation was too untractable to be of any practical interest in solving ‘al problems, Liu et.al (1985) introduced a new implementation scheme for the perturbation based Imire tlement: method. 1 expansion was carried ont " ‘of the finite element assembly uations, however, are identical with those obtained he more standard method mentioned previously. Liu et.al applied their approach to analyzing random nonlinear and transient problems (1985, 1986, mm whose para el coupled results fr with the fin pin an inference statistical an ey used a first on real phy ne perturbation 1.2, REVIEW OF AVAILABLE TECHNIQUES. 7 1988), All the metho mn of the rand discussed above are based on @ perturbation type ex- pa quantities involved. ‘Their validity is restricted to ceses where the random elements exhibit small fluctuations about their mean values, in which case a sr approximation yields fairly good results. Also, a solution to the problem based on the perturbation approach does not provide, without a prohibitively large amount of additional analytical and putational effort, higher order statistics. ‘Another more elegant method, not very popular in connection with the st Of ck Bharruche-Reid, 1968). The method is bas moments of the output and of the system as functions of ments. Consider, for example, equation (1.1) where the stochastic operator ‘Ais split into a deterministic part L and @ random part I of problems treated herein, is the hierarchy closure approximation on expressing joint statistical L+T)u Solving for w results in L'Itu 1.3) Note that on averaging equation (1.3), the moments of w di of Tw, To obtain these latter moments, apply the operator IT to equation 1,3), to derive Ty = ML“y — 1M 1.4) Substituting equation (1-4) into equation (1.3), gives web y—- Lonny + Lton oe 1.5) and so cess, an approximation has to be made of the form on, At some point in the substitution pr = <(Lm)"> 1.6) <(u“n)” The unearte fied intuitively jan (1983) he same denotes the operation of mathematical éxpeatat pling in equation (1.6) has no rigorous basis. It is often j by a local independenc wagested by Ac that closure at certain level of the hierarchy is equivalen order of perturbati bation based solution. The e argument. It was8 CHAPTER 1. INTRODUCTION enough to cases of small random fluctuations. latin it the applica higher order lity of the sure approximations are substant ethod t and Grigoriu of Closed form lutions were very restrictive Howe VF sesumptionso ristics ofthe beatn make the method of a rather limited interest. ssfenana - Adomian (1964) introduced the concept of stochastic Green's function the solution of diferential equations involving stochastic coefficien The concept was further developed by Adomian and Malakian (1980) ond Adomian (1988) whe cd using decomposition method for solving nonlinear stochastic differential equations. Accordingly, the inverse of the es sr ae ance of MACSYMIA snced in ation packagt salties were expe wo term in the expansion, a fact thet handicaps the Further, Shinozuka and ain a Neumenn expansion media. They implemented only for the case of troduced ¢ utes a plend The » for the soluti vith sizable randomness the delta met Nomoto 1980} deve! emazaki et.al (198% absequently 2 of proble ised on the Neumann expansion chod vw for the inverse of the stochastic coefficient matrix whereby digital simnlation J togenerate therandom matrix. The method was applied with J Rclents of variation and good agreement, was ved with the @ Carlo simulation. However, the method ral times in order to prodice reliable echiniques a vor THE MATHEMATICAL MODEL 9 results, especially when sizable random fluctuations ate involved raat 1.3 The Mathematical Model /O2.-< The class of problems dealt with in this study’ if not of the conventional engineering kind in that it involves concepts of a rather abstract and math- ematical nature. It is both necessary and instructive ta introduce at this point the mathematical concepts which are used in the sequel. The de- gree of abstraction is kept to a nec minimum to avoid obscu the engineering aspect of the problem. Also a notation is in simplifies, to some extent, the mathematical complexity. It first, that a mathematical abstraction of the problem is believed, however, that such an approach is vital for a complete and mature coduced that may seem, at aitous. It is understanding of the problem as well as directing any f at extending the present formulation Tho Hilbert space of functions (Oden, 1979) defined over a domain D, with values on the real line R, is denoted by H. Let (9,42, P) denote a probability space. Lot i PD and é be an element of & Elements of H and © are Capital letters are used to denote algebraic structures and spaces as well as operators defined on these spaces, with greek lottors referring again to those operators defined om spaces of random functions. The inner products over Hi and over © ate defined u saeure and the pr cloments f(x) and (x) in HL their inner product ( hi(x) defined ing the Lebesgue That is, for any two s(x) ) is lity measure, respectively (x)dx 17) (hs) hyo) =f iba The domein D represents the physical space aver which the problem is defined, Similarly, given any two elements a(@) and 3(0) in ©, thei inner product is defined as 0(0),3(8) = f a()(@)aP 18) Jn ral conditions, the the integrand with where dP is a probability measure. Under very is equivalent to the average of integral in equation (1.At Ln lant Any two elements of the Hilbi ‘thogonal if their inner product vanishes. om defined on the prod defined above are said to be or- random process may then be space D x 9. Viewed from rded as a curve in either of H t spaces erspective a random process can be re The physical mo Dit random spatial fluctuations and which is subjec The mathematical representation of this problem involves an opetator equation el under consideration involves a medium whose prop nal excitatior A(x. O)ulx.8 1.10) where A(x,9) is some operator defined on Hx ©. In other words, A is a differential operator with coe s exhibiting random fluctuations with respect to one or more of the independent variables. The aim then is to for the response u(x, )) 98 « funetion of both its arguments. With no of generality, A is assumed to be a differential operator, whose random tricted to being second order random processes. This is not ‘ction for practieat problems, since most physically measurabl Then, each one of these coef pinistic component and a coefficients are processes are of the ax(x,4) cam order type, nts component in the form o%E) = dale) + an) 1 where d 0 the math matical expectation of the process ay(x. ocess, having the same ¢ n (1.10) can then be written as and a(x, 0) is fumetion as the process a(x. 4) L(x) + Mix. @ju(x.a) = (x0) a2 ic differential operator and TI(x,6) is a differen- al operator whose coefficients are zero-mean random processes, Before ssential to clarify what is meant solution to by such a solution 1.3, THE MATHEMATICAL MODEL As was mentioned above, @ denotes the Hilbert space of functions defined on the o~field of events generated by the physical problem, with renge the interval [0,1]. In other words fle realizations of the random [0.1], and these numbers (0.41, tt \were numbered continuously on 1 igned to the variable 6, then ay(x, 8) is a deterministic function of x, a realization of the proce observing a finite number of realizations of the process, distribution theory 1e used to construct the distribution of the process along the # dimen: sion. For a given x, az(x,6) is a random variable with su Obviously, for a complete description of the process, the joint distribution all x € D is required. However. if the process is assumed to be Gaussian h a distribution all the finite dimensional distribution functions are also ian, Clearly the usually limited number of observed realizations of @ random process cannot, in general, suggest any definite distribution, However, invoking the central Innit theorem, the Gaussian distrib likely candidate for many physi nee the coefficients in the 10) have been defined through their probability distribution functions, the main question rex as to what is meant by a solution to the problem. Obviously, the di sion and commen he coefficients proce: to the solution process as well, A conceptual modificatic introduced. Specifically, a quite general form of the solution process cant be on appears to be the 1 applications. ator equation ( just made concernin expressed a8 w= glanlx.6), Fl 8).x) (1.13) Clearly, a com- ponse would involve the prescription of its joint ibution with the various processes appearing in equation (1.13). TI information could form the basis for a rational reliability and ris However, given the infinite dimensional structure of the random proct 3), stich a ta A finite-dimensional description of th where gf plete description of th 's some noulinear functional of its arguments. appearing in equation (1 of currently used met pro- cesses involved is required if the solution is to proceed in a computational etns to exceed the capability Given the abstract nature of the finetional spaces over which rendom processes are defined, a finite dimensional represen be achieved through partitioning of these spaces as is usually done with the deterministic finite element analysis. Alternatively, an abstraction of the troduced which is mathematically equivalent tion cam discretization process can b2 CHAPTER 1. INTRODUCTION © a discretization with respect to a spectral measure. Indeed, a mumber of spectral representations are introduced in the text which permit the algebraic manipulation of random processes through that of an equivalent diserete set of random variables. As the title indicates, this monograph is a study in the application of spectral discretizations, taken in the context just explained, to problems of mechanics treatable by the finite clement method. This is, indeed, the main impetus of the book. Chapter 2 : 1.4 Outline wnweauey REPRESENTATION OF wwe’ STOCHASTIC PROCESSES Ia Chapter, available techniques for addressing problen in Chapter 1, relevant elements of the tepretentation theory for con- ovo representations used in the ensuing development of the stochastic finite ve dbase element method are discussed in deta. These are the Keshunen-Loeve |’ FT Preliminary Remarks expansion and the Homogeneous Chaos expansion a aud Chapter IT includes a review of dctorminite finite element method plasiaing the features that can be extended tothe stochastic case, Further, a polescioes poo ems er number of the most widely used techniques for treating the model developed, fotos ae represented by a denuerablectof parameters consitng of in this chapter are reviewed. Attention is focused on mathods based on |e, he values ofthe funtion and is derivatives a hen 8, the problem the spectral representation of stochastic processes as outlined in Chapter II. “| ~* (eucountered)in the stochastic case is that mdiom Process ‘These methods are shown to provide a sound theoretical extension of the deterministic finite element inethod to problems involving random 5) prop Tn Chapter IV reliability methods in structural engineering are briefly reviewed. Further, itis shown how the developments of the previous chapters In the deterministic domain has a physical can be used in a mimber of approaches to compute approximations for the #t#S-»(appeal) The domain in t however, have a physical srobetiity density function cre ee ees meaniig that permits a sensible discretization, In this context the abstract, In Chapter V detailed numerical example ate given which demonstrate the finite elemen the usefulness of the spectral approaches introduced in Chapter III it can be extended to(deal}with random functions. The requisite x Similarly to the case of the deterministic finitg element method, (wherebs by a denumerable set of random he proces: random processes in such a sulation of the In a more applied sense, it is soug! manner that they can b physical problem, stic case does foundation c ethod becomes useful as Finally in Chapter V1, pertinent conclusions and a _ matical rigor for a coniplete mniderstanding of the theory of representation the prablam at —— for stnchastic + is heyond|the,szope of this book. However, a brief review of the theory will ptovide)more(isight into the problem at hand and a(broad)perspective of the relevant ma the heuristic local averaging representati ntation are reviewed in the next hematical concepts, Tn t and the move rMM CHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES 22. REVIEW OF THE THEORY 6 eo rowebthists Je prrasue ACL Gn 2.2 Review of the Theory” yo ojo monbie dt involved. ‘This fact necessitates either the use of an independent mesh for . Juke pee ttles off olin THETfalation, or the use of a mesh size such that both stresses and material Aithough (te Reutsti, local averaging is probably thenot widey)used properties ae dente nod fo ‘and consistently represented. In either case, the dimension of the problem, as reflected by the number of random variables random pro es. A continuous random process is Tnden(belonging) used to represent the underlying processes, is quite large. The associated : be approximatec computational problem is, in general, of probibitive dime as closely as esived)by resteicting the index to a/se¥lense}in the indexing set Alternatively to the heuristic arguments associated with the local aver- mathematical rigorfurthej, a random process can be represented aging approach, a rigorous exposition of the basic concepts of the theoty of representation for random processes can be formulated (Patzen, 1959), This theory is a quite rich and mature mathematical subject. ‘The developinent of element analysis of the theory parallels that of the modern theory of random processes, and has abs on by random had its origin in the neod for more sophisticated models in applied statistics. as to coincide with Some local avérage of the ‘Most of the related results have been derived for the class of second order It is expected, however, that the result would processes. Perhaps the most important result is the spectral representat pend to @ notable extent on the averaging method used. ‘Local averageS> of random processes (Gel'fand and Vilenkin, 1964), which, in ts most general 2 usually of tw k gived average) over a Subse form, can be stated as set. and the(ollocation averagejover each such subset whereby eh the indexin the process over the subset is replaced By its value at some point in the 2.1) subset. It is clear that, in general, the first approach(smoothes|the random process, whereas the second approach introduces additional irregularities. we (1, X2) ‘This suggests that the respectively, for the randor » approaches provide lower and upper bounds, behav’ In the limit, as the size of the proces | on im tes § ofeach subset becomes vanishingly small, the representation resulting from saya) = fol) gle} jo" / the two approaches should converge to the exact process. Its obvious that a Faction on te | telatively lange numberof random vasiables is required to representa random Tn equation (2.1), 4(4) is a deterministic function. Further, %> process in this fashion. It is noted that local averaging) parallels(pointwise) orthogonal sot funel st orth approximations of deterministic functions, ‘The size of the individual used for the averaging process does in general depend on the (frequency spectral decomposition oceurs if the process w(x, 8) itwide) content of the process. That is, the broader) the frequency content of the In this case, equation (2.1) ean be shown to reduce to process, the smaller the region over which ‘the process shows a. definite relation (Yaglom, 1962) and the following equations hold (Gatterr, ond thus the smaller the size of the necessary subset to meet a in precision criterion. This problem with local averaging is particularly cee on the o—field W of random events. An important specializ stationary. © dulw,8) cial in the context of the finite element analysis of str Joos a The finite ele ly and te Tequires recoar ved elements an form spacing of t] Cowl ibs: points. {The shape and size of the finite clement mesh i es < distribution within the structure. /This stress distribuitia Here, the symbol 7 aEUMIF tdependent of Aa dame de LEME con CF ak fevtlion dete las " BP aRintan oe Wie mate eh dn then oot dSabitinds tate) hy ean eR ernst dag apo t wEN YA an Yl Oyen jenotes vector transposition, S(w) is the usual spectral density of the stationary process, and w is the wave number vector fainiy of the random para16 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSE: The p sequent development of the theory of 0 had a randomly excited deterministic s ‘eceding representations rong impact on the sub- However, thei ‘em: lve applications have been r This is largely attribute: differential: mensional space, not readily amenable to computational algorithms. erative formulation of the GReasively used in the se fF these rep of random functions, and ere therefore set in an infinite di- An ‘and one which spectral represent a random process w(x, 0} ean be snumerable set of prthogonal randot va hoger oes (25) where {§us(0)} is a set of arth e random process ye regarded as her, it is important tation of the process fanned by the set {&(0)}. The 1d 98 a direct sum of orthogonal projections by the magnitucles of the projections on successive igenvalues of the covani- w(x,8) . Given the erucial role mnsion has in relation to the methods discussed to note that # sion of the random proce: is equation can be viewe x,6) @6 a curvo in the Hilbert space that the Karhunen-Loove exp in this monograph Collectively, t linear operators of (Doob, 1953). Interest can be generalized to allow for the representation of nonlinear functionals of the orthogonal stochestic measures dy(Q). The theory of nonlinear functionals was developed by erra (1913). He gene Taylor expansion of functions to the Te was Wiener, however, who first applied Volter hastic analysis, and developed whet is now known as the H mogeneous Cizos, Based on Wiever’s w meron mn (1947 (2.13) hunen-Loeve expansion achieves. The expansion was = ¥¥ VI Aw Fale) fino) by a number of investigators (Karhunen, 1947; Loeve, sans) 1948; Kae and Siegert, 1947) oaesiool?} Let he position vector x def ace of random events le real ‘Then, multiplying both sides of equation (2.12) by fe(xz), integrating ove: the domain D, and making use of the orthogonality of the eigenfunctions, he yields over the don ii(x) denote the expected vali zations of process, and C(x1,x2) denote its covaria definition of the covariance function, it is bounded, symmetric and positive definite, Thus, it [Obe.%) faba) d= 2.13) has the spectral decomposition (Courant and Hilbert, 1953) ‘D> cpvahotios ae = = iu Roetmn To C1.) = 2 An fale) fale! ) “Ladin de nectonen yen Multiplying once more by fi(s) and integrating over D, gives where 2y and jn(x) are the eigenvalue and the eigenvector nce vere) respective ave the solution to the integral equatior : x i) ds (1) fi dada = OE > VInNe bt es eae Due to the symmetty and the positive definiteness of the covariance ker eee cee cae wean Loeve, 1977), its eigenfunctions are orthogonal and form a complete = dedu = VIE (> 9.15) bbe normalized according to the following eriterion j : 7 Equation (2.15) can be rearranged to give > = bux (2.16 where Sam is the Kronec Thus, the random process w(x, 8) can be written as rejolx. 6) is a bri . RD Fels \ where, I yn prt iaton) ‘20 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES and A, equation (2.17) at the fx(xe) are so}tiod to equation (2.8). ‘Trmcating the series in i / Mt erm, gives x) +O $10) Vin fale) xplicit expressionfor-$q{4}-can-beobtained by multiplying equation ey fs and era he domain D. That is, w(x,6) = even nian) ean DP, Indeed, it can be shown (Parzen, 1959 ) that if e funetion f can be represented ther bmp toc AKI omepmding wh at C30, \ (0) = ables {n(6)} ‘with the covariance function of the process that }. Another point operations used to represent f in terms of {C(.,x2) , x2 © of practical importance is that the expansion given by equation (2.19) can be used in @ numerical simulation scheme to obtain numerical realizations of the random process. In fact, this simulation procedure is used in conjunction with the Monte Catlo method for one of the illustrative examples in Chapter IV. It is optimal in the Fourier sense, as it minimizes the mean square error g from truncation after a finite number of terms. ‘The expansion is in the fields of detection, estimation, pattern recognition, and image proc an efficient tool to store randorn processes (Devijver and Kittler, 1982), Tr is we ig at this point that the Karhunen Loove expansion was independently derived in connection with stochastic ‘turbulence problems (Lumley, 1970). In that context, the associated eigen- functions can be identified with the characteristic eddios of the turbu field. AO ed da an pomsis a oeve age 23, KARHUNEN-LOEVE EXPANSION 2 It is well known from functional analysis that the steeper a bilinear form decays to zero as a function of one of its arguments, the more terms are needed in its spectral representation in accuracy Noting that the cder to reach a prest isform operator is a spectral x it may be concluded that the faster the autoca zero, the broader is the corresponding spectral density number of requisite terms to represent the Karlunen-Loeve expansion, epresentation lation function tends to and the greater the ne underlying random process by For the pecial case of a random process possessing a rational spectrum, the integral eigenvalue problem can be replaced by an equivalent differential equation that is easier to solve ( is reminded thet a neces: Trees, 1968). In the same content, st ry and sufficient condition for a proc 1 (Kree ins for the integral finite dimensional Markov realization is that its spectrum be ration and Soize, 1986). Further, note thet analytical solu ‘equation (2.8) are obtainable for some quite important and practical fo of the kernel C(x1,x2) (Juncosa, 1945; Slepian and Pollak, 1961; Van Trees, 1968), some of which are treated later in this chapter It may seem that any complete set of functions can be used in-lieu of the eigenfunctions of the covariance kernel in the expansion (2.11). However, it will now be shown that the Karhunen-Loeve expansion as described above, has some desirable properties that make it a preferable choice for some of the objectives of the present approach, Properties ing Property ed coordinate system defined by the eigenfunctions of the co nal in the sense that the mea n of the process w(x, 8) a8 man square error resulting ‘This property can be proved as follows. Given a complete orthonormal set of fanctio can be approximated in a convergent seri nl), the process w(x,8) wv(8) = So dn E(B) fn) 2.2%22 CHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES Truncating equation term results in an error ¢4¢ equal to ultiplying equa ~ ff u(x. 8) Pls) a 2.28) n(x). Substi where use is made of the orthogonality property of the tuting equation (2.23) for Em (8) back into equation (2.22), the mean-square ervor €, can be written as Integrating equation {hi(x)} yields xy dx ibject to the orthonormality o he functional the problem, then, is to minimize Jp (x). In other words, the solution minimize: the functions given by the equation Fixe} = YS fpr en nxt de Jb = dm fh Bm ()Pm() a — 2.98) 2.3, KARHUNEN-LOBVE EXPANSION 23 Diff nn (2.26) with respect to fe(x) and setting the resul equal to zero, gi qu which is satisfied when f, tone) Uniqueness of the Expansion dxg = A, hilo) o (2.28) by equat only if” part, equation (2.12) ean be used with )> = By 10 obisia 7 Clea) = Sand . 2.29) Multiplying both sides by fn(x2) and integrating over D gives = wn fol) 2.30) In the context of this last theorem, it is interesting to nove that some given by equation (2.11) with orthogonal random variables and orthogonal deterministic functions that nvestiga 7) have used an expansion kind Jo not satisfy equation (2.8). It is obvious that such an expansi orm a basis for the representation of random processes. 1 of Gaussian Processes Lt with covariance function C(x, x2)- Then ae aposition given by equationXR 2. REPRESENTATION OF STOCHASTIC PROCESSES n vector, That is, any subset 1e these random variables are uncorrelated, heir independence. Some important conse- ifically, their Gaussion property implies and 8) (8)> ie imanation extends over all the partitions of the set {&(9) f two elements, and the product is over all such sets in a given wermore, it ean be shown (Loeve, 1977) that for Gaussian jon is almost surely convergent. Other Properties In addition to the mean-sq disconsed any further. A detailed study of the properties of the Karhunen Loeve expansion is given by Devijver and Kittler (1982) - a 2.8.38 Solution of the Integral Equation 44 . i i foes fo Chxisxe) floea)dxn = AFG) 2.33) > ion (2.93) is a he- Kind The theory investigated and is graphs (Mikhlin, 1957). Being an tance function, the ke bounded, symmetric, and This fact simplifies the ensuing analysis considerably in where Cl 269) i iogenevus Fredholm integra qu underlying this kind of positive def 2.3, KARHUNEN-LOEVE BXPANSION that it guarantees a number of properties for the eigenfinetions and the eigenvalues that are solution to equation (2.38). Specifical 1. The set f(x) of eigenfunctions is orthogonal and complete. 2, For each eigenvalue \y, there correspond at most a finite number of linearly independent eigenfunctions. 3, There are at most a countably infinite set of eigenvalues, 4, The eigenvalues are all positive real mumbers, 5. The kernel C(xi,x2) admits of the following uniformly convergent expansion Ox = DAs fale) filma) (2.34) Si ‘The Karhunen-Loeve expansion of a process was derived based on the preceding analytical properties of its covariance function. These properties are independent of the stochastic nature OF the process involved, which allows the expansion to be applied to a wide range of processes including nonstationary and multidimensional processes, ee sh Of special interest in engineering applications is the class of one-dimensional random processes that ca be realized as the stationary output of a linear Alter to white noise excitation, These processes have a spectral density of the form N(u?) S(w) = 7 Dw) where V(.) an 1d D(,) are polynomial operators of order n and d respectively ‘The interest in this class of processes stems from the fact that a necessary and sufficient condition for a process to be realizable as a finite dimensional Markovian process is that its spectral density function be of the form ex- pressed by equation (2.35) (Kree and Soize, 1986). Loosely speaking, the ‘Markovian property of a process implies that the effect of the infinite pastCHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES (ble. That is, the process has a finite memory. For 3) becomes, oon the present is negli stationary process, equatiox how xa = d f(x) 2.30) he whole real line When the domain D covers tbove equation is equiv- alent to the Wiener-Hopf int fe solution of which may be \d explicitly (Paley and Wiener, 1934; Noble, 1958). The case where D fever, is more relevant to the context of the monograph, and the associated integral equation is next detailed, Taking into consideration the one-dimensional form of equation (2.4), and substituting for S(w) from equation (2.35), equation (2.36) becomes ¢ solutior Differentiating equation (2.37; multiplying the integrand by w?. Thus, applying the differential operato: Djd2/dz3| to this equation yields rv 1 f(x) = 22) 6(a1 — 2) dg (2.8) sshere 6 the dirac delta function. Equation (2.38) can be viewed as a ref on of equation (2.96) as a homogeneous differential equation. This equation may be solved in terms of the parameter \ and of 2d arbitrary s which are calculated by backsubstituting the resulting solution into equation (2.36). Note, parenthetically. that explicit expressions f the transcendental characteristic equation associated with the differential uation (2.38), for a number of kernel functions, are given in Youla (19 In the remainder of this section, the preceding treatment of equation (2.31 is applied to the important kemel represt process, This kernel has been used extensively to model processes in a variety of fields (Yaglom, 1964). Further, it is noted that higher order Mathoviau. ns linear combinations of ing the first order Markovian kemels may be expr 2.89) 2.3, KARHUNEN-LOEVE EXPANSION with the same units as 2 and is often termed the correlation length, since it reflects the rate at which the correlation decays between two pe sumed that the process is d over the one dimensional interval [-a,a]. Clearly, (ry, 12) can be made rapidly attenuating versus |21— 2a) by selecting a suitable value of the parameter b. In case the domain D of the problem is the one-dimensional segment {~a,], the eigenfunctions and eigenvalues of the covariance function given by equation (2.39) are the solutions to the following in (Van Trees, where b is a parameter ' of the process. It is al equation dry = d f(x) (2.49) can be written as A fe) =-e Differentiating once more with respect to xi, the tained lowing equation is ob- 2d fle (2.43) Introducing the new variable ay equation (2.43) becomes To find the boundary conditions associated with the differential equation28 CHAPTER 2. REPR. ATION OF STOCHASTIC PROCESSES ), equations (2.41) and (2.42) are evaluated at 2 = —a and 2 = +a, After rearrangement. the boundary conditions become ef(a) + F(a) = 0 (2.46) (2.47) is transformed into Thus, the integral equation given by equation (2.4 the ordinary differential equation (2.45) with appended boundary conditions sgiven by equations (2.46) and (2.47). It can be shown that «? > Os the only range of wgfor 445) is solvable, the solution being given by the equation hich equation + a2 sin (wa (248) Further, applying the boundary conditions specified by equations (2.46) and 2.47), gives a; (¢ — w tan (wa) + ww +e “)) = 0 (2.49) (ce — & tan (wa) = 0. the determinant of the homogeneous system ing this determinant equal to zero Nontrivial solutions exist onl in equation (2.49) is equal to zero. St he following transcendental equations (2.50) *, the resulting (2.52) 23, KARHUNEN-LOBVE EXPANSION 29 Figure 2.1: Eigenfunctions aciance, Correlation Length=1. XS = 2.53) a § ~e x 24 2.50). Thus, a. process Si 39) can be ¢ L tok sar clawne tir | Le Gy Wy Me a ms as detintéd by-eqitations (2.51) 2) shows the30 CHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES 23, KARHUNEN-LOEVE EXPANSION 31 WW | go J" ey Zo. 14 | & I ‘ “ | | ‘ 7 ~ EXPONENTIAL COVARIANCE Various Values of onk Correlation Length = 1 In| 0, 2 =1,2 Another process that is useful in engin ance kernel given by the equation ng applications has a covari- , sin(Q(x1 ~ e2)) C(z1i.22) = Soa runeated white noise process, whos spectral density vanishes outside a certain interval where it is con associated eigenfunctions and eigenvalues were obtained in explicit form by Slepian and Pollak (1961). They ere the angular prolate spheroidal wave functions and the radial prolate spheroida functions respectively. They have some quite interesting properties that are, however, beyond the scope of the prosent work (Slepian, 1964, 1965, 1968, 1977; Landan et.al, 1961 196: ‘A final remark is in order concerning the choice of the domain D of definition of the random process being investigated. Taking D to be the finite domain over which the process is being observed may often be the most obvious choice. Clearly, this choice does not induce the ergodic assumption which involves observing infinite length records. is by no means & handicap of this approach since the ergodic assumption usually introduced for convenience and is not necessary for the present study. If ergodicity is needed for some particular problem, then it may be recovered by extending the limits of integration in equation (2.33) to infinit40 CHAPTER 2. REPRESENTATION OF STOCHASTIC PROC SSES This modification may be convenient numerically if, for instance, an explicit olution of the integral equation is available for the infinite d nain and not nite one Numerical Solution nt of the Fredholn ection, a Galerkin typefprocedure is described for the solution equation (2.33) In Chapter IV, this procedure will be illustrated through its application to a curved geometry two-dimensional problem. Let Ay(x} be a complete set of functions in the Hilbert space H. Bach eigenfunction of the kernel C(xe1,x2) may be represented es fis) = Soa (2.74) with an error ey resulting from truneati This error is equal to the dif nation after the N® between the left hend sid 38). Substituting equation ( term, the righ srence and hand side of equation into equation (2.88) yields the following expression for the error = al | f cra.) helo) dee — dy abs) | ) Requitin, the follov rthogonal to the approximating spac ‘ing form, Equivalently [Clea x2) rien a] hslor Icons lt Ales) Ay (2) da] = 0 Denoting 23. KARHUNEN-LOEVE EXPANSION 41 HGENVALIE, Figure 2.15: Ayan: Exact and Approxi Eigenvalues Aq, of the Exponential Covariance Kemel, 1 ate values 4 (2.79) (2.80) and (2.81) equation (2.77) becomes CD=ABD (2.82) clements are Fanation (2.82) represents a generalized where C, B and D are three N-dimensional matrices whose given hy aqnations (2.7R)-(2.79) algebraic eigenvalue problem which may be solved for the matrix D and the eigaHVENRS Ay Backsubstinitmng bito equation (L72) VME the eigenfunc- tions of the covariance kernel. The preceding procedure can be imnplemented using piecewise polynomials as the basis for the expansion. With this choice42 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES of basis functions, the columns of the matrix D become the eigenvectors computed at the respective nodal points of the induced mesh, and the ¢j"* element of the matrix C becomes the weighted correlation between the process at nodes i and j. Note that both matrices C and B are symmetric positive definite, a fact that substantially simplifies the numerical solution. Figure (2.15) shows the exact eigenvalues of the kernel given by equation (2.39) and the results from the algorithm described above; note the excellent agreement. Obviously, the eigenvectors and eigenvalues computed based o1 the above scheme pr ent estimates to the true values. Certain properties of these particular estimates make this scheme computationally attractive. Specifically, the Galerkin scheme deseribed above can be shown to be equivalent -atment of the problem. This property ensures that the computed eigenvalues are a lower bound of the cortespond~ ingly numbered exact eigenvalues. This implies that the convergence of each eigenvalue is monotonic in N. Further, note that the accuracy in estimating genvalues is better than that achieved for the eigenfunctions (Delves Mohamed, 1985) oa variational & 2.4 Homogeneous Chaos 241 Preliminary Remarks It is clear from the preceding dis Karhunen-Loeve expansion requi process being expa verned, efficie sussion that the implementation of #! 1s mowing the covariance function oi ded. As far as the system under consideration is con- ‘implies that the expansion can be used for the ran However, it cannot be impleme for the solution process, since its covariance funetion and therefore the corresponding eigenfunctions are not known. An alternative expansion is clearly needed which circurnvents this problem. Such an expansion could involve a basis of known random functions with deterministic coefficients to be fuuudl by minimizing some of the error resulting fram representation. This should be construed as similar to the Fourier series solution of deterministic differential equations, whereby the series coefficients are determined so a3 to satisfy some optimality criterion. To clarify this important idea further, # useful modification of the problem suggested by in the operator equation. ite 2.4, HOMOGENEOUS CHAOS 13 equation (1.13) is noted. For this, equation (1.13) is rewritten as w= hiGlO).2 (2.83) where h{] is nonlinear functional of its arguments. In equation (2.83), the random processes involved have all been replaced by their corresponding Karhunen-Loeve representations, It is clear now that; what is required is @ nonlinear expansion of hi. in terms of the set of random variables €,(8). If the processes defining the operator are Gaussian, this set is a sampled derivative of the Wiener process (Doob, 1953). In this case, equation (2.83) involves fanctionals of the Brownian motion. This is exactly what the concept of Homogeneous Chaos provides. This concept was first introduced by Wiener (1998) and consists of an extension of the then obscure Volterra’s work on the generalization of Taylor series to functionals ( Volterra, 1913 ). Wiener’s contributions were the result of his investigations of nonlinear functionals of the Brownian motion. Based on Wiener’s ideas, Cameron and Martin (1947) constructed an orthogonal basis for nonlinear functionals in terms of Fourier~ Hermite functionals, Wiener’s Homogeneous Chaos was subsequently refined by Ité (1951) into what is known as the “Multiple Wiener Integral”, About the same time that this analytical and measure-theoretic development of the theory was being pursued, Murray and Von-Neumann (1936) were es- tablishing the parallel algebraic structure of rings of operators. Wiener’s ‘theory was further dovoloped through research efforts that led to a series of reports (Bose, 1956; George, 1959). Numerical implementation of the basic ideas as well as convergence properties were addressed in these reports. ‘This theory has drawn the interest of investigators in the fields of communication ‘Yasui, 1979), neuro-science (Palm and Poggio, 1977), e Jahodi and Ahmadi, 1983), statistical physics (Imamura et.al, 19 mathematics (Hida and Ikeda, 1965; McKean, 19 1978, 1979; Kallianpur, 1980; Engels, 1982) ngels (1982) and Kallianpur (1980) hav treatment of the Volterra series, the Wier ab) and ; Huang and Cambanis, Tn particular, Yasui (1979), attempted to develop a. unified series, the Cameron-Martin expansion and the It6 approach. They have concluded that the last three spproaches are equivalent and that they are superiar to the Volterra series n terms of their convergence properties and their applicability. In the same manner that the Homogeneous Chaos can be viewed as an orthogonal development for nonlinear functionals with Gaussian measure, the Discrete Chaos (Wintner and Wiener, 1943; Hida and Tkeda, 1965; Ogura, 1972) is,44 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES an orthogonal development with respect to the Poisson measure. Extensions to general measures have been investigated by Segall and Kailath (1976). jamal po $c d 24,2 Definitions and Properties (Let {6i(0)}32; be a set of orthonormal Gaussian random variables. Consider | the space ff of all polynomials in {&(@)}2, 'p Tepresent the set of all polynomials in f'p orthogonal to Py—1. Fi Fy, be the spacefspanned) by I',. Then, the subspace T, of © is called the of degree not exceeding p. Let ly fot) * Homogeneous Chaos, and 1 is called the Polynomial Chaos of order p. Based on the above definitions, the Polynomial Chaoses of any order p consist of all orthogonal polynomials of order p involving auy combination of the randoii- WTIBTES {6(0)}7E)- It is clear, then, that the mumber of Polynomial Chaoses of order p, which involve a specific random variable out es with p. This fact plays an important role in the sequel. Furthermore, since random variables are themselves functions, it becomes clear that Polynomial Chaoses are functions of functions and are therefore fumetionals ‘The set of Polynomial Chaogodis a linear subspace of the space of square- integrable random variables @, and is a ri snufieation Too) = Ty0)T (a). with respect to the functional must be is context, square integrabil ued to be with respect to the probability measure defining the random variables. Denoting the Hilbert space spanned by the set {&:(8)} by O(), the resulting ring is denoted by Soc), and is called the ring of functions generated by @(6). Then, it can be shown that under some general conditions, the ring @g(q) is dense in the space © (Kakutani, 1961). This means that any square-integrable random function (@ — R) can be approximated as closely as desired by elements from ®eqq. Thus, any element (0) from the space © admits the following representation, =o (2.84) DL apt py VplEos()s---$o0(8)) » “Pe SS mt tnem 01 where P(,) is the Polynomial Chaos of order p. ‘The superscript n; refers to the number of occurrences of &,(0) in the argument list for Tp(.}. Also, the double subscript provides for the possibility of repeated arguments in the argument list of the Polynomial Chaoses, thus preserving the generality of 24, HOMOGENEOUS CHAOS 45 the representation given by equation (2.84). Briefly stated, the Polynomial Chaos appearing in equation (2.84) involves r di of the set {&(0)}32;, with the ** random variable &(#) having multiplicity nig, and such that the total number of random variables involved is equal to the order p of the Polynomial Chaos. The Polynomial Chaoses of any order will be assumed to be symmetrical with respect to their arguments. Such a symmetrization is always possible. Indeed, a symmetrical polynomial can be ‘obtained from a non-symmetrical one by taking the average of the polynomial over all permutations of its arguments, The form of the coefficients appearing in equation (2.84) can then be simplified, resulting in the following expanded expression for the representation of random variables, inct random variables out + DS vali) (8) 85) aT a(E:,(0).2(8)) acatsT a{Gir (0), Gia(O), Sig) $ig(9)) + where T,(.) are successive Polynomial Chaoses of their arguments, the ©: pansion being convergent in the mean-square sense. The upper limits on the summations in equation ( ject the symmtery of the Polynomial Chaoses with respect to their arguments, as discussed above. The Poly nomial Chaoses of order greater than one have mean zero. Polynomials of different order are orthogonal to each other; so are same order polynomials with different argument list. At times in the ensuing developments, it will prove notationally expedient to rewrite equation (2.85) in the form where there is a one orate correspondence between the functionals W[}] and Tj, and also between the eoefi .. appearing in equation 5). Implicit in equation (2.85) is the assumption that the expansion cients a, and ay,46 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES 2,85) is carried out in the order indicated by that equation. In other words, the contribution of polynomials of lower order is accounted for first Up to now, and throughout the previous theoretical development, the syinbol d has been used to emphasize the random character of the quantities involved. It is felt that, although somewhat cumbersome, this notation underlines the fact that a random variable is a function defined over the space of events of which @ is an element. Having noted this. the symbol 0 will be deleted in the ensuing development whenever the random nature of fa certain quantity is obvious from the context, 'As defined above, each Polynomial Chaos is a function of the infinite sot {Gi}. and is therefore an infinite dimenstonal polynomial. In « computational setting, however, this infinite set has to be replaced by a finite one, In view seems logical to introduce the concept of a finite dimensional Polynomial Chaos. Specifically, the n-dimensional Polynomial Chaos of onder p is the subset of the Polynomial Chaos of order p, as defined above, which is a f the uncorrelated random variables &. AS n— ce, the Polynomial Chaos as defined previou: ered. Obviously. ‘he convergence properties of @ representation based on the n-dimensional nial Chaoses depend on n as well as on the choice of the out-of the infinite set. In the ensuing analysis, this choice based on the Karhnnen-Loove expansion of an appropriate random Since the finite dimensional Polynomial Chaos is a subset of the ( dimensional) Polynomial Chaos, the same symbol will be used for both, with the dimension being specified. Note that limit on the summations in equation (2.85 smials invalved: is rewritten ubset ill be for this case, the infinite upper is replaced by a number equal to For clarity, the in a fully expanded form, 26 2.87) ani Po(Gas61,62) + aaa Talay $2061) In view of this last equation, it becomes clear that, except lin a different indexing tionals U{,) and T[ are identical. In this regard, equation (2.87) can be recast in terms of [as follows mn, the f HG) = Goll + 22 + GoVs + GsVs + arts 24, HOMOGENE + GpWo + ary + Gols + from which the core dence between W[] and I ‘the term aziiPs(E2, 1,1) of equation (2.8% of equation (2.88) is evident. For example, identified with the term Wy Construction of the Polynomial Chaos | \ A direot ap ch to construct the successive Polynomial Chaoses isto start with the set of homogeneous polynomials in {&(8)} and to proceed, through a sequence of orthogonalization proced: a constant and it ean be chosen to be 1 The zeroth order polynomial is That is 9) ‘The frst order polynomial has to be chosen so that it is orthogonal ¢o all zeroth order polynomials. In this context, orthogonality is understood to be ‘with respect to the inner-product defined by equation (1.8). Since the set {G3} consists of zero-mean elements, the orthogonality condition implies ‘The second arder Polynomial Chaos consists of second order polynomials in {6;} that are orthogonal to both constants and first order polynomials. Formally, e second order palynomial can be written as Ta(Gyr€e) = da + an, + Make + Aynbibie 2.91) where the constants are so chosen as to satisfy the orthogonality conditions. The second of these requires that 2.92) 2.98) Allowing ig to be equal to iy the coefficients ai, and ai, as nnd ip successively, permits the evaluation of 2.94) ‘OUS CHAOS a i | |48 CHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES orthogonality condition results in ‘an be normalized Equation ‘with conditions of being ortho and second order polynomials. The fir GinSa)> = 0 2.100) on ot + iy + 0 2.101 The second condition implies & (2.02) Ei Gy Sis Sie (2.103) airdiy Binty + aindisny + Sraiaso Ein Sia Sis Sic fia) Gu & (2.104 Ts(GusSiargig) 6 & 24, HOMOGENEOUS CHAC 49 which results in 9 bixis Fis =O. (2.105) The above equations can be normalized by requiring that Aint = 1 (2.106) ‘Then equation (2.103) becomes Aiding + Aidiaie + Cabin + = 0 (2.107) Due to the Gaussian property of the set {G}, the following equation holds riadise + Gixisdinig + Site (2.108) Substituting for the expectations in equations (2.107) and (2.106) yields Andie + Aigdigig + @indinig + Biiabists + Sits Bints + Sirti 2.108) and axis [ Sitadians + + SinssFias F dizts (Bniadiats + Sirtadigig + Suis Sins + ding [ SinaSiate + Siatadise Sisey | = 0. (2.110) Substituting for ap from equation (2.101), equation (2.110) can be rewritten taints +5insaGinis | + Minis [ bintadi +oins { Sintints + Siisdits | = 0 (a1) From equation (2.111), the coefficients aise, iiss Ad @ing can be eval rr) aus = 0 )50 CHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES | 24. HOMOGENEOUS CHAOS eat squation (2.110) again, it is found that y & < TT &> a = 0 (2.13) alicia) aL tare juation (2.109) can be rewritten as z Equation (2.109) can be oe : am Guia tig + Seats) + Bisel is + Sits) + dary + (2.4) 7 fom which the coefficients aj,, aia, and ai, are found to be. i me | nod a, = ~Sats where 7.) denotes a permutation of its arguments, and the summation is aig (2.115) | over all such permutations such that the sets {Gj ,..n€i,} is modified by the ay permutation. Note that the Polynomial Chaoses as obtained in equations (2.80), (2.00), (2.98), (2.117) and (2.118) are orthogonal with respect to the Gaussian probability measure, which makes them identical with the corresponding noltidimensional Hermite polynomials (Grad, 1949). These polynomial ave boen used extensively in relation to problems in turbulence theory . fa Sa — Gi Sats ~ Ga bis ~ Gq Sisg» 2116) | (Imamura et.al, 1965e-b). This equivalence is implied by the orthogouality f ws) CS ofthe Polynomial Chaases with respect to the inner prochct defined by equ tion (1.8) where dP is the Gaussian measure e~ 2 €ag where Edenotes tie . iptlations, the fourth order Polynomial Chaos | Yetor of random variables {6j,}f_1. This measure is exactly the weighing Alter laborious algebraic manipulations, the § function with respect to which the Hermite polynomials are orthogonal in the I, sense (Oden, 1979). This fact suggests another method for constructing the Polynomial Chaoses, namely from the generating function of the Hermite polynomials. Thus, the Polynomial Chaos of order n can be obtained as The third order Polynomial Chaos can then be written a can be expressed as (2.7) Tel€iss See 8:0» 64) Sy Sia Gia Sie Gis Gia Gigtg — Six Sin Sie ~ Sis Sis Sint | — Eig Sig Fig — Siz Sia Finis — Gia Sis Si | | | 38 19 te (2.119) + Gixiabisie + Saabs + Sti Bauiation (2.119) can be readily evaluated symbolicall bolic manipulation program MACSYMA (1986). ‘Tables (2.1) - (2.4) diss play expressions for the one, two, three and four-dimensional Polynomial Chaoses up to the fourth order along with the values of their variances. The term W, in these tables refers to the quantity appearing in equation (2.86). It is readily seen that, in general, the n"* order Polynomial Chaos can beCHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES | 2.4 HOMOGENEOUS CHAOS Onder of th ynomial C jomogencous Cl v pao i | pel \ i {| Il Pop Onder of the [J Polynomial Cha 1 |" | Homogeneous © so Dimensional Polynomial Chaoses and ‘Theit VHOMOGENEOUS CHAOS | p Onder ofthe | Polynon | Homogeneous Chacs ¥ | ed ee + aT | 38 | ao] ‘ | | 40 | 2 i | 41 i | fia] | | 43] i | |) ram | |45| 6 16 | 2 Pog : | i | 6 Table 2.4: Four Dimens! sasional Polynomial Chaoses ancl Theit Variances; n= 4TIC PROCESSES 56 CHAPTER 2. REPRESENTATION OF STOC! ere obtained the two-dim ions corresponding to these polynomials MA programs shown in Figure (2.16) f e (2.17) for the four-dimensional polynomials, The exp 1p MACSY polynomials and Figu jon (2.85) represent the Gaussian component of the red aG ). Therefore, for the coefficient nussian process, this expansi 28 ‘ai, being the coefficients in the notion. ts probability swation (2.83). For a given non-Gaussian pro cd by : ‘distribution function, @ representation in the form given by equation (2.85) ‘an be obtained by projecting the pro This hieved by using the inner product defined by to determine the re 1 in devising efficien a ne Monte Carlo simulation method (Chorin, 191 3 on the successive Homogeneous Chacses, can be 2 This concept has been sto be {eltz and te coefficien rariance reduction technic 24, HOMOGENEOUS CHAOS oT * kronecker : *\ kdelta(x,y) ::=builag({x,y], if xey then 1 else 0)$ \s average : *\ average(p) ::= buildg( tpl, ( logexpand: super, norder:0, for i: 1 thrun do ( z:coeft(Log(p) ,og(x14])), if 2#0 then for j:1 thru z do yLjrorder]:xfi], order:ordertz ), hom:p/prod(y[4],4,1,order), if oddp(order) then 0 else (if order=0 then hom else if order=2 then homtkdelta(y [1] ,y(2}) else homtsun (kdelta(y [1] ,y(41)*prod(y (i ,k,2,0rder) /yli],4,2,0rder))))$ Table 2, MACSYMA Macro to Generate the Two-Dimensional Polynomial Chaoses and Evaluate ir Varian58 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES \s here? : *\ Gato} si for i:i thru 2 Gné:ind+t, 2 [ind] :expand( for 4:1 thru 2 do (for j:4 thru 2 (nd: ines, G2[ind] :expand(-dits (-ait# (gen, x(4]) ,x[5])/gen)))8 for a:l thru 2 do (for j:i thru 2 do Gor k:j thru 2 do (ng: ind+t, expand (-aif¢ (-aiff( itt (gen,x[1]) (JI) ,x[e))/gem))))8 2/2,4,1,20)8 sn, x(2]) /gen) )$ att (-aite ‘gen, x[8]) ,x{3)) x0) ,x(2])/gen 10) he Two-Dimen MACSYMA Macro to G J Chacses and Evaluate Their Variance. Table 2.5, (continued) Polynon 24, HOMOGENEOUS CHAOS 59 \w squared : *\ square? (nun) 2:2, load (Kronecker) , load (average), Load (h for j (pp:expana(2tj]6265]), plo) pp, sa:0, for u:1 while inte pla] :if Length(p rp (part (pla-t] ,2))=true plant] ,1))91 en average (pla-1]) lee nap(average,pla-t)), var2Cj] :ptae , print (j,var2(33 pp) ) 028 Table 2.5, (continued):, MACSYMA Macro Polynomial Chase o Gene ‘and Evaluate T ate the ir Verianee.60 CHAPTER 2, REPRESENTATION OF STOCHASTIC PROCESSES \+ kronecker : +) kdelta(x,y) : smbuildg( Ex, 4 x=y then 1 else 0)$ \w average : *\ average (p) builda( (pl), ( logexpand: super, 1 thraa do ( zicoefs(log(p) ,og(x(i])), if 280 then for j:1 thru z do ylj+order] :x{i], ordertordertz ), hom:p/prod(y[i],i,1,order), Af oddp(order) then 0 else (at order=0 hoz else af orde: en hom+kdelta(y 1] .y else homvun tay £1] ,y [4] eprod (y Del ,x,2,order) /y(i] 4,2, order))))$ Table 2, MACSYMA Macro to Generate the Fotr-Dimensional Polynomial Chaoses and Evaluate Their Variance, 24, HOMOGENEOUS CHAOS 6 \e hermes 18 for i:t thr Gind:ind+1, G4Lind] :expand(-dit#((gen,x[4])/gen) )¢ 4 a0 for i:1 thru 4 do (for j1i thru 4 do Gnd:ind+t, GéLind] :expand(-aaze (~aiee ( for i:1 thru 4 do (for j:4 thru (or ¢:j thru Gnd:inas1, [41),x€j]9/gen) >) 2D sexpand (ais (~aise (-aize ( “dif (gen, x[4]) ,xff1) 5x02) 5 /gen) )))8 Table 2.6 (continued):, MACSYMA Macto t Polynomial Chaoses and E Generate the Four-Dimensional Juate Their VariancCHAPTER 2. REPRESENTATION OF STOCHASTIC PROCESSES msquared: *\ squared(aus) ::=(builég( [nun] , Load(kronecker), prexpana(G4{j)*64(5]), then ay else map(aver: var4(}] :plun print(j,var4{j),pp) > ) ‘Table 2.6 (continued):, MACSYMA Macro to Generate the Four-Dimensional Polynomial Chaoses and Evaluate Their Variance. Chapter 3 STOCHASTIC FINITE ELEMENT METHOD: Response Representation == (4 3.1 Preliminary Remarks .ddressing the problem of, 's have involved isties of the response or implemen Second-Order (FORM/SORM) (Der-Kiureghian et.al, 1987a) algorithm conjunction with & finite element code. Monte Carlo simul have also been investigated towards improving their efficienc and Astill, 1972; Shinozuka and Lenoe, 1976; Polhemus and Cakmak, 1981; Cakmak and Sherif, 1984; Shiiozuka, 1987, jlet, 1987). However, due to sstem stochasticity eand order Spanos and stantial requisite computational the Monte Carlo simulation method is still used mainly to verify her approaches. The perturbation method (Nakegiri and Hisada, 1982: Liu, Besterfield and Belytschko, 1988) and the N Shinozuka and Nomoto, Ador ‘ak, 1987; Shinnanka umann expansion method jan and Malakian, 1980; Be re heen n top acceptable results for small random fluctuations in the material properties. Typical results from these two methods have been restricted to second order statistics of the response quantities. Obviously, this information does not risk assessment and failure aulfice for a mean analysis, In a different 6364 CHAPTER 3, SFEM: RESPONSE REPRESENTATION context, a stochastic finite clement method has been developed that couples response surface techniques with deterministic finite element formulations. ‘The cimphasis in these methods has been on developing efficient search algoritims for locating the poi ned the design point, at which the response surface is to be expanded in a first or second order Taylor series (Der-Kiureghian et.al, 1986, 1987, Cruse et-sl, 1988) and on formulating the problem using equivalent Gaussian variables (Wu, 1987; Wirshing and Wu 1987; Wu and Wirshing, 1987). Following this approach, a number of finite element simulations are performed, the solutions of which define From this approximation, the ‘Once located, the distance from an approximation to the response surface, design point is subsequently determined. this point to the origin can be used to obtain a rough approximation to the probability of failure, ‘The accuracy of this approximation deteriorates rapidly as the dimension of the space increases. A number of stochastic finite element codes have emerged from these research efforts, ‘They all ize a deterministic finite element code which is coupled with either a standard optimization algorithm or a search algorithm that is customized to implement the specific reliability concept being adopted. Of interest ate the codes NESSUS/EXPERT, being developed at the Southwest Research In- stitute in collaboration with NASA-Lewis Research Center (Millwater et.al, 1988; Millwater et.al 1989; Cruse and Chamis, 1989), and CALREL-FEAP. developed at UC-Berkeley (Linet.el, 1989). Note that alternative .e response surface that do not rely on a Taylor series expansion proposed in the Kterature, Of these, the polynomial approximation suggested by Grigoriu (1982) is worth mentioning, Ik relies on fitting « polynomial to a number of poluts on the response surface. Eacit se points, however, is obtained as the solution nulation. ‘Alternatively, methods based on the optimal feature extraction prop- the Karhunen-Loeve have been recently added to the arsenal of approaches to address the associated class of problems. The Karlunen-Loeve decomposition has been coupled with either a Neumann expansion scheme (Spanos and Ghanem, 1989} or a Polynomial Chaos expansion in conjunction swith a Galerkin projection (Ghanem and Spanos, 1990) to achieve an efficient implementation of the rendomness into the solution procedure. In this chapter the deterministic finite element method is first brief Then, the non-spectral methods for solving class of stoc! reviewed. 3.2, DETERMINISTIC FINITE ELEMENTS 65 tie problems are discussed. Finally, two spectral methods involving the Karhunen-Loeve expansion and the concept of Polynomial Chaos for stochas- tic finite element analysis are examined in detail. It is noted that the prime theme of this chapter is an expeditious representation of the system response itself; the determination of the statistics of the response will be addressed in Chapter IV. 3.2 Deterministic Finite Elements 3.2.1 Problem Definition In the deterministic case, the space 2 of elementary events is reduc gle element, coinciding with lis obvious, then, that this is a special case of the problem defined in Chepter I For convenience, and omitting the randomness argument @, equation (1.12) is rewritten as tual realization of the problem L(x) (u(x)] = f@), xed (1) In the following, two equivalent formulations for deterministic finite ele- ‘ment analysis are presented. They are the variational formulation and the Galerkin formulation. The stochastic finite eleme duced In the sequel is based on a Galerkin projection in the space © of random variables, However, the variational formulation is presented to draw ‘a phiysically appealing analogy between the mechanical energy of a system as given by its potential energy, and the uncertainty energy of a system, as given by its information-entro is meant to outline some formulation as intro: Note that the exposition in this section he basic concepts of the deterministic finite clement method that are relevant to the stochastic case; it is not meant to provide an account of the state-of-the-art of finite element techniques. Equation (3.1) can be viewed as a mapping from the space over which the response u(x) is defined to the space over which the excitation is defined. All the excitations dealt with in the sequel, es well as in most engineering applications, have finite energy. That is, 3.2) [, £007 ax <0 lb ‘The range of the operator L(x) is thus the space of all square-integrable functions, The domain of L(x), that is the space spanned by the solution, is elli-Be66 CHAPTER 3, SEM: RESPONSE REPRESENTATION obviously determined by the special form of L(x), as well as by the boundar conditions and the initial conditions associated with the physical problem. Assuming that L(x) is an m** order differential operator, let ©” denote t space of all functions thet are m times differentiable. Then, the solution space is some subspace of C™ whose elements satisfy the associated hoino- geneous essential boundary conditions of the problem. The exact nature of this subspace depends on the specific finite element formulation employed fand will not be dwelled upon any further. The solution function u(s) ean be expanded along a basis in this space, and equation (3.1) becomes uw [Ease] = YE) [91001] = fx, (3.3) svhere us the component of the solution u(x) along the basis funetion a), In the sequel, the above summations are trunc vated at the N** term and the rslom then is t0 compute the coordinates {u} ofthe response a(x) with ve-dimensional besis {o(x)}8 ional Approach, Variational prhiciples were introduced and studied well before the introdue- tion of the finite element method. The theory has become an integral part of functional analysis with @ solid mathematical foundation. The physical meaning of the corresponding minimization problem, although (quite helpful, is no longer necessary, as long as the operator deseribing the system satisfies certain conditions of jointness (Rektorys, 1980). In fact, it may be Shown that to every self-adjoint operator equation, is associated a quadratie functional whose stationary value coincides with the solution of the equation Physically speaking, a self-adjoint operator results from situations where reciprocity, as given by Betti’s law (Shames and Dym, 1985) for example, ‘Such is the case with most linear differential operators of ‘even order. Then, the solution to equation (3.1) is that function u(x) which ne functional Tex) = (LEO bod]. wld) — 204). HG), BA) ) denotes a si Indeed, setting the ble inner product 3.2. DETERMINISTIC FINITE ELEMENTS or fanctional derivative of T[v(3)] with respect to v(x) to zero, results in L(x) [ex] — fx) = 0 (35) Physieally, the functional Jv(x)] corresponds to a measure of the energy in the system, which assumes its stationary value at the true response. Substituting equation (3.3) into equation (3.4) gives Tpl)] = OY Le ey (LG0) [ore]. gi0 0] (36) 2D w ( F). gile)) Setting the variation of I{u(x)] to zero, and recognizing that u(x the stationary point of I[] gives, (x) at 2 (LG) lov 9562) ) = 2( 16) = 05 a7 Equation (3.7) can be expressed as a system of algebraic equations Lu=f (38) where Ly = (LG) igi) ) (3.9) f= (F(x). gi(x)) (3.10) Note that as given by equation (3.9), the basis {g:(x)}¥ must span an N- limuisiusel sulmpace of ©, Ab this point, may be noted that performing the integration indicated by equation (8.9) by parts, the operator L(x)[] can be split into two lower order operators asin Tay = (Lal) [ose) La(x) (9)(20] ) @u68 CHAPTER 3, SFEM: RESPONSE REPRESENTATION where now Ly[,] and Le! L(x} Whether equation (3.9) or equation (3.11) is used to compute Lj, the result is an equation of the form (3.8). Ordinarily, the set gi(x) is chosen to be the set of piecewise polynomials of order n+ discretization of the yproach just outlined is used later im this chapter to generate a set of algebraic equations from the governing differential equation. ‘The main point to carry over from this section to the stochastic case is thet the solution of the problem can be obtained as the element of a certain Hilbert space of admissible functions that minimizes a certain norm of the energy of the system. In Chapter IV, an analogy is drawn to this principle for the stochastic case, where the energy norm is replaced by an uncertainty @2.3 Galerkin Approach) A condition associated with the variational formulation of the finite ele- rent method is that the operator L{x)[] be self-adjoint. ‘This restriction excludes a clas of important practieal problems. Further, in sone case, the variational principle fora given problem may lack the intuitive physteal association with the energy principle. Thus, it may be advantageous or even necessary to resort to the Galerkin formulation of the finite element method (Zienkiewiez and Taylor, 1980). The method consists of expanding the response function along a busis of @ finite dimensional subspace of an admissible Hilbert space (Oden, 1979), and requiring that the error resulting from taking a finite number of terms in the expansion be orthogonal to another Hilbert space, the “test space”, whose functions are called “test functions". Usually, the test space is chosen to coincide with the admissible space, In other words, expandin dimensional basis {oi0x) eto of the form are two appropriate operators of lower order than ‘This choice for the basis nnain D, ‘The variational (xx) in equation (3.1) in terms of a finite of a subspace of the space ©” introduces an ex = SH) lol] = Fx) @. 3.2, DETERMINISTIC FINITE ELEMENTS 69 Requiring this error to be orthogonal to the subspace spanned by {9i(X) yields a set of V algebraic equations Cen, 0) ) = OF =1 (3.13) ‘This equation is equivalent to DY w(LGe}fou()], 95620) = (FO0)-95080)s N (3.14) which is simailar to equation (2.8). Here again, integration by parts can be used to enlarge the admissible space. The Galerkin method will form the basis for the spectral extension of the finite element method to problems involving stochastic operators. 3.2.4 p-Adaptive Methods, Spectral Methods and Hierarchi- cal Finite Element Bases It is customary in fini basis functions as repres element analysis to regard the cocflicients of the 1g physically meaningful nodal variables. degrees of freedom are it d through mesh refinement with the new coefficients representing the physical quantities at the new nodal points. Ob- viously, as more degrees of freedom are introduced, the approximation error diminishes. This is the so-called hemethod whereby the approximation error is reduced through successive mesh refinement. This mnethod constitutes the ost widely used version of the finite element method. Recently, however, fa new error reduction techniques with better convergence properties has been actively investigated. The so-called p-method consists of reducing the approximation error by using higher order interpolation within each element (Babuska, I. et.al, 1986). Thus, the seme discrete mesh is used in successive approximations. The successive interpolation bases over each element are .duced in a hierarchical manner permitting results from previous approx- imations to be efficiently used in computing higher order approximations. In the limit, as the number of finite elements is reduced to one, a global approximation is obtained. ‘This limitin rectal approximation. It has some obvious advantages when dealing with functions that are defined over domains that are either too difficult to discretize or too abstract for such a discretization to be intuitively feasible. Indeed, the spaceCHAPTER 3. SPEM: RESPONSE REPRESENTATION of random variables is just such a space. The measure associated with this space is a probability measure that, cannot be made to correspond with the Lebesgue measure used in spatial discretization procedures. Still, functions defined over this space ean be discretized using an adequate spectral measure. context, discretization refers to expanding the functions using a set of that are globally defined. An additional appeal of the spectral cient use of basis fune formulation is the fact thet global approximations can make 1a) basis functions. Interestingly, the set of basis functions used in the ensuing spectral stochastic finite element formulation is hierarchical and .ssociated solution procedure can be optimized to make use of this fact 3.3 Stochastic Finite Elements 3.3.1 Preliminary Remarks In this section three of the methods used for treating problems involvin random media SFP ftst discussed, Then, pug recently developed method are presented. The three methods have been selected due to their popul ‘with various investigators and their compatibility with the finite eTement method. The theoretical foundation and the numerical implementation for each of these mnethods are discussed. ining Tealizations of a random process, two important features can bbe observed. The frequency content of the random fluctuations and the magnitude of the fluctuations, The former feature can be used to situate the random process with respect to white noise. The broader the frequency content, the closer is the process to white noise (Lin, 1967). This feature reflects the level of corzelation of the process at two points in its domain The second feature reflects to, a certain extent, the degree of uncertainty the process and can be related to its coefficient of variation. he zange of applicability of riations and the range of AI] three methods to associated wit In view of that, # meaningful definition of ¢ fa given method is the range of coefficients of frequency content that the method can accommod bbe discussed are hase! on performing direct operations on equation (1.12) ‘To show explicitly the dependence of I] on the random process a(x, 6) this equation can be rewritten as (3.15) ‘atx 8) « x} ] [ufalx 0). xd} = Fe, 8) 3.3. STOCHASTIC FINITE ELEMENTS n 3.3.2 Monte Carlo Simulation (MCS) The usefulness of the Monte Carlo Simulation method (MCS) is based on the fact that the next best situation to having the probability distribution function of a certain random quantity is to have @ corresponding large population, The implementation of the method consists of numerically sim- lating a population corresponding to the random quantities in the physical problem, solving the deterministic problem associated with each member of that population, and obtaining a population corresponding to the random response quantities. This populstion can then be used to obtain statistics of the response variables, role in introducing the method tothe field of engineering mechanics. They with the MCS to achieve a more efficient implem ation of the simulation stems. ions, and various other2 CHAPTER 3, SFEM: RESPONSE REPRESENTATION translate a two-dimensional random plate problem into 2 format compatible with the finite clement method. Subsequent applications to problems of statics and dynamies involving random media were again carried out by Shinoauka and a number of co-workers (1980, 1985, 1986, 1987). Over « period of years the application of the MCS to problems of struc tural mechanics has involved one-dimensional Gaussian processes. Recently the digital simulation of non-Gaussian processes hiss been investigated (Yor mazaki, 1987). Further, results have also been obtained for the sirmulatior yeesses (Mignolet, 1987). As far as using the MCS in Jblems involving random media is concerned, the computational cost of the approach is apparent. Ordinarily, the MCS is used as a brute force technique for assessing the validity of other approaches. It involves numerically gener ating realizations of the random processes ax(x,@) and f(x,@) appearing, in equation (3.15), and proceeding with a deterministic equation to be solved for u(a(x,),x), for a fixed value of @. The procedure is repeated a mumber of times for different values of @ € 9. This is obviously a collocation scheme in the space Q of elementary random events. Realizations of the p are obtained, usually, either as local averages as discussed in section (2.2) or by representing the process as the response of a linear filter to a white noise excitation. The first approach is usually used when spatial randomness is involved, whereas the second approach is preferred for temporal random 3.3.3. Perturbation Method tion approach as applied to problems of random media is an jethod used in nonlinear analysis (Nayfeh, 1973; Jordan and Smith, 1977). Given certain sunoothness conditions, the functions and ‘operators involved can be expanded in a Taylor series about their respective mean values, To outline the method, it is assumed that the random process representing the system variability has boon modeled by r random variables and that the excitation is deterministic. In this ease, equation (3.15) becomes {L(x} + Ha(6),x}] [ula(o),x)] = Fe) (3.16) Expanding T1(a(9),x) and u(a(9).x) about their mean values and noting 33, STOCHA! IC FINITE ELEMENTS B thet ag(9) is a zero-mean random vector, leads to D(a).x) = Yo ald) gE Male).x) (3.17) fei a > x oH) 05) Sera O29 + and ula(@),x) = Y alg pulato).x) (3.18) + SE al) ao) pe ) u(a(0).x) + ae Far Byoajaj (9) Assuming stall random deviations of the variables a,(@) from their mean values, the contributions from all ax(0) are small compared to that of the average quantity. Then, substituting back into equation (3.16), a multidi- mensional polynomial in a4(9) is obtained. The right-hand side is obviously polynomial of order zero in o. Equating same order polynomials on both sides yields a set of equations to be solved sequentially for the successive derivatives of the response. The first two terms of this succession are L(x) [also] = F(x) (3.19) Lx) bgegpulalllex)) + Raat fa(x)] = 0 (8.20) The larger the magnitude of the random fluctuations is, the more terms should be included in equations (3.17) and (3.18). This task is considerabiy complicated, thus greatly limiting the applicability of the method. Further secular terms appear in higher order expansions. These are terms the mag- nitude of which increases with increasing approximation order, thus causing the expansion to diverge. The method has been applied extensively in recent years to problems involving random media. Nahgisi aud Hisud (1982) aad Hisada and Nakagiri (1985), and Liu et.al (1985, 1986, 1988) have applied the method to linear and nonlinear problems of statics and dynamics. Good results have been obtained for narrow band magnitude random fluctuations of smallSPEM: RESPONSE REPRESENTATION CHAPTER 3. Once several of the sequential equations (3.19) and (3.20) have been solved, the response process ean be written formally as, : Olen wlalO).x) = ax) + Dale) 5 eynlal0).x) + EE amas we u(a(@).x) + 1a(8) where all the partial derivatives are evaluated at the mean value (zero) of the random variables {a;(6)}{.1. The case where the random aspect cof the problem is modeled as a random process can be reduced to the previous case by recalling the discussion in Chapter II concerning local ‘ation of random processes. Namely, a random process can y ables corresponding to its spatial average over a number of subdomains in its domain of definition. Note that the number of random variables involved in this case is equal to the number ‘ubdomains over which the averaging is performed. Usually these sub- domains are the same as the elements used in a finite clement analysis. The amount o! ons required to perform the indicated operations becomes prohibitively large rather quickly as more terms are included in equations (3.17) and (3.18). For implementing the perturbation method into a finite element analysis, either the variational principle or the Galerkin approach can be used directly on equations (3.19) and (3.20), yielding a sequential system of algebraic equations to be solved for the successive Gerivatives of the response, Alternatively, these methods can be applied to equation (3.16), prior to performing the perturbation expansion, which when carried out, yields the same sequential algebraic equations as before In terms of the comparison criterion defined at the beginning of this chapter note that the higher the frequency of random fluctuations, the moze random variables are needed to represent the s, That is, more elements are required and the problem can become quite large, It is noted at this point the convergence of the series (3.21) is not guaranteed even for small levels of sauclom uctuations due to the presence of ocoular terme in higher order expansions, ‘The first term in equation (3.21) is obviously the average response, The second moment of the response process ean be obtained, at that least conceptually, upon multiplying equation (3.21) with its conjugate, by averaging. ‘The difficulties associated with that computation are such th 3.3. STOCHASTIC FINITE ELEMENTS only first order second moment statistics are usually computed; order ere refers to the order of the polynomial in 0;(x,8) included in the expansion, It is obvious that the method cannot be readily extended to compute the probability distribution function of the response process. 3.3.4 Neumann Expansion Method From an operator-theoretic perspective, equation (3.15) rellects the prablem of computing the inverse of a given operator. It is well known that when it exists, the inverse of an operator, called the resolvent, can be expanded in a convergent series in terms of the iterated kernels (Mikhlin, 1057). The theory was developed by Neumann and was further investigated by Fred- holm (1903). Subsequent extensions of the theory inchide the concept of generalized inverse (Nashed, 1976). The concept was applied to the solution of stochastic operator equations by Bharrucha-Reid (1959). Subsequent contributions were made by Adomian (1983) and Adomian and Malakian (1980). The treatment was largely theoretical until Shinozuka and Nomoto (1980) introduced this concept to the field of structural mechanics, The method proved the implementation of higher order terms in the expansion to be quite laborious. Later, Shinozuka and co-workers coupled the New- mann expansion with the Monte Carlo simulation method to produce an efficient algorithm. The Neumann expansion method consists of expressing the solution to equation (8.15) as the series jous gut ) [L7XGx) W(atx,8).x) | 169]. (8.22) ‘To guarantee the convergence of serie ing criterion be satisfied, (8.22), itis necessary that the follow- || L(x) Mat It is apparent that the algebraic manipulations incurred by including h order terme, even cocond order terme, in the expansion are quite involved Benaroya and Rehak (1987), analyzing a single-dogree-of-freedom system and using the symboiie manipulation program MACSYMA, were not able fo include more than the second term in the expansion. Yamazaki et.al (1985) suggested applying a Monte Carlo approach to equation (3.22) to76 CHAPTER 3, SFEM: RESPONSE REPRESENTATION Jmulate the processes a(x, 8) and hence the operator T(a(x,9),x). At this stage, the successive orders in the expansion can be obtained numerically. Note that with this implementation sequence, only the deterministic average operator L(x) needs to be inverted. However, the method still requires simulation, a fact that necessitates several runs of the simulated problem in order to assess the reliability of the results. Further, it can be seen that it js quite difficult to extend the method to obtain higher order moments than the first two. With regards to representing the random process, the same discussion presented for the perturbation method is pertinent, Namely, the processes are replaced by their local averages over the finite elements. In terms of implementation into a finite element code, either the variational or the Galerkin method may be used in equation (3.15). Specifically, replacing the random process by its local average over each finite element yields [L+m)U=F (324) where L is an n x n deterministic matrix, IT is ann x n random matrix function of the random variables {cx (@)}/.,; and F and U are n-dimensional vectors representing the excitation and the response, respectively. Here, m Genotes the number of degrees of freedom in the finite element mesh, At this point, a Neumann expansion of the inverse operator may be performed, leading to uo = [(L+my'F = yo [iin} ite (3.25) i=” 7 Further, simulation can be used to numerically generate realizations of the iatrix IT, producing a statistical population for the response vector U, from which various statistics can be obtained, Equation (3.25) is valid provided that Wnomy) ]- The aj" component of this matrix represents the cross-correlation function between the processes w(x.) and w(x.) taken at x = x; and x = xp respectively Let the domain D be subjected to a general oxternal excitation denoted by (8) In general the response vector u(x, #) is related to the external excitation through a transformation defined by the medium D and its physical prop- erties, resulting in equation (3.15) which is rewritten here for convenience Lx) + M(a(x,a), x) ] fuatxd), x= fe. 8). Further, let the domain be subjected to a set of boundary couditions scribed by the operator equation B(x.8) fulx9)] = 0, (3.28) where © is a random operator and OD denotes the boundary of the domain D, It is obvious that u(x,0) belongs to the Hilbert space of functions satisfying equation (3.27) in D and such 8) is satisfied on AD. However, in view of the discussion of section (3.1), this space may be extended, through integration by parts to include functions not smooth enough to be operated upon as required by equation (3.27). To avoid obscur- the notation and to be able to proceed without digressing to a specific problem, it is assumed that only one parameter of the medium is random fand that it appears as a multiplicative factor in the operator TI(a(x, 0), x) Also it io acoumed 1 the medium is subjected to a set of deterministic boundary conditions. Then, the covariat a single function Cyw( 1.32) become x2) reduces to and (3.28) [Lee + a(x,8)78 CHAPTER $. SFEM: RESPONS! 3.3. STOCHASTIC FINITE ELEMENTS 9 and (3.34) would involve a random matrix with fully correlated elements. ‘This Dix) {uex.#]] = 0 (3.30) approach forms the basis for the methods discussed at the beginning of this chapter. | pter. In sad, a(x,9) is exprossed in its Karhunen-Loeve expansion, as ee “ P ae - 3.3, in the form tdinssible space corresponding to the operators L(x) and R(x), as discus in section (3.1). Specifically, lot ¢i(x) be the set of piecewise polynomials enough to allow it to span the admissible space. This choice | a(x) = So dn ba aba 3:35) induces a discretization of the domain D and makes the coef : fhe generalized coordinates g(x) directly relate to the components Sor-0) at the nodes of the induced mesh. The response can be written as | where Ay , a(x) are the eigenvalue/eigenvector doublets corresponding to the covariance function of a(x), and & 38 a set of orthonormal random variables. Substituting equation (3.5) for a(x, #) into equation (3.34) give jescribed in sect Sk, + Se KI | (2.36) Substituting equation (8.81) into equation (3.29) and dropping the argument | . | Ry = [ [ 1G) glx) | gil ax m7 Sw (L(x) + a(x) RG) ] g(x) = FO) (3.32) Muliplying equation (8.82) throughout by gi(e) and integrating over D_ | dnl) (G6) i) | 9) x (338) [fies +e Ro) waco de} we — GH) | ol «a.09) bo £0) 108) de Note that with the choice of the set gi(x) as piecewise polynomials, the ntegrals aver D in the previous equations reduce to integral sits, Writing equation (3.36) for all values . set of N eqiations to be solved for the random reaponse at the L(x) ex) ] gi(eodx (334) nodes, x i , a6 1 : ' | fm+Saxmlunt (3.0) [. atx) (RG) 0s) } gi) ax | ws = f f U = J Db Jo where the matrices are defined by equations (3.37)-(8.89). At this stage the Note that if, at this point, a(c.8) is expanded along the same basis es | boundary conditions may be imposed on K and each of the K(? matrices ux. 6), in terms of ils pointwise realization, the left hand-side of equation | individually. A detailed account of this procedure is deferred to the next80 CHAPTER 3. SPEM: RESPONSE REPRESENTATION ion where the problem is again encountered. Equation (3.40) can be tiplying throughout by K~ to obtain [r+ Daa (3.41) Qh) = RK (a2) (3.43) Equation (3.41), governing the behavior of the discretized system, can be lly written as fo I+ Wi{e}])u= 8, (3.44) its arguments. Symbolically, equation -esponse vector 1 is a nonlinearly filtered where W[{G,)] 5s some functional (3.44) suggests that in general, the version of fy. A straightforward from equation (3.44) relies on performing » Neumann expansion for the se operator to derive ‘heme for obtaining the respouse vector r = [¥ i) wes] g (3.45) In terms of QU"), the response vector takes the form u Yas (3.48) Expanding equation (3.46) gives enn Y fm & QQ” L ast 7 an) EY Ge Haan + | 3.8. STOCHASTIC FINITE ELEMENTS a1 Note that in the limit as Mf — oo, the results obtained by the preceding approach are identical with the results obtained from direct Neumann expansion solution as described in section (3.3.3). The present formulation, however, provides an expression which is computationally tractable and amenable to automation, 3.3.6 Projection on the Homogeneous Chaos The development in the previous section may be considered as a useful modification of existing techniques. Yet, it is limited in its applicability by equation (3.26). Clearly, this restriction is not prohibitively severe However, given the range of applications encountered in practice it is de- sirable to devise a formulation which is more versatile. In addition, all the methods described previously lack the geometrical appeal that underlies the deterministic finite clement method. In this section, a formulation of the problem exhibiting this attribute is introduced as a natural extension of the geometrical concepts of Hilbert spaces that form the basis for the deterministic finite element method 2s are describ uations. ing point. B Two approac! embody these concepts and lead to identical final e Equations (3.29) and (3.30) constitute the a(x,0) in a Karhunen-Loeve series gives x) + 3 Ea gu(x) R(x) | u(x,e) = f(x,0) (3.48) Assuming, without loss of generality, that u(x, ) is a second order process, it lends itself to a Karhunen-Loove expansion of the form ax,8) = Yes x3(0) 660) (3.49) where fi Coe (a5) bla) cea = 65 Bf), (6.50) lb ad = 2 funni a sn edi Sh |82 CHAPTER 3. SPEM: RESPONSE REPRESENTATION Obviously, the covariance function Cy(2t1,22) of the response process is wot | known at this stage. ‘Thus, e; and 6;(x) are also not known. Further, u(x,0), not being a Gaussian process, the set )(6) is not » Gaussian ‘veetot. Therefore, equation (3.48) is of little use in its present form. Relying on the discussion of section (2.4) concerning the Homogeneous Chaos, the second order random variables ;(8) can be represented by the mean-square ‘convergent expansion x4(0) = of? To + where al? , are deterministic constants independent of @ and Pp( 4 «+ iy) is the p" order Homogeneous Chaos. Equation (3.52) is truncated after the P* polynomial and is rewritten for convenience, as discussed in equation | (2.86), in the following form, - (0) = Daf? wile] 8) where 2() and ,{{é,}] are identical to al!) ,, and Dp(Ea.-.-£i,). respectively, In equstion (3.53)- P denotes the total tumber of Polynomial Chaoses ‘used in the expansion, excluding the zero" order term. Given the number MY of terme used in the Karhunen-Loeve expansion, and the order p of Homogeneous Chaos used, P may be determined by the equation Table 2.1 shows values of P for combinations of p and M Substituting equation (3.53) for xx(9), equation (3.49) becomes wu(e6) = SS a0 Willer] be) (355) 38. STOCHASTIC FINITE ELEMENTS 3 OT ponder of the Homogeneous Chaos 7 15 1 70 20 sponding to selected values of M and p order of the Karhunen-Loeve expansion. der of the Homogeneous Chaos expansion, where esl) = ej bj(x) (3.56) Changing the order of summation in equation (3.55) gives water 3° 28? a6 u(x,0) = = jst B = DY WG deo) (357) = where, a) = (x) Substituting equation (3.57) 2,0), equation (3.48) becomes a 12 Lex) + DO Se an(x) ROX) | SO WAKE dex) = Foo), (8.59) where reference t o the parameter @ was eliminated for notational simplicity. ‘The response u(x, 6) can be completely determined once the functions d(x) are known. In terms of the eigenfunctions 0;(x) of the covariance function of u(x.8), d(x) ean be expressed as8 CHAPTER 3, SFEM: RESPONSE REPRESENTATION | 3.3, STOCHASTIC PINITE ELEMENTS s = Yl (3.60) | | + 1 yields Note that, if u(x, 0) satisfies homogeneous deterministic boundary conditions | con some section of the boundary Dy € OD, then Cyy(x1,%2) = 0 for et 7 * dey | WAKE L(x) guelx) gr(x) dx x 1. Thus, according to equation (3.50), bj(x) = 0 for | [PME f, LO) g(x) a) x si be deduced that the set by (x) satisfies the homogencous : boundary conditions imposed on the problem end s0 do the elements of the | + [R60 940%) a0 ax | set. di() by virtue of equation (8.60). Equation (3.59) may be written in 8 | Fa Jp “ J more suggestive form | =f fale, ts iy, N 3.65) | > 5 L 3 SS GSH LOO dbo + OY & WEN ROO doo = Fox). (8.61) 7 J, 0) aut) ats) dx (3.66 This form of the equation shows that dj(x) belongs to the intersection a ‘equation (3.65) becomes Then, equation (3.61) becom Mw WHER bu + So 610) Wy ICG) Rew | e L(x) aut) = fy tate 5.60) + ES GO) lle ES dy RO aoe) = £60 Note that the index j spans the number of Polynomial Chaoses used, while { a & the index A spans the number of basis vector Multiplying i equation (3.69) by Vma[{-}], averaging throughout and noting that i Equation (3.63) may be rearranged to give } . _e aa i : SUiLLE}] WnlGrH> = Bim 3.70) 1 FS ay (miter e9 once ne can derive q eee Al balan +e Say cel oe i S G00) 8/161 RO) onG0 | f@ (3.64 aay & ; fi UmifeH>, b= 1 = Law A 3.71)Table 3.2: Coefficient Two-Dimensional Polynoinial Chases; equation (3. Gijm = 1 For a large number of index combinations the coefficients cjym are iden tically zero. Ei implemented using the symbolic manip ‘am MACSYMA (1986). The non-zero values corresponding to L l 34 Table 8.8: Coeficent cf! = < & W, > Four-Dimen: 2 al Poly mil Chaoses; equatCHAPTER 3. SFEM: RESPONSE REPRESENTATION 3.3, STOCHASTIC FINITE ELEMENTS 89 t ip 5 | ra 1 12] fi | B 2 | 1 (4 u | B 1| 16 | n 2 : a | 20 1 20 | | 2 2] 21 23 1 25 9 26 2 26 10 28 2 or | nt 20 14 28 i {12} 31 6 29 | | [as] 2 2 30 i | fu B la 36 6 37 6 30 16 40 2 | f jaz 42 4 u 18| 43 2 45, 19 | 46 46 | 20 48 | v 21 49 1 4 22 6 | 23 | 2 | 24 2 a | 2) 58 le | 59 | 29 | 62 2] 60 | 20 63 2 au Fo fac 6 24 e lee fs2]| 66 6 63 33 @7 4 mf ule Le] ‘Table 8.3 (continued): Coeficiente] = <6 ¥. = olf Table 8.8 (continued): Cooficient = Four-Dimeusional Polynomial Chaoses; equati Four-Dimensional Polynomial Chaoses; equation (3.90 CHAPTER 3, SFEM: RESPONSE REPRE’ 3.3, STOCHASTIC FINITE ELEMENTS m1 the first four Homogeneous Chaoses are shown in Tables 3.2 and 3.3 for ‘two and four terms in the Karhunen-Loeve expansion (K = 2 and K’ = 4) respectively. Forming equation sf m, produces a sot (3.74) re G and R are block matrices of dimension NV x M. Their mj** blocks are N-dimensional square matrices given by he equations Gmy = bmj and Ray = Sem In equations (3.75) and (3.76), L and Ry denote N-dimensional square matri- ces whose Ai element is given by equations (3.66) and (3. In equation (3.74), h signifies the , respectively. N x M vector whose m** block is given 6 by the equation ealtrrll hig = (rr | |28| | 9 The N-imensional vectors dy can be obtained as the subvectors of the | 20 | solution to the deterministic algebraic problem given by equation (3.7: 2 2 Once these coefficients are obtained, back substituting into equation (3 | 22 | yields an expression of the response process in terms of the Polynomial 133 jet Chaoses of the form a j Phwoses of the form » 2 | ts | al 6 | v= Da Hit (3:78) | 2 50 I ia | \a] 60 ie | Equation (3.74) can also be derived by a different approach, which is « a 2 2 simple vatiation of whet has been presented above. In fact, if the compu | 29 63 2 | tations leading to equation (3.41) are carried out in the same manner as | \30] 6a 6 bore, an equation is obtained invelving a st of unknown random variables | |] | | pretnting the repane a tu todl plu of te Ante elewent meat ne ea im convenience equation (3.41) is sewtitten as | jaa] 6 ow | js] ne) j1+ Dawa” |usg <& WWD. cil = ‘equation (3.72) ‘Table 3.3 (continued): Coefficient cf) Four-Di nsional Polynomial ChaosSTATION 92 CHAPTER 3, SFEM: RESPONSE REPRES! Further, each element of the veetor u, being a second-order random variable, 53) to obtain can be expanded as in equation (3.5 = Xe WUE (3.80) or ec) whore ¢, is a vector of the seme dimension av, Substituting equation (8.51) back into equation (3.79) gives x l ZY 60) Q" | oy Wiel = & (3.82) Forming the inner product over © of equation (3.81) with {Vmi{§H]} Po + OV om Qe = Ft ng out the P indicated inner products yields an equation identical to equation (3.74) with @ =I. 3.3.7 Geometrical and Variational Extensions Geometry of the Solution Process ‘The development in the last section hinges on the fact that the Polynomial Chaoses form a complete basis in the Hilbert space ©, or equivalently thet the Homogeneous Chaos ring &gi¢) defined in Chapter I is dense in ©. The orthogonality property of the Polynomial Chaoses is not a requirement for the Galerkin procedure and is the result of the construction procedure followed in section (2.4.3). Unlike the deterministic case, whore the set ¢4(%) has to satiafy certain smoothness and boundary conditions, any complete basis in © can be used as an alternative to the Polynomial Chaoses in the development of the stochastic finite element method, Obviously, one such set is the set of the homogeneous polynomials formed by simple produets of 3.3. STOCHASTIC FINITE ELEMENTS 93 the oloments of the sot &. These polynomials are not orthogonal but they easier to generate than the Polynomial Chaoses. the previous development would yield an expansion of t u(x,0) in th Using such a set in response process ul, 8) Equation (3.84) is similar to equation (8.47) obtained in section using the improved Neumann expansion. In the N however, the optimality eritesion, in terms of some norm of the error, is not well defined. The Neumann expansion can then be vie case of a more general procedure, whereby the response process is exp in terms of a complete set in the space © of second-order random vari ables. Viewed from this g the methods presented in this Chapter lend themselves to interesting interpretatio the perturbation approach is a Taylor series expansion, the validity and nuce of which depend upon the behavior of the response process hborhood of its mean value, The Neumann expansion is an expression of the response in terms of a complete non-orthogonal besis in ace ©. Finally, the Homogeneous Chaos approach is an orthogonal Clearly, each of these methods is based on a different error minimization criterion. Unlike the deterministic ere non-orthogonal bases may have the merit of being flexible so as to satisfy the smoothness and boundary conditions, there is no merit fo sing a non-orthogonal expansion in @, since such conditions 1s fact 3.3.4) metrical perspective ‘will be observed in the next chapter by comparing the resulting expressions for the covariance matrices of the response, orthogonal exp: yields much simpler forms for the response statistics. Also it Chapter V thet the Homogeneous C b properties than the Neumann expat aos approach has better ¢ sion approach. In this context worth pointing out the similarity of the approximation scheme employed horo to the p-method of the deterministic finite element mot earlier. Indeed, the interval along the random dimension is not discretized, and a single element is used in the approximation. Instead, the refinement in the results is achieved by increasing the order of the approximation, disouesod netional polynomial94 CHAPTER 3. SFEM: RESPONSE REPRESENTATION ‘Variational Extensions It may be observed that the development of the Homogeneous Chaos expan: sion may be considered as an extension of the Galerkin finite element method to the stochastic case. Indeed, the error resulting from a finite exp: made orthogonal, in the sense of the inner product in , to the subspace spanned by the approximating basis. A question that arises in this context involves the possibility of developing an analogous theory based on the extension of the variational principles to the stochastic case. The variational approach in the deterministic case usually expresses the constraint that the true solution to @ given problem is that function which minimizes a certain norm or measure of the energy released by the system. The stochas parallel to this energy norm would be a measure of the stochastic energy dissipated by th Stochastic energy may be interpreted as bei the uncertai the system under consideration. Such an ‘uncertainty is well known in the context of information theory and it has an associated measure, That is the information-entropy of the system. Indeed, a related variational principle can be formulated based on the concept of maximum entropy. It may then seem plausible to develop a stochastic variational principle based on the concepts presented above. At this point, however, there is no indication as to how to assemble these ideas into a coherent theory Chapter 4 STOCHASTIC FINITE ELEMENT METHOD: Response Statistics | J. th nappa antoinenale 4.1 Reliability Theory Background An important objective of a stochastic finite element analysis of an engineer- ing system should be the determination of a set of design criteria which can be implemented in a probabilistic context. In other words, it is important that any method that is used for the numerical treatment of stochastic systems be compatible with some rationale that permits a reliability analysis. Traditionally, this problem has been addressed by relying on the second order statistical moments of the response These moments offer useful ical information regarding the response of the system and equations vill be developed in an ensuing section for their determination. However second order moments do not suffice complete reliability analysis. For this, higher order moments and related information are required. In this context, the response of any system to a certain excitation may be viewed as a point in the space defined by the parameters describing the system. In other vrords, for every act of auch parameters, (lie sesporise is uniquely determined 8 @ function of the independent variables. When these paramoters are garded as random variables, the response function may be described as 8 point in the space spanned by these variables, defining @ surface that corresponds to all the possible realizations of the random parameters. The96 CHAPTER 4. SFEM: RESPONSE STATISTICS ere the parameters are modeled as r din C ion of a given system, it is usually andom processes may be reduced, Il, to the case of random variables. For the safe quired that the response, or some oper measure of it, be confined to a certain region in this space, termed the sale region, Formally, this condition ean be expressed as otal 2 0 4d) (42) is the equation of the failure surface in the €-space. Of special interest in reliability assessment is the failure probability, Py ‘that equation (4.1) is not satisfied. ‘This equals th some probability measure, of the region in the not satisfied. That is, That is, the probability volume, with respect to pace over which (4.1) is ee. (43) where Jg(g) is the joint probability distribution of {61 ion (4.3) in engines &). The difficulty ng practice is obvious, Specifically, the -gration is zarely available in an analytical form. Further, the 202 is usually quite large to the extent that the requisite intogration can seldom be carried out either analytically or numerically eller and Stix, 1987). Finally, realizations of the set {&;} are often searee and do not permit inference about their joint distribution beyond thetr joint second order moments. Grigoriu et.al (1079) have suggested ng the selection of this joint distribution as a problem of decision theo ond proposed a number of optimality criteria to resolve this issue. Tl problem remains, however, of efficiently impiementing these distributio into numerical algorithms. This is a matter that persiste limitations on the amount of usable probabilistic information in using equ domain of in poses seve nese computational and theoretical difficulties, it hecame losin. able to introduce a standard measure for reliability analysis, Such a measure 1 be based on readily available second moment properties of the random variables involved and would reflect a safety measure associated with the tion of the system, The reliability index was introduced as an attempt 41. RELIABILITY THEORY BACKGROUND to meet these needs The basic idea behind the reliability index concept, is the fact that the less uncertainty there is concerning the limit state of a system, the safer the system is. In other words, the greater the mean of the limit state relative to its standard deviation, the less likely is the response to wander beyond this state. «As an intuitive measure, the reliability index was first defined as (Freudenthal, 1947; Freudenthal et. al, 1966; Cornell, 1963; 1981; Thoft-Christensen, 1982, 1983; Melchers, 1987; Madsen B= (44) wwhete Mf denotes the safety margin whieh isthe difference between resistance and load effects, denotes its expected value, and gas its standard Note that, if the random variables {6} are Gaussian and the limit state fonction given by equation (4.1) is linear in its arguments, the margin would have a Gaussian distribution. Further, 9, as defined by equation (4.4), would be equal to the distance from the mean value of M to tho hyperplane representing the limit state. For nonlinear limit states, the indard deviation of the safety margin is related to higher order moments of the system random variables which often are not known. In order to be able to use 8, as given by equation (4.4), the limit state surface is usually incarized using the first term in a Taylor series about some point in the Espace, In this context, it ie well known (Ditlevsen, 1981) that the mean value does not represent the best linearization point. Also note thet if function g{] satisfies equation (4.1), so does any function that is a power of gf], as well as any function that shares the same zeros with g{]. This so called lack of invariance problem was resolved by introducing a new reliability index (Hasofer and Lind, 1974) defined by the equation deviation, Bus, = myn mit 6 x )? Of! (x= ) where Ly denotes the limit surface and Cy is the covariance matrix of the design variables {&;}. Given the set {6:}, zz is uniquely determined as ¢ shortest distance to the limit state surface. Thus, it is seen that the optimal (45) point for the Linearization of the limit state function, is the point on that surface that is closest to the mean value after the design variables have been transformed to a set of uncorrelated variables. A deficiency of the Hasofer- Lind index is a lack of omparativeness (Ditlevsen, 1981) associated with the fact that any surface that is tangent to the failure surface at the design point98 CHAPTER 4, SFEM: RESPONSE STATISTICS has the same Gz. ‘Transforming the design variables to a new set having the Gaussian distribution (Rosenblatt, 1952) resolves this issue. When the failure surface is a hyperplane, manifested by the linearity of equation (4.1) he Hasofer-Lind index coincides with the previously defined index given 1y equation (4.4). The design point obtained in this way can be shown to be the point of meximum failure likelihood when the design variables are uncorrelated and Gaussian (Shinozuka, 1983). Given the reliability index, for Gaussian variables, the failure probability may be approximated by the expression Py % O(-Bue) (46) where @ is the normal probability function. Equation (4.6) is exact for linear limit states and becomes approximate for nonlinear limit states. ‘The the first order second moment method as reviewed herein icity and ease of implementation. However, the method has some serious shortcomings, (Schueller and Stix, 1987), associated with the replacement of the failure surface by a tangent hyperplane: the magnitude of the error is hard to estimate but is known to increase drastically with increasing number of design variables. main advantage of is its sf 4.2 Statistical Moments 4.2.1 Moments and Cummulants Equations From the preceding review it becomes clear that the statistical moments of the response not only can provide bounds about its expected range, but they can also provide meaningful information about the reliebility of the random system. Thus, it becomes desirable to develop equations for the determi- notion of the statistical moments, up to an arbitrary order, of the response process. In this context the general series representation of the response, associated with the stochastic finite element method presented in Chapter ILL, is quite well suited for an efficient computation of the corresponding statistical moments. 3 vllaed eithe figs the improved Neumann expansion or from the Hom resented by the jogeneous Chaos formulation, the response process ve may be rep- w= ay + ay Ma (6) + aaa, Marnl€) + - 4.2, STATISTICAL MOMENTS = YS Mier Mon rslG) » (a7) & ‘where Tly..19(€) isa k order polynomial in the variables {E45 €a4) and ay.g are deterministic vectors. Specifically, in the improved Netmann expansion, the polynomials TI are simply the homogeneous polynomials given by the equation Thay .rv() = TT &, a Chanses the Homogeneous Chaos formulation, these are the Polynom) Tyg del) = Pryare (ror on Gas) (49) In equation (4.7), indicial notation is used, and a repeated index implies summation with respect to that index over its range. MeCullagh (1984 1987) deseribes methods to obtain the cummulant generating function for a polynomial of the form given by equation (4.8). The r** component of u, ut?) can be written as uO = POE 0) here PC) is the operator indicated by Then, following McCullagh (1987), the e 0? is given by the equation tation (4.7) as applicable to ul") ant generating funetion for Kyo) = FP ng, aan) where ng represents cummmulants of g and eX?" is an operator yielding upon expanding, Pe 1a yy PO ad yay POO 4 lL (4.12) ing the Specifi- The result of the operation of e*"” on ng is obtained by compou action of the individual components appearing in equation (4. it is noted that Ing = 0 (4.13)100 CHAPTER 4, SFEM: RESPONSE STATISTICS (4.14) a mg = of? of), al?) RIMM, eto, (4.16) where Ait =F Sh test Fest ~ Hein ig the generalized cummulant of order (x Relying on equation (4.12), the generalized cummulants o can be obtained. Then, ordinary cummulants and moments may be deduced. Note that symbolic manipulation packages can be quite helpful in carrying ‘out the symbolic algebra involved in these equations. Alternatively, crass moments between elements of the response vector U For example may be obtained by direct manipulations of equation (4.7 fhe m-+-n order moment involving ul” and ul) is F oa Ma ay Hhoa(€) | > 7 Note that equation (4.17) involves deterministic tensor multiplication and momials of orthonormal Gaussian variates, which may be the recursion verages of pol expedited by usin fon the right hand side of equation (4.18) in, the algebraic manipulations involved The expectation of the prod cariables. A involves (& — ulation progran for the symbolie m 4.2, STATISTICAL MOMENTS 101 1d products of two-dimensional Polynomial Chaoses up to a certain .ows the code corresponding to the four-dimensional Polynomial Chaoses, It is noted that a trivial modification of these codes js required to permit the evaluation of averaged products of Polynomial Chaoses of arbitrary order, By including additional terms in the expansion (4.7), and random variabies in the set {&;}, any level of accuracy for any order. Figure (4.2) order moment can be achieved.102 CHAPTER 4. uilag(O if xey then 1 else 0)$ yl, \x average : *\ (pl, goxpand: super, puil order:0, i: dthrua do ( zicostt(log(p) ,log(x[i])), Af 20 then for j:1 thru z do hon:p/prod(y[1] ,4,1,order), (order) then 0 olse jeorder] :x{i], acy Ctl yl prod (y Ce] ,k,2,ordes i],4,2,order))))¢ ‘Table 4.1:, MACSYMA Two-Dimensional Polynoxial Chaoses. acro to Average Products of 42, STATISTICAL MOMENTS \« nerm2 : *\ ind:08 ato} :18 gen:exp(-sum(x(1]°2/2,i,1,2))8 for i:1 thra 2 do nd: ind+t, G2[ind] :expand(-ditf (gon,x{i])/gen) )$ re i:t thru 2 do (or j:i thru 2 do Gndsinaet, Galina) :expand (-dis# (-dif# (gon,x{4]) ,x{J])/gen)))$ for 4:1 thra 2 do for j:i thru 2 do k:) thru 2 do nd:ind+1, G2[ind) : expand (-ditt (ait? ( wait (gen,x[1]) ,x[3]) ,x0K))/gem))))8 for ist thru 2 do or j:d thru 2 do (for k:j thru 2 do (for 1:k thru 2 do (nd: indét, 2[ind] : expand (-dift (-aift aie ( ditt (gen, x[1]) L5}) x{k]) ,x[21)/gen ))2)8 (t Table 4.1, (continued):, MACSYMA Macro to Average Products of ‘Two-Dimensional Polynomial Chaoses, i |104 CHAPTER 4, SPEM: RESPONSE STATISTICS | \s product? : *\ product 2 (aus) :=(ouildg( Coun] , Load( ker), load (average), Load (hera2), for 4:1 thru mum do (for j:i thru nun do Gor ij thru mun do (pp expand (G2(i] +6205) G21K)), pio] for ail vhile integerp(ph (aan, J:if Length (pln and invegerp(part (pla~1] 2) )*true and length (part(p[n-1} ,2))=1 en average(p{n~1]) else map(average.? 1)*false do af c2[i,j,k]#0 then print (i,j,k,c2l ‘Table 4.1, (continued):, MACSYMA Macro to Aver ‘Two-Dimensional Polynomial Chaoses. sdpp) ))) 8 Products of 42, STATISTICAL MOMENTS 105 \« kronecker : *\ kdelta(x,y):=buildg( [x.y], Sf x+y then 1 else 0)$ \x average : *\ average(p) ::* builag( Cp], (Aogexpand:super, order:0, for i: i thru a do ( z:coett log(p) ,2og(x{11)), 1f 20 then for jit thru z do y{jrorder] :xLiJ, order:order+2 ), hon:p/prod(y {il ,i,t,order), Af oddp(order) then 0 else (Gf order=0 then how else Sf ordere2 then nedoita (y12 ,y (23) else hontsun (kdeltaCy [1] ,y[4)) wprod(y [kl ,k,2, order) /yLi] ,4,2,0rder)))) ‘Table 4.2:, MACSYMA Macro to Average Products of Four-Dimensional Polynomial Chaoses.SFEM: RESPONSE STATISTICS [APTER 4. 1,428 thru 4 do nd:ind+4, GéLind] :expand(~ for i:i thru 4 do (for j:i thru 4 do Gnd:indet GALinal sexpand (-dite (-aift (gen, x[31),x 1 thru 4 do (tor J: thre 4 do (tor ie: thru 4 do Gnd:tnd+t, 4nd) :expand(-ait# (-ai#¢ £4 (gen,x[1]) ,()]) » x04 )/gen))098 for i:1 thru 4 do (for j:i thra 4 do (for k:j thru 4 do (gor Lik thru 4 do Gind:ing+t, G4 ind] exp: (nats malts gen,x[i)) ,x1j1) sxe) aff (gen, x[4])/gen) 8 1)/gen)))8 £11)/gen)}2))8 ‘Table 4.2, (continued):, MACSYMA Macro to Average Products of Four-Dimensional Polynomial Chaoses. 4.2. STATISTICAL MOMENTS 107 \x product : *\ product4 (mua) Q oad Ceronecker) doad(everage), oad (herad), for i:1 thru nun do (for j:4 thru nun do (for k:j thru mun do (pprexpand (G4L1] 64 (j]+G400), pl0l:pp, 22:0, for a:1 uhile integerp(pla-i pla) :if Lengeh(pte-1])=2 and invegerp(part (p [n-1] ,2))=true and Length(part(p(e~2] ,1))=1 then average (p{a-11) else aap(average,pta-11)), Ali, j,4] :plnal, {i,j,4]¥0 then print(i,j,x,04Ci,j,4] pp) 2) 1008 =(ouildg( {nun , ‘Table 4.2, (continued): MACSYMA Macro to Aver u-Dimensional Polynomial Chai ge Product 4.2.2 Second Order Statistics As it has already been discussed in section (4.2.1), the second order statistics of response quantities are of particular importance in reliabi ‘These can provide preliminary estimates of the values of the reliability index: and the probability of failure. ‘The stochastic finite element method as described in the Chapter ITT can be used quite efficiently to produce accurate approxima statistics. Specifically, incorporating the improved Neumann expansion method described in section (3.3.5) in the procedure of ‘he previous seetion, one obtains the following equation for the mean of the analysis.108 CHAPTER 4. SFEM: RESPONSE STATISTICS response process, s = [r+ QQ? (4.19) MMM +E EEE , am ae G = ge” (4.23) Expanding equation (4.22) yields Re =| 14 ¥ SF [Q’aXe +.Q%Ga" +6Q"Q] l =i Moa Mat . EEE SS agar [aa Qhare 29 QQ" GQ” + Qt Q” GQ" Q! +2 GQ” gH Q® + 6Q® Q4 Q® Q”| + (424) 43, APPROXIMATION TO THE PROBABILITY DISTRIBUTION 109 Equation (4.24) can be simplified by relying on equation (4.18) which, for the the sixth order term becomes, (4.25) Alternatively, certain properties of the Polynomial Chaoses can be used to simplify considerably the calculation of the associated response statis- tics, Specifically, recalling that these Polynomials have zero-mean, the mean 1 is given by the equation response vect f= = dy (4.26) where the notation conforms to that introduced in section (3.3.6). Further, relying on the orthogonality of the Polynomial Chaoses one can derive the following simple expression for the covariance matrix of the response, Rea = Do <¥s[{eH) Wl}! > dy a (427) where the superscript H over a vector denotes its hermitian transpose. In this regard recall that values of the variances of some of the Polynomial Chaoses, appearing in equation (4.27) are displayed in Tables (2.1)-(2.3), ‘whereas MACSYMA symbolic manipulation programs to compute them are shown in Tables (2.5) and (2.6). 4.3 Approximation to the Probability Distribution ‘The moments computed according to the previous section can be used either directly to estimate probabilities of failure, or in conjunction with some expansion, stch as the Edgeworth expansion (McCullagh, 1987), to approx- imate the probability distribution function of the response process. An. alternative approach to obtaining these probability distributions consisno CHAPTER 4, SFEM: RESPONSE STATISTICS of divectly estimating the coefficients appearing in their expansion, thu bypessing the moment calculation stage. As a consequence of the gener nm into which the response process was cast, equation (4.7), it is com ceptually possible to obtain approximations to the probability distribution ‘a, Grigoriu (1982) has called attention to methods for estimating the probability distribution of polynomials as weighted avorages of some judi. distributions. A different approach is followed that the entire probabilistic information he response process is actually condensed in the coefficients of mn (4.7). Note that using the Homogeneous Chaos formulation the polynomials in equation (4.7) are orthogonal. Further, assume without es of generality that they are normalized. Then, multiplying equation and averaging gives Dy (428 WIP = Woody where w ig used to denote any element of the vector u. The expected value jn equation (4.28) may be formally written as feu (4.29) syl6) Pu gli €) du dE = 35. here py gl g) an [K-+1) mensional joint dstsbuton density of Err tfsiy function can be expanded in 4 multidimensional Edgeworth need Eben tantee eee, — a where (2, €) is the (K+ 1)-dimensional Gaussian distribution derived fron hhe covariance matrix of the vector €, augmented by the response w. That (0.8) waren, (431) : 2 = 432) and 7 | (433) 43. APPROXIMATION TO THE PROBABILITY DISTRIBUTION 111 Clearly g represents a first order Polynomial Chaos, itis orth other Polynomial Chaoses. Consequently 32) can b oe ygonal to all the ce matrix in equation (434) Ik where Ix is the K** order ide ity matrix, oy is the stands 3s, and a; is the coefficient of g; in equation (4.7) equations (4.7) and (4.30) into eqnation (4.29) gives leviation of the (4.35) Faia, SE baons es [0 Pasnral€) Poa d(€) Pron aplte§) (08) dud = a. Note that the integrands in equation (4.35) involve products of multidimen- sional polynomials with the Gaussian distribution. Thus, the value of the integral is equal to sums of moments of the multidimensional normal dis nition given by equation (4.31). This distribution is completely determined. by the joint second order moments of the response process and the variables {Ei}. These moments can be readily obtained by multiplying equation (4.7) by & upon averaging. Equation (4.35) can be cast in a matrix form Ab= (4.36) ‘where b is the veetor of unknown coefficients in equation (4.30), a is the vector of Imown coefficients in equation (4.7), and A is a rectangular matrix Note that the polynomials used in equation (4.90) have K+ 1 variables to éecommodate the response u, whereas the polynomials used in the expansion (4:30) have only K variables. Therefore the number of unknowus is greater {tian the number of available equations. Tn this regard, additional equations uuay be obtained by requiring the vector § to be jointly normmal. That is, | wo Paglteg) du = (6) 5 (437)ne CHAPTER 4. SFEM: RESPONSE STATISTICS where (€) is the K-dimensional normal distribution of independent ides tically distributed, zero mean Gaussian variables. Using these addition: conditions, equation (4.36) can be augmented to involve a cnefficient ma the augmented matrix equation, the vector b can be found and used in equation (4.30) to express p, g(t §). Finally, note that 1g the vector b, the marginal distribution of the response iso be obtained from equation (4.30) by integration over the which is square. Solvi E-space, That is, pa glin$) dé (4.38) 4.4. Reliability Index and Response Surface Simu- | lation roductory section to this Chapter, the generalized reliability index is the shortest distance to the limit state surface from the mean design variables. The point on the limit surface corresponding to this Shortest distance was deseribed as being the most probable failure point. Equation (4.7) with the left hand side replaced by taure may be regarde ‘as an explicit expression of the failure surface in the space spanned by the $+ ‘This surface is obviously nonlinear. Standard optimization algorithms may bbe used to compute the reliability index and the corresponding probable failure point. The Karhunen-Loeve expansion may then be used to recove the physical value of the failure point in the space spanned by the random Ss liself ‘Computing the reliability index, however, is unnecessary, as the calew- lation of the probability distribution of the response is itself within reach ‘The latter is best evaluated by simulating points on the response surface Clearly, once the coefficients for the representation of the response process in the form (4.7) have been computed, it becomes possible to obtain point ton the response surface tsing simulation techniques. In fact, to each real &,} there corresponds a unique realization of the As defined in the in ization of the set {Eo vaponse process (0), Clearly, this procedure yields » serople population for the response process which can be used to estimate its probability den function, In this context, Non-parametric kernel density estimation methods (Beckers and Chambers, 1984) are particularly useful, Chapter 5 NUMERICAL EXAMPLES 5.1 Preliminary Remarks ‘The methods introduced in sections (3.3.5) and (3.3.6) are exemplified in this chapter by considering three problems reflecting some interesting appli- cations from engineering mechanics. These problems involve a beam with random rigidity, @ plate with random rigidity, and a beam resting on a random elastic foundation and subjected to a random dynamic excitation In addressing these problems, it is reminded that the ultimate goal of a stochastic finite element analysis is the calculation of certain statisties of the response process. These statistics can be in the form of either statistical moments, or probability distribution function, or some other measure of the reliability of the aystem. As a first step in the solution procedure, the varia- tional formulation of the finite element method is used to obtain a spatially diserete form of the problem. Following that, the Neumann expansion for the werse, as given by equation (3.46), and the Polynomial Chaos expansion, a8 given by equation (3.81), are used to derive a representation of the response process. Statistical moments and probability distribution functions are then obtsined as discussed in Chapter 1V 11314 CHAPTER 5, NUMERICAL EXAMPLES Figure 5.1: Beam with Random Bending Rigidity Under Uniform Loads Exponential and Trisngular Covariance Models. 5.2 One Dimensional Static Problem 5.2.1 Formulation wulli beam shown in Figure (5.1), of len static transverse lo Consider the Euler Bi clamped at one end and subjected to a determinist P. It is assumed that the bending rigidity w = BI of the beam, which nnvolves the modulus of elasticity E and the inertia I is the realization of a Gaussian random pr occupied by the beam. It is also assumed that the randon s w is represented by its mean value W and its covariance spatial domai bending rigid function C(x, x2). The Jn energy V’ stored in the beam can be written as [| o(2) etx) de (61) -e o() and e(:r) represent the 1d the strain as a function of the beam. Assiming the material to be linear, and making 52. ONE DIMENSIONAL STATIC PROBLEM 16 of Hooke's law, equation (5.1) becomes (52) where 1n(z, 8) is the random bending rigidity as described above, and u(x, 9) is the random transverse displacement of the beam. It is reminded here that 40, where 9 denotes the space of random events. At this point, u(7, 6) may be expanded, as a function of x, along a basis of piecewise polynomials as follows . ule.) = Yo veld) plM(e) (3) where JV is the total number of degrees of freedom in the discrete model. In this equation, p{")() is a vector of polynomials that are identically zero except over a subdomain {* where they are of order ne such that certain compatibility conditions are satisfied on the boundaries of i. Further, w.(8) is a vector of unknown random coefficients, as indicated by the argument 4, representing the transverse displacement and slope at: some poiit in (© The subscript ¢ on the superscript n provides for different order polynomials over different subdomains. These subdomains represent the induced finite flement mesh over the beam. Note that the ord used in this problem equation (5.2) to remais 5.2) wives rm of the polynomials ds to be at least equal to two for the integrand in finite. Substituting equation (5.3) into equation Sy / Ina similar manner, the work pet the form 1d by the applied force may be put in Vim Ya) ff vie) He) de (55) ot ‘The Minimization of the total potential energy (V — V’ ) can be expressed116 CHAPTER 5. NUMERICAL EXAMPLES mathematically as av-V) Bul) Substituting equations (5.4) and ( into equation (5.6) gives Emo f (Geto ) wee (Golo) fe Me) Ha) dey = yy N 5.7) Evaluating equation (5. or all s € [1, N], leads to Ku-f (58) where Ka =f ( Zn ce) in co) uy = Ue(4) (5.10) and b= f we) f0) ae (6a) () have bounded support, the inte- laced by integrals over ti x, these integrals involve the -e equations may be \ding supporting subdomains. Furd fandom process to(zr,8), Whos tion is not suitably defined for e purpose of carrying out the This problem can be resolved by expanding w(2,@) in its truneated Karhunen-Loeve series. Specifically corres. =) + Y &l6) Ve fale) 5.19) where §:(8), Az an defined in section (2.3). Substituting fox (2,8), equation ( fala) are 8) becomes Y a@QKuse, (5.13) 5.2. ONE DIMENSIONAL STATIC PROBLEM uy where 5.14) and 5.16) In the remainder of this chapter, the argument @ will be dropped except where it is needed to emphasize the random nature of a certain quantity The integration in equation (5.14) may be performed either analytically if the eigenfunctions of the covariance kernel are known, or using a quadrature scheme if the eigenvectors have been computed numer tial points. To expedite the calculation of the above vatinble is introduced involving the local coordinate where 2; denotes the coordinate of the left end of the clement (e). Then, equation (5.14) ean be viewed as a combination of integrals of the 1 = where n is some integer. Further, the covariance kernel of the random process nodeled using both the exponential model defined by equation (2.39) and the triangular model defined by equation (2.56). For the exponential model. g for je(2) from equation (2.51) and (2.52), equation (5.18) after substitatus CHAPTER 5. NUMERICAL EXAMPLES cos( zy (rie dr, keven 5.19) where a = L/2 and ¢ = 1/b, and b is the correlation length. he triangular kernel, the eigenfunc |, and equation (5.18) becomes are given by equations (2.62) and (2.6 Relying on equations (5.19) or (5.20), closed form expressions for the ele- ments of the matrix K defined by equation (5.14) are derived. To simplify the algebraic manipulations, equation (5.13) is rewritten as } 5.2, ONE DIMENSIONAL STATIC PROBLEM ug Due to the boundary conditions imposed at the clamped end of the beam, the first two elements of the vector u are equal to zero. These elements represent the displacement and the slope at the fixed end, respectively. Substructuring the matrices in equation (5.21) to refle Jag UW oe it ESL ab af he substructuring is performed such thet up represents the displacements and the rotations of the beam at the free nodes, The 2x 1 vector gt corresponds to the reactions at the fixed end and may be recovered once the vector uy is computed. Expanding equation (5.24) gives [r+ Daae]u = T+ PeQe)m =e Upon solving equation (6.25) for uz, equation (5.26) yields a direct expres- sion for the unknown reactions. In the numerical implement preceding analysis which p duced the following numerical results, piecewise cubic polynomials were usec That is, ne =3, Further, the beam was discretized into ten finite elements, resulting in twenty two ( sedom, 2 Results Improved Neumann Expansion Following the discussion in section (3.3.5), the response uz may be ay | Qe | ee | @ average response may be expressed as in equation (4.19) and the covari- ce matrix of the response as in oquation (4.22). Figure (5.2) shows theNUMERICAL EXAMPLES CHAPTER 5. jon Normalized Standard Ds onential Covariance: o7.maz Figure 5.2: Beam Tip Defie ation vers Bending Rigidity Standard Devia 0.297; Neumann Expansion Solution. nt Along the Beam Exponential Figure 5.3: Standard Deviation of the Displ Numerical Values Determined Only for the Ten Nodes: Covariance; Neumann Expansion Solution; oy = 0.2, 5.2, ONE DIMENSIONAL STATIC PROBLEM 121 - eT Bae | SORE en | om 4 | 002 Standard Deviation of the Displacement Along the Beam; for the Ten Nodes; Exponential Figure 5.4 Numerical Values Determined Only Covariance; Neumann Expansion Solution; ogy = 0.3. BENDING RIGIDITY, Beam Tip Deflection Normalized Standard Deviation v Triangular Covariance; or, Bending Rigidity Standard Deviation; 0.298; Neumann Expansion Solution.mo | | | re 5; Standard Deviation of the Displacement Along the Beam; | Numerical Values Determined Only for the Ten Nodes; ‘THangular Covariance; Neumann Expansion Solution; 7; | Standard Deviation of the Displacement Along the Beat: | Triangular | merical Values Determined Only for the Ten Nodes: Covariance; Neumann Expansion Solution; gr = 0.3. ONE DIMENSIONAL STATIC PROBLEM 123 Figure 5,8: Beam Tip Deflection Normalized Standard Deviation versus Bending Rigidity Standard Deviation; Exponential Covariance: 0.297; Polynomial Chaos Solution. standard deviation of the displacement at the tip of the beam, 0, against the standard deviation of bending rigidity of the be several values of the order p of the Neumann expansion and for various values of the order M of the K-L expansion. Figures (5.3)-( 1 deviation of the response along the beam, again for various values of p and ‘These figures show the results corresponding to the exponentially decaying nce model. Figures (5.5) Its for the triangular model, plotted oer, for 5.4) show the standar 5.7) show the corresponding re Projection on the Homogeneous Chaos As discussed in section (3.3.6), the response process may also be expanded as im equati elt + ¥ af! Multiplying equation (6.28) by [{&i] and taking the mathematical expec24 CHAPTER 5, NUMERICAL EXAMPLES 4.2, ONE DIMEN SIONAL STATIC PROBLEM 125 10 ] i Seer) | Pinar of he Homey | } 5 | w+ ssa es LE | lene | ys | | Boo | | Figure 59: Beam Tip Deflection Normalized Standard Deviation versus | Figure 5.11: Linear Interpolation of the Nodal Values of the Vector e; of Bending Rigidity Standard Deviation; Triangular Covariance; o7mae = Equation (5.29) for the Beam Bonding Problem, ¢ = 0,...,5; Displacement 0.298; Polynomial Chaos Solution. Representation; 2 Terms in K-1. Expansion, p= 01,2 "EXPONENTIAL COVARIANCE 7 -XPONENTIAL COVARIANCE Figure Linear Interpolation of the Nodal Values of the Vector ¢; of Equation (5.29) for the Beam Bending Problem, i Representation; 2 Terms in K-L Expansion, Figure 5.10: Linear Interpolation of the Nodal Values of th Equat for the Beam Bending Problem, i = 0,1,2; Displaceme Representation; 2 Terms in K-L Expansion, M = 0,1. 9; Displacement26 CHAPTER 65. 1 5.1% Equation (5. presentation Linear Interpolation of the Nodal Values of the Vector ¢, of 14; Displacement 4: Line of Equation (6.2% Interpolation of the Nodal Values of the Vector ¢; for the Beam Bending Problem, ¢ = 0,1,2; Slope ation; 2 Terms in K-L Expansion, M = 2; p = 01 5.2. ONE DIMENSIONAL STATIC PROBLEM 5.18: of Equation Representation Figure 5.16: of Equation Representai | EXPONENTIAL COvARIARICE | a“ DDD Meanteechon Lincar Interpolation of the Nodal Values of the Vector ¢: (5.29) for the Be nding Problem, i = 0,..,5; Slope 2 Terms in K-L Expansion, M = 2; p= 0,1,2. Linear Interpolation of the Nodal Values of the Vector (5.29) for the Beam Bending Problem, 0,49: Slope on; 2 Terms in K-L Expansion, M = p = 0,1,2,3:128 CHAPTER 6. NUMERICAL EXAMPLES of Equation 5.29} for the Beam Be 2 Terms in K-L Expansioy Problem, § = 0,034 12,34 polation of \ding Problem, 1 = 1; Disp 1e Nodal Values of the Vi =2; P=3,6,10,15, Linear Interpolation of the Nodal Values of the Vector ¢; ) for the Beam Bendin Expansion, M Slope accent, 129 | Figure 5.19: Linear Interpolation of the Nodal Values of the Vector e of Equation (5.29) for the Beam Bending Problem, i = 2; Displac: Representation; 2 Terms in K-L Expansion, M = 2; P =3,6,10, 16 | wo 1 | Figure 5.20: Linear Interpolation of the Nodal Values of the Vector ¢ of Equation (5.20) for the B ing Problem, ¢ = 3; Displacew Representation; 2 Terms in K-L Expansion, M = P =6,10,15.130 onal. va Figure 5.21: of Equation Representation; 2 Terms Figure 5.2 of Equation Representation; 2 Terms in K-L Expansion, Mf (CHAPTER 6. = ‘ci a L NUMERICAL EXAMPLES Linear Interpolation of the Nodal Values of the Vector ¢: 5.29) for the Beam Bending Problem, i = 4; Displacen 1 K-L Expansion M 2; P = 6,10, 15 EXPONENTIAL COVARIANCE Line 99) for the Beam Bending Problem, 7 = 5: Displacement ______J Interpolation of the Nodal Values of the Vector 6,10,15, 5.2, ONE DIMENSIONAL STATIC PROBLEM ne TAL COVARIANCE 131 Figure 5.23: Linear Interpolation of the Nodal Values of the Vector ¢: ‘of Equation (5.29) for the Beam Bending Problem, i = 6; Displacement Representation; 2 Terms in K-L Expansion, M = 2; P = 10,15, 008 | TRIN ute puna ot oS 7 | oo 4 a | Figure 5.24: Linear Interpolation of the Nodal Values of the Vector ¢: of Equation (5.29) for the Beam Bending Problem, i = 7; Displace Representation; 2 Terms in K-L Expansion, Mf = 2; P = 10,15,132 CHAPTER 5. UMERICAL EXAMPLES tation, or equivalently requiring the representation error to be orthogonal to the space spanned by the approximating Polynomial Chaoses, gives iD [35 3 a® | os = «2 limb = (6.29) where & = 1 and Q‘ = I Recalling the discussion in section (3.3.6) it becomes clear that equation (5.29) leads to a set of algebraic equations to be solved for the vectors ¢,. Figures (5.8) and (5.9) show the standard deviation of the displacement at the tip of the beam versus the standard deviation of the bending rigidity. Comparing figures (5.2) and (5.5) to figures (58) and (5.9), it is observed that a third order approximation with the Polynomial Chaoses achieves a convergence comparable to that of a fourth order Neumann expansion. The vectors e: represent the magnitude of the projections of the response process u(z@) onto the spaces spanned by the successive Polynomial Chaoses W[féi}]. Figures (6.10)-(6.13) show these projections for the displacement process, corresponding to the exponential covariance model, for a value of the coefficient af variation of the bending rigidity equal to 0.8 and for a corxelation length equal to 1. Note the rapid se in the magnitude of the projection on the higher order polynomials. Figures (5.14)-(5.17) show similar results pertaining to the slope process along the beam. Due to the term <éq Wil{é-}] Ws[{E-}]> in equation (5.29), the resulting system of equations, although sparse, is coupled. This fact causes the projections to require updating as more terms are included in the summations of equation (5.29). The convergence of these projections with increasing order of the Polynomial Chaos used is shown in Figures (5.18)-(5.26) for the displacement process and in Figures (5.27)-(5.85) for the slope process. Observe the excellent convergence achieved by the leading projections with only a second order expansion. Note that these leading terms provide the largest contribution to the response process. The figures suggest that a reasonable approximation may be obtained by assuming, for higher order polynomials, equal contributions from polynomials of the same order. This assumption may then be relaxed sequentially as more accuracy is sequired. Once the cosficients in the Polynomial Chaos expansion have been computed, a statistical population corresponding to the response process can be readily generated upon noting that to each realization of the set of random variables {€,}, there comesponds a realization of the r vector. Realizations of these random variables are obtained from realizations 5.2. ONE DIMENSIONAL STATIC PROBLEM | | 133 Figure 5,25: Linear Interpolation of the Nodal Values of the Vector ¢: of Equation (5.29) for the Beam Bending Problem, i = 8; Displacement Representation; 2 Terms in K-L Expansion, M = 2; P = 10,15. of the random process representing the material variability using equation (2.20). For Gaussian processes, these variables are uncorrelated zero-mean } Gaussian random variables with unit variance. Using non-par: estimation techniques, as noted in section (4.4), the probabili fimetion of the response can be approximated. Figures (5.36)-(5.43) show the probability distributions, obtained as described above, using two, and four terms in the Karhunen-Loeve expansion for the variability of the system. Up to third order Polynomial Chaos is used in the associated computations, y distribution Monte Carlo Simulation To assess the validity of the results obtained from the analytical methods, and to test their convergence properties, the same problem is treated by the Monte Caslo sizmula hod, Realizations of the hencling rigidity of the | been are numerically simulated using digital filtering techniques (Spanos and Hansen, 1981; Spanos and Mignolet. order twenty is designed to match the spectral densi | expected, only the fst coffins ofthe filter fs non-negigible, effectively34 CHAPTER 5, NUMERICAL EXAMPLES 135 Linear Interpolation of the Nod of Equation (5.29) for the Beam Bending Pro Representation; 2 Terms in K-L Expai inear Interpolation 29) for the Bean Bending Problem, Expansion, M = 2; P = 3,6,1 the Nodal Values of the V Jal Vaiues of the Vector ¢; oblem, i = 9; Displacement 10,15. = 1; Slope Representation 0,15. | | | | Figure 5.28: Linear Interpolation of the Nodal Values of the Vector ¢; of Equation (5.29) for the Beam Bending Problem, i = 2; Slope Representation: 2 Terms in K-L Expansion, M = 2; P = 8,6,10,15. EXPONENTIAL COVARIANCE ™”==" oon os ‘odal Values of the Vector ¢ of Equation (5.29) for the Beam Bending Problem, i = 3; Slope Representation; 2 Terms in K-L Expansion, M = 2; P = 6,10,15,136 CHAPTER 5, NUMERICAL EXAMPLES ONE DIM ‘SIONAL STATIC PROBLEM 137 6.908 EXPONENTIAL COVARIANCE: | ooo 4 EXPONENTIAL COVARIANCE ~ sae | an-{ \ = | =| | | = 2 po) Be > | ‘ camenel a Nove Nope 32: Linear Interpolation of the Nodal Values of the Vector ; of Equation (5.29) for the Beam Bending Problem, 0,15. 2 Terms in K-L Expansion, M 10,15 Figure 5.30: Linear Interpolation of the Nodal Equation (5.29) for the Beam Bending Problem, i 2 Terms in K-L Expansion, M = 2; P 6; Slope Representation: vy EXPONENTIAL-COWARIANCE falues of the Vector ¢; of [F Figure 6.33: Linear Interpolation of the Nodal Values of the Vector ¢: of 5; Slope Representation 1 (5.28) for the Beam Bending Problen Expansion, M = 2; P = 10,15 Equation (5.29) for the Beam Bending Problem, 2 Terms in K-L Expansion, M = 2; P = 6,10, 5. i=7; Slope Representation138 CHAPTER 5. NUMERICAL EXAMPLES | 5.2, ONE DIMENSION Equation 1 Bend 2 Terms in K. 10,15. Figure 5.35: Linear Interpolation of the Nodal Valu Equation (5.29) for the Beam Bending Problem, i 2 Terms in K-L, Expansion, M =2 P = 10,15. polation of the Nodal Values of the Vector e of oblem, = 8; Slope Representation ues of the Vector ¢; Slope Representat L STATIC PROBLEM 139 | | | | | | o | S| Figure 5.36: Probability Density Function of the Displacement at the Tip of the Beam Using 10,000-Sample Monte Carlo Simulation and up to Third Order Homogeneous Chaos; Two Terms in the K-L Expansion, Monte Calo Simaitin Fig ‘Tai of the Probability Density Function of the Displacement at the Tip of the Beam Using 10,000-Sample Monte Carlo Simulation and up to Third Order Homogeneous Chaos; Two Terms in the K-L Expansion.NUMERICAL EXAMPLES CHAPTER 5. ago | | | | vey j ‘Senda Devon | | | 2 ol Figure 5.98: Cumulative Distribution Function for the Displacement at the Tip of the Beam Using 10,000-Sample Monte Carlo Simulation and a Third -L Expansion. Order Homogeneous Chaos; Two Terms in the on | | | om] Toor BEB os | | | oss 4 et Figure 58% Tail of the Cumulative Distribution Function for Displacement at the Tip of the Beam Using 10,00Sample Monte Simulation and « Third Order Homogeneous Chaoss Two Terms in the KL 5.2. ONE DIMENSIONAL STATIC PROBLEM ML z Figure 5.40: Probability Density Function for the Displacement at the Tip of the Beam Using 10,000-Sample Monte Carlo Simulation and up to Third Order Homogeneous Chaos; Four Terms in the K-L Expansion. Figure 5.41: Tail of the Probability Density Fimetion for the Displacement the Tip of the Beam Using 10,000-Sample Monte Carlo Simulation and up to Third Order Homogeneous Chaos; Four Terms in the K-L. Expansion.we CHAPTER 5. NUMERICAL EXAMPLES yielding a first oF This agrees with the fact that the process being imulated, which c .ce function given by equation 930), is a first order Markov process. The filter is then used to simulate five thousands realizations of the bending rigidity. For each of the simulated realizations, the associated deterministic problem is solved and a data ban for the response process is generated to compute its statistics. The process 's repeated for various values of the standard deviation of the bending rigidity opr. For cer in the neighborhood of 0.3, the problem of deal ‘vith realizations involving negative values of the bending rigidity becomes crucial, This problem is a consequence of the Gaussian assumption which wr the process EZ. During the simulstion, # Email fraction of the realizations with negative BY is excluded. In essence, @ Jan distribution is used for BI to the extent that equation matel 1 for the corresponding &. In physical ‘est, the coefficient of variation corresponding to a 1 0.2, s0 that the Gaussian ass truncated Gau 2,20) is only approx problems of related in medium is usually less ¢ plots as the results obtained ‘om the simulation are shown on the same opriate, The result ‘om the analytical anal 5.3 Two Dimensional Static Problem 5.3.1 Formulation In this example, one edge and is subjected to a uniform in-plane tension along its opposite edge. The problem is depicted in Figure (5.44), of the plate is assumed to be the realization of a two-dimensional Gaussian fan value E and known covariance function he external excitation is deterministic and of unit magnitude. ‘The same problem is then formulated and solved for the plate with curved ‘own in Figure (5.45), thus demonstrating the applicability of this stochastic finite element, method to problems with 1 thin square plate is considered that is clamped along The modulus of elasticity random process with known C(x.x2). Further, itis assumed that arbitrary geometry In treating nt variational formulation is ved than the one used in the previous example. This is done to demo: strate the floxibility of the Karbunen-Loeve expansion, and its compatibility jrith various formulations of the finite element method. Let the domain 4 tf the plate be discretized into N’ four-noded quadrilateral finite elements, his problem a slightly dif 5.3. TWO DIMENSIONAL STATIC PROBLEM 143 wf — + 4 / | | SthiudBaiscin= 03 Figure 5.42: Cumulative Distribution Function for the Displacement at the Tip of the Beam Using 10,000-Sample Monte Carlo Simulation and a Third Order Homogeneous Chaos. Four Terms in the K-L Expansion, each element having eight degrees of freedom. The strain energy V* stored in each element A® can be expressed as - 5 |, O70) €00) aa, (5:0) where dA is a differential element in A*. Further, o(x) and e(x) denote the stress and the strain vectors respectively, as a function of the | he location x within each element, Assuming linear elastic material behavior, the stress ay be expressed in terms of the strain as, o=De (5.31) where D® is the matrix of co titutive relations. Here o and ¢ are the vectors cof stress and strain as given by the equation oF = [on Or Tea} (5.32) f = [en tam | (6.3)uaa CHAPTER 6. NUMERICAL EX. Figure 5.43: Tail Displacement at the Tip of the Beam Using 10,000-Sample Monte Simulation and Expansion. jure 5.44: Plate with Random Rigidity; Exponential Covariance the Cumulative Distribution Function for 4 Third Order Homogeneous Chaos. Four Terms in the K- C Mode 5.3. TWO DIMENSION: Figure 5.45: Plate with Random Rigidity, Exponential Covariance Model. where oz, is the stress along direction 2; and ¢,, is the strain along that same direction. For the plane stress problem considered herein, D® is given [1 be wet = E(%0)PP, (6.8) oo ere P* is a deterministic matrix, 1¢ is the elemental Poisson ratio, and 6) is the elemental modulus of elasticity. The two dimensional dis placement vector u(x.) representing the longitudinal and transv placements within each element may be expressed in terms of the nodal displacements of the element in the form dis. u(x,0) = HE 0) 5.35) the local interpolation matrix, U*(é) is the random nodal sponse vector, and ry and ry are local coordinates over the element. For146 CHAPTER 5. NUMERICAL EXAMPLES the rectangular elements used herein (x = te) n= BP ne (6.36) ecranglat denen long the ortho ys the dimension ofthe side ofthe etang us A lietions and ie eepaon aatix P(r, 7) ean be expressed 7 Tm 0 M0 Ns 0M 0] (at) Heir) = |g xy 0 Na 0 Ns 0 MJ Ni = ni(l—72), Ne = ra(1—") Ng = viva Ny = (1-n)(l—7) (5.38) Substituting equation (5.34) into equation (5.80) gives (5.39) 1 r 1 e(x,8) dat ee 5 (x, 6) eh (x,8) PE elx,6) da wes fe hin an element is related to the displacements, longitudinal The strain w ‘and transverse, through the relation | (5.40) | nan Lan da J Using equation (5:35), equation (5.40) is rewritten as e(x,@) = Be US (5.4) 5.3. TWO DIMENSIONAL STATIC PROBLEM ur Substituting equation (5.41) back into equation (5.39) leads to 1 1p , ve FUT ff Blr1.12,8) BT rr) PE BE rir) lies dra U (6.43) where |J¢| denotes the determinant Jacobian of the transformation that maps an arbitrary element (c) onto the four-noded square with sides equal to one. ‘The integration is then performed over this master element. In performing the numerical integration, a four-point quadrature rule is used, and a procedure identical to the one described by Akin (1982) is employed to compute the value of the Jacobian at each integration point. The total strain energy V is obtained by summing the contributions from all the elements, This procedur ve SUT LL Btrry 0B ar) Pew rr) dr dre" The local representation of the response is related to the global repre tion through the following transformation Mgr (5.45) ‘where C? is a rectangular permutation matrix of zeros and ones reflecting the connectivity of the elements and the topology of the mesh. Using equation 545), the following expression for the total energy stored in the system is obtained UTKU, 5.46)48 CHAPTER 5, NUMBRICAL EXAMPLES K=So7Ke (647) 7) may be performed without computing the matrices C*, by ), The next procedure, as described by stage in the computations involves solving the integral e associated with the covariance kernel. That is, Cler,m:zasua) falaarve) dx din (5.48) s example is defined by the equation nifty ~ \ea-val/b2 5) here b, and by are the correlation distances in the x and 22 direction Square Plate - Analytical Solution assuming thet Substituting equation (5.49) for the covariance kernel, (Px (5.50 lution of the equation reduces to th the solution of equation an) dyn where | 5.3. TWO DIMENSIONAL STATIC PROBLEM The solution of equation (5. 52) is the product of the individual solutions of the two equations AP ha) Au) dr 5. and (ya) dye Differentiating each of equations (5.54) and (5, 8) twice, two second order differential equations are obtained along with their associated boundary conditions. The solution of the first of these equations produces the follo eigenvalues and normalized eigenfunctions, and In equations (5.56)-(5.58}, the symbol w; refers to the solution of the follow: ‘ranscendental equation 0 fori odd and (5.59) wy + cy tam (wia) = 0 for i even The solution of equation (5.55) is identical to equations (5. with150 CHAPTER 5. NUMERICAL EXAMPLES cq replaced by 2. Note that if ex = ¢2, then to each eigenvalue there corre, spond two eigenfunctions of the form given by equation (5.50). The second function being obtained from the first one by permuting the subscripts. In this case, therefore, the complete normalized eigenfunctions are given by equation + faye] 660) In the expansion of the random process, the terms are ordered in descending order of the magnitude of the eigenvalues An. Curved Plate - Numerical Solution subdividing the domain A ofthe pate ito 2 Anite elesents AP, equation (ats) becomes Ae fatan) = Sof Cleuussensn) farm) dat. 6) Interpolating for the value of the unknown function within an element in terms of its nodal values results in the following expression y) = HY(rim) fe (5.62 where H¢(rs, 12) is the interpolation matrix in terms of local coordinates "1 ‘and ra, and ff is the vector of nodal values for the unknown function asso- tiated with element (¢). For this particular problem, bilinear interpolation fs used over four-noded quadrilateral element. The matrix H*(r1,72) is then given by the equation Hr) =t | G-nG-m) Gtr 5.69 i (endtn) G@-na+n)] Substituting equation (6.62) and peeforming # transformation from global to local coordinates, equation (5.61) becomes x dn frlerus) = Of Cleswse AP, (5.04 5.3, TWO DIMENSIONAL STATIC PROBLEM 181 e |J*| is the Jacobian of the coordinate transformation. A system Of algebraic equations is obtained from equation (5.64) by requling the The result is a ae equation identical to equation (2.82) cD =ABD where now the j** column of D nodal points and snfunction calculated at the is the j** ig (66) Mattices © and B are obtained by assembling matrices Czy and By; where ey = ff, Ceenmiea ue) HP.) Hl rasra) dat da’, (6.67) [, Clanmies T (rasra) HE (rar) da® (5.68) ere ( hates ofa okt in AX ‘The assy procure jt metloned consis ol combining entries conespondng tothe ste node (Alin, 1982) Onc te Szeto an gents bmn cue teen re be combined from iareene ‘ortho sare plese peas " The Karhinen-Loee expansion & the modulus of elasticity ssform equation (5.46) nto a L a@Ku 7 (5.09) The integrations indicated in equation (5.44) may be per Iytically o using some numerical quadrature scheme. ‘The w ly the externally applied forees is k performed x we fi, wt.) to0) aor = UT yct f wet sjad = UTE,182 CHAPTER 5. NUMERICAL EXAMPLES “ (7) Minimizing the total energy (V — V") with respect to U, leads to the equation . | Ae |KO + (6.72) From this point on, the treatment of this two dimensional problem is identical to that for the previous one-dimensional beam problem. Specifically, the sponse vector U is expand (7 or coefficients ¢; evaluated as the solution to the following of algebraic equations wil Hwsltshl> K! 5.3.2 Results Analytical Solution ided into sixteen elements, four along each direction. The nodes are numbered in increasing order away from the clamped edge and counting from left to right. Figures (5.46) and (5.47) show the standard Jeviation of the response at the free corner of the plate versus that of the for the improved Neumann expansion and for the Ho- The square plate is ¢ modulus of elasticit mogeneous Chaos expansion, respectively, Observe that better is achieved with a third order Homogeneous Chaos expansion than with a Pigures (5 48)-(5.50) show the projections fon the Homogeneous Chaos for various orders of the expansion. These are the vectors ¢, appearing in oquation (5.74). Note thet the fluctuations of " reflects the fact that the two-dimensional fourth vider Neumann exponei the nodal values of th of the by a one-dimensional figure. Further, 53. TWO DIMENSIONAL STATIC PROBLEM 153 LATE CORNER LONGITUDINAL i Figure 5.46: Normalized Standard Deviation of Longitudinal Displacement at Comer A of the Rectangular Plate, versus Standard Deviation of the Modulus of Elasticity; mar = 0.483; Exponential Covariance; Neumann Expansion Solution.154 CHAPTER 5. NUMERICAL EXAMPLES 5.3, TWO DIMENSIONAL STATIC PROBLEM 158 alized Standard Deviation of Longitudinal Displacement | sree Pla, versie Standard Deviation of the | 38: Exponential Coverance; Polynomial Figure 5. at Comer A of the Recta Modulus of Elasticity Chaos Solution = Noes crest a small amount to 1)-(5.55) show the that the third order Chaos contribu * ge compare to the fst two. Figures (6 er jual projections os the order ofthe expansion 1 3) is nezeased. Figures | ove igure 548: Linear Interpolation of the Nodal Values of the Veetor¢; of uation (5.73) for the Rectangular Plate Stretching Problem, = 0, 1,2 Lagoa Dileep 2 eosin EL Esonin, M= total variat ronvergence of 5 © 6 increased, that is ¢s the value of P in equation (5.20)-(5.68) show the results corresponding to the probability distributot ‘he response variable at the free corner of the plate, The method described in ction (4.4) is used here again. own in Figure (5.45). The curved side is a ninety degree are of of unit radius. The length of the straight edges is 35, ‘The standard deviation of the longitudinal and transverse ‘1. using two terms in the K-L expansion for the ms seal stochasticity, is shown in Figures (5.64) ond (5.65), plotted against the indard deviation of the modulus of elasticity. The results co Jing to fcums in the K-L expansion are shown in Figures (5.66)-(5.67) G0), respectively. Note the excellent convergence. Also note. | ‘the number of terms used in the Karhune equal displacements at nod four and and (5.68) that for this example, the effect Locve expansion of the material stochasticity is more pronounced than fo156 CHAPTER 5. NUMERICAL EXAMPLES 49; Linear Interpolation of the Nodal Vatu he Vector ¢; of Equation (5.73) for the Rectangular Plate Stretching Problem, i = 0,5 itudinal Displacement Representation; 2 Terms in K-L Expansion, 53, TWO DI. NSIONAL STATIC PROBLEM Figure 5.50: Linear Interpolation of the Nodal Values of the Vect Equation (5.73) for the Rectangular Plate Stretching Problem, ¢ Longitudinal Displacement Representation; 2 Terms in K-L Expasion, 4 2 p=0,1,2,3.58 CHAPTER 5, NUMBRICAL EXAM 5.3. TWO DIMENSIONAL STATIC PROBLEM 159 | /\ | —— | = | /\ | 2 | | f ie | 2 | } | | see i \ LA | 2 00 5 \/ V/s | 0s | Hees | {77 | [esa 4 | a — 5 F 2 | 5 0 is 20 smear Totnpeaton of the Nodal Vales of the Vector ¢; | Fue 5.52: Linear Interpolation of the Nodel Values of the Vector: Fagune 9.1: Linea Interpoiat ler Pate Stotching Problem, «= % | of Equation (6.79) forthe Rectangular Plate Stretching Problem, = 2 ‘tion; 2 Terms in K-L Expansion ngitudinal Displacement Representation; 2 Terms in K-L Expansion, = %; P=3,6,10. ngitudinal Displacement Represe =3,6,10.160 CHAPTER 5. NUMERICAL 5.3, TWO DIMENSIONAL STATIC PROBLEM (AMPLES [EXPONENTIAL COVARIANCE 1 ped | po ooo 4 | . TI Nioairpemnn pan ic = — om 4 Figure 5.53: Linear Interpolation of the Nodal Values of the Vector ¢ ae of Equation (5.73) for the Rectangular Plate Stretching Problem, # = & Longitudinal Displacement Representat = 2; P=6,10. igure 5.54; Linear Interpolation of the Nodal Values of the Vector ¢ | of Equation (5.73) for the Rectangular Plate Stretching Problem, i = 4 ongitudinal Displacement Representation; 2 Terms in K-L Expansion, P=6,10, on; 2 Terms in K-L Expansion,| 162 CHAPTER 5, NUMERICAL EXAMPLES | 53. TWO DIMENSIONAL STATIC PROBLEM 163 | T TIA. COVARIANCE ~ | Mgacniosimanios SCY coe exroNenTIN | z ‘Sac Hemozenons hos | 5 2 ABI Hieron Ce oo 4 Serr ppt no | g | tir tnagensas Choe g | “re z | | Z ooo 4 | | Contato Lena = | 7 | 5 | Stniurd Bavenon 02 | go] | 2 08 | LL ( NobE DISPLACEMENT Figure 5.55: Linear Interpolation of the Nodal Values of the Vector ¢, | Figure 5.56: Longitudinal Displacement at the Free End of the Rectangular of Equation (5.73) for the Rectangular Plate Stretching Problem, 7 = 5: Plate; Probability Density Function Using 30,000-Sample MSC, and Using K-L Expansion, | Third Order Homogeneous Chaos; Two Terms in the KA Exponential Covariance. Longitudinal Displacement Representation; 2 Terms 4 Expansion:164 CHAPTER 5, NUMERICAL EXAMPLES 5.3. TWO DIMENSIONAL STATIC PROBLEM pf pase | 7" [sexes 7 “| / | " | b=) Cee | — a | Figure 5.58: Longitudinal Displacement at the Free End of the Rectangular Plate; Cumulative Distribution Function Using 30,000-Sample MSC, and Using Third Order Homogeneous Chaos; Two Terms in the K-L Expansio Exponential Covariance. Figure 5.57: Longitudinal Displacement at the Free Bnd of the Rectangular Plate; Tail of the Probability Density Function Using 30,000-Sample MSC, and Using Third Order Homogeneous Chaos; Two Terms in the K-L. Expansion; Exponential Covariance166 CHAPTER 5. NUMERICAL EXAMPLES | | = Homogeneous Chaos | osrs — _ | Figure 5.59: Longitudinal Displacement at the Free End of the Rectangular | Plate: Tail of the Cumulative Distribution Function Using 30,000-Sample MSC, and Using Thitd Order Homogeneous Chaos; Two Terms in the K-L | Expansion; Exponential Covariance 5.3, TWO DIMENSIONAL STATIC PROBLEM Carlo Simulation AN Sons ane a Figure 5.60: Longitudinal Displacement at the Free End of the Rectangular Plate; Probability Density Function Using 30,000-Sample MSC, and Using Thitd Order Homogeneous Chaos; Four Terms in the K-L Exp: Exponential CovarianNUMERICAL EX. 168 CHAPTER 5. Figure 5.61: Longitudinal Displacement at the Free End of the Rectangular Plate; Tail of the Probability Density Function Using 30,000-Sample MSC. and Using Third Order Homogeneous Chaos; Four Terms in the K-L Expansion; Exponential Covariance. TWO DIMENSIONAL STATIC PROBLEM 169 | | | Figure 5.62: Longitudinal Displacement at the Free End of the Rectangular Plate; Cumulative Distribution Function Using 30,000-Sample MSC, Using Third Order Homogeneous Chaos; Four Terms in Exponential Covariance ‘and ne K-L Expansion;170 CHAPTER 5. NUMERICAL EXAMPLES | 5,3, TWO DIMENSIONAL STATIC PROBLEM a7 the square plate example. This behavior is attributed to the fact that in | this case the random field involves intricate geometric boundaries: therefore more terms are required for its adequate representation. ‘The probability distribution functions correspondin, A are depicted in Figures (5.70)-(6 to one of the response variable at node Results corresponding to up to six | terms in the Karhunen-Loeve expansion and third order Polynomial Chaos | 1000 { } 8 | 2 | | 5 ee] Pe. —— Monte Carto Simul | a | z : oo -! ous ots amas a md * ” Figure 5.64: Normalized Standard Deviation of Longitudinal Displacement eae | et Comer A of the Curved Plate, versus Standard Deviation of the Modulus of Elasticity; Two Terms in the K-L Expansion; Exponential Covariance Figure 5.63: Longitudinal Displacement at the Free End of the Rectangular Plate; ‘Tail of the Cumulative Distribution Function Using 30,000-Sample MSC. and Using Third Order Homogeneous Chaos; Four Terms in the KL. Expansion; Exponential Cov Monte Carlo Simulation The two dimensional process represen plate is simulated as follows. First, th the ijt element of which corresponds to the correlation of the process at points é and j. Following that, the Cholesky decomposition of the positive definite covariance matrix is obtained (Golub and Van Loan, 1984). The colimns of the Cholesky decomposition factor are then used as the basis for simulating the process. In other words, the process is obtained b ng the modulus of elasticity of the ‘covariance matrix is constructe y premulti-CHAPTER 5. NUMERICAL EXAl NER TRANSVERSE DISHLACEMENT Figure 5.65: Normalized Standard Deviation of Transverse Displac Corner A of the Curved Plate, versus Standard Deviation of the of Elasticity; Two Terms in the K-L Expansion; Exponential 20.4. MPLES ‘Modulus arian PLATE CORNER LONGITUDINAL DISPLACEMENT, Figure 5.66: at Comer A of the Cu of Blastiei Omer = 194. TWO DIMENSIONAL STATIC PROBLEM 173 Normalized Standard Deviation of Longitudinal Displacement ved Plate, versus Standard Deviation of the Modulus y; Four Terms in the K-L Expansion; Exponential Covariance:i 74 CHAPTER 5. NUMERICAL EXAMPLES Figure 5,67; Normalized Standard Deviation of Transverse Displacement at 4 of the Curved Plate, versus Standard Deviation of the Modulus Four Terms in the K-L Expansion; Exponential Covariance Comer of Elasticity Sinan = 20-4 5.3. TWO DIMENSIONAL STATIC PROBLEM [Besser romero os —————— wr os 4 Figure 5.68: Normalized Standard Deviation of Longitudinal Displacement at Comer A of the Curved Plate, versus Standard Deviation of the Modulus of Elasticity; Six Terms in the K-L Expansion; Expont 9.4 ial Covariance;CHAPTER 5. NUMERICAL EXAMPLES [peter ati Homogeneaus Conor 5 os 4 reocrerieL expansion . | | os 4 | eet | oe la SE o2s ‘PLATE: CORNER TRANSVERSE DISPLACEMENT Figure 5.69: Normatized Standard Deviation of Transverse Displacement at Comer «A of the Curved Plate, versus Standard Deviation of the Modulus of Elasticity; Six Terms in the K-L Expansion; Exponential Covarianee; 20.4. So | — Rss Fou \ 20 : Figure 5.70: Longitudinal Displacement at the Free End of the Curved Plate; Probability Density Function Using 30,000-Samples MSC, and Using Thitd Order Homogeneous Covariance, Chaos; Two Terms i the K-L Expansion; Exponential plying the Cholesky factor with a vector consisting of Gaussian uncorrelated ‘variates, For the example involving the curved geometry problem, a different simulation technique is used. It is prompted by the fact that the nodal points are not uniformly distributed over the domain of the plate. The issue fs addressed by using the Karhunen-Loeve expansion to simulate @ truly continuous random field. To this end, the eigenvalues and eigenfunctions of the covariance kernel are computed as described in the previous section. The random field is then simulated using equation (2.6) with the number of terms equal to the mumber of nodes in the system. ‘The orthogonal random variables appearing in that equation are obtained as pseudorandom computer generated uncorrelated variates, with zero mean and unit variance. The resulting simulated random field is not as sensitive to the mesh size and ‘edal point distribntinn as the field obtained nsing the more conventional procedure described for the previous example. The results from using the Monte Carlo simulation method are superimposed on the same plot as the analytical results, Observe the good agreement; between the analytical and the simulated results even for large values of the coefficient of variation.CHAPTER 5. NUMERICAL EXAMPLES = os ———— = Seo, Io Ss coos 4 Onder Hmogeneous Chaos | aa Secets | sito 572 Longedinal Dspace t tho Fre End ofthe Curved Pate oon | / Probability Density Function Using 30,000-Sample MSC, and Using Third Zs Order Homogeneous Chas; Four Terms in te KL Expansion; Exponential o et Covariance | : . 5.4 One Dimensional Dynamic Problem DISPLACEMENT Figuro 5.71; Longitudinal Displacement at the Free End of the Curved Plate; Tail of the Probability Density Function Using 30,000-Sample MSC, and Using Third Order Homogeneous Chaos; Two Terms in the K-L Expansion: Exponential Covariance 5.4.1 Description of the Problem The last example to be considered involves an Euler-Ber in Figure ( of length L, modulus of elasticity inertia I, and mass density m. Let the beam be supp: foundation having a reaction modulus &(:r,9) which is considered to be a one-dimensional Gaussian random process with mean value i and covariance function Ce(2,t,0). The beam is assumed to be acted upon by a zero-mean. stationary random process f(,t,6), featuring both temporal and spatia random fluctuations, a given by the equation ‘flected by its spectral density fanetion which is nw) = job) > 01 where 2, and 2 denote two locations on the beam, w is the wave mumber cssociated with time, and # is the correlation length of the excitation process.MPLES 180 CHAPTER 5. NUMERICAL EXAl ONE DIMENSIONAL DYNAMIC PROBLEM Is IONFUNC igure 5.74: Longitudinal Displacement at the Free End of the Curved Probability Density Function Using 90,000-Sample MSC, and Using Third | Order Homogeneous Chaos; Six Terms in the K-L Expansion; Exponential Covariance In the present analysis, the correlati length b is assumed to be a constant, although any dependence on frequency can be accommodated. DISPLACEMENT 5.4.2 Implementation tre 5.73: Longitudinal Displacement at the Free End of the Curved Plat a nee he Probability Density Function Using 30,000-Sample MSC, and | The differential equation governing the motion of the beam with constant ng Third Order Homogeneous Choos the I-L Ex bending rigidity is ial Covariance, # ot 7 0) + 6) +81 Zoute,t,0) 5.76) ae ont + K(x.) ule, t,8) = flat, re ¢ is a coellicient of viscous damping. Adopting the discretization ocedure of the previous exemples, the heam may he divided int N' finite elements, Then, an equation involving 2 x 2/V matrices is obtained of the M U(i,6) + C Ujt.8) + K Ult,é) + Ky Ult,8) = £(t,6)Figure 5.75: L Tail of the Probability Using Third Order Ho: Exponential Covariance ;tudinel Displacement at the Dens CHAPTER 5. NUMERICAL EXAMPLES i Free End of the Curved Plat 4 80,000Suinple MSC, 2 Pune! 8 Chaos; Six Terms in the K-L Expansio 5d. ONE DIMENSIONAL DYNAMIC PROBLEM 183 where a dot denotes differentiation with respect to time and M, K, Ky are generated by assembling the elemental matric M = [HT m He ate 5.78) K = er [BY Bat 5.79) Further, the damping matrix C is assumed, for simplicity, to be of the form C = ey M + ex K 5.81) and ex are constants, Replacing the proce 6) in equation (5.80) by its Karhunen-Loeve 2 ky = f Raye eae + Sg 2) HT HE ai® , (5.89 where f(z) is the £" eigenfunction of the covariance kernel and \y is the corresponding ** eigenvalue. Equation (5. M Ot) + CUlt) +K Ut) + Kp Ut (5.83) +> ax in which the argument 9 is deleted for notational simplicity. Taki transform of equation (5.83) yields184 CHAPTER 5. NUMERICAL EXAMPLES 5.86) - H 2M + wO+K + Kr] 5.88) 1+ Saal] vw) = Pw 5.80) | | wher 5.20 Hw)! Fw way Equation (5.80) has the seme form as equation (5.28). The spectral density of the response process can be written as pew) 1+ 0 & QP | i = [r+ Deo] pt H denotes hermitian transposition. Using the Ne ation (6.92) ean be put in the y [eae y. 99 5.4. ONE DIMENSIONAL DYNAMIC PROBLEM Alternatively, using the Homogeneous Chaos approach, the of the response can be represented in the following form Svols) = SS NUE TEE where a is a vector that is obtained as the solution to a system of linear Note that the spectral density matrix of P(w) is related to the spectral density matrix of F(w) by the equation equations as indicated earl Spp(w) = Hw) Syj(w) Hl)" (5.95) In addition, Spy(a1,2r2500) is as a frequency dependent 5 fen by equation (5.75) which may be regarded on, Thus, it is numerically efficient to expand Sjj(on1,2 al series and perform the 5 cretization on the eigenfunctions of the expansion, Indeed, this procedure is, adopted in the finite eleme ti ode developed to solve 1 cally, Spplzi, asic) is expanded 0s Sppleyeriw) = almiw) gears (5.96) In equation (5.96), 4s refers to the eigenvalues associated with the kernel Sip(2t1,22;w), while gi(2su) denote the corresponding eigenfunctions. Results In the numerical irnp mel is used for the rei equal to 1, The numerical v sm, the exponential covariance mentation of this modulus k(c, 6 |, with a correlation le ues of other parameters are included in Figure shot isplacement he spectral density Sy. (w) of the st one end of the beam along with pertinent Monte Carlo simulation re 4 frequency increment equal to 0.1. Note the excellen wo solutions until the exact damped natural frequency of the beam causing the norm of the matrix [r+ S34, & Q | toNUMERICAL EXAMPLES CHAPTER 5. bast oN RANDOMELASTIC FOUNDATION Figure 6.76 s1 Random Elastic Foundation Subjected to 8 Random Dynami Exponential Covariance Model roblenn is not encountered with the Homogeneous Chaos approach, a com ce .¢ of the completeness of t haos basis in the space of Fi ssults corresponding to yandoin variables. Figure (5.78) a results corresponding he Inmproved Neumann expansion successive orders of approximation x rad the Homogeneous Chaos methods respectively. The Monte Carlo simu ion ed, us describe ‘ction, using the Cholesky noted that using more than the covarian ie i 1 ex four terms in the spectral exp fie derived solutions. Clear articular problem depends on the magnitude Tength of the excitation process. 5.4, ONE DIMENSIONAL DYNAMIC PROBLEM 487 tral Density of the Displacement at the End of the Beam Foundation, Figure 5.78: Spectral Density of the Di on Random Elastic Foundation rement at the End of the Beam] | CHAPTER 5, NUMERICAL EXAMPLES | Displacement at the End the Bear Chapter 6 SUMMARY AND CONCLUDING REMARKS 6.1 This monograph has considered a class of havior is governed id as ran ting gystems whose be- x differential equations with coefficients the random variability in the material proper a material property ¢ pes been 1 response, Specificall The first st ed random variables; nsion has the quite desirable feature of being valid over the entize f definition of the problem. ‘The end product from the first stage matrix: equation involv daterministic matzicee multiplied by rando: age in the solution approach co ‘wo options hav xplicit expression for the response process as a multi=NUMERICAL EXAMPLES Figure 5.79: Spectral Density of the Displacement at the End of the Beam Chapter 6 SUMMARY AND CONCLUDING REMARKS 6.1 ‘This monograph havior is governed by linear differential equations with coefficients that are modeled, as random processes. These coefficients can be associated with random variability in the material properties of the system. In this regard, ‘ material property of an actual system can be thought of as a realization of an appropriate second-order stachastic process, 8 considered a class of engineering systems whose be- phasis has been given to a spec system response. Specifically, the disct ‘al approach to determining the solution approach involves two s. ‘The first stage hinges upon adequately representing the stochas- processes corresponding to the random system properties. This has been achieved by means of the Karhunen-Loeve expansion. This expan provides an optimal mean-square convergent representation of a stochastic process in terms of a denumerable set of uncorrelated random variables; the expansion has the quite desirable feature of being valid over the entire domain of definition of the problem. The end product from the firs is a matrix: equation involvi variables. The second stag the res Jetorministic matrices multiplied Ly 1audui in the solution approach consists of solving ulting set of equations. ‘Two options have been presented to this end. The first one relies on utilizing the Neumann expansion fREMARKS 190 CHAPTER 6, SUMMARY AND CONCLUDIN dlimensional homogeneous polynomial in the uncorrelated random variables. The second option consists of a Galerkin scheme whereby the solution process is approximated by its projection onto a complete basis in the space of andom vatiables. Such a basis has been identified as consisting of the Polynownial Chiaoses. These ae orthogonal multi-dimensional polynomials in ‘he uncurelatel vandom variables. An explicit expression fr the response 2 a surface in the process is thus obtained. The response is represented as e ‘multi-dimensional space defined by the uncorrelated random variables. This form for the solution process is particularly suitable for reliability analysis of the system being considered. plied to typical problems of statics The spectral approach has been applied to typical pi with data generated from Monte Carlo simulation solutions. ‘These results pertain to second-order statisties of the response process, as well as, to its A somewhat detailed description ofthe content, density function. proail Seah cate on poll ste Te cme mal the maf ie ines from in net intbeca Shae they constitute the majority of mathematical models sed f Hind of represent hearetical developments of analyzing issue of representing stochastic process pivotal in the kind of representation « systems with random parameters, it is environment where only countable and digerote quantities can he imple. mented. Addressing this issue, another kind of representation of stochastic processes was introduced involving @ denumerable set of random variables, ‘This representation utilized expansions in terms of a finite dimensional basis in the space of random functions. In this connection a stochastic process 6.1. SUMMARY AND Ci INCLUDING REMARKS 191 was viewed as a curve in this Hilbert space. As a particul and usefull example, the Kathunen-Loeve expansion was introduced as an optimal representation of this kind. The expansion was considered as a representation of the stochastic process along an orthogonal basis in the space of random variables in terms of the eigenfunctions of the covariance funtion of the process, This fact was used to view the expansion as an expression of a congruence between the Hilbert space spanned by the random variable basis and the Hilbert space defined by the covariance function. The Karhunen-Loeve expansion was shown to be well suited Purposes since it i for computational optimal, convergent, and invalves a demumerable set of random variables, Crucial to implementing the Karhunen-Loeve expansion is a specification of the covariance function of the process to be represented. Such information is known in relation to the processes describing the material properties, butis not available, in general, far the unknown response process. This issue was addressed by introducing the Homogeneous Chaos expan sion. This is a mean-square convergent representation of multidimensional functions of Gaussian variables in terms of the Polynomial Chaoses that are orthogonal with respect to the Gaussian measure. It can be viewed as the discrete equivalent of the Wiener-Hermite expansion and may be obtained from the Cameron-Martin expansion by a limiting process, ‘The Successive Poiynomial Caoses were shown to be readily computable either algorithmically or with the use of symbolie computation programs. It wes shown that any random function admits a. mean-square convergent expansion along the basis provided by the Poiynomial Chaoses. ‘The importance of suck a representation in connection with a finite element analysis is great. Furthermore, it was shown that using the Kathunen-Loeve expansion to Tepresent the rancom system paramoters circumvents the problem of making the size of the finite element mesh depend on the level of random fluctuations of these parameters. An abstract mesh was thus provided along the random dimension that wes uncoupled from the physical finite element mesh. It ‘was indicated that this abstract mesh is actually a discretization along the random dimension with re Polynomial Chaos basis spect to the spectral measure associated with the In Chapter UIT the deterministic finite element method was reviewed first. Next, some of the difficulties associated with extending the analysis to stochastic systems were identified. The inadequacy of non-spectral tech- niques to effectively deal with these issues was also indicated. Specifically102 CHAPTER 6, SUMMARY AND CONCLUDING RE} [ARKS nber of random variables resulting from the pointwise repre- f the stochastic processes involved, an the resultant random matrices were pointed out. In this context, porturbation methods and Neumann expansion methods were presented ag the major tools available to address the issue, The methods have been proven successful in dealing with problems involving randomness agnivude. They have been found quite eiffcul to implement for Siaable randomness requiring a significant number of terms to be included in the series representation of the solution process. In the remainder of the ciuapter, two other methods wore discussed that overcome, to @ large extent, the dificulties involved. Both methods rely on the Karhunen-Loeve expan- sion presented in Chapter III to recast the problem in terns ofa denumerable sot of uncorrelated random veriables, When coupled with deterministic finite element methods provided matrix equations that were ‘much simpler than the equations obtained by the other methods. As a fist eppronch to solving the resulting equations, the Neumann expansion for the inverse was applied. The procedure was much eesier to implement, hhaving introduced the K-L expansion, than it would have been otherwise This was due to the fact that the matrix to be inverted consisted of the sum of deterministic matrices with the random variables appeati ts multiplicative factors. ‘This yielded a series expression for the process 8s 6 multidimensional homogeneous polynomial in the uncorrelated Tandom vatiables. ‘Tho second approach to solving the equations resulting fom coupling the finite element discretization and the K-L expansion was to expand the random response process along a basis in the Hilbert space random vatiables and to compute the coeficents in the expansion with the use of some error minimizing criterion. ‘The Polynomial Chaoses defined iu Chapter HIT were used 8 such an expansion end the Galerkin scheme tras invoked to make the error orthogonal to the approximation space. The csult of the procedure was a convergent expansion of the response process ems of multidimensional orthogonal polynomials the large 1 sentation o In Chapter IV the usefulness of the representation of the stochastic ssed from a perspective response of the system. lu umponse as derived in Chaptor HE was ad isties of of determining the probabilistic chera In this context elementary concepts from reliability theory were reviewed, ‘The validity and limitations of the approximations involved were discussed analysis wes introduced that can be coupled with the A method for reliability 6.1. SUMMARY AND CONCLUDING REMARKS 193 system response representation pr ented in Chapter IIL. The method was based on viewing the Homogeneous Chaos expansion of the response process as providing an explicit expression of the limit state surface. Reliability related results could be readily obtained from this expression, One option in implementing the proposed method involved computing the moments of the response. These are obtainable, to any order, in a closed form. Alternatively, the cummulants of the response ean be used. Another option was based on a multidimensional Edgeworth approximation for the probability distribution function in terms of multidimensional Hermite polynomials. ‘The coefficients in this expansion could be obtained via # Galerkin scheme, Finally, a thitd option consisted of numerically simulating the Polynomial Chaoses resulting in a corresponding statistical population for the response process which was used in conjunction with nonparametric density estimation techniques to obtain approximations to the probability den ity function of the re ‘The solution methods presented in Chapter III were applied in Chapter ‘V to three problems of engineering mechanics. The first problem was that of a cantilever beam subjected to a static load. ‘The modulus of elastic- ity of the beam was considered to be the realization of a homogeneous Gaussian process, The beam was discretized into ten finite elements and a variational formulation over the discrete mesh was used to obtain the associated algebraic random equations. Both solution procedures described in Chapter III were implemented to solve the resulting equations. The second problem involved a plate subjected to a static in-plane tension. Two eases ‘were investigated. The firet case involved a square plate; tl involved a plate with curved boundary. Sixteen finite elements were used jn both cases. Again a variational formulation was employed to obtain the governing diserete random equations. For both the beam problem and the plate problem, it was found that for the response pracess to achieve a certain accuracy, the terms needed in the final expression using the Ho- ‘mogeneous Chaos expansion were fewer than the terms required using the Neumann expansion approach. The method described in section (4.4) was employed to obtain approximations to the probability density funetion of the response of the beam and the p second case Good agreoment was observed ‘with the corresponding probability distributions that were deduced based on a large number of Monte Carlo simulations of the solution process. The third problem that was investigated involved a free-free beam lying on a inearly elastic foundation and subjected to a random excitation possessing194 CHAPTER 6, SUMMARY AND CONCLUDING RE} ARKS both temporal and spatial random fluctuations, The modulus of resction of the underlying foundation was also cousidered to be the realization of homogeneous s the other examples, a discrete set of equations involving random variable obtained. Again, the spectral approach was applied to derive an expression for the spectral density function of the response p ‘The analytical result in good agreement with Monte Carlo simulation data except in the neighborhood of the natural frequency of the system where the Neu mann expansion approximation seemed to diverge. This phenomenon was attributed to the fact that the matrices being inverted tended to be ill conditioned at the vicinity of the reson xy and the requirements for the applicability of the Neumann expansion method were not satisfied However, results from the Homogeneous Chaos approximation were well behaved and matched remarkably well the simulation data. Some possible extensions of the concepts in this monograph are particu larly notable, specifically in terms of improving the computational efficiency of the proposed method. As mentioned in Chapter V, the coefficients in the Homogeneous Chaos expansion obtained from lower expausions provide ood initial guesses when used in the context of an iterative computational scheme, This fact suggests the possibility of implementing concepts fron multigrid techniques ( Hackbusch and Trottenberg, 1951 ) to increase the computational efficiency of the proposed methods. Another potential exter sion consists of implementing different basi he Hilbert space of ‘andom veriables. Specifically, rational fun ‘uncorrelated random variables are an appealing choice. Of course, simulation methods will have to be applied in order to compute the coeffictents ciym appearing in equation However, these coefficients need ticular set of basis functions. The additional computational cost may be warranted by possible convergence acceleration, Tt is believed that the material presented in this monograph constitutes ‘a meaningful tool for accelerating the evolution of the field of Stochastic Finite Elements. Irrespective of the nature of future development in this fh as static or dynamic problems, steady ‘J ur fluid media, the pivotal concept of any analysis method wil in an expeditious and tractable representation of the randomness ix the excitation, the medium, and the response; the spectral representation of random processes used herein is a viable option meriting further attep rchastic process. Following the same procedu to be calculated only once for a tate or transient solutions, Bibliography [1] Adomian, G., “Stochastic Green's functions", pp. 1-39, Proceedings of Symposia in Applied Mathematics, Vol. 16: Stochastic Processes in Mathematical Physics and Engineering, Edited by Richard Bellman, American Mathematical Society, Providence, Rhode Island, 1964. Adomian, G. and Malakian, K., “Inversion of stochastic partial differential operators - The linear case”, J. Math, Vol. di Anal. Appl. 77, pp. 309-27, 1980. [3] Adomian, G., Stochastic Systems, Academic Press, New York, 1983. 4] Akin, J.B. Application and Implementation of Finite Element Methods,, Academic Press, New York, 1982 i Aronszajn, N., “Theory of reproducing kernels”, Transactions of the American Mathematical Society, Vol. 68, No. 3, pp. 337-404, May 1950. [6] Babuska, I., Zienkiewiez, O.C., Gago, J. and de A. Oliveria, ER., Editors, Accuracy Estimates and Adaptive Refinements in Finite Element Computations, John Wiley and Sons Ltd., New York, 1986, Beckers, Richard A. and Chambers, John M., $: An Interactive Environment for Data Analysis and Graphics, Wadsworth Statis- tics /Probability Series, Belmont, California, 1984. iS] Benaroya, H., “Stochastic structural dynamics a theoretical basis and selective review of the literature”, Weidlinger & Associates, Technical Report No.1, New York, April 1984 195196 BIBLIOGRAPHY [0] Benaroya, H. and Rebak, M., “Parametric excitations I: Exponentially cortelated parameters”, Journal of Engineering Mechanics, ASCE, Vol. 113, p. 861-874, June 19% {10} Benaroya, H. and Rebak, M., “Finite element methods in probabilistic structural analysis: A selective review", Applied Mechanics Reviews Vol. 41, No. 5, pp. 201-218, May 1988, 11] Benjamin, J, and Cornell, C., Probability, Statistics and Decision for Civil Engineers, McGrave-Hill, New York, 1970. “On random operator equations in Banach Ser, Sci. Math. Astr. Phys., Vol. 7 (12) Bhasvuche-Reid, A-T. space”, Bull. Acad. Polon. Sci pp. 561-564, 1959, [13] Bharrucha-Reid ed., Probabilistic Methods in Applied Mathematics, Academie Press, New York, 1968, [14] Bose, A. ‘Tech A Theory of Nonlinear Systems, Massachussets Institute of RLE Report No. 309, July 1956, 15] Boyce, BW. and Goodwin, B.E., “Random transverse vibration of clastic beams", SAM Journal, Vol. 12, No. 3, pp. 613-629, Septembe 1064. (16) Brockett, RW Automatica, Vol “Volterra series and geometric control theory”. 12, pp. 167-176, 1976. [17] Cakmak, A. and Sherif, R., “Parametric time series models for earthquake strong ground motions and their relationship to site parameters”, Proc. 8th World Conf. Barthg. Bng., San Francisco, pp. 581-588, 1984. [18] Cameron, R.H. and Martin, WT, “The orthogonal development of nonlinear functionals in series of Fourier-Hermite functionals”, Ann ‘Math, Vol. 48, pp. 385-392, 1947. [9] Chorin, AL 4. Journal of Ce Ilermite expansions in Monte Cazlo computation” maputational Physics, Vol. 8, pp. 472-482, 1971 0] Chorin, A. J. Fluid Mechamies, Vol. 87, Patt 4, pp. 7 ‘Numerical study of slightly viscous flow”, Jou 35-796, 1973, BIBLIOGRAPHY 197 {21] Collins, J.D. and Thompson, W.T., “The eigenvalue problem for vactural systems with uncertain parameters”, ALAA Journal, Vol No.d, pp. 612-648, 1969. [22] Comell, C.A. Ami “A. probability-based structural code”, Journal of rican Concrete Institute, Vol. 66, No. 12, December 1969. (23) Courant and Hilbert, Methods of Mathematical Physics,, Interscience, New York, 1 [24] Cramer, H., “Stochastic processes as curves in Hilbert space”, Theory of Probability and its Applications, Vol. IX, No. 2, pp. 169-179, 1964. [25] Cruse, T.A., Wu, YT., Dias, J-B. and Rajagopal, K.R., “Probabilistie structural analysis methods and applications’, Computers and Structures, Vol. 30, No. 1-2, pp. 163-170, 1988. 6] Cruse, T.A. and Chamis, O.C., Editors, Probabilistic Structural Analysis Methods (PSAM) for Select Space Propulsion Compon Manuscripts and reprints by Southwest Research Institute, San Antonio, TX, May 1989, } Dendrou, B. and Houstis, E., “An inference finite element mode! for field problems”, Appl. Math, Modeling, Vol. 1, pp-109-114, 1978. [2] Denérou, B. and Houstis, ‘Technical Report CSD-TR , Uncertainty Finite Blement Analysis, Purdue University, 1978b. [29] Der-Kiureghian, A., and Liu, P-L., “Structural reliability under ine complete probability information”, Journal of Engineering Mechanies, ASCE, Vol. 112, No. 1, pp. 85-104, 1986 (30) Der-Kiureghian, A., Lin reliability approximations Vol. 113, No. 8, pp. 1208-1 HAZ. and Hwang, $-J lournal of Engineering Mechanics 1987, , “Second-order SCE, “the stochastic method in structural reliability”, Proceedings, U.S. Austria Joint Seminar on Stochastic Structural Mechanics, Florida Atlantic university, Boca Raton, Florida, May 1987b, Anite elem [31] Ver-Kiureghian, A.,198 BIBLIOGRAPHY P.A. and Kittler, J., Pattern Recognition: A Stat Prentice-Hall, Englewood Clifls, New Jersey, 1982, '33] Ditlevsen, O., Uncertainty Modelling, McGraw-Hill, New York, 1981 [Si] Dood, J.L., Stochastic Processes, Wiley, New York, 1953. [35] Engels, D.D., The Multiple St Memoirs of the American Mathematical Society, Vol. 38, No, 265, 198 36] Fredholm, I., “Sur une class d’equations fonctionnelles”, Acta Mathematica, Vol. 27, pp. 365-890, 1903. [97] Freudenthal, A.M., “The safety of structures”, ASCE, Vol. 112, pp. 125-180, 1947 [38] Freudenthal, A.M., Garzelts, J.M., and Shinozuka, M. of structural safety", Journal of the ASCE, Vol. 92, No. ST!, pp. 267- (99] Gel'fand, IM. and Vilenkin, N.Ya. Academic Press, New York, 1964. [40] George, P., Continuous Nontinea! ‘Technology, RLE Report No. 3 (41) Ghanem, R., Spanos, P, and Akin, E., “Orthe Beam Variabilit Engi 1988 ygonal Expansion for Probabilistic Va, pp. 156-159, May 25 Discrete Eler ember 1988 i42] Ghanem, R.,, Analysis of Stochastic Systems PhD Thesis, Rice University, Houston, TX, N lerkin Based Response Surface Ap- as of the bility, San Francisco, CA [43] Ghanem, R. and Spanos, P., proach for Reliability Analysis”, Proceed Conference on St pp. 1081-1088, August 8 1 cctural Safety and Reli 1080. [44] Ghanem, R, and Spanos, P., “Polynomial Chaos in Stochastic Finite Elements”, Journal of Applied Mechanics, ASME, Vol, 57, No.1, pp. 97-202, March 1990. BIBLIOGRAPHY 99 [45] Ghoniom, A.F. and Oppenheim, A.K., “Numerical solution for the problem of flame propagation by the random element method”, ALAA Journal, Vol. 22, No. 10, pp. 1429-1435, October 1984. M6] Golub, G. and Van Loan, C., Matric Computations, Johns Hopkins University Press, Baltimore, 1984. [47] Grad, H., “Note on N-dimnen tions in Pure 1949, onal hermite polynomials”, Co nd Applied Mather Vol. 2, pp. 325-830, December [48] Grigoriu, M., Veneziano, D. and Comell, C.A. Modeling a8 Decision Making’ ASCE, Vol. 105, EMA, pp. 585-671 ‘Probabilistic Journal of Ei (49) Grigoriu, M. roximate reliability analysis 1982. Hac vusch, W. and Trottenberg, U. Editors, Multigrid Methods, Proceedings of the Conference Held at Koln-Porz, November 23 Springer-Verlag, Berlin, 1981 Hans, O., “Random operator equations", Proceedings of t Berkeley Symposium on Mathematical Statistics and Pro Neynian J. Editor, University of California, Vol. Il, pp. 185-202, 1961 Pourth 52] Harichandran, R. and Vanmarcke, E.H., “Stochestic earthquake ground motion in space and tine”, Mechanics, ASCB, Vol. 112, No. rariation of 5] Hart, G.C. and Collins, J.D. finite element modelling”, SAE Angeles, CA, pp. 2509-2519, Oct. 1970. [4] Hasofer, A.M. and Lind, N.C., “Exact and i code format”, Journal of Bn BML, pp. 111-221, 1974. fariant second-moment anics, ASCB, Vol. 100, Hasselman, TK. and Hart, G.C., “Modal analysis of random structural systems", Journal of Engineering Mechanics, ASCE, Vol. 98, No. EM3, 1972, pp. 561-