oman Sirti, Theory ond Proce Uaay ATLAS, Sevont Bon, Mander . Ges Angie P Arey eapyrght 2008 Ts Wiley & Sos, Fe SUNG: 0471992515 Ularbucky BATE26638-8 (Users) Kalman FilteringKalman Filtering: Theory and Practice Using MATLAB Second Edition MOHINDER S. GREWAL California Slate University at Fullerton ANGUS P. ANDREWS Rockwell Seience Center ® A Wiley-Interseience Publication John Wiley & Sons, Inc. NEW YORK» CHICHESTER e WEINHEIM « BAISBANE « SINGAPORE « TORONTO.Copyrighs 2001 by Jobm Wiley & Sous, ae lis served ‘No past of his peblietion nay be repre, sed ina eral warn orasmie inary em or by aly means, eleewocie oF near, inline uploacing, desniodieg ping, éesomtig, _eociding or obese, exept tv petted wade Sections 10? o¢ 108 of he 1976 United Stes CCopyrghe At tu he pr ston perro oF the Pie. Reyuewsto she Pier for enrissoe sul be aldose othe Peisions Depstueat, Join Wily & Ses, ne. 608 Third ‘vas, New Yook, NY IOISS-UbL2, (12 SHAH. fax (212) $5000, [eMac PLRMILLO @, WILLY.COW “Thi publication is designe in psi accra atid ahora nro nega ode sje notes covret Is ald witht understanding ther Ue push snc ensnged in rnderne peso seyvises If profssona avies of aber csper assistance i required the Servisss oF 8 compet ‘poten er sh be eh ISBN o-a71-26688-8 ‘Thi ig sho walle prot oe SAW 9-471 392848 or more iefaanarion hoor Wily prodet, vst our web se at swe ileconContents PREFACE, ACKNOWLEDGMENTS General Taformation 11 On Kalinan Filtering 1.2 On Estimation Methods 1.3 On the Noution Used in This Book 14 Summary Problems, sar Dynamic Systems ‘Chapter Focus Dynamic Systeins Corsinuous Lincar Systems and Their Salutions Diseeete Lincur Systems and Their Solutions Observability of Lincar Dynamie System Models Procedures for Compuiing Matrix Exponentials Summary Problems, Random Processes and Stochastic Systems 3.1. Chapter Focus 3.2. Probubility and Raruom Vaviables 3.3. Statistical Properties of Random Variables 56 56 58 66conrens 3.4 Statistical Properties of Rendon Processes 3.5 Linear System Models of Randora Processes anel Sequences 3.6 Shaping Filters and State Augmentation 3.7 Covariance Propayation Equations 3.8 Onthogonality Principle 3.9 Sumamary Problems Linear Optimal Filters and Predictors 4.1 Chapter Poets 42 Kalman Filer 43° Kalman Bucy Filter 44 Optimal Linear Predictors 45. Coztelated Noise Sources 4.6 Relationships between Kalman and Wiener Filters 4.7 Quadratic Loss Functions 48 Matrix Riccati Differential Equation 49° Matrix Riccati Equation in Discrete Time 4.10 Relationships betwcen Continuous and Disercte Riveati Equations 4.11 Model Equations for Transformed State Variables 4.12 Application of Kalman Biltexs 4.13 Staoothers 414 Summary Problems Nonlinear Applications 5.1 Chapter Focus 5.2 Problern Statement 5.3 Linearization Metiods 5.4 Linearization sbour a Nominal Trajectory 5.$ Linearization about the Estimated Trajectory 5. Digerete Linearized and Extended Filing, 5.7 Digerete Extended Kabnan Filter 5.8 Continuous Lincarized and Extended Filters 5.9 Biased Errors ia Quacnitie easements 5.10 Application of Nonlinear Tlters 5.11 Summary Problems. Implementation Methods 6.1 Chapter Focus 6.2 Computer Roundat 6.3 Elleels of Roundoll Exxocs on Kalmaa Filiess 6.4 Factorization Methods for Kalman Fiteriag 6 % 58 0 102 104 us nd 16 126 128 129 130 131 133 148 153 154 155 160 let 165 169 169 170 I7L IL 175 176 178 18) 182 184 198 200 202 202 204 209 216conan: 65 66 6 Square-Root and UD Filters Other Alternative Implementation Methods Suiminary Problems, 7. Practical Considerations nM 72 73 14 Chapter Pocus Detecting and Correcting Anomalous Behavior Prefilieriag, and Data Rejection Methods Stability of Kalman Fillers Suboptimal and Redueed-Order Fillers Schmidt-Kalman Filring Memory, Throughput, and Wordlength Requirements Ways to Reduce Computational Requirements Error Budgets and Sensitivity Analysis Opsimizing Measurement Selection Policies Application t Aided! neti Navigation Summary Problems. Appendix A MATLAR Software Al Al A3 Aa AS AG AT AS Ag Notice General System Requirements Diskette Directory Steucture MATLAB Software for Chapter 2 MATLAB Software for Chapter 4 MATLAB Software for Chapter 5 MATLAB. MATLAB Softw Other Sourves of Sofiware Appendix BA Matrix Refresher BL B2 B3 Ba BS Be B7 BS Be Matrix Forms ‘Matrix Operaions Block Mauris Formulas Functions of Square Matrices ‘Norms Cholesky Decomposition Onthogonal Decompositions of Matrices Quadeatie Forms Derivatives of Matrioes REFERENCES INDEX 238 252 265 266, 270 270 27 294 298, 299 309 316 326 332 336 342 346 M7 350 350 350 351 351 352 352 353 353 355 359 363 366 370 373 375 377 379 381Preface The first edition of this book wa shed. by Prentice-Hall in 1993, With this second edition, as with the frst, our primary objective is tw provide aur readers 1 working familimity with both the reoretical and practical aspects af Kalman filesing by including “real-world” problems in practice as illustrative examples. We are pleased to have this opportunity to incorporate the many helpfull corrections and suggestions ftom our colleagues and students over the last several years for the ‘overall improvement of the textbook. The book covers the historical background of Kalman fiitering and the more practical aspects of implementation: how to represent the problem in a mathematical model, analyze the performance of the estimator as a function of mode] parameters, implement the mechanization equations in numeri cally stable algorithms, assess its computational requirements, test the validity of resus, and monitor the filter performance in operation, These are important auributes of the subject that ore often overlooked in theoretical treatments bul are necessary for application of the theory te reakworld problems. ‘We have converted all algorithm listings and all sollwaze to MATLAB™,, so dat users cau take advantage of its excellent graphing capabilities and a programming interface that is very close to the mathematical equations used for defining Kalman ‘ltering and its epplications, See Appendix A, Section A.2, for more information on MATLAR, ‘The inclusion of the software is practically a matter of necessity, because Kabnant ‘iltering would not be very useful without computers to implement it. Tr is # better learning experionoe for the student t discover hhow the Kelman filter sworks by ‘observing it in action, The implementation of Kalman filtering on computers also illuminates some of ‘the practical considerations of finie-wordlength arithmetic and the need (or alter MATLAB i a teiserd trademark of The Matis, Inex paerace native algorithins t preserve the accuracy of the results, If the student wishes to apply winat she or he learns, then it is essential that she or be experience its workings ard failings—and leare to recognize the difference, “Tae ook is organized for use as a text for un infuductory course in stochastic processes af the senior level and as a first-year graduate-level course in Kabnan filtering theory and application. It could also be used for self-instruction or for purposes of review hy practicing engincers and scientists who are not intimately famifiar with the subject. The organization of the meterial is illustrated by the following chaplerevel depencency graph, which shows how the subject of each chapter depencs upon material in other chapters. The arrows in the figure indicate the recommended onier of stxly. Boxes above anorher box and connected by arrows. indicate that the material represeated by the upper boxes is background materia! for the subject in the lower box. ‘Chapter { provides «informal Smeoduetion to the general sukject matter by way of ity history of development and application. Chapters 2 snd 3 and Appendix B cover the essential uelwround maeral on linear systenis, paububility, stocheste processes, and modeling. These chapiers could le covered ita senior-level oourse ia aleetrical, computer, and sysieras engineering Chapter 4 covers linear opdimal fillers and predicioes, wilh detailed examples off applications. Chapter 5 is devoted to aonliness estimation by “extended” Kalmaa T-GHNERAL, LINDA DYNAMIC SYSTEMS INFORMATION 3. RANDOM PROCESSES & STOCHASTIC SYSTEMS £ “OPTIMAL LINEAR FILTERS AND PREDICTORS | | 'B.NONDINFAR APPLICATIONS (6. MPLEMERTATION METHODS: ‘1. PRACTICAL CONSIDERATIONS:paeeace x ‘ters, Applications of these teclniques to the identifiction of yaknown parameters of systems are given as examples, Chapter 6 covers the mone miedern iinplementae tion techniques, with algorithms provided for computer implementation, Chapter 7 deals with more practical matters of linpleinentation and uve beyond the umerical methods of Chapier 6, These mates include memory aid lnrouphput requitements (2nd melnods (© reduce them), divergence problems (and effective remedies, and practical approaches v9 suboptimal Gilering and measurement selection, Chapters 47 cover ie essential material for a fist-year graduate class in Kalmaa filtering theory and application or as a basic course in digital estimation theory and application, A solutions manual for each chapter's problems is available aot; MoHINDER S. GRLWAL, PHD, PE alr Sate Universo Pallet ANGUS PB. ANDREWS, PHD. Rocket Scions Car, Tews Ob, CaosAcknowledgments The authors express their appreciation to the following individuals for their ‘contributions dusing the preparation af the frst edition: Robert W. Buss, E. Richard Coben, Thomas W. De Vries, Reverend Joseph Gaffney, Thoms L. Gunckel 1, Dwayne Heckman, Robert A. Tubbs. Thomas Kailath, Rudol E. Kalman, Alan J. Laub, Robert F. Nease, Jolm C. Pinson, fol Mi, Richardson, Jorma Rissanen, Gerald , Runyon, Joseph Smith and Donald F. Wiberg. We also express our appreciation to Donald Kanth aud Leslie Lamport for TEX and LATEX, respectively. 1m addition, the following individuals deserve special recegnition for their careful review, corrections, and suggestions for improving the second edition: Dean Dang sand Gordon Taverarity ‘Most of ll, for their dedication, support, and understimding. through both ccditions, ve dedicate this book to Sanja Growal and Jeri Andrews, MS.GABA,oman Sirti, Theory ond Proce Uaay ATLAS, Sevont Bon, Mander . Ges Angie P Arey eapyrght 2008 Ts Wiley & Sos, Fe SUNG: 0471992515 Ularbucky BATE26638-8 (Users) General Information te things ofthis workd eat be made known withou mates, Roger Bacon (1220-1292), Opus Majus, rausl R. Burke, 1928 1.1. ON KALMAN FILTERING 1.1.1 First of All: What Is a Kalman Filter? Theoreticully the Kalman Filter is am estiznavor for what is called the fneaa-quadvanic problem, which is the problem of estimating the instantancous “state” 4a concept that will be made more precise in the next chapter) of a linear dynamic eystera perturbed by white aciss—by usiug measurements linearly related to the state but ccorruptec! by white noise, The resulting estimator is statistically optimal with respect to any quadvatic function of estimation error. Practically tis certainly one ofthe greater discoveries inthe istory of statistical cessation theory snd possibly the greatest discovery in the twentieth century Te hus cabled mankind to do many things that could ne have bon done without it and ft has become as indispensible as silicon ia the makeup of many elvehamie syste. is most immediate applications have been for the contre] of complex. dynamic systems such as continuous manu/aciuring processes, airerall, ships, or spacecral To control a dynamic system, you must first keow what it is doing. For chese applications, ic is not always possible or desirable to measure every variable chat you ‘want to control, and the Kalman filter provides a means foe inferring ce missing information frorn indirect (and reisy) mneasurements, The Kalman filter iy also used for predieting the likely furure courses of dynamic systems that people are not likely tp control, such as the flow of rivers during flood, the wajectories of celestial bodies, ‘or the prices of traded commodities. From a prletical standpoint, these are the perspeetives that chis book will present:2 enenAL NrOREATION © fe 1s onle a tool, Wt does not solve any problem all by itself, although it eam make it easier for you to do it, I is not a physical tool, but a mathematical one, eis made from mathematical models, which arc essentially ools forthe mind. They make mental work more efficions, just as mechanieal tools ake physical work moze ellcieet. As with any tool, iL 4s important 10 understand ils use and lianction belore you eaa apply it elleclively. Tae purpose of this book is 10 sake you sufficiently familiar with and proficient in the use of the Kalman filter that you can apply it eoxtectly and efficiently # L4is a computer progran. tt hs beea called *iseally suited to digital computer implementation” [21], in part because it uses a finite representation of the cstimatioa problem—by a finite number of variables. does, however, assume hat these variables are real munbers—with fafinite precision, Somme of the problems encountered in its use arise from the distinefion between finite dimension and finite information, and the distinction hetwven “finite” and “manageable” problem sizes. These arc all issuoy on the practical side of Kalman filtering that must be consickresl along with the theory. © [risa complete starvical characterization of an estimasion problem. Ikis much more than an esrimazor: because it propagates the cee probability distribution ol the variables itis tasked lo estimate. This is a complete characterization of the current state of keowledge of tae dynamic systern, including the influence of all past measurements. These probability distributions are also usefil for statistical analysis and the predictive design ef sensor systeras, © fr @ liniwa contest, ies a learning metiod. W uses a model of the estimation problem that distinguishes between pheozena (what one is able t observe), houmteno (what is really going op), and the state of knowledge about the owmacra that one ean dladuee from the phenomena. That state of knewvledye is rypreseated by probubility distbutons. To the extent thue thase probability Uiswibutions represent Fovndedge of the aval world and the cumulative processing of knowledge is Jeuirring, this is a learning process. Its a lazly simple one, but quite elective ta many applications 1 these answers provide the level of undesstanding that you weve seeking, then there is no need for you to tead the zest of the book. If you seed to understand Kalmaa filters well enough to use them, then read ont 1.1.2 How It Came to Be Called a Filter Ir might seem strange that the term “filer” would apply t an estimator, More commonly, a filer is a physical device for removing unwanted fractions of mixtures, {The wort fel comes fiom the same medieval Latin stein, toe the material was used asa flor for iguids,) Originally, « filter solved the problem of separating anwarned ‘components af gas-liquid-solid mistupes. In the emu of exystal madios and vacuurn lubes, the tenn was applied to anslog circuits that “ler” electronic signals. These1 OW RALNAN FtLreRiNS 3 Karan ‘farny Teas Sate con, | Som Toest | Proatiiy | Oyearie cecavs | hooy” | ete Mathematical fourations: Fig. 1.1, Purdistional conospis fn Kalman fering. signals are inistures of different frequency cormponents, and these physieal deviees ‘preferentially alenuate unwanted frequencies ‘Tas concep was extended in the 1930s snd 1940s to the separation of “signals” from “noise,” both of waich were characterized by their power spectal densities. Kolmogorov and Wiener used this statistical charneterization of their probability Alistsbutions in forming an optimal estimate ofthe signal, given the soma of the sigmal and noi With Kalman filtering the term assumed a meaning that is well beyonel the coriginal idea of separation of the components of a mixture, It has also come 1 include the solution of an imersion problem, in which one knows how te zepresent the measurable variables as Tinctions of the variables of principal imezest, a essence, it invers this fuactioaal relationship aad estimates the independent variables as inverted fonctions of the dependent (measurable) variables. These variables of interest are also allowed t0 be dynamic, with dynamies thar are only pastally predictable, 1.1.3. Its Mathematical Foundations Figue 1.1 depicts the essential subjects forming the foundations for Kalinan Altering theory, Although this shows Kalman filtering as the apex of a pyramidk itis iselF but pit of the foundations of another discipline—"modern” control theory—ane a proper subset of statistical decision theory. ‘We will examine only the top three layers of the pytarnid in this book, and a kde of the underlying mathematics! (matrix theory) in Appeadi B. 1.1.4 What It Is Used For ‘The applicaiions of Kalman filtering encompass many fields, but ils use as « too] is almost exclusively Tor two pimposes: estimation ond performance anaiysis of estimators, "ice est that ote no exarrne the Haina ayers ese malhemote! Kudaion lo ery sunnes. They every rst on uma iat, the Roundations of wish ae ot 36 well widest4 enenAL NrOREATION Role i: Hsrimating the Stare of Denaile Systems What is a dynamic system? Almost everything, if you are picky about it, Except for a few fundamental physical constants, thea: is hardly unything in the universe that is. ‘constant. The exbital parumncters of the asteroid Ceres are aot constant, ard even ike “Iised” slacs and continents are moving. Neatly all physical systems are dysamic lo some cegzee, I one wanls very precise estimates of theit characleristies over time, then one has 10 take their dynamics into considera tion, The problem is thar one does not always know their éynamics very precisely either. Given this state of pastal ignorance, the best one can do is express our ignorsice more precisely —using probabilities, The Kalmau filter allows us to estimate the state of dynamic systems with certain types of random behavior by using such statistical intoemation, A few examples of such systeras are listed in the Sconnd column of Table 1.1 Rote 2: The Analysis of Estimation Systems, The thie column of Table 1 lists some possible sensor types that might be used in estimating the state of the corresponing dynamic systems. The objective of design analysis is 1 Selermine how best io use these seasor types for a given set of design criteria, ‘These ctiferis are typically related to estimation accutacy and syste cost ‘Tae Kalman filter uses a complete description of the probability distribution ofits cstimation ez1ors in determining the optimal filtering gains, and this probability distribution may be used in vssessing its pevfonmance as a fimotion of the “desig paramcters” of an estimation system, such as f# the types oF sensors to be used, © the locations and orientations of the various sensor types with respect to the syste to he estimated, TABLE 1.1 Examples of Estimation Probloms ‘pplication Dynamio Sysiom Sena Tynes Provess cons! Chantel plant Pras Temporatrs low rate Gas analyeor Flood presioton Wer system Water lev! Fan gauge ‘woather rex Tracking Spacenrat Facer Imaging system Navigation Shp Sextent og Syroscape ARovelarometor Global Fostionng System (GPS) recaver12 owEsrinarion teTHOOS 5 «© the allowable noise characteristics of the senso + the prefltering methoc's for smoothing sensor noise, the data sampling rates fr the various sensor types, and ‘the level of model simplification to reduce implementation requirements ‘The analytical capability oF the Kalman filter formalism also allows a dosizer to assign an “error budget" to subsystems of an estimation system and to trade off the budget allocations to optimize cost or other measures of performance hile achieving « required level of estimation aceuraey. 1.2 ON ESTIMATION METHODS: We consider here jus a few oF the sourves of intelectual material presented in the ‘remaining chaplers and principally those contibuvors* whose Kiefines are shown. fn Figure 12. These cover only 500 years. and the study and developmen of ‘mathematical concepts goes back beyond history. Readers interested in more detailed histories of the subject ace referred (o the survey articles by Kailath [25, 76), Lainiotis [192], Mendel and Geiseking [203], sod Sorenson [47, 224] and the personal accounts of Battin [135] and Schmidt [216], 1.21 Beginnings of Estimation Theory ‘Tae first method for forming an oprimal estimate fiom noisy date is the method of least squares. Tes discovery is generally atibuted to Cust Friedrich (777 1855) in 1795. The daevitabilily of measurement cetors had been recognized since the time of Galileo Galilei (1564 1642) , but this was the Grst formal method for dealing with them, Altiough it is moze commonly used (or linear estimation ‘problems, Gauss fist sed il lor 2 nonlinear estimation problem in mathematical asteoaomy, which was part of a dramatic moment ia the history of astronomy. The following narative was gleaned from muy sources, with the majority of the material from the account by Baler and Makeinson [97 On January 1. 1801, the fist day of the sinciceuth ezmtuy: the Talian astronomer Giuseppe Pinz2i wns checking an eniy in a si catalog. Uibeauoom co azz the sentry had eon sad erwoncousiy by the printer, While sewing For the “missing™ sur, Paves discovered: inslead, # new plat IL was Ceres ihe largest of the minor ‘plies an! dhe Brat i he disecnened—Bu. Paz did not kw ha ye. He was ale ‘vock and measure is apparent motion against the “Zxed” star background duving 4L sights of viewing from Palermo befors his work was intsrmpred, Wihem he retemod 79 bis worl, lowever, he was unable to find Ceres again, tony commuter aller RF. Kats his i 6 Ge. Beran. at ey apes. voeale of munerivaly sce extn. rhode, Oer sea coining ve ackomledgod ia Chapter 6enenAL NrOREATION 100 10 100 100 10 me Cohan Lependee Sale Gass Vesa Mawel teal, ie Higgs Cree Now Wns Berea Kolm tate 1300 tan m0 18 1900 200 Fig. 1.2. Lioines of materencod historical tires and RE, Kaban, (On Januaty 24, Maes’ had weillen of his discovery to Johann Bode, Bode is best Jnowa for Bode’ tw, which sttes that tie distances of te planets fom the sua, 2 astronomice! units, ae given by the sequence A= Ge BM) fora —99, 0.1204 aan Actually ie was wot Bode, but Johann Tietz who fst proposed this fora, Sm 1772. Ar ‘ha time shore ware only six Known ples. hy 1781, Fridrigh Herschol disceners Uns which Gt nicely ino this formu for 9 = 6. No ple had bees diver fe 5. Spurred on by Ble, an association of Eumipean asiranomrs had been seuelling for he “missing” eight planet for nearly 3D yeacs.Pigeti was nol past of ‘hus association, but te id inform Bade of his nnincendes: discovery. Piazais keer diel wot vsaeh Bode ust March 29. (Eletronic tual was discovered sush lats:) Bode suspected thar Pia’ disoxery mit be the essing planet, but there wes insalficient da For eecrining iy onbital elemsenis by the mets hen availble, 19 2 problem in monlineseequstions that Newton, hime, hae etaned as ‘aeiig among the mos ilficul in mathematical astronomy. Noboxy had solved it at, 9 9 soaul, Cares ss lost in space again Pigs discoveries wore not published unél the suum of 1801, The possible iscovory and subsequent loss oP anew planet, coinciding withthe hoginning of « eww ceniury, was exciting news. IL contradicted a philosuphical jscaion for there ‘being cnly seven planels—dhe womber knew before Ceres and a aumber defended by the respected philosopher Geers Hess, among, ers. Hegel had tecesy publishes! a ‘pool in which he chestised the astrosomers for wasting tei tue in searching for an cighth planet wher there was w sound philosepticaljusifetion for there bing ony seven. The new planet became 2 subject of eomersation in inlets civies nearly the problem cauht tke stlemion ef @ 24-yearold mathemse12 owEsrinarion teTHOOS 7 ons al uy wilh he enti detecmination problem afew weeks earlier fl ha sel it aside Gor vlher interests, He nine devoted most of hit tame Ur the prabler, produced an estimate oF te orbit of Ceres bx December, and sent his ssulls to Piaze, ‘The new planet. which had been sighted onthe fits day of the yea, ws Hound agin ly ts discoverer om the lst cay of the year Gauss didnot publish his ovbit_dcsemination methods until 1809.° In this ublicton, he abo described the meboad of leas squares that he ha discovered in 1795, a dhe age of 14, and hud used iL refining. is estas of he enbit af Ceres, Although Ceres pleved a significant role in the history of discovery and it still reappears repularly in the nighttime sky, i hus fided into obscurity as sn ohject of {intellectual interest, The method of least squares, om the other hard, has been mn object of continuing interest and henefit to generations of scientists and rochnol- ogists ever since its iaoduction. JL hes had a profound eflet om the history of scieave, It was the List oplimal estimation method, and i provided an important connection between the experimental and theoretical sciences: It gave experimen tolists a practical method for estimating the unknown parameters of theoretical models. 1.2.2 Method of Least Squares ‘The following exanple of a least-squires problem ix the one mext often see although the method of least squaces may be applied to a muck greater range of problems, EXAMPLE Ll: Leas¢Squares Solution for Overdetermined Linear Systems, Gauss discovered that iF he wrote a system of equations in matrix frm, a5 fy tha ass de TP ty te tt [Ps fay the Bay dy [Pas | = (12) fn ti go bn fe te ua the manlin, the na of at squares Bad sen discovered independeody and publi By “Andvice Maric Logon (1752 1433) ia Pance nd Rabeat Advi (1778185) inthe Llestea Sass [17 eed ss bose casoesed cade boone Gass oe bean y thc Getana-Svs sii oss oinvch Taner: (1726-177) Suh Jung synchronicity ie. the phenome of mle, neue silane dearer a8 toe epee lore treater rey, 28 in the Wiener fer and the Kalan fle8 enenAL NrOREATION ther he could consider the problem of solving for that value of an estioare & {pronounced "xshat”) that rainimizes the “estimated snegsurement error” JF — 2, He could characterize that estimation error in terms of is Euclideat vector nr [1% — 2), or, equivalently, its square: aay as) which is @ continuously dillerentiable fimetion of the m unknowns .f.%3.8+ This fiction :7G) > 96 as aay eomponent i, —> too. Consequently. it will achieve its minimum vaine where all its derivatives with respect to the 3 are zero, There are 7 such equations of the form 0.8) = = 28% 5 - y | (7) toe = 1.2.3... Noe thal in his lst equation the expression Eni 08) the mh row of HE — 2, and the ovrermost summation is equivalent tothe dovproduet ‘of te kik column of H with A — =, Therefore Eqvation 1.7 eam be written as o=2nM r=] 09) 2H" HS — 2's (1.10) Hi where the matrix amspose His defined as Fey day bey Irn fiz dng he faa P= | ay fag ty fas, ay Fy Bay tay Frow12 owEsrinarion teTHOOS 9 The normal equatton of the Uneae least squares problem, The ecuation Wes 13} iscalled the norma equation or the normal form of the equation for the linear east squares problem. It has precisely as many equivalent sealar equations as unknowns. The Gramion of the near least squares problewy, The voral equation has the solution ary Us, provided that the matrix gait aay is nonsingular (ie.. invertible). The mattis product = HAF in this equation is, called the Gruman monte.* The determinont of the Gramion mais cheravieriass ‘whether or not he column vectors of H are linearly independent. Lf its determinanc is zero, the column vectors of H ace linearly dependent, aud £ cannot be determined uniquely. F its determinant is aowzero, then the solution # is umiguely determined, Leasiesquares solution. In the cose that the Grarian matrix i swvertible (je., ‘nonsingulay), the solution £ is called the least-squares solution ofthe evendctermined linear aversion ptoblem. IL is an estimate that makes 0 assumptions about the notre of the unknown measurement ertors, although Gauss alluded 10 that possibibiy in his description of tne method. The Formal treatment of wieertainty ‘in estimation would come later. This form of Ge Gramian matrix will be used in Chapter 2 t0 define the observability matrix of a linear dynamic system model in discrete time, Least Squares in Continuous Time. The following example illustrates how the principle of least squares can be applied to fiting a vector-valued parameuic mode! to data in continuous tae. Ir also illustrates iow the issue of determinacy ie, whether these is a unique solution 10 the problem) is characterized by the ‘Grsmina mateix in this context. “Mars fr fhe Danish rafratiean Jongen Peden Gram (ISD (916, This mato ao ead ‘va aed ne ans Fr inhravnn mare, mn afar the Exes tian Ronse Ae ‘ishr 1890-1962) Altcugh inforawian momice ad Granian mamices ae ¢ifeeat definitions and shes, hey srt simon hee ing mi clarence. The form sii ftom tthe ler dfurmaion wee repent dhe evel aban for smile aes fn ot ‘pki distin. Wt corespoads 10 ssled version of the Ciamian wate when De esse exrs ia ive ajlnt Gassan sisubuea, withthe scaling rtd to be uassctiny of the meat ‘Ue. The inforation mauris sa gaan seas chawerizon uf Be ofrtion™ (i 30me ers) fal in hed se etal, Te Gn fee hae edd cn gla gsirie ehaaceiatio ofthe enunses of te sahon10 enenAL NrOREATION EXAMPLE 1.2: LeastSquares Fitting of Vector-Valued Data in Continuous “Time Suppose that, for each value oftime fon an interval fy < F< 4y.=(0) ivan f= dimensional signal vector Out is modeled as 2 funetioa of an unknown n-veetor x by the equation a) = Hid. she H(i) isa known & x matrix. The squared ero inthis elation stench time 1 will be 2) = [2 — Hees s'LAHU: = 2x + EEO! ‘The squared integrated ervor aver the inverval will then be the integral itr = | ene [foment] ae[f toca] [sora which bas exevily the same array structure with respect (0a the algebraic lenst- squares problem. The Teast-squares sofution for x ean he found. as before, by taking the devivatives of |e" with respect to the components of x and equating them t0 7er0, The resulting equations have the solution z [I | [I woxn| |, , provided thar the costespouding Gramian matrix [ tenow is nonsingular, This form of Ge Gramian matrix will be used in Chapter 2 t0 define the cobservabitie motrix of a linear dynamic system model in coutimuons time. 1.2.3. Gramian Matrix and Observability For the examples considered above, observability does not depend upon the ‘measurable data (2). Tr depends only on the nonsingulsity of the Gramian matrix (9), whieh depends oly on the linear constraint mantix (IP between the unknowns. and kaowas.12 owEsrinarion teTHOOS 1" Observabiiqy of a set of unknown variables is the issue of whether o° not their values are aniguely determinable from a given set oF consindints, expressed as equations involving functions of the unknown variables. The unknowa vasiables said to be observable if their values sre uniquely determinable from the given constrainis, and they aze said (o be imobserveble i Usey are not untiouely determine able Icom the given eonsiaints The condition of nonsingularity (or “fill sank") of the Gramian matcis is an algebraic characterization of observability when the coastaining equations are Jinear in the unknown variables. I also applies to rhe case that the constraining equations are nar exact, dive to errors in the values of de allegedly known parameters (ofthe equations, ‘The Granwizn matrix will be used in Chapter 2 to define observability ofthe states fof dynamic systems in continuous tine and discrete tne. 1.2.4 Introduction of Probability Theory Beginnings of Probability Theory. Piobabilities represent the siete of knowl edge about piysical phenomena by proviing something more useful than “1 don’t know” (@ questions invotving uncertainty. One of the mysteries in the history of science is why it 100k so long for niatiematicians 10 formalize a subject of such practical importance. The Romans were selling insurance and annuities long before expectancy snd risk were concepts of serious mihematial interest, Much later, the Iislians were issuing insurance policies agains. business risks in the early Rensis- sanee, and the List known allempis at a theory of probabilities for games of chance occurred in that period. The Iulion Girolamo Cardano’ (S01 1576) pestormed an accurate analysis of probabilities for games involving dice. He assumed thar successive tasses of the dice were statistically independent events He and the contemporary Indian writer Brahmagupta cated without proof that the accuracies af empirical statistics tend to imprave with the number of tals. This ‘would liver be formalized as a Faw of large numbers. ‘More gencral treatments of probabilities were developed by Blaise Paseal (1623— 1662), Piene de Fermst (1601-1655), and Christizan Huygens (1628-1695). Fenmats woik on combinations way taken up by Jakob for James) Bernoulli (1654 1705), who is considered by some historians to be the Founder of probsbility theory. He pave the fist rigorous proo! of the Iw of laxge aumbers lor repesied independent trials (now called Bemnouli tials), Thomas Bayes (1702-1761) detived his famous rule for statistical infeceace sometime after Bernoulli, Abrabam de Moivre (1667-1754), Pieste Simon Marguis de Laplace (1749-1827), Acvien Marie Legenclre (1752-1833), and Carl Friedrich Gauss (1777-1893) continued this developraent into the nineteenth century. *Ciuouo was prctcnsplsisinn a Milan sho also ots books on maitemztcs ls book De Ludo os, on te mecha arly it mes of chawe (pineal ie gunveh was psa ney ‘emi lees eh Cand 8 alk the rven a ee! em EO wT Ks utolies smetines elle he Canon joa Cann sh2 enenAL NrOREATION Between the early pineteenth century and the inid-twentith cennry, the probe abilities themselves beyan to take or more meaning as physically significant aributes. The idga that the laws of nature embrae random phenomena, and that these are testable by probabilistic models beyua to emerge in the nineteenth century ‘Tae development and application of probebilistic models lic the physical world expanded rapidly ia that period. It even became an impoztant par af sociology. The work of James Clerk Maxwell (1831 1879) in statistical mechanics established the probabilistic treatment of natural phenomena as a scientific (aad successful) iseipline, ‘An important figuse in probability cheocy and the theory of tandom processes in the prenteth century was the Russian academician Andrei Nikolgeovich Kolno- gorov (1903-1987), Starting around 1925, working with H. Ya, Khinchin and others, he avestablished the foundations of probability teary on measurement theory, whieh Docaune the aecepred mathematical basis of povbability und rmdom processes. Alongs with Norbart Wiener (18041964), he is rodited with founding ame of the theary of prediction, smoothing ane filtering of Markov processes, and the general theory of ergodic processes. [lis was ihe fast formal theory of optimal estimsiion For systems involving randam processes. 1.2.5 Wiener Filter Norbert Wiener (1894-1964) is one of the snore famous prodigies of the exely ‘twentieth century. He was taught by big futher until the age of 9, when he entered high seligol. He finished high school at the age of 11 and completed iis under- uaduate degree fn mathematics in three years at Tufis University. He then entered sduate school at Harvard University at the aye of 14 and ceanpleted his doctorate degree in the philosophy of mathematics when he wast 18. He studicel abyiad ard teied his hand at several jobs far six moze years. Then, in 1919, he obtained a leachieg sppoiniment al the Massachuseuls Institute of Technology (MIT). Te remained on the faculty at MIT for the rest of his Ife. 1a the popular sciensific press, Wiener is probably more famons Zor aarning and promoting cvberneties dan for developing the Wiener filter, Some of his greatest ‘mathematical achievements were i generalized harmonic analysis, in which he extended the Fourier trausform to functions of finite panier Previous results were resbicted (© fictions of finite exer: which is an wreasonable constraint for signals on the real Line. Another of bis many aebievements involving the generalized Fourier tansform was proving that the transform of white noise is also white noise.” Wiener Fitter Development, tn the early years ofthe World War I, Wiener was involved in @ military project to design’ am automatic contraller for direeting ansiaireraft fire with tadar information, Because the speed! of the wirplame is a “eis alse cred withthe dicavary dhs the power pour dase of signal aula the Four irnshinm nie ouloomtation Keto, along i sas hier dcowert aL Fie Hed kr efor bm12 owEsrinarion teTHOOS 13 nonneyligible faction of the speed of bullets, this system was required to "shoot into the future.” That is, the controller had to predict the fiture course of is target using rnniyy nada tracking dats. Wiener devived the solusion forthe lcast-mean-sgpaired prediction error in terms fof the aulecorzelation functions of Ise signal and the noise, The solution is ia the form of ax integral operator that ean be synthesized with analog eizeuils, givea cetlain constiaints on te segularity of the aulovortelation functions ot, equivalently, their Fourier transforms. His approach represeats che probabilistic name of random phenomena in terms of power spectral densities. ‘An analogous detivation of the optimal linear predictor for discrete-time systems ‘as publisied by A. N, Kolmogorov in 1941, when Wiener was just completing bis ‘work on the continuous-time predictor Wiener's work was not dcelusificd uotil the bale 1840s, in a report titled “Extrapolation, interpolation, snd smnaething of stationary time series” The 6le ‘was subsequently shortened! 0 “Time series." A curly edition of the report hal a yellow cover, and it came to be called “the yellow peril” It was loaded with ‘mattiematical details beyond the grasp of most engineeting undergraduates, bul it was absorbed and used by a generation of dedicated graduate stxlenis in electrical engineering. 1.2.6 Kalman Filtor Rudolf Emil Kalman was born on May .9, 1930, in Budapest, the son of Orto and Ursula Kalman, The family emignned fom Hungary to the United States doring ‘World War TT. Th 1943, when the war in the Mediterranean was essentially over, they traveled through Turkey and Africa on an exodus that eventually brought them to ‘Youngstown, Ohio, in 1944. Rudolf aulended Youngsiown College there lor three years belore entering MIT. ‘Kalmea reovived his bachelor’s and master's degrees in electrical engineering a MIT in 1953 and 1954, respectively. His graduate advisor was Etust Adolph Guillemin, amd his thesis opie was the behavior ef solutioas of second-order difference equations [114], When he undertook the iavestigation, it was suspected that seconk-order cifference equations might be modeled by something analogous to the describing fictions used for seconc-order differential equations, Kalman diseovered tht their solutions were not at all ike the solutions of diffesential equations. In fact, they were found to exhibit chaotie bchavior, Tin the fall of 1955, afer a year building large analog control system for the ET DuPoat Company. Kalman obained an appointment as lecturer and graduate student al Columbia University. At tbat time, Columbia was well known [or the work in control theory by John R. Ragazzini, Lotil A. Zadeh,” and others, Kalean taught at Columbia vniil he complered the Doctor of Science degree there in 1987, For the next year, Kalman worked at the research Inboratory of the International Business Machines Corporation in Poughkeepsie and for six years after that at the cls pecaps mone Famous asthe “Fair” of Fuzzy syste Eicon’ at anceps easing,4 enenAL NrOREATION research center ofthe Glens L, Martin coingany in Boltinove, the Revearet Institute for Advanced Studies (RIAS). Early Research Interests. The algebraic nawwre of systems theory frst became of interest 1o Kalman ia 1953, when he read a paper by Ragszzini published the previous yew, It was on the subject oF sampled-data systems, for which the time ‘variable is discrete valued. When Kalinan realized that linear discrete-time systenis could be solved by transform methexs, just like linear continueus-time syaters, the idea occurred to him that there is no fimdamental difference hetween continuous and aliserete linear systems. The two must be euvalent in sou sense, even though the solutions of linear differential equations cannot go to zero (an stay thers) in ints lime and those of discrete-time systems can, That slarted his interest in the ‘connections between systems theory and algebra. In 1954 Kalman began smdying the issue of coatroltability, which is the question ‘of whether there exists an input (contro!) fimetion to a dynamic system that will “sive the scae of Gar syscein to Zero, He was encouraged ond aided by the work of Robert W. Bass during this period, The issue of eventual interest to Kalinan was whether there is an algebraic condition for contrsllability. That conditian was eventually found as the rank of a matrix.’ This implied a convection between algebra and systems theory, Discovery of the Kalman Filter. 1n late November of 1958, not long alter coming t RIAS, Kalman was returning by ‘rain to Baltimore from a visit 1 Princeton, At acound IL PM, the train was halted for about an hour just outside Baltimore. Tt was late, he was tired, and he hud # headache. While he was trapped there om the trata for that hour, an idea oveursed to him: Why wot apply dhe notion of siaie variables? to the Wiener filtering problem? Te was (oo tired 10 (ink much ‘more about il thal evening, but it marked the beyinning of a great exercise lo Go just that, He sead cirough Loves book on probability theory [68] and equated expectation wich projection. That proved to be pivotal in the derivation of the Kalman filter, With tte additional assumption of finite cineusionality ite was able to derive the Wiener filter as what we now call the Kalman filer, With the change ta stave-space fon, the mathematical background needed for de derivation became ‘much simpler, and she pron were within dhe mathematical reach of many under uaduates, Introduction of the Kalman Filter. Kalman presented his new results in talks at several universities and research laboratories before it appeared in prin,"° Hig ideas ‘were met with some skepticisin among is peers, and he chose a mechanical Se comtetaitis mans, come dened im Caper 2. “Altnous fonction space mashods Ker ths the prefered approach to he Flin pblom, tbe use of stusapace model fx tnesayiag yas ad also bes nrodiee (O° Late und ta [67] in 1956), "memento fhe sere fe Kanan ler ad en pce y Sein [2271 1959 and Sratonomch [5,26] Ln 18512 owEsrinarion teTHOOS 5 engineering journal (rather chan an electrical engineering journal) for publication, because “When you fear stepping on hallowed ground with entrenched interests, its best to yo sideways.” Hlis second paper, om the comfinuousetime ease, was once rejected because—as ene referee put tone step in the proof “cunt possibly be true.” (It was (rue.) Te persisted in presenting, is lilter, and! dere was more ‘nnmediate scceplance elsewltere. Il soon became the basis for research lopics at many universities and the subject of dozens of doctoral theses in electrical engineering over He next several yeas, Early Applications. Kalan found a receptive audience for his Stern the fal of 1960 in a visit ro Stanley K. Schunidt at the Aries Research Center of NASA in ‘Mouptain Vow, California [118]. Kalan deseria his revert result and Schmidt recognized its potential applicability to x problem then being sted at Amex —the trajectory estimation and eantrol problem for the Apallo project. & planned mined ‘mission to the moon and back, Schmid’ begin work immediately on what was ‘probably the fis fll impplementaton of ths Kalman ier. Te soon discovered shat ‘snow called “ester Kalan filtering.” which kos been used ever snes for most real-time noolinesr applications of Karna fering, Entbused aver is own suceess ‘withthe Katman filter, he ser about proselytizing o‘hers involved in similar work, Ia te carly part of 1961, Ssimidt describe his results to Richard H, Babin fora the MIT Instrumentation Laboratory (later renamed the Charles Stark Draper Labora tory), Battin was already using state space methods forthe design and mplesnentae tioa of astzonautical guidance sysiems, and ke made the Kalman filter past of the ‘Apolla onboard ausdance.' wise was designed and develope si tke Intrumenia- tion Laboratory. In the mid-1960s, theougis the influence of Schmidi, the Kaloxan ‘ter became past of dhe Noxthrup-built navigation system forthe CSA air anspor, then being designed by Lockineed Aircraft Company. The Kalman filter solved the data fision problem assovated with combining radar data with inertial sensor daa 0 arive a an overall estimate of te aircraft trectory and the daw vejetion problons associated with detecting exogenous errors in measurement daa. It has been a integral part of acurly every onboard wajcetory estimation and eontel system dosigned singe that time, Other Research Interests. round 1960, Kalmen showed that the related noxion of observability for dynamic. systems had an algebraic dual relacionship with controllability. Thar is, by the proper exchange of system parseneters, one problem could be transformed into the otter, and viee versa Richard S, Buey vats also at RIAS in that period, and it was he who suggested Kalman that the WicuerHopf equation is equivalent to the matrix Riceati equa- "thie aut segments in hie paragraph arena ak om Sure Tew ocd Present ven by Kelenan athe Unveniy of Califamia 2t Lob Angeles (UCLA) ee Apel 17, 1491 un 2 specu ongataed sa hosted by A. Balkibtnea ot UCLA a sponsored jo by UCLA an he Natio ‘Aarau aud Spee Adiriniseton (NASA) iden Takats. "mater andar paver Kalo ie mplenerition mes at mae Sm ler Wy Jas E. Powe at MIT Tastuentaon Laborato. This il be eseussed ia dhe nox sabsceon.16 enenAL NrOREATION tion if one assumes a fnite-dimnensional state-space model, The general aature of this relationship between integral equations and differential equations first became apparent around that time. One of the more rmasluble achievements of Kalman ard ‘Bucy in taut periad vats proving that the Ricealf equation cas bave a stable (steady slai2) solution even if the dynamic system is unsiable provided that the system is observable and controllable ‘Kalman also played a Jeading zole in the development of realization theory, witich also began to face shape around 1962. This theory adresses the problem. of finding a system mode! to explain the observed input-outpur behavior of a system. This line ‘of vestigation Jed to a uaigueness principle for the mapping of exact (ie noiseless) dsta fo linear system snodels. ny 1985, Kalman was awarded the Kyoto Prize, considered by some to be the Japanese equivalent of the Nobel Prize, On his visit to Japan to aecepe the Kyoto Prize, he relaned to the press un epigram that he had frst seen in « pub in Colorade Springs in 1962, and ic had made an impression on him, Tt said: inte people diseussothar people Average pene divers events Big peuple discus ideas His own werk, he felt, hud boen concerned with Mess. Tn 1990, on the oceasion af Kalman’, sistioth birthday, « speetal ternational symposium was convened lor the purpose of honoring his pioneering achievements in what has come 10 be called maiiemiotica! system theory, and a Fespschrift with that title was published soon after [3 Impact of Kalman Filtering on Technology. Krom the stwndpoint of those involved in estimation ark! control problems, at least, this hus to be considered the ucwatest achievement in estimation thoory of the sventioth century. Many ef the achievements sinee its imauduetion would not bave been possible without it. Te was ‘one of the ensbling teeanologies lor the Space Age, in particular. The precise and cllicient navigation of spacecrall through the solar system could not have been done without i “The principal uses of Kalmaa Altering have been in “modesn” control systems, ia the tracking and navigation of all sores of vehicles, snd in predictive design of estimation and control systems. Tliese technical activities were made possible by the introduction of the Kalman filter. (If you need a demonstration of its impact on technology, enter the keyword “Kalman filer” in a technical literature search, You will be overwhelmed by the shoev number of references it will uenerate,) Relative Advantages of Kalman and Wiener Filtering L. The Wiener filter implementation in analog cloetonies can operate at much higher elective throughpur thi ike (digital) Kalman tier 2. The Kalman filter is implementable in the farm of an algorithm for a digital computer, which was replacing axalog eizcuitzy for estimation and eanteol at12 owEsrinarion teTHOOS 7 the time that the Kalman filter as introduced, This implenentation may be slower, but itis capable of much greater accuracy than had been achievable with analog ft, 3. The Wiener filter does not require finite-dimensional stochastic process models for the signal and noise 4. The Kalman filer does nat require that the deterministie dymamies or the rindem processes have stationary properties, and mimy applications of {importance include nonstationary stochastic processes. 5. The Kalman filter is compatible with the state-space formulation of optim: controllers for dynamic sysiems, aad Kalman was sble to prove uselil dul ‘properties of estimation end vonteol far these systems, 6, For the modern contvols engineezing student, the Kalman filter requizes less additional snathematical preparation lo leara and use than tne Wiener fille. As fa result, tae Kalman Eller can be aug at the undergraduate level in engineering custicula 7. The Kalman filer provides the necessary information for mathematically sound, statisticaly-based decision methods for deveoting and rejecting anom- lous measurements, 1.2.7 Square-Root Methods and All That Numerical Stability Problems. The grat success of Kalman filtering was not without its problems, not the Teast of which was marginal stabifity of the numerical solution of the associated Riccati equation. In some applications, small roundolt crrors (ended to accumulate and eventually dewrade tte performace ofthe ltr: La the decades immediately following the introduction of the Kalman filter, there appeared seycal better mumecical implementions of tae original formulas. Many of these were adaptations of methods previously devived for the least squares problem. Early ad hoc Fixes. 1 vas discovered early on'* that forcing symmetry on the solution of the matrix Riceati equation improved its apparent monerical stabiliy—a. ‘phenomenon that was lster given x more theoretical basis by Verbsegen and Van Dooren [232], It was also found that the influence of roundoff errors could be ameliorated by artificially inercasing the eovarianee of process noise in the Riveati ‘equation. A symmetrized form of the diserete-me Riceati cquation was developer by Joseph [15] and used by R. C.K. Lee at Honeywell in 1964, Tis “structural” relormulation of the Kalman [ier equations improved robustness against roundolt errors in some applications, altaough later methods have performed better on some problems [125} "thew Mes were uppity conor induperdanly by seve people, Schmid C115] wal is legge a NASA had acosent Re Reo Gnesi ar “peelonose” Lo cover nem ftbes and cite R C.K, Lae a Honcewall wih the indopendet eatery of Ue syennety cet18 enenAL NrOREATION ‘Square-Root Filtering. These methods can also be considered es “seuetural”™ reformulations of the Riccati equation, and they prelate the Buey-loseph form, The first of these was the “square-riot™ implementation by Pole and Stern [208], fist published in 1963 and successfully implemented for space navigation on the Apollo ‘manned lunar esplosation progeam. Poller and Steea snuoduced tae idea of faciosieg. le covariance matrix into Chofesky factors, in the Format Pe ct, ay and expressing the observational update equations in terms of the Cholesky ficror C, rather than P. The result wus better numerical stability ofthe filter plementation at the expense of aod computational complexity. A generalization of the Potter ard Stern inetd to fuanlle vector-valued measurements was published by onc of the authors [130] in 1968, hur & more efficient implementation—in terms of wrimgutar Cholesky factors—was published hy Bennet in 1967 [138], Square-Root and UD Filters. There was a rather rapid development of faster algorithmic methods for square-root filtering in the 1970s, following the work at NASA/JPL (then called the Jet Propulsion Laboratory, atthe California Institute of Technology) in the late 1960s by Dyer ancl McReynolds [156] on temporal update methods for Cholesky factors, Extensions of square-root covariance ind information filters swore intuced in Kaminski’s 1971 thesis [115] at Stanford University, The fin of the trimgular factoring algoritams for the observational update was duc to Agee and Turner 106}, ia & 1972 repore of rather limited circulation, These algorithms have roughly the same computstional complexity as the conventional ‘Kana fiter, but with better numerical stability, The “fast triangwolae” algorichma of ‘Carlson was published in 1973 [149], followed by the “squareerootefree™ algorithm of Bierman in 1974 [7] and the associated temporal upcate method introduced by ‘Thornton [124]. The computational complexity oF the square-root filter for fime- inveriant aysteras was greatly simplified by Monf and Kailath [208] soon after that Specialized parallel processing architechies for fast solution of the square-root filter equations were developed by Jover and Kailath [175] and others over the next decade, and much simpler Getivations of these and earlier square-root. implementa lions were discovered by Katoh [26], Factorization Methods. The square-root methods make use of matrix decom position"® methods that were originully derived for the least-squares problem. These "4a aquaze wor 8 ofa ents sais the equation ?~ 88 Cc, withthe ranspose on exe scone ‘cin. Poti an Str's extn usceta spacial ype of sree mai cule n lenient a. ‘They fcr am elemertary mi as sare ther elerenary ra Te hy ct, the fatory were bray sgre ret he Fetred mats, This quare-rool appli has sack wi extern Per aint Sens apres, eve tsigh che ates red are Chey factors, eatin acre wok "tbe tema “decemvositn” rfsst the epreseruion of wis (ni eae, a covariance atid) 258 peice of mizios ving mot ese eomputional prope, seh os apraenes Cr wangue Taser) ot amid namenl sb (li organ flor). The ke “Tavira” used By Bice [7] for suc epson,12 owEsrinarion teTHOOS 19 include the socalled OR decomposition of a matrix as the product of an orthogonal rateix (Q) and a “trizngular™'® matrix (R), The mate ® results From the appBeation of oxthogonal transformations of the original matrix. These orthogonal tsrnsfirmae tions tend to be well conditioned sume ally. The operation oF applying these ‘wansloreations is called the “Isiaxgulatization”’ of the origiaal matrix, and isa aulacization methods derived by Givens [164], Householder [172], and Gentleman [163] are used 10 make Kalman filtering more robust against roundoll estors. 1.26 Beyond Kalman Filtering ‘with impunity—ard considerable suceess—to many nonlinear problems. These extensions generally use parlial derivatives as linear approximations of nonlinear relations. Sebmnici [118] introduced the idea of evaluating these partial éerivatives at the estimated value of the state variables. This approach is generally called the extended! Katman filter. bul was called tae Kumar Sefvmid filter in some easly publications. This aad otter methods for approximate linese softtions to nonlinear problems are discussed! in Chapter 5. where it is noted that these will not be adequate for all vonlinear problems. Mentioned here are some vestigations that lave addressed estimation problems from a miore general perspective, although they are nor covered in the rest of the book. Nonlinear Filtering Using Higher Order Approximations. Appreaches using higher order expansions of the filter equations (ie, beyond the linear terms) have been derived by Stratonavich [78], Kushner [191], Bucy [47], Bass et al [)341, and others for quadratic nontinearties and by Wiberg and Campbell [235] for terms through third order. Noniinear Stochastic Differential Equations. Problems involving nonlinear and random dynamic systems have been studied for some time in siatstical ‘mechanics. The propagation over time of the probubitig: diuiinaion of te sate of a nonlinesr dynamic system is described by a nonlizesr partial diffeential equation called the Fokker-Plonck equevion. It bas beea studied by Einstein [187], Foiker [160], Planck [207], Kalmogorow (187), Sttatonovich (78), Baras and Mireli [32], and others. Stratonavich modeled the effect on the probebility distbution oF information obtained Oirough noisy measurements of the dynamic system. an effect called conditioning The partial differrsial equation that includes these effects is ealled the conditioned Frkker-Planek equarion. Te has also hoon sugied by Kushner [191], Buey [147]. and others using the stochastic enfcutus of Kiyosi H6. also called the “IW calculus” Nis © non-Riemannian calculus devel- oped speciieely for stochastic different systems with noise of infinite bandwith Tais general approncl resuls i a slotastc pasta dillerenial equation describing "ice Chapkr 6 and Amped 8 fr clacusios of wing fe20 enenAL NrOREATION the evolution over time of the probability distribution over a “state space” of the dynamic system under study, The resulting medel does nor enjoy the finite representational chamicteristies of the Kalman filter, however. The computational complenity of obtaining a solution fir excouds the alzeady considerable burden of lke conventional Kalman lites. These methods are of significant isiezest and ulity ‘but are beyond the scape of this book. Point Processes and the Detection Problem. & point process is « type of Tandom provess for modeling events or objects that are distrhuted ever time or space, such as the arivals of messages a a communications switehing center or the Toeations of stars inthe sky. Ii also a mode forthe intial sates of systems in-memy estimation problems, such asthe locations ofaioraf or spaceeralt under surveillance by a radar installation or the Tocations of submarines in the ocean. The detection protien for these surveillance applications mist usually be solved before the estimation problem (i. tacking of the objects with a Kalan flee) ean begin ‘The Kalman filter vequires 2m initial state for each object, and that inital state estimate must be obtained by detecting it, Those initial states are distributed aecocding to some point provess, but these are no tcebmically mature methods {comparable to the Kalan filt:) for estimating the state of @ point process. A unified approach combining detection and tracing into one optimal estimation rnettiod was detived by Richardson [214] and specialized to several applications, ‘The detection and tacking problem for a single object is represented by the conditioned KokkerPlanck equation, Richardson derived from this one-ohject model an infinite hierarchy of pactial differential equations representing object ddonsives andl teuneated this hierarchy with a simple closure assumption about the relationships between orders of densities. ‘The result is a single partia) diffevential ‘equation approximating the evolution of the density of ebjeets. It ean be solved numerically. It provides « solution to the difficult problem. af detecting, dynamic ‘objecis whose initial states are represented by a point process. 1.3 ON THE NOTATION USED IN THIS BOOK 1.3.1 Symbolic Notation “The fundamental problem of symbolic natation, in almost any context, is that there are never enough symbols to go around, There are not enough letersi the Roman alphabet to represent the sounds of standard English, tet alone all the variables in Kalman filleriag andl is applications. As a resull, some symbols must play aulliple roles. In suclt eases. their roles willbe defined as they awe introduced. It is sometimes, ccanlising, but unavoidable. “Dot" Notation for Derivatives. Newton's notation using j(0), {2 for the first tnvo devivatives off with respect to # is used where convenient 10 Save it,19 OW THE NOTATION USED Wy This BOOK a TABLE 1.2 Standard Symbots of Kalman Filtering Symbois — symeel e we ie Detintion F F A DBysamis coetient matrx of continuous liner frei ‘squation defining dynam system 6 ' B Coupling matrix between random process raise ard state ot Thvear dynamic systom # M e Monsurement geneity ma defining linear relations Danvean state ofthe dynamic system and measurements . that ean be made x a K Kanan galn matibe Pe Pe Cavarianoe main of sale estimation uncesainy Q @ Covariance max ef process noxs9 in tna systom stato ynamies a ° Covariance main of observational (messiternert; ‘uncer ny x x State vari of a linear dynamic systom r Vecor (or scalar) of measuned valves ° State ranshion matty of a dere near dynamic system *Thslok f 19. 16,21) * Kaman (2 178° Ofer sours (10, 18,88 Standard Symbols for Kalman Filter Variables. Tnere appear to be wo “standard” comventioas in technical publications for the symbols used in Kalman filtering. The one used in this book is similar to the original notation of Kalman [179]. The other standerd notarion is sometimes associared with applications of Kalinan filtering in control theory. It uses the first few letters of the alphabet in place ‘ofthe Kalman notation, Both sess of symbol usages are presented in Table 1,2, along. ‘with the original (Kalman) notation, State Vector Notation for Kalman Fittering. ‘ne sats vector x has been lomd wit all sorts of othor appondages in the usage of Klin firing, Table 1.3 fists che notation ws in tis book (et elute) along wit notations found Sn some other sources second colunsn), The site vector wears a “hat as the extmeted value. § end subseripling to denote the sequence ovals tha he estate assumes ‘over time. The problem is that it has wo values at the same time: the a priori"? value ‘etore she measoremen: at te current tie has been used i refining te estimate) and the « postertaré vate (ater the curreat measnremeat hss been vised in reining the esimte). Ties dstneions are indicated by the sigrum. The negaive sign (—) indicates thea pri’ vate, and the postive sign 4) indicates the a posterior’ wae. "ibis uae of the fll Lie pnsos 0b sUlctivs forthe pior aud poweror ste ie as uafornwate ‘nice snd oan, Daisey wa ws ten Even th nial aeons are ‘ere Hho who Fe hs notaon had ow how sormrenpie wold ome, hey HE Juve aed tern or2 enenAL NrOREATION TABLE 1.3. Special State-Space Notation The Other book scuroze Detnitan of Notatonal Usage x Vector % The kth componert of tha vector x 5 AIK] The kt element ofthe sequence Kaa Ke OF YOEIOIS ® Fix) An esate a the value of x Rye» Apo state of x, sonettanad on a4 prior % measurements excopt the one at tne Hine postin estimate of x, condones 5. ofall avsllabie measurements atime t x Dative of with ragpact to # 4a} aio ‘TABLE 1.4 Common Notation for Array Dimensions Dimensions ‘Symbol__Vectar Name Dimensions Syritol Mat Name Row Galurnn «System stale a ® Slate anetion > oe w Process raise r G Process noise couping 9 Gamat nput © Process noise covariance 4? 2 Measurement ‘ HA Measurement sonsiivily Measurement noise Measurement reise oot Common Notation for Array Dimensions. Syinbols used for the dimensions of the “standard” arrays in Kulman filtering will also be standardized, using the notation of Gelb et al. [21] shown in Table 14, These symbols are not used exclusively for these purposes. (Otherwise, one would. sown ru out of alphabet) However, whenever one af hase arays is used in the discussion, these symbols will be used for their dimensions, 1.4 SUMMARY ‘The Kalman filter is an estimator used to estimate the stare of @ linear dynamic system perturbed by Goussion white noise using measurements that are lear fmetions of the system state but eorupted by additive Gaussian white noise, The ‘mathematical model used in the derivation of the Kalman filter is « reasonable representation for many problems of practical interest, including control probleme asbe suimuany 23 ‘well as estimation problems, The Kalman filter model is also used forthe analysis of measurement and estimation problems, “The method of least squares was the first “optimal” estimation method. 1 ys iseovered by Gauss (and others) around the end of the eighteenth ecntury, and 3x sill much in use today. I'he associsted Grarmian matrix is nonsingular, the method of least squares detetmsines the unique values ofa set of unkstown variables such that the squared deviation from a set of constraining. equations is minimized, Observabiligy ofa set of unknown variables isthe issue of whether or not they are uniquely determinable from a given set of constraining eguattons. If che constants are fincar functions of the unknown vasiables, then those variables are observable and only ¢ the associated Gramian matrix is nonsingulas, Ifthe Gramian onatrix is singular, ten the unknown variables ave amobsorvahle, ‘The MienerKolmoyoruy filter was derived in the 1940s by Norbert Wiener using 4 model ip continuous time) and Andrei Kolinogoroy (using a model 3m isevete time) working independently. Ih is a searlslout estimation merhod. It estimates the sale of a dynamic process so as to minimize dhe meun-sgreared estimation enor IL can take advantage of starivtical knowledge about random processes in terns of their power spectzal densities in the frequency domain The “starespace” model of a dynamic process uses differential equations (or difference equations) to represent both deterministic and randorn phenomena. The stare variables of this model are the variables of Snterest and their derivatives of, interest, Runcam processes are characterized in terms of their staistical properties in the rime domain, rather than the frequency domain, The Kalman filter wus derived as the solution the Wiener filtering problem using the state-space model for dynamic and rendoma processes. The result is ensier to derive (ond (0 use) than the Wiener Kolmogoroy filter, Squcre-root fitering is a reformulation of the Kalman filer for beiter numerical stability in finite-precision arithmetic. It is based on the same natuematical mode] but it uses an equivalent statistical parameter that is less sensitive to roundof errors jn the computation of optimal filler gains, It incorporates many of the more numerically stable computation methods that were originally derived for solving the least-squares problom, PROBLEMS L.A Jean Baptiste Fonsier (1768-1830) was studying the problem of approximaing 2 finction f() on the ciscle 0 << 2x by a linear combination of trigono- netic fanctions: FO 05 + ¥ [ap cos) +4; sint 0) (115)enenAL NrOREATION See if'you can help him on this problem, Use the method of least squares to demonstrate that the values ty z[ ro aa J, fees sta 1 1 [[ pes ia fof the coeffivionts a amd dy fin 1S approsimation error 2 give the lowst integrated squared F(ab) = i Jta, Bs, =f fo-roy a | 4g + Say cost (0) + 6, sin( Ot a Bi 2 I, {= + ¥ tla cost 8) + 6, sint safe +[ Pua. You may assume the equabiies ir [fstinsoanan=[& 144 p= i O jek [ Sint 0) inti) 48 {° i [ cost j@)sink@jd =0, OS7Em Sk en y as given,oman Sirti, Theory ond Proce Uaay ATLAS, Sevont Bon, Mander . Ges Angie P Arey eapyrght 2008 Ts Wiley & Sos, Fe SUNG: 0471992515 Ularbucky BATE26638-8 (Users) Linear Dynamic Systems What we espertenee of nature i tx model, en all of natures models are so beaut BR, Buckminster Fuller (1895-1983) 2.1 CHAPTER FOCUS Models for Dynamic Systems. Since their inroduetion hy Tease Newton in the seventeenth century, differential equations fave provided concise mathemotical models for many dynamic systems of importance 0 humans, By this device, ‘Newton was able to model the motions of tne planeis in our solar system with @ small number of varisbles and parameters. Given a finite mtanberof initial conditions (ihe initial positions and velocities of the sua and planets will do) and these cquations, one can uniguely determine the positions and velocities of the planets for alltime, The fnite-dimensional representation of a problem (inthis example, the problem of predicting the Future course ofthe planets) isthe basis forthe so-called state-space approach to the wpresentation of differential equations andl tht solutions, which is the focus oF this chapter. The dependent variables of the differential equations become state variahfes ofthe dymemie system. They explicitly represent all the imporant characteristics of the dynamic system at any time, ‘The whole of dynamic system theory isa subject of considerably mare scope than ‘one needs forthe present undertaking (Kalman fering), This chapter will stick to just those camcepts sharare essential for that purpose, which isthe development of the sate- apace representation for dynamic sysrems described by systems of linear differential equations, These sve given a somewhat heuristic trestment, without the mathematical rigor offen accorded the subject, omiting the develapment and use of the transform methods of functional anelysis for solving differential equations when they serve no purpose in the derivation of the Kalman filter. The interested reader will find & more ‘formal and thorough presentution in most upper-level and yeuduate-level textbookson "om incre quot by Cabin Takin a "Fm nse out orn” Thy Naw Tbs Kaman 8 1986. 2526 mean ane SYSTEMS ordinary dlifferential equations, The objective of the more engineering-oriented trgatments of dynamic systems is usually to solve the contots problem, whic is the problein oF defining the imputs (.c., contwol stings) that will bring the stase of the dynam spate tv a desirable condition, That is not the objective here, however. 241.1. Main Points to Be Covered ‘The abjeetive in this chapter is to characterize the measurable ouputs of dynamic systems as flinctions of the internal sfates anc! inputs of the sytem. (The italicized torins will be defined! more provisely further along.) The reatmeat here is deter nisti, in arder to define functional relationships between Saputs and ousputs. Tn the nest chuptoy, the inputs ave allewed © be nondeteceinistie ¢Le., candom), and the objective of the following chapter will be 10 estimate the states of the dynamic system in his context. Dynamic Systems and Differential Equations. 1s the context of Kalmen filtering. 9 dioume spvtem bas come to be symonymous with @ system af ontinary Jilferentisl equations deseribing the evoluvon over time of the state of a physical system. Tais movhematical model is used to derive ics solution, which specifies the fanevional dependence of the state variables on ther initial velues ard the system inpus. This Solution defines the functional dependence oF the measure outpits on te inputs and the coeFcions of the mode. Mathematical Models for Continuous and Diserete Time. The principal dynamic system models are summarized in Table 2.1. For implementation in digital computers, ihe problem representation is tansiormed from smn enclog model (Fime- tions of continuovs time) 1 & digital model (functions defined ar diserete times). Obvervabilty charicterizes the feasibildy of uniquely determining the stave of a sven dynaraic system if its outputs are known, This characteristic of a dvaamic system ig determinable from the parameters of ify mathematical mee 2.2. DYNAMIC SYSTEMS 2.2.1 Dynamic Systems Represented by Differential Equations A sysiens is an assemblage of intezrclated entities that can be considered as a vole, If the attnbutes of interest of a system are changing with time, then itis called dhnanic stem, process is the evolution over lime of s dycamie system, ‘Our solar system, consisting of ike sun and ils planets, is a paysical example of a dynamic system. The motions of these bodies are governed by laws of mation that i Es Bray po 3 = Hr+ Du Fig. 24 Bock agram ofa frear dynamic system. 2.3.2 Dynamic Coefficient Matrices and Input Coupling Matrices The dynamics of linear systems ate represented by 9 ser of first-order Tinear Jifferential equations expressible in weeror form as a su = Sate) a = Fenda) + Clas) 23) ‘where te elements and components of the matrices and vecrors can be functions of time: AM Fiat Fal es Fin) FD fl fl Lud Fy = | Al fol fo Lyd) Fal fol fl o> Flt eu 44 eA elf ex th cep=| one es ctl |, eal Ga) Ga Gl MO= WA WO we ako The matrix Fe) is called the dynamle coefficient maivts, or siomply the dynamic muarrde. Ts elements axe eallad the direamie enerficienzs. Phe matrix C() is ealled the hype coupling matric, and its elements are called input coupling coeficienss. The pevectar 1 48 ealled the inyrut vector.2 mean ane SYSTEMS EXAMPLE 2.4: Dymamig Equation for 4 Heating/Cooling System Consider the temperature 7 in a heated enclosed room or building as the state variable of a dynamic system. A simplified phint model for this dynamic system is the Enea xq to KITE) = F,G)] + dye) where the constant “cooling coefficient” &, depends ow te quality of taermal insulation fom the outside, 7, is the temperature outside. & is Ove beating /eooling tate coeffiefent of the heater or cooler, and wis an input funetion that is ethers {oll} ar v= 1 (on) and can be defined as a fiction of aay measurable quaatiies. The outside temperature Z,, on the olher hand, is an example of ax input function which may be dizecily messurable at any time but is not predictzble in the future. ILis effectively a raadom process, 2.3.3 Companion Form for Higher Order Derivatives Tn general, the uth-order linear differential equation 290), ft HO Gp thle bet a0 Ohm — 06 ‘can be rewritten as « system of'n Lirst-order dillerentisl equations. Although the state variable representation as @ IrsI-order system is not amique [S6], there is a unique ‘way of representing it called the companion form, Companion Form of the State Vector. For the ntionler linear dynamic system shown above, the companion finn af the state vector is d c a a Me) [0. Sy. SMO. oo Sw) - en Companion Form of the Differential Equation. The atir-order linear ditfer- ‘ential equation can by rewsitin in terms of the above state vector x(0) as the vector dlfferental equation af o 1 0 0 f9M)] Po (0 o 6 1 o lfam) Jo ape Je : : ail | |e fate. wy Ee Frail) ° ° a 1 ° aio | Lone “nel Lain) Lt 28)25, COWPNUOUS LINER SYSTEES AND THE SOLUTIONS 33 When Houation 2.8 iy compared with Fquation 2,5, the inarices #G@) and CUA) are easily identified, The Companion Form is iil-conditioned. Although it simplifies the zelation- ship between higher order linear dilleeential equations and Irstorder systems of dlifferential eciations, the companioa matrix is aot tecommiended for iaypleamenta- tion, Studies by Kenney and Liepni [185] have shown that # is poorly conditioned for solving differential equations. 2.8.4 Outputs and Measurement Sensitivity Matrices Measurable Outputs and Measurement Sensitivities. Only tbe inputs and ‘oulpuls of the sysiem ean be measured, and il is usual practice to consider the sariables 2; as the measured values. Far linear problems, they are related to the state variables and the inputs by « systema of Linear equations that can be represented ia vveetar form as st) = HOOK + DA, 2.9) where a0 Ce AG) Aylt) hislt) BGO Figg) Bayh) tgs) o> Ing) y= | al) rl) asl) Aye) alt belt) helt) Belt) ial) igh) esl) oo yl oy) dsl) xl) oo dof) Diy = | aul) doin) dytry 4 |, dy dott) dali) s+ dy) The vector 27) 3s called the merement vecier, oF the otput veeror of the systema, The coefficient Ay(s) represents the sensitivtgy(measuremear sensor scale facto of the Ah measured output tothe fh internal tae, Toe mates 1/0) oF these values is called the measwremcnr senstivigs matrix, ard D() is ealled the tapas ouput coupling mani. The measwenent seasiedes hg(0) ard input/output eoupling evtickons dy(0). 1515 4.1 £) <7, are known hunetions of time. The sce equation 2.5 and the output equation 2.9 together form the dynamic equations of the system shows in Figute 2.1oe mean ane SYSTEMS 2.3.5 Difference Equations and State Transition Matrices (STMs) Difference equations are the discrete-time versions of diferential equations. They are usually written in terms of forward differences s(¢,.)) — a(/,) of the state variable ‘ue dependent varinble), expressed as a fanccion y ofall independent variables or of the forward value x(¢,,)) asa function 9 ofall dependent variables (cluding the previous value as an independent variable), Mia) Hig) = Mt ay) ED Mis Prog otal Oa MH) HED, (2.10) 1) + Wests e€Dh ‘Tae second a these (Equation 2.10) bas tae same general form of the recursive relation showin in Fquation 2.4, which is the one that is usually implemented for discrete-time systems For linear dynamic systems, the functional dependence of a) 07 2144) and tg) eam be represented by matrices wlan) x1) = Me) + COG MED), ny + Cm, ea +E where the matriees and replace the fnetions and ¢, spectively. The mais 4 is called the sae ounsition maurix (SPM). The mais 2 is ealled the dserete-rime Input coupling muwrie, ov simply che input eonpling marl —IF the diserewe-time ccantest is alresdy established. 2.3.6 Solving Differential Equations for STs ‘A slate teasition matia % 2 solution of what is called lhe "homogeneous"? amaitix feqnation associated with a given linear dynamic system. Let us define first what homogeneous equations are, and then show how their solutions are related to the soletions of a given lines dynamic system, Homogeneous Systems. The equation si?) = Flt?) is called the homoge- neous part ofthe linear differential equation its) = P(E. + Cleat. The solution ‘ofthe homogeneous part can be obtained more easily than that ofthe full equation, and its solution is wed to define the solution o the general (nonhomogeneous) Envear ‘equation, “This tering somes fom the nalon thal exer lem A the exten Teele nine Ie ependeat variable. Thats the expression is famgumere Wik resp to dhe depend Vale25, COWPNUOUS LINER SYSTEES AND THE SOLUTIONS 35 Fundamental Solutions of Homogeneous Equations. An 1 x 4 matrix. valued function «(t) is called 4, encamental solution of the homogeneous equation 50) = Foha(2) om the interval £€ [9, TF O(e) = F(NO(N und OO) =F, the wm jdentity matix. Note that, for ainy possible initial vector x(0), the vector (6) = ((eIa(0) satislies the equation sn =4 wenao ea) d = [Sao 13) = [Furano an colori} 215) = Fuster 216) O(Fx(0) is the solution of the homozencous equation + = Fe with {nial value x(0) EXAMPLE 2.5 Tae unit upper triangular Toeplitz mavcix vede hs ore be woel@o rt 000 4 ooo Oo 1 is the fimdamental solution of & = Fx for the strictly upper ciangular Toeplitz dynamite coefficient matrix: 010 0 en) Papi : 0006) 000 0 which can be verified by showing that (0) = 7 and b= Mb, This dynamic coeflicieat mux, in un, is the companion mairix far the vh-order linear hhomagencous dileseaial equation (d/d"s() = 9.36 mean ane SYSTEMS ote a9 wo OF oo * om > tO 6 t ¥ Fig. 22. The STAs as a compositon of turlumental slutan manices, Existence and Nonsingularity of Fundamental Solutions. the elements of the matrix #(2) are contineus functions on goine iaterval @ 27 T, then the fundamen] solution matrix (7) is guaranteed 10 exist and to be norsingular on un interval 0 7 < + for some > 0. These conditions ulso guarantee that 4} will be ‘nonsingular on same interval of nonzero length, as eansequeree ef the continuous dependence of Lae solution @(*) of the satsix equation on iis (nonsingular) initial canditioas [*(0) = 1] [57]. Stato Transition Matrices. Note thot the fondeemental solution matrix 47) eunstorms any initial sate x0) of the dynanic system to the eoresponding stale 2) ot time fF OL) ig nonsingular, ther the produets "(2jx(?) =.) ard {0 0"! (43() = a2). That is the mate product O67, = OO“) ean ‘wunsforms a solution fiom time #% the corresponding solution at Gime 0, lnere 7 isan 17x 1 identity matrix, ”~" is the inverse Laplacian operator, md s isthe Laplace transform variable 3. The “scaling and squaring” method cambined with a Padé approximation is the reeornmiended genenil-purpese method. This method is diseussed in prealer detail in Section 2.6.0 mean ane SYSTEMS 4, Numerical integration of the homogeneous part of the different | equation, Fauna ro ean will initial value 40) = systeras.) (This method also works Jor time-varying ‘Taore are many other methods,* bur these are the most important. EXAMPLE 2.7: Solution of the Damped Harmonic Resonator Problem with Constant Driving Funetion” Consisky again the damped roxonator mode! of Examples 2.2, 2.3, and 26, The model ean be written in dhe form of « sevond- corer differential equation OCD + Dem, diO + MRL) = WD, where fk ‘Tae parameter £ ix a unitless damping eocflicient and wy the “natural” (i, undamped) frequency of the resonator ‘Tais sevonc-order linea diffevential equalion ean be zewrillen ia stalzspace form, with siates xj = 0 and 32 4 aad paramelets ¢ and 04, 88 aeo)-[se ae IST [3] ‘with inital conditions ‘As a numerical example, let al) so that the eoefficieat matrix r-[4 a] See, or exp, Hine [86], Desa ot, 58), or Kaci ad Saraki [18924 OSCHETE UNEAH SYSTEMS AND THEM SOLUTIONS a a Therefore, GIF) (Fy! sel 1 = =e or Fy! sl 1 vet Festi 2.3.9 Time-Varying Systems Af F(A) is nor constant, the dynamic system is called rime-varying. If F(@) is a piecewise sonooth function of , the m x 0 homogeneous ina differential equation 2.24 ean de yolved numerically by the fourth-onder Runge-Kutta smethoe.® 2.4 DISCRETE LINEAR SYSTEMS AND THEIR SOLUTIONS 2.4.1 Discretized Linear Systems Af one is only interested in the system stare at discrete times, thea one can use the formula HG) = OU ted + | Oe. ACO) 2.25) to propagate the state vector between the times of interest "Nur after the German matherainns Kil Dav Tolne Range (1886-1927) ard Wiel Mestn Kone (1867-194,a2 mean ane SYSTEMS Simplification tor Constant 1. If ig constant over the interval [f,_1. Als the above integral can be simplified to the forma ate) = Oe. yt) + ta det) (2.26) TH -f Pg. 0)Co)ao. 27 Shorthand Discrete-Time Notation. Fo: discvete-time systems, the indices k in the time sequence 4f,} characterize the times of interest. One ean save some ink by using the shorikand. notation: oy Bay DE i OH ! a(t). a Zuo. Hy OK Ty ETE HUG) for discrete-time systems, eliminating + envrely. Using this notation, one cna represent the discrete-time state equations in the more compact form Oy ET (2.28) Mx + Dou 2.29) 2.4.2 Time-Invariant Systems: For continuous time-invariant systems thet have boon discretized using fixed! tine intervals, the matrices @, I, Hf, and D are incependeat of the discrete-time index as well, In that case. the solution can be written in closed lonm as a Hy LO Ty, 230) where is the Ath power of . The matrix @ can also be computed as OF = % or—ay'y), 231) Where 2 is the 2-lranslirm variable and 2~! is the inverse z-transform, 2.5 OBSERVABILITY OF LINEAR DYNAMIC SYSTEM MODELS Observability is the issue of whether He stutwe of a dynamic system is uniquely sdotermiinable thom its inputs aed oucputs, given a model fur the dynamic system. Tis essentially a property of lhe given system model A given linear dynamic systern25. OBSERVABLITY OF LEA DYALSC SYSTEM MODELS 43 mode! with a given lines input/output model is considered observable if and only ity state is uniquely determinable from the model defnition, its inputs, sd its outputs. IF the system state is not uniquely determinable from the system inputs and outputs, ther the system model is considered unobservable How to Determine Whether a Given Dynamic System Model Is Obser- vable. If the measvrement sensitivity mawix is invertible at nay (continous ot ligerete) time, then the system state can be uniquely deterinined chy inverting i) as 2, Inthis case, the system model is considered’ to be completely observable at tht time, However, the system can still be ofservuble over a tine fmerval even 11's nor invertible at aay tne. Tn the later ease, the unique solution for thy systern staze can be defined by using the levst-aquares methods ef Chapker 1, including those of Sections 1.2.2 and 1.2.3. These use the so-called Grumian matrix to characterize whether or not a vector vaciable is determinable Eom a given linear model. Whea applied to the problem of the deletminacy of the stale of a linear dynamic system, the Gromian matrix is called the vbservabiliy matris of the given system model. “The observabiliry matrix for dynamic system models in eontimucus time has the form 64. F. t= | OAH MODI as 2.32) for a linear dynam system with fundamental solution mateix (¢) and measurement sensitivity matrix /7(7), defined over the cortinuousetime inerval 4, < # <4. Note that shis depends on the interval over which the inputs and outputs are abserved but fot on the inputs and outpers per se, la fact, the observabilry matrix of a dyaamic system model does nor depend on the inputs 1, the input coupling matrix C, or the inpur-ourput coupling matrix D—even though the outputs and the state vector depend on them, Because the funcamental solution marrix @ depends only on the dyamie coefficient mattis F the observability mauix depends only on Hand F. ‘Tae observability matin of a lincay dynamic system model over a discrete-time interval f= FS 4, has the general form Alt.) kk) = ( Ste. )en{iie.J} ea where His the observability matrix at tine é and o, is the state transition matrix from time i, to time ff 0 = & = &, Therefore, the observability of diserete-tine aystem models dapenels only an the values of J, and &, ever this inser. A in the feontinuus-time ease, observahility daes not depend on the system inputs “Tae derivations of these formulas ate leas exceeises for the zeader.4“ mean ane SYSTEMS 2.5.1 Observability of Time-Invariant Systems ‘Tao formulas defining observability are simpler when the dynamic coefficient matrices or state transition matrices of the dynamic system model sre time invariant. In that case, observability can be characterized by the zane of the macrices it wt o's" on) e234 for diseretestime systems anc Meat ra tur wy ry 2.35) (Or continuous-time systems. The systems are observable if these have rank, the dimension of the system slate veetor, The first of these motrives can be vbiained by representing the initia! stare of the linear dynam system as n fonetion of the system “nnputs ind outpurs, The initial scete ean then be shown to be uniquely determinate if and only if the rank condition is met, The derivation of the latter mnatex is not as strsightforward, Ogata [38] presents a derivation obtained by using properties of the characteristic polynomial of Practicality of the Formal Definition of Observability. Singularity of the ‘observability mantis is 2 concise mathematical characterization af observability. This eur be tov fine a distinction for practical application especially in dinite-prevision aritunstio—because abitraly small changes in the clemeats of a singular matrix ‘ean render it nonsingular, The following practical cousiderations shoud be kept in mind when applying the formal definition of observabi ity: «Tris important fo remember thatthe model is only sn approximation toa re system, ane we ar primly interests inthe properties oF te rel system, not the model. Differenees between the real system andthe noel are called meade ‘rancation errors, The an oF system modeling depends en knowing where 40 ‘cuneate, but there wil almost suwely be some iruncalion error in ary model Computation of ihe observabiliy mauris is subjeet to model truncation errors and sonndoff errors, which could make the dillzvence bevseza siagularity acd nonsingular of tie resule. Even if te computed obsecvability matex is efose vo being singular, i is cause for concern. One shonld consider a system as poorly observable if is observabilty matrix is close to being singular. For tat purpose, one can use the singular-rulue decompasition or the condition nusther of the observability marix ta define « more quantiecie measure of uunobservabiliry. The reciprocal af ts condition number measures how clase the system isto being unobservable Real systems tend to have some amount of unpredictability in their behavior, due to unliaawa or neglected exogenous inputs. Although Such offers cannot be modoled deturministically, they are not always nogl gible. Furtacrmore, the process of messucing the oulpuls wi physical sensors inutoduces some25. OBSERVABLITY OF LEA DYALSC SYSTEM MODELS a amount of sensor noise, which will eause errors in the estimated state, Ie would be better to havea quantitative characterization of observability that takes these ‘yes of uncertainties into aeeount. An approach to these issues (pursued i Chapter 4) uses a statistical characterization of observability, based on a satistcal model of the uncertainties in the measured system outputs cad the system éynamies. The degree off uncertsiniy in the estimated values of the system slates can be characterized by an dformation mmatrin. wiaich is a statistical generalization of die observability matrix, EXAMPLE 28 Consider the following continuous system: on] fe x)= [ pro-[ pe o of? AN =[1 Osi. ‘The observability matrix. vsing Equation 2.35, is ere, AF has renk equal io the dimension of x}, Therefore, the system is observable, EXAMPLE 2.9. Consider the following continuous system: 01 o an) [ Jol Jo 00 1 0 LK. ‘The observability matria, using Equation 2.35, is [}s] ew Here, Af haw ak less than the dimension af x(@). Therefore, the system is not observable,6 mean ane SYSTEMS EXAMPLE 2.10 Cor te following diserete system: coo 1 swafoo olnit] tia rio 0 foo My ‘The observability matrix, using Fquation 2.34, is o10 M=|0 1 0], amkera 100 ‘Tao rank is less than the dimension of x,. Therefore, the system is not observable. EXAMPLE 2.11 Consider the following diserete system: {kl} [vb ‘The observability matrix, using Equation 2.34 is, ‘Tae system is observable, 2.8.2 Controllability of Time-Invariant Linear Systems Controllability in Continuous Time. The concept of observability in estima- tion theory has algebraic celationships to the concept of conimllabiity in eontrol theory. These concepts and their relationships were discovered by R. K, Kalman as ‘what he ealled the duafty and separabiltse ofthe estimation ard cartwol problems for linear dynamic systems, Kulmans? dual concepts are presunted here and in the next subseetion, although they axe nov issues fe the estimation problem. "oe dul elfonsip broom sintoa und contre elven sere thes crisnally dined Kanan, ‘These ence he heen refine rete by ier investors clude concep mecha 1 snes an well Te nlrese ear i relered the re ree lente on “wer” onl theory fr fares exposing af toss er sc."25 ossenvaur or tneon ortuc SYBTEM MOORES ” A nae sytem defied othe five interval sy <5 4 by the lnesr model Hf) = HD + Cat, 340) = He) + Dut) (2.36) and with initial state vector s(f:) is said to be conmrollable at time 1 = y if, for aay esired final sate x{¢), there exists « piecewise continuous input finetion s) that drives t9 state x(j)). IF every initial state of the systein is controllable in gonte finite time interval, then the gestem is suid to be eomollable. ‘Tae system given in Equation 2.36 is controllable iF and only if anatix S has 3 linearly independent columns, sale FO Pe. PG, ean Controttability in Discrete Time. Consider the vime-invariont system model sven by the equations ready +My (238) I, + Duy. 239) slae xp €0 a given final stale xy, in N sarapling instants, where Nis fi integer. This condition can be shown to be equivalent to the matrix s=r or vr o'r 2.40) having rank m. EXAMPLE 2.12 Determine the controllability of Example 2.8. The controllabile ity max, using Equation 237, is [ts] Herc, Shas rank cqual to the dimension of x0). Therefore, de system is eonmllable. EXAMPLE 2.13 Determine the controllability of Example 2.10. The controll abilily max, using Equation 2.40. is hoe s=]1 00 020 ‘Tae system is not contollable. rank of this condition i ak ealed vastly, with eorolabily vse toy =O.4“ mean ane SYSTEMS 2.6 PROCEDURES FOR COMPUTING MATRIX EXPONENTIALS, Ina 1978 journal article titled “Nineteen dubious ways to compute the exponential ‘of a matrix" [205], Moler and Vian Loan reported their evaluations of methods for coniputiag, suai exponentials. Maay of the methods tested had serious short cantings, and 20 method was considered universally superior The one preseeted hngre was reconsneaded as being mote reliable thaa-most, It combines sevetal ideas. ‘due to Ward [233 including setting the algorithm parameters to mee: a prespecified error bound. It combines Padé approximation with a technique called “scaling and squaring” (© maintain approximation errors within prespecified bounds. 2.6.1 Padé Approximation of the Matrix Exponential Padé approximations. Tucse approximations of functions by rational functions ‘eatios of polynomials) date stom a 1892 publication [206] by H. Padé.” They have been used in deriving solutions of difevenial equations, inluding Rigcall equae tions” [69], They eun also be applied 19 melons oF matrices, including the sate exporentigl, nthe matrix ease, the power series is approximated as a “snatrix fraction” of the fom: "A", with the numentor mats (1°) and denominator matrix (99 represented as polynornisls with matrix arguments. The “one” oF the Padé epproximation is two dimensional It depends on tie orders of te polynomials in the numerator and denominator of ie rational funciion. The Taylor series is the special ease in which the order of te denominator polynomial of che Pade approximation is ero. Like the Taylor series approximation, the Pade approximation tends to work best for small values of is argument, For matt arguments, it wil be some mats norm of che argument that will be required to be smal Padé approximation of exponential function. The exponcatial function with argument 2 has the power series exparsion Promina pak DAT "the ose the nunerator and demi th was Hation ore ever ere Rte re ase in ncavsig te Rieat caution in Chap 425. PROCEDURES FOH COMPUTING MATPIX EXPONEN TIALS 49 are the nuineeator aud Uenominator polynonials, respectively, of the Padé approx imation of e* The key feature of the last equation is that there are no tenng of order fein ®._4 isthe Xn state transition matrix forthe system at time fy uy is the input vector to the system i time 4, and Ty 3s the encresponding input coupling sate Time-Varying and Time-tnvariant Dynamic Systems. If F and C (or @ and ©) do not depend upon 1 (or 4), then the continuous (er discrete) mode! is called sme invariant. Otherwise, the model is time-varving, Homogeneous Systems and Fundamental Solution Matrices. The equa- ion. a Gn = Feomtn is called the fiomoyencous port of the model equation, d Gt) = Peony + Cone). A solution (0) tothe corresponding Xn marx equevion ad Say = FD) con an interval starting at ime 7 = fy snd with initial condition tig) =1 (the identity mates) is called a fimdamenial solurion motric v0 the homogeneous equation on that interval. I hs the property that, if the elements of Fe) are bound, then 4) ceimnot become singular na finite interval. Ferermore, for amy inital value xi). a aAated is the solution to the covresporsling homogeneous equation, Fundamental Solution Matrices and State Transition Matrices. For a homogenous system, the sate transition matrix %4_ from time ¢, te time g, ean be expressed in fesns oF he fundamental solution (2) as = (HY Ny) foe times fy = oy = bs27 suimeasy 83 Transforming Gontinuous-Time Models to Discrete Time. The model for a dynamic system in continuous time ean be transformed into a inodel in diserete time using the above formu for the state taunsition anatrix und the following, formula for the equivalent dixerote-time inputs: ne =otifl cocina Linear System Output Models and Observability. Ax ouput of a éyaaraic: system is something we can measure diecily, such ss directions of he lines of sight to the planets (viewing condiins peemiting} or the temperate al thermocouple. A dynamic system model is said ta be observable from a given set of outputs if is feasible to determine the stare of the system froma those ourputs. the deperslence of a outpat= on the system stte «is linear, it can be expressed in the form 2=He, where H is called the measurement sensitiviey mutrix. It can be @ fimetion of continuous time [H(p)] or discrete time (#,). Observability can be characterized by the rank of an ohservability mamix associated with a given system model. The observability matrix is defined as [ PTET HAVA for continuous-time models, x(¢ ‘The system is observable if and only if its obyervability matrix nas full rusk (#) For some integer m2 0 or lime # > fy. (The tes for observabilly ean be simplifies for time-invariant systems.) Nole that the dewermination of abservabihity depends on the (continuous or diserete} iniervel over which the observsbilily mots is determined. {ler} tie) | for discrete-time models. Reliable Numerical Approximation of Matrix Exponential. The closcd- form solution of a system of Listorder differential equations with constant coeflicients can be expressed symbolically in terms of the exponential fumetion of 4a matrix, bur the problem of auuncrical approximation of the exponential function of a matris is notoriously ill-conditioned. PROBLEMS ow 2.1 What i a siale vector model for the linesr dynamic system 20), expressed in terms af y? (Assume the eampanion form of the dynamic coefficient matrix.)27 28 29 mean ane SYSTEMS What is the coimganion matix for the mtkonler differential equation, (afaty's10) = 0? What ae its dimensions? Whar isthe companion matrix of the above prolem when # = 1! For n = 2? What iste fimdamental solution mattix of Exercise 2.2 when = 1? Whea ‘What is the state tnstfon matrix of the above problem when w= 1? For Find the fundamental solution macix Oy) for the system eps] _p Oya] pt alvol-Ls alto} fi] and also the solution wf) for the initial conditions x)= 1 and x(0)=2. Find the total solution and state tasition snares for the systema adeel-[o SIR Gl with initial conditions (0) = 1 and x,(0) = 2 The reverse problea: from a discrete-time mudel 1 a continuens-tbme model For ihe cisereie-time dynamic system model oo[t ooh find che state (ransition matrix for continuous time and the solution for the continuous-time systein with initial conditions o-[] Senta a of]. and [o J] =U ea ()27 suimeasy 25 240 Detenning the contratla given below: ty and observability of the dynamic system model icol-L ollvedeLs “1L) _ s(t) wone off] Devive the stuns transition matrix of the time: a= [i the 2.12 Find the sate wansition matrix for ain) 2.13 For the system of three frstorser differential equations jing system (@) What is the companion inatrix 72 (bj What is the fundamental solution matix ¢() such that (f/doy(e) = FOU) and DW) = 12 2.14 Show that ike matrix exponential of mats, antisymmetric mattis is an orthogonal 2.18 Decive the formula of Equation 2.32. for the obsorvability matrix of @ Hineae dynamic system model in continuous time. (ine Use the approach of Example 1.2 fir estimating the inital state of a system and Equation 2.19 lor the sate of a systema as a linear fiction of its initia state and its taputs,) 2416 Derive the formula of Equation 2.33 for the observability matrix of a dynamic system in discrete time, (it: Use the method of least squares of Example 11 for estimating the fitiad sue oF a sysoomn, ane compure the resulting Gramian maeris to the observability matrix of Equation 2,33.)Kelman Birla, Theory ond Proece Usa ATLAS, Second ton, Mokinder Ces Angi Are upyneh 200) dm Wey & Sis, Ie ISDN: O471-39254-9 ULadbiey, WATT 206308 dLloceoa Random Processes and Stochastic Systems 4 cumpleely sutifactary definition af random seauence is yet to be discover G. James and R. C. James, Mathemaries Diedonars, 'D. Van Nostrand Ca, Princeton, New Jersey, 1959 3.1 CHAPTER FOCUS ‘The previous chapter presents methods for representing a class of dynamic wystenis with teltively small nuinbers of eomponents, such as w harmonie resenator with one sass and spring. The results are models for deterministic mechonties, in which the slate of every component af the system is represented and propagated explicitly. “Another appzoach lias heen developed for extiernely large dynamic sysiems, such as the exsemble of gas molecules in a reaction chacubet. The state-space apyeoach for such: large systems would be impractical. Consequently, this other approach ‘focuses on te easemible sfaristicut properties of the system and trea's the uadedlying dynamics as a random process. The resulrs are models for statistical mechanics, in hich only the ensemble statistical properties of the system sve represented and. propagated. explicitly. Tn this chapter, soine of the basie notions and mathesnetical models of statistial and doterministie mechanics axe combined in « stochastie system model, which represuns the sta af Inawledge about a dynamic system. These models wypresent what we dow about a dynamic system, including a quantitative model for our uncertainty about what we knovs In the next chapter, methods will be derived for medilying the state of knowl edge, based on observations related to the stare of the dynamic system. 5631 cHanen Focus o7 3.1.4. Discovery and Modeling of Random Processes Brownian Motion and Stochastic Differential Equations. The British osunist Robert Brown (1773-1888) xeported in 1827 a phenomenon he had observed while studying pollen grains of the herh Clarkiu pulchella suspended ia ‘water and similar observations by earlier investigators. The particles appeared to rave about erratically, as though propelled by some unknown force, This phenom enon came ty be called Bresimian movement or Brovmian marion. Tehas bon studied extensively—hoth cmpirically and thooretically—by mimy eminent scientists {including Albert Einstein [157] for the past certary. Empirical studies demon- siraled that no biological forces were involved amd eventually established that individual collisions with molecules of the surrounding uid were causing the motion observed. The empirical results quatitied how some statisties] properties of the random motion were indluenced by such playsical properties as the size and mass of the particles aud the temperature and viscosity of the suzrouading uid. Mathematical models with these statistical properties were derived in terms of shat has come to be called stochastic differential equations, P. Langevin. (1872— L346) modeled the velocity r of @ particle in terns of a differential equation of the form ae Ba bet ale, en where fis a damping coefiicient (due to the viscosity of the suspending medium) and (2) is called a “random force.” This is now called the Lamgevie equation. Idealized Stochastic Processes. the random forcing function a?) of the Langevin equation has been idealized in two ways fram the physically motivated example of Brownian motion: (1) the velocity changes imparted to the particle have been assumed to be statistically independent from ane collision to smother and (2) the effective time betwcen collisions hs been allowed to shrink to zero, sith the magnitude of the imparted velocity change shrinking avcordingly. This model ‘uanscends the ordinary (Riemann) calculus, because 2 “white-noise” provess is zo. integrable in the ordinary calculus, A special caleulus wes developed by Kiyosi ro (called the 16 cafculus or the stochastic ealetlus) to handle such factions, White-Noise Processes and Wiener Processes. A more precise mathema- tical characterization of white noise wis provided by Norbert Weiner, using his generalized lartnonic analysis, with a result that is dificult to square with intuition. Ir has a power spectral density that is uniform over an infiuite hardvideh, implying that the noise power is proportional to handvidt and thatthe total powee is infinite Alf “white Kght" had this property, would we be able so ses?) Wiener pastored ta facus on the mathematical peoperties of v(e}, which is now called x Wlencr process lis mathentatical ptoperties are more eniga than those of white-noise processes.58 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS 3.1.2. Main Points to Be Covered ‘Tao theory of random prooesses and stochasti systems repeesents the evolution over me of the unceraimty of our knowledge about physical systems, Tais representation includes the effects of say measurements (ar odservarions) that we make of the physical process ane the effeers oF uncertainties about the measurement processes and dynamie processes involved. The uncertainties in the measurement and dynamic processes are modclod by rimdam processes und stockastic systems Properties of uncertain dynamic systems an: characterized by statistical param ters such 1s means, correlations, and covariances, By using only these nomerical parameters, one cam obiain ¢ Finite representation of tae problem, which is important for implementing the solution on digital computers. This representation depends upon such statistical properties as orthogonality, stationarity. etgodivity, and Mario iamness of the random processes involved and the Gaussianity of probability distributions. Gaussian, Markov, and uncorrelated (white-noise) processes will be used extensively in the following chapters. The autocorrelation finitions and power spectral densities (PSDs) of such processes are also used, These are important in the development of Fequency-domain and time-domain anodels, The time-domain ‘models may be cithe? continuous or diserete. Shaping. filles (continuous and diserete) are developed for randon-eonstant, random-walk, and ramp, sinusoidally correlated andl exponentially correlined processes, We detive the linear covariance equations for continuous and discrete syslems to be used in Chapter 4. The orthogonality principle is developed and explained with seslar examples, This principle will be used in Chapter 4 to derive the Kalinan filer equations, 3.1.3. Topics Not Covered Tris assumned that the reader is already flay withthe imachematieal foundations of probability theory, as covered by Papoutis [39] or Billingsley [53], for example, The treatment of those concepts in this chapter is heuristic and very brief. The reader is relested to textbooks of this type for more detailed background material. ‘The 106 csleulus lor the inicgration of otherwise aonintepreble fumetions (white noise, in particular) is not defined, although it is used. The interested reader is refezted to books on the mithemarics of stochastic differenti equations («-¢., those by Amold [51], Baras aud Miselli [52], IG and McKean [64], Sobezyk [?7}, or Stratorovieh [78D 3.2. PROBABILITY AND RANDOM VARIABLES ‘The elationships between unkiowi physical processes, probability spaces, and. random variables are illustrated in Figure 3.1. The behavinw ofthe physioal processes is investigated by what is called 2 siaiuical experiment, which helps w define 1 ‘model far the physieal process as a probability space. Stielly speaking, dhs is not a60 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS A ramon variable assigns real mambers 10 outcomes, There is an integral number of dots on each face of the die, This defines x “dot funtion” ds = Yon the sample space %, where d[@) is the numnber of dots showing for the outewne © of the stitistial experiment. Assign the values ae) AEs) 2 de, ae) Hed=4 dC) ‘Tais fiero is om exaraple ofa srndome variable. The usefil slatistial properties of tins random. variable will depend upon the probability space defined by statistical experiments wih the die, Fronts and sigma olgebras. The statistical properties of the random varuble d depend on the probsbilites of set of outcomes (called events) forming what is called a sigma algebra’ of subsets of the sumple space 7, Any collection of events that inchuies the sample space itself, the empty ser (the set with no elements), and the set unions and set complements oF al its members is called a sigma algebra ence the sample space. The set of ail susets of SF ix a sir algebra with 2° — 64 events, The prohubility ypuce for a fair die. \ diz i considered “fair” i in a lagge number of 1osses, all outcomes wend to occur with equol lrequency. The relative Frequency of any outcome is defined as the ratio of the number of occurrences of Cat ‘outzome co the number of occurrences of all outcomes. Relative fiequencies of outcomes of a statistical experiment are called profailities. Note that, by this cfnition, the sum of the probabilities ofall outeomncs will aways be equal 10 L. This defines s probability p(6) for every event & (a sct of outcomes) equal 0 #) me wey whore ae) is the cantinalize of 4, equal to the numer of outcomes ( € 6. Note that this assigns probability zero to the empry set and prabability one to the sample space, The profability diseiusion of the random variable d iy a nondecreasing function Pa) defined for every real number x ay the probability ofthe event for whiel the score is less than x, Ithas the formal deLiaition ror’, 90, 0015 (a2). MalO , 1) = ptitiatey <1) = ll) (he empty sex) =6, PAL 0-01) = pldM(20, 10-+-01))) = alley = 10-01) = A@al PAGV-—-O1) = AH) as plotted in Figure 3.2. Nove that P, is nor a comeianons fimetion in this paricular example, 3.22 Probability Distributions and Densities Random vaviables are required to have the property that, for every veal a and f+ such that 00 , Here we have used a probatility- preserving form ofthe Fourier transform, defined withthe factor of 1/\/2e in front ‘of the insoyoal. HW ethoe forms of the Fouticr oanstems are used. che sult is not a probability distribution bur a sealed probabisy distrbution 3.2.3.1 Vector-Valued (Multivariate) Gaussian Distributions. The formmila for the a-cimensional Gaussian distribution 1°, P), where the mean ¥ is an a= vector and the covariance P is sn a > e symmetric positivedefinite matrix, is sake 86 wm orn eu6 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS ‘The multivariste Gaussian generating function has the form = ginitew, 35) VOnria ea) po) where v9 1 an nvector, This is also a mulivariate Gaussian probability distribution (0, P71) if the scaled form of the Fourier transform shown in Equation 3.11 is used, 3.2.4 Joint Probabilities and Conditional Probabilities ‘Tae joint probability of wo events 4, and fy $8 the probability of their ser intersecsion p(t, 164), which is the probability that bosk events accut. The joint probability of independen: evens is the praduct of their probabilities ‘The conditional probabitiey of event J, given that ever J, has occurred, is defined as the probability of & in the “conditioned” probability space with sample space &,. This is 8 probability space defined on the sigma slyshra AO, = (60 E18 € ol B16) of the set intersections of all events &€ of (the original sigma algebra) with the conditioning event 4, The probability measure on the “conditioned” sigma algela <6 , defined in terms ofthe joint probabilities in the original probability space by the mile BENE) ie, ar) ein where P(E 6.) de joint probably of end, Equation 3.17 is called Bayes’ rule. EXAMPLE 3.2: Experiment with Two Diee Consider a toss with ro dee 51 which one die bas come to rest before the other and just enough of its face x visible to show that it eomtains either four or five dots. The question is: What is the probability distribution of the seore, given that informacion? The probability space for nea dice. “This example Mustrates just hows eapidly the sizes of poobubility spaces grow with the “problem: size” (in this case, the aumber of dice). For a single die, the sample space has 6 outcomes and the sigma algelara has 64 events. For (wo dice, the ssmple space inas 36 possible ouicamtes (6 indepeedlert ‘ovtzonies for ench of ro dee) and 2°° = 6, 719, 476, 736 possible evears, IF each “Discovered by he nih clowymi as msibomstician Thon Bows (1702-1761), Coudonine oa pvesile even smo Stine Noe Te endins prbubity Ts bed te simon a Ina aera Ts ld arnt ol hal 6 een wh nse proba, nich oe CN expos foun ptesaypicatons oF Rapes” mi.39. STATSrigaL PROPERTIES OF aANGON YARIAELES or The nth cena! moment of « is defined as be) Ble — Bay? G19) [ito esr mea a 8.20) The fist moment of +i elle ity mean? n= Be= [Papaya 620 In gener, a fmetion of several arguments suck as fry) has fist moment Pena) = IJ fla, y. ater) ded de. 623) Array Dimensions of Moments. The first moment will be a scalar or a veer, ‘depending on whether the function x, » =) is sealar or vector valued. Higher ordet moments have tensorlike properties, which we ean characterize in terms of the rrumber of subscripts used in defining them as deta structures, Vectors are singly subscripted data structures. The higher order moments of vector-valued variates are successively higher onder data structures. That is, the sccond moments of vector vahted RVs are matrices (doubly subscripted data stuuctres), and the third-order ‘moments will be miply subscripied data structires These definitions of a moment apply to disexete-valved random variables if we simply substicure smmmnotions in place of integrations in the desinitions 3.3.2 Functions of Random Variables ‘A function of RV xis the operation of ussigning to each value of x another value, for ‘example y according te rule or function. This iv represented by =f. 6.23) where x and y are usually called inpet ane oulput, respectively. The statistical properties of y in terms of ave, fe example Fol) de, Ey= B24) ryt [Errore as when y’is scalar, For vector-valued linctioas x; similar expressions ean be shown, ke hore rset the oder of th mene 9 the postive inegen. The zeeth.arter momar woul tbeize ai clit 168 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS ‘The probability density of y ean be obtained ftou the density of x. IF Equation 3.23 can be solved for s, yiekfing the unique solution 0), 8.25) ‘Ten ve have » (v) — 2a p= e5) Te lens sshere pa ,(x ae he densi ston ofv and x respectively Afton oF two RWS x, 3 the povess of assigning to each pair of x.» another vl, lor fxampl, 2 aconding to ie same mle 2=fl.9. 627) and similarly functions of» RVs. When « any in Equation 3.23 are dimensional ‘vectors and if'a unique solusion for x in terms of » exists, r= 46) 6.28) Equation 3.26 becomes ale 629) ‘where the Jacobian | defined as the determinant ofthe aay of pantial derivatives dpa we ay my OB he th [yj seer] Oe a, (3.30) a ae an yy 3.4 STATISTICAL PROPERTIES OF RANDOM PROCESSES 9.4.1. Random Processes (RPs) A RV wus detined as a funetion x(y) defined for cach autoome of an experiment Idemtifed by s. Now ifove assign to eneh outcome w/a tine fimetion tes}. we abstain34 STATS rigaL AROPERTHES OF AANGOH PROCESSES 68 a family of finotions called random processes or stochastic processes. random process is called discrete if its argument is a discrete varinble set as sks}, KE 1,2 631) Ty clear that the value of a random process 2(2)at any particular time ¢ = Zp, namely {fps}, is random variable [or a random vector if Mp. 8) is veetor valued), 3.4.2 Mean, Correlation, and Covariance Let x(@) be an aevector random process, ls mea Extth= fr. xr phd) xi), B32) which can be expressed elements a8 f= Parodie a), 1 tam Fora random sequence, the integral is replaced by a sun “The correlation of the vetor- ‘ahd process 2) i defied by Eisley lay) EX iy Bligh Elete W76)) = : . (3.33) Ele Fd) Wel) whew Baar yies) = J [seomeapisco. al se yt). 634) “The covariance of xf7) is defined by Ext) ~ Bx ots) ~ Eee") eas = Els{t be!) — Bt DE la). When the process xif} has zev0 mean fie, Ex(e) = ceovavianee are equal. ‘The correlation matsix of G40 RPs.atA), an n-veetor, and yf), an z-yeetor is given Dy: an xm mats fr all 2), ts eorekation and Extn’ 6.30)7» rawext mocesses any srocmsre sors whew sadoniid =| fatnbtemtsenhnilasinraeey 63 Similar tbe cosecorarianse x me mati Ele) — ExtegMott) — Foe. 838) 3.4.3 Orthogonal Processes and White Noise ‘Two RPs x(/) and y() are called uncorrelated if their cross-covatiance matzix is identically zero forall f and F Pista) — Bieta bt) ~ ENT 8.39) ‘The processes 47) and s{ are called orrkogonal if their correlation matrix is identically zero: Eeeieyy" (ey) = 0. G40) ‘The random process 2(7 is called! uncottelaved i Eee) ~ Bit VA} — BG) = Oy. GH —}. BALD where dl) is the Ditne delta *fanction’* (actuslly. a generalized function), defined. by * np ft Pasese sa Pao [) ae ee Similarly, a random sequence xy is called aontelated if Els, — Edel; — B4G)]") = OUk../) Al —/), 43) where A() is the Kronecker delta fimetion’, defined by Atk) = | R=0 Baa) © |G. otherwise, A white noise process or sequence is an example of aa uncorrelated process oF sequence. ‘Nard Jor the Ph physi Poul Arion Maive Dis (102 19. Nc! so the Gnman mathomatean Lenpold Kronceker (INP 184.34 STATS rigaL AROPERTHES OF AANGOH PROCESSES 1 A process xf) is considered independent if for any choice of digtinot times fistasee fos the random variables f0,},.(}, ff) are iadependent, That is, Poa sPaa gts Treats 645) Independence {all of tie moments} implies no eoetelation (which restricts attention tw the second moments), but Ge opposite impBeation is aot wie, except in such special eases as Gaussien processes (sce Section 3.2.3). Note that whiteness means uncorrelated in time rather than independent in tine (.2., including all moments), although his distinction disappeass for the important case of white Gaussian processes (see Chapler 4} 3.4.4 Strict-Sense and Wide-Sense Stationarity ‘The random process sf) (or random sequence .,) is called sirct-sense stationary if all its statistics (meaning p[sir,), #()...,) are invariant with respect (o shits of the 40) whe HEHE ‘The sandom process x(0) (or 44) is ealled wide-sense stationary (WSS) (or “wea seuse” slaionary) if Fiat) {2 constant) B47 and Body ln) = Vit, — 4) = OO), B48) where Q is » mais with esch element depending only on dhe dillerence fy — 4, = «. Therefore, when x() is staionary im the weak seas ib implies that ils rst- and second-order statistics are independcat of time origin, while stict stationarity by 6.90) FOF +e. Bo because the variance E(w} ) “The complete model is then and the parameter G tj hav Tey a sith Eton) =O and Eto 9,3 = Athy hy ‘The dynamic process model detived in Example 3.6 is called a shaping filter, Block diagrams of this and other shaping lilecs are given in Table 3.2. along with their diffecence equation models. 8.5.8 Autoregressive Processes and Linear Predictive Modals A Pinca predictive model foe « signal is « representation ia the form Dpatias t he Gx where th is the prediction error. Successive simples of the signal are predicted as linear combinations oF the 7 previous values. ‘An autoregressive process has the same fomulaiion, except that fig 38 @ white ‘Gaussian noise process. Nove that sis Formula for an autoregressive process can be ‘evricten in state wansition matrix (STM) form as He 10 ooo ag o 4, oo o olf 2 lylol gon end. ooo 1 OSL a aye oO sgn = Dag ba 624 Where the “tate is the a-vector of the last 7 sumples of the signal and the ccovatianes matrix Q, of the associated process aaise a, will be filled with zoms, except be the termTABLE 32 Stochaste System Model of Discrete Random Sequences Process 1 Asaconoianon Block Diagram Sie Space Mesl Fandom al ae constant pfs pe Random +400 vale 4 Exponential heh ores _ssoNsn038 oh sa86I00N4 HOSA 0 STSCN NAISKS BUSHES Seoe RaNboU PROCESSES AKD STOCHASTIC SYSTEMS 3.6 SHAPING FILTERS AND STATE AUGMENTATION Shaping Filters. The focus of this seetion is nonwhite models for stationary processes, For mimy physical systems encountered in paustice, it may not be justified lo assume that all noises ate white Gaussian noise processes, It con be useful to generate an aulocorrelativa function or PSD from seal daia and then develop ea appropriate noise model using dilleremtial or difiereuse equations, These models are called shaping filers. They ae drivea by noise with a flat spectrum (white-noise processes), whieh they shape to represent the spectrum of the actual system. It was shown in the previous section that a linear cime-invariant system (siaping filter) driven by WSS white Gaussian noise provides such a model. The state vector en be “augmented” by appending to it the state vector components of the shaping dlter, withthe resulting mode! having the form of a linear dynamic system driven by white 3.6.1 Correlated Process Noise Models Shaping Filters for Process Noise. Let « system mesic! be given hy HO = FIOM GE. 21) = C+ 09) B95) where 1d) is nonwhite, for example, cormelated Gaussian noise, Aw given in the previous scetion, of) isa zens-anean white Gaussian noise. Suppose: tat w {2} ean be ‘modeled by w Bnear shaping fter!™: ltl = Popltheg (th + Gerhtdnat?) PV) = sO 6.96 Where SF denotes the shaping (iter and w3(¢) is ze mean white Gaussian noise, ‘Now deline a uew augmeuted stzte vector XG) = [RO ago”. B97) Equations 3.95 and 3.96 can be combined into the matrix form [] (" evn | [ ° | , = y|t fe. 3.98) Syl) 6 FA Siew” Leg, M8 = FON + Grohe, 8.99) "Seo eample ia Seton 3.7 foe WSS process3 SHAPING FILTERS ANC STATE AUGHENTATION 85 and the output equation ea be expressed in compatible format as a 2) = 1) af ‘ Joo @.100) wp) = HDX + ot. B10) “This total system givea by Eemations 3.99 and 3.101 is a inear differential equacion ‘mode! driven by white Gaussian noise. (See Figure 3.7 for a nonwhite-noise model) 3.6.2 Correlated Measurement Noise Models, Shaping Filters for Measurement Noise. ‘ similar éevslopment is feasible for the case of time-conlated measurement noise (0) M0 = Fanta + Gunnin, 2 = DKO-+ (0). 8.10) In this case, let ey(f) he zero-mean waite G noise (2) be modeled by fan noise amd let the measurement gpl = FoolOegelt) + Goole. 8.103) FO) = Mag 2 Fig. 3.7 Shapirg titer ode! or nonwhite neice.25 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS ‘The total augmented system is given by [wl at) (" 0 1) [“ a ye . G10) 0 KOlLsO] Le esol. até) x90) ‘This is in the form of a linear system model driven by white Gaussian noise and ‘ontput eqeation with no input noise ‘These systeins can be specialized to the WSS process for continuous and diserete ceases as by shaping filters shown in Tables 3.1 and 3.2 EXAMPLE 3.7. The “midpoint scecleration™ error for am acceleration sensor {accelerometer} is defined as the elléctive neveleration error al the midpoint of the sapling period. The associated error model for an accelerometer in terms of ‘unknowa parameters of the sensor is as follows: by OE + by + HaBhe + HR(RE — ANID + OB, where Syn = the midpoint acceleration error (he cross product (Ibe 3 — veetors) 43x 1 vector representing attitude alignment errois between “platform” axes and computational axes be = 031 secior of unknown accelerometer bisses, normalized to the magnitude of gravity Si 82 by OS ay oo &3 SHAPING FILTERS ANC STATE AUGHENTATION 7 and ‘9; = unknown accelerometer seale factor erroxs (7 = a, own accelerometer axes nononhogenalities a ber ervor terns, some of witich are obseevable; for reasou of practicality in our example they are not estimated, only eoinpensated with factory-calibeated values 3.x 1 unknown acceleration-squared nonlinearity for acceleration along the accelerometer inpu axis PX = a3 x 1 unknown acceleration-squared nonlinearity for acceleration normal fo the accelerometer input axis By = 0.3 X lvector (f), By, As)" of midpoint components of acceleration in platform coordinates B88 Pa=|0 0 oo & The 12 x 1 accelerometer stare vector x* is composed of the components by Ss ‘The 12 unknown parameters will be modeled as random walks (see Table 3.1) forthe ‘parameter identification problem to be discussed in Chapter 5 (Example 5.4). EXAMPLE 3.8 A gyrseope drift for model is given as follows: By Phin FU BAK f+ [lol], + hgt + Cyt whese 5, = 03 x | vector of unknown gyroscope fixed drill parameters f= a3 x3 mauris containing umknown scale factor (Sy) and linear axes alignment eros (Ay) 25 components (j= 1.2.3) Sa Aa Ma Ay Se An ay Ay Sp 3.x 1 vector of unknown nonlinear yyrosoope tarquer seals faetwr eras, with elements 3S,88 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS fen = 3x 3 diagonal marrix composed of absolute values of the components of (platforer snertial angular rate) on the corresponding diagonal clement 3 3 matrix of unknown wyroseope rise unbalance parameters (dl) fy doy sy fee dg dy doy dey des indices 4,0, and § denoting input. output. and spin axes, respectively. for cach gyroscope 1, 2 and 3. K, =43 x 6 matrix of umknown gyroscope compliance (g-squared) errs feta est Fro essa Asso Ky Renz Hise Bou Broz Foy esx ies Kos ban Fist 2x 1 vector eP unknown gyroscope fxed-dift trend parameters 2 6 matrix of unknown gyroscope mass unbalance trend parameters % 1 vector of vertical direction cosines (normalized gravity) (fy, fa. Bs)* x 1 vector with components (7. 83. 63. BBs. Pips. Pais)’ Bad Ox1 Oxt xt det et out]! v0 [% UL KL Ty vn ‘The 4¥ vakuown parameters will be modcled as mudom walks and randam rungs {S0e Table 3.1) for the parameter identification problem to be discussed in Chapter $ {Example 5.8) 3.7 COVARIANCE PROPAGATION EQUATIONS ‘The second moments of the state «(2 (a sundom process) und fonving function (2) {anouet random process) can be described in terms of covariance matrices. Let us elime the m x covariance matrix, PO} = Ella — Leatnylste) — Blo). 3.108) Te ve replace Bjx(e)) with the estimace of x(f) defined by &(F) in Chapter 4, then Pig will be called the error covariance mals27 COVARIANCE PROPAGATION ECUATIONS 39 3.7.1 Propagation in Continuous Time Ler 5 Fle + Girl. Eni) = 0. , 6.106) Eves (9) = Ole. 1 ls — 4) ‘The solution of this equation with 2(4,) as initial condition and fe, ) as state transition mnatrix has been discussed in Chapter 2, = Fy elta) + a4, t)G(tw{t dr. G07) ‘Take the expected value enact [ aveireoena et Then Q(t, Gtedn(7) de GB.109} Ln) — EbsCH] = Ee, ty )hetg) — Etaieh] + ‘The covariance matrix A) is given by Phe) = Ed) — EHC) — Bete] - x [> fake) — Botan + | Ot HK HEE ] >| flatig) — Elan) + [ow AGED «] ) = BE. f VE \Lr4g) — Efategh Neala) — EGKt@))] IO (ip) + 0th if Elst) — Botonduint IONE 9 ae + J oe cacne bute — Bean" 0". ae +f foe potatoe ratio) 6 OMe ddr90 AALDOU PROCESSES AND STOCHASTIC STSTENS The undertned quantities are re since 17) and sig) ae uncorrelated foe # > i PA) = BU AIP UDHE G9) +[ [oe , that minimizes the mew-squared error eS Els —a2P) 143) in meh that 2,2 i orthogonal to 2, Tht is Kea =, G46) and the value of the minimum mena-squaved error is given by dhe Sommula ou = Elle mb) 6.87)102 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS 18.9 Oniogonaty diagram. 3.8.3 A Geometric Interpretation of Orthogonality ‘Consider al sanclom veriables as veciors in abstract vector spaces. The inner product ‘of x and c is taken as the second mament (x2). Thus B(x) = Biss) Glas) is the square of the Length of x. The woctons x, 2, 22, and x— 22 are as shown in Figure 3.9. ‘The mean-squared estor E((s — 97) is the squave of the length of x —22- This Jengt is mininm if x ~%42 is ortiogonal {pecpendicular} to 2, EQ ay21e) =0. 8.149) We will apply the exthogonality principle to derive Kilinaa estimatons in Chapter 4. 3.9 SUMMARY 3.9.1. Important Points to Remember Probabilities are measures. That is, they are fuuetious whose arguments are sets of points, not individual points, The domain of a probability measure P is a sigma algebra of subsets ofa given set S, called the measunrble sets of S, The sigma algebra of measurable sets has an algebraic structure under the operations of ser union anxl set complement. The measurable sets alsvays ineluale the empty set { and the sot S, and the probability P(S)= 1, PUY) = 0, PUES B) + PUB) = P(t) + PCB) forall measurable sets 4 anal B. A prohabitity space is characterized by .asct $, sigma slgchra of ts measurable subsots, and a probability measure defined ‘on the meastible subsets Events Form a Sigma Algebra of Outcomes of an Experiment. 4 suarisrical experiment is an undevtaling, with an uncertain outeome. The set of all possible outcomes af an experiment is elled 2 sumple space. An event is said to ‘occur ifthe oaleome of an experiment is aa element of the eves.39 suimensy 103 Independent Events. A collection of events is called mumally independent ifthe occurrence or nongecurrence of ay finite number of them: has no influence ou the possibilities for oocurrenee or nonnecurtunee of fhe others. Random Variables Are Functions. 8 sealar random variable is » real-valued uustion defined on the sample spave ofa probability space such thal, for every open interval (2, 6). 96 f, giver its state at time f, is the sane as its probability distribution given its slate a all Bones © < 6 A Gaussion process if the probability distribution of its possible values at any time is a Gaussian clstibution, Startonaey if certain statistics of its probability distributions are invariant under shifts of the time origin, only its fist and second snoments are invariant, iis called sridensense stationary or weak-sonse stationary, WF all iy statistics are ‘nvr, se called strict sense stationary. Eroodic if the probability distribution of fs values st any ane time, over the ‘ensemble of sample fimetions, equals the probability distribution aver all ime of the values of randomly chosen member fimtotions, Orthogonal 1 anether rardom process if the expected value af their poinewise product is 220,108 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS 3.9.2 Important Equations to Remember ‘Tae density fiction of an m-vector-valued (or multivariate) Goussian probability disiribudion A°(&, PY bas the fonetional form 1 g uxt aw 9, Vorraetr po where & is the mean of the distribution and P is the covariance matrix oF deviations for the mean, ‘A linewr stochustie process in contimeons time with stale x and stale covariance P thas the madel equations 348) = PUAN) + Glo, 26) = THOM) +140, PU) = FNP) + PLETG) + GNOGT EN), where O14) is Uae covatiance of zero-mean plant nvise wid). A discrete-time firew stochastic process hins the model equations Dy we it Ge Me Fe +0 Pea My Py ME HG, whore x is the system stu = isthe system output, wis che zen-meun uncorrelated plant wolse, Q., is its cosmianes of v4, adv is the zere-mcan uncorrelated measurement noise. Plant noise is also ealled process noise. These models may also hove known inputs, Shaping fers are models of these types tat are used to represent rundom processes with certain types of spectal properties or tentporal correlations. PROBLEMS 341 Let dock of $2 cards he divided info four piles (abelod North, South, Bast West), Find the probability thar euch pile contains exactly ame ace. (Thene are four aces in all.) Show that (r)-()*(;Prostsks 105 3.3. How many ways are there to divide & deck of $2 car’ into four piles of 13 cach? 3) fa hal of 13 card are denn fom a deck of 52, what i the probability that exactly 3 cards are spaces? (There are 3 spades in all) 3.8. Iethe 52 canis ate divides into four piles of 13 each, aad i we are told that ‘North as exactly three spades, find the probability that South has exactly three spades, 3.6 A hand of 13 canis i dent fron a well-rindomized bridge dock. (The dock contains 13 spades, 13 hearts, 13 diamonds, and 13 elubs.) (@) Whot i the probability that the nnd conuans exacly 7 hens? {) During the deal, th face of ne of the ear inacrtntly exposed and it fs seem to be a hear, What is now the probability hot the hand contains exsully 7 heans? ‘You may Teave the above answers in ers of fcioiats 3.7 The random vatables same variance of. We de Xi, are independent with unean zor and the the new zandont variables ¥, Y,..0.Y, by Find the correlation cncficient r betwen Y,. and ¥, 3.8 The random variables W and Fave independent and uniformly disteibared Detween 0 xd | Geetangular distbution). Wind the probability density funtion of % = | — ¥| 3.9 Two random variables X and ¥ have the density function Co-st), Usy % elsewhere Pre) whete the constant C < 014s chosen. 10 nomialice the diseibution. (2) SKeteh the density faneron inthe x, pane (©) Determine the value of € fer normalization. (6) Obtain ro marginal density fumetions (@) Obsain ECF 9} (©) Discuss the mature al use of the relation » As. 3.10 The random variable x’ has the probability density function my, Ose], ie) 0, elsewhere.106 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS Kind tre following: {a) The cumulative function # (x) (b) The median. (©) The mode. (a) The mesa, 61%). (@) The meanesquave value £(47) (The variance o[X]) ant ude-modulated signal is specified by Hb = [1 Famxteif costar + a), Here s(/) is a wide sense srationary candom process independent of 2, which is a riadom variable uniformly distrTsuted over (0, 2a]. We are given that 1 Wi) = aT (a) Verify that (2) is an autocorcelation. () Lots) hive the autncorlation given shove, Tsing the direct method for computing the spectsl density, calculate , 3.2 Let ACP) be an acbitrary autoeorlation function fora mcan-sguare contin uous stochastic proces x) and let #40) be the power spoctal density fo the poses fa). Bit rue that lin Yto) vette Justify your unswer, Find the state-space models for longitudinal, vertical, and lateral turbuleace for the following PSD of the "Drysien” turbulence model: vo-«(5) (iam) where ‘e = frequency in radians per second @ = root-mean-square (RMS) turbulence inteasity L = scale length in feet ¥ = airplane velocity in feet per second (290 fi/see)Prostsks 107 35 a7 a) For longitudinal turbulence: (©) For lateral webulence: La ont a, = 0.15 mean crassewind (knots) Consider che random process (0) = oso? + 1,)e0s(eye + Ua), where 6, and 8 ane independent random. variables uniformly divribused between O and 2x {a) Show that x() is wide-sense stationary. (b) Calculate ¥,(0) and (0) 6) Discuss the ergodicity of 27), Ler yi) he the autocorrelation of a widessense setionary randem process. Is the real port of y(t) necessarily alsa an autocorrelation? If your answer is affirmative, prove it; fF negative, give a counterexample Assume xi) is wide-sense stationary: JO) = MP) costar +0), where is a constant and P is a uniformly disibuted [0,22] andor phase Find (0) The random process (0) has mean zero and sutocorelation function b=, Find the autocorrelation function for 0 [snd a108 39 3.20 an RaNboU PROCESSES AKD STOCHASTIC SYSTEMS Assume 5) is wide-sense stationary with power spectral density Lasse, Ylon= sus Z {i otherwise Sketch the spcetral ders ty of the process NO = sinh e0s{ + 0), Where 6 is a uniformly distributed random phase and O > « (a) Define 2 wide-sense stationary random process. (h) Define o striclesense stationary random process. {6) Define s realizable near system. {@) Is te following am wutocorrclition function? | Lop lish 0D ee Explain, Assuune xf ny dom process with autoeorcltion funetion I-|, -1ees1 YA Tg otherwise Find the spectral density ¥,(o) for HO =axt)eos(ene + ad hen ea is a constant ang 2 isa random variable uniformly distibuted on the interval [0,2]. A random process x() is defined by sts) = cose +8) where 0 is a random vasiable uniformly distsbuted on the ieterval (0, 2] Calculate the amtocoreelation function ¥, (9) for v= feo ae Let and gr, be two arbiaesy continuous, absolutely integrable autococrela- ion Tunctions. Are tse following necessarily autocazselalion functions?323 324 3.26 108 your answer (a) thy de (b) hy +0 () Wve (A) 6 py (the convolution of yy with Wh) Give 8 short reasoa for each answer: a) ICF(Z) ond G(P) are autocortelation functions, /°(é) + x(t) i (neeessae sly, pesbaps, never) an aulocostelation function (by As in (a), F716} — gtr i (necessarily, perhaps, never) an autocorrelation fionoton, () 0x40) is 0 strictly stationary provess, (0) + 2x(r — 1) is (necessarily, perhaps, never) stiely stationary: (@) The ficnction ole) [" ~fxer 0 otherwise, 5, is not) atv autocorrelation funetion fe) Lot x() be strictly stationary and ergodic ancl a be a Gaussian random variable with mean zen and variance one and 2 is independent oF x(). ‘Then 3{¢)= axle) is (necessarily, perhaps, never) ergodic. Which of the following fwetions is an autocorrelation function of a wide sense stationary process? Give a brief reason for each answer, @ @ o (b) Vea rey © ro={h ree wo Discuss each of the following: (a) The distinction between stationaety and wide-sense stasionarity. (b) The periodic chameter of the cross-cortelation fiction of two processes that are Ghernselves periodic with periods mand nZ, respectively. A system trauslee fianetion can sometimes be expecimentally determined by iajecting while noise n(i} and measuring the ctoss corvelation between the system output and the white noise. Here we consider the following system: Sav ” eet, -10 328 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS We assume (01) known, Sf) eid aff) independent, and (0) = 1, Walon). Hints Write ¥(2} = 68) + yal) where ys andy, ave the parts of the ‘ouput duc to S and m, respectively. Let S{) and af) be res stationary uncorrelated random processes, exch with “et bie 0 z | Hore, IFC Poro), Th f2xe0). ond HC 2x01) are transfer funetions oF time- invariant linear systemis and (2) is the output when aff is zero and mg is the ouput wiaen St) is zeso, Find Whe oulpul signaltonnoisezaio, defied as E(SHONEOR(O). A single random dt source 38 measuned by ns diferent ransdueers, ad their outputs axe suitably combined into a final measurcracat y). The is as pietaod below: OES}. Assuine that 4)(¢) and 1,(") are uncorrelated random processes, data and. noises are uncorrelated, filter T has transfer function 16)/s, apd filter 2 has transfer function I — 14s). Suppose that it is desired to determine the mesa square error of measurement, witere the error is defined by e(4) = 3409 — v4 Calculate the mean-squave value of the error in terms of FIs) and the spect dcasities ¥, ¥,,, and Y,. Let x{/) be the solution of Ftx ani, where n(d) is wisite noise with spectral density 2, (a) Assuming that she above system has been operating since ¢ = —o0, find Wilt) Investigare whether s(/) is wide-sense stationary, and if so, express i, accordingly. (bp Instead of the system in (a), consider _ fa. 1209, where x{0) = 0. Agate, compute yj.)Prostsks m (6) Lety)= [ited (ah and (6). (4) It is desired to predict x(¢+ 2) fom (0), that is, a ture value of the process Rom its prescat value. A possible predictor St +27) fs ofthe fore ind Yar. f2) fOr both of the systeins described in SG +9) = ain. Find that a that will give the sinllest mean-square pred st rninimizes jon error, that Ei +2) ale +P), where x() is as in part (a) 330 Let x{/} be the solution of Shy ant, wit initial earftion x(0} =p, Te i assumed that ne} is white noise with speeral density 2 and 3s turned on at 7 =0. Tae initial eandition 3y 58 a random varigble independent of nf} smd with zero mam. (@) Isr bas variance 07, what is (4, £)? Derive the resul. {(b) Find that value of @ (ell i 95) for whic, Y(t) is the same For sf 120. Determine whether. with ¢ =o. (6.6) 8 a fimetion only of ity (6) Ite white noise had been turned on al ¢ = —s¢ and he iia] eondlitian has vero mean and variance 2s above, is x?) wige-seuse saionary? Justify your answer by appropiate ressoning andor computation, 331 Let where x, is zero-mean random veriable witk covariance matrix P, and Epwigh =O ve, Exo( wth) = Qt) 9) View Lista") = 0 {a) Determine the mean mi) and covariance Pe, #) for the process x(F) {b) Detive 2 differential equation for M0.ne 333 337 338 RaNboU PROCESSES AKD STOCHASTIC SYSTEMS ind the covariance inatrix PQ) and ity steady-state value Poo) for the following continuous systems; Lo ot [2 Shee[ Epo [s Lio ro-[, °] where we 10,1) and white Find the covariance matrix P, and its steady-state value 2, forthe following discrete system: where wy €4'(0. 1) and white Find tie steady-state covariance for the state-space model given iy 34, Show that the continuous-time steady-state algebraic equstion #P(0) + Poo)" + GGT has no nonnegative solution for the scalar case with Equation 3.110.) Show that the diserete-time steudy-state algebraic equation Q=G=1. Gee Py = OPO" +O has no solution for the scalar ease with = Q= 1, (See Equation 3.112.) Find the covariance of xy a8 a fimetion of & and its steady-state value forthe system, tee 2a bh where Big ,=0 and Elm) =e"! Assume the initial value of the covariance (P,) is Find the covariance of x) asa function of f and its steady sta value for the system i= Dein) +08) sshere Bae) = Band Eur ote)) = 6 Assume de intl value ofthe sovavianee (Py) 8 LProstsks 13 3.39 Suppose that x has autocorrelation function yg(x) =e, Tes desired t0 pevdict al¢+2) on the basis of the past and present of s(0, that is, the pudictor may use xfs) for all ¢-< a) Show that the minimum mean-square ertor linear prediction is see rtd {b) Find the mean-square error corresponding to the above, Hue: Use the orthogonality principle.Kelman Birla, Theory ond Proece Usa ATLAS, Second ton, Mokinder Ces Angi Are upyneh 200) dm Wey & Sis, Ie ISDN: O471-39254-9 ULadbiey, WATT 206308 dLloceoa Linear Optimal Filters and Predictors Prudicton & diicnds—expectally ofthe nur Anibured to Niels Henrik David Bohr (1885-1962) 4.1 CHAPTER FOCUS 4.1.1 Estimation Problem ‘ais is the problem of estimating the state of a linear stochastic systema by imeasuvements thar are linear factions of the state We suppose that stochastic systems can be represented by the types of plant and measurement models (for continuous and discrete tine) shown as Kquations 4.14.5 iTable 4.1, with dimensions of te vector and matrix quantities as shown in Table 4.2. The syinbols A(k— ¢) and dU — s} stand for the Kronecker delta funevion and. the Dirue detta function (actually, 4 generalized function), respetively. ‘TABLE 41_Linear Plant and Measurement Models ‘oa Caninuous Tine Dseee Tee Equation Numbar Plant MFM Messuement 20) = Hitt) + 4c) 2 Plane noise Eivatii = 0 43) Evialeay = Mt — 907 aay Observation nolse Evin Eqvitwl sy) 41 at 45h 11841 cramer Focus 5 ‘TABLE 42. Dimensions of Vectors and Matrices in Linear Model ‘Syme0 Bimensiors Syma! Dimensions x et 7.0 axe ay ext 8 feo a ei Ao scalar ‘The measurement and plant noise r, and wy sare assumed to be 7ero-mean Gamssian. process, and the initial value 3p is Gaussian variate with known meen xq and known covuriance mattis Py. Although the noise sequences wy ane vy are sumed tn he uncorrelated, the derivation in Scetion 4.5 will remove this restrietion 1d moslify the estimator equations accordingly, “Tae objective will be lo find am. estimate of the r stave veclor x, represented by Fy, a linenr funetion of the mensurements 2, 2 that minimizes the weighted meu squared error Ex, — I" MIx, — els 4.6) shore AY iy any sammerric nonnegativesteinite weighing matrix, 4.1.2. Main Points to Be Covered Linear Quadratic Gaussian Estimation Problem. We are now prepared to rive the mathematical forms of oplinal linear estimators for the s(aies of linear stochastic systems defined in the ptevious chapters, This is called the fixear quadratic Gaussian (LQG) estimation problem, The dynamic systems are linear, the pesfonmance cost functions are quadratic, aad the random processes. are Gaussian. Filtering, Prediction, and Smoothing. Thete ate tee general types of estimators for the LQG problem: «Predictors use observations siviily pelor tw the time that the stare of the dynamic systein is to be estimated ae < bow «© Filters use observations up to und inchuling the time that de state of the dynamic system is to be estimated ave S be16 Lmiens GpriaaL RLTERS A0IO PREDICTORS # Smovthers use observations beyond the time that the state of the dynamic syste is to be estimated Fee > fw Orthogonality Principle. & staightforward and simple approach using the ‘onhogorality prineiple i used in the derivation! of estimators. These estimators will bave minim variance and be unbiused sad consistent. Unbiased Estimators. The Kshran filer can be characterized as an algorithm ‘or computing, the conditionel mean snd eovasiance of the probability distibution of the state of & linear stochastic system with uncorrelated Ganssian process and mensurement noise. The conditional mean is the unique smbiased estimate, Te is propaggoied in Reechack form by a sysiem of linear dierent equations or by the corresponding discrete-time equations, The conditional covariance is propagated by nonlinear dillerential equation or is discrete-time equivalent, This implementation aucomatically minimizes the exnected risk associated with any quadratic toss function of the estimation exvor, Performance Properties of Optimal Estimators. The statistical performance cf the estimator can be predicted a priori (Hat is, before it is actually used) by solving the nonlinear differential (or difference) equations used in computing the ‘optimal feedback gsins of the estimator. These are ealled Riccati equasions,* ane the behavior oftheir solutions ean be shown snalytically in the mast rrivial cases. These ‘equations also provide & msans for verifying the proper perfomance of the seta estimator when it is running, 4.2 KALMAN FILTER Observational Update Problem for System State Estimator. Suppose that a measuvement has been made at time 4, aud thatthe information it provides isto be “Faeroe reralisly oriented deriestons, sunslt my afte refines nach as Andes and More UJ. 8vake [9h une and Siting [9 Bras LT 1 Been mud Ho [Lh Boe’ a Josh LSI atin Hoh, Chand Chen [18 Gel otal. 21. Jarwaski 23), Kall [2, Macc [8 31. Monde [3 45) ok (36. ayn on Soong (42), nd Sorenson (47) 2Nared 1763 by Jom e Ror ' Alene (1717 (ORS) Rr Coat Ia Francesco Rise (1675 1754), wh ad i a sted scala illest (213), 2M tte on tha wo aver [5,210 Kalan gives ert wo RichirdS, aes frshoriag i tat Ue Rissa sail ‘sions analoyons w social fettizaton for dining odin yin. a Rise oguaton also ales Setirliy in the problem oF sqnmnon of aublsy omtnary eirenul exumone ond 9 ARE trian oC unin bounary-vile pls ia-alu problems [SS42 Raussen ruven "7 applied in updating the estimate of the state » of a stochastic system af time f, Its assumed that the measurement is Finearly related to the state by an equation of the form 2, = Thy, +4, whore IT is the measurement sensitivity matric and eis the Estimator in Linear Form. The optimal lincar estimate is equivalent wo the ‘general (nonlinear) optimal eslimalor ithe varistes x and = ae jointly Gaussian (see Section 38.1). Therefore, it sulices to seek an updated estimaue {(+) based on the observation 2;—thar is a fincer function of the a priori estimate and she measurement &: GH) = KE) + Rize an where ,(—) is the a prion estimate of x, and £,(-+) i the a posterior’ value of the estimate Optimization Problem. Tac mastioes K} and K, me as yet unknowa, We seck those values of KY aad Ky such that che new estimve §)(-+) will satisfy the corthogonslity principle of Section 3.8.2. This erthogoualty condition cau be waiten in the form Elly, — 34H) = 0, (48) Fels, — COED = 0 (4.9) Ione substitutes the formula foe, from Equation 4, (ia Table 4.1) and for 244) froin Equation 4.7 into Equation 3.8, them one will observe fom Equations 1 and 42 thatthe data =, ...,% do not invelve the nove toon ws. Therefore, because the rrindorm sequences Wy, andr, are unearrchited, it follows that Eng? = 0 for Lf i ofcily elle 3 debut fen clea “squigghe”42 Raussen ruven ng Substituting for KE, 2, and R(—) and using the faet shat (ef result can be modified as follows: 0 this last 0 BUD, HIF Reh) — Bul, — ghd) Hhesy = 0?) = Eile eC = EWU () + Kyl —) — Kyle “dl AW — Kathe — Ctalcte—) — 8] al). By definition, the a priori covarimce (the crror covariance matrix before the updated is 2A) = BIH. 1 satislies the equation UK PM — RR, = 0. and therefore he gain ean be expressed at Ry = PMU PATE + RT (4.19) ‘whieh is the solution we seek forthe gain a a funetion of thea priori covarianes One ean derive simi formula for the @ posterion covariance (tke eter covariance muttix aller update), which is defined as PCE (EHEC (4.20) By substituting Equation 4.13 into Equation 4.7, one obtains the equations R= u- FAH) = 4 Riker — OD e2y) Hil) + Ezy Sullrgct fiom both sides of te later equation to obtain the equations rl) + Bytes + Rey — Ruth) — a0, -) Ei + Katy. = Ruth yat-) + Rate 623) By suipstituting Equation 4,22 itp Equation 4,20 and noting that (yy obtains Gone Pye) = EW — RMS (ME — RAT! + Reet Re =U RHP — KA) + RRR, (4.23)120 Lmiens GpriaaL RLTERS A0IO PREDICTORS ‘This lust equation isthe so-called “Joseph form” of the covariance upelate equation derived by P.D. Joseph [15]. By substituting for K, from Equation 4,19, itean be put into the following form Pal) = Pe) —Rathat) — PWR, + ET, P OER] + RRR, Ut -Rtts Pa) — + Rh PAE + RRL PE =U Rh Ph. (4.24) the last of shich is the oue most offen used! in computation. This implements the affect that constiioning on the meosirement as on he covariance matix of estimation uncertainty. Evvor covariance extrapotation moxiels the effects of time on the covariance matrix of estimation wieerinty. which is reflected in the a priori values of the scovarianee and state estimates, PO) i 4.25) respectively, Subtract 5, from botk sides oF the last equation to obtain the equanions A) as BL) 2) 4 BAH Wy {or the propagatioa of the estimation error. Postmuliply i by 24(—) on both sides of die equation) and (ake the expected velucs. Use ibe lact al ZX, yx! = 010 obisin the results Pf EKO) a, AMR HEE OT, + El, et HO, OE +O (4.26) which gives the a prion value of the eovarlanee mati of estimation uneertaimy as faxsiion of ike previous a posterior value.42 Raussen ruven 124 4.2.1 Summary of Equations for the Discrete-Time Kalman Estimator ‘The equations derive in the previocs section are sutamarized in Table 4.3. la this formulation of te fier esiations, Chas been combined with the plant covariance by imuliplying G_, and C7, for example, Q1 ig El IG] jn Oe-sGh “The relation of the fille to the system is illusttaied in ihe block diagram of Figure 4.1. The basic steps of the computational procedure for the discrete-time Kalmaa estimatar are as follows: Compute P,(—) using Py y+), ®,.. and Qe) Compute K, using Po(—) (computed in step 1), ,, and Ry. Compute F4(+) using Ry (compute in step 2) ene Py(—) (fom step 1). Compute successive values of % (4+) recursively wing the computed valves of , (from step 3, the given inital estimate Ry and the Fur daa =, TABLE 43. Discroto-Time Kalman Filtor Equations ‘System dynamic model: Meets Hs NOG) Measurement mod: ses Heaney ANNOY Ina coma tons: Eval = fe Esky) — Po Independence assumgtion Eom —@ fr alixana) Sate etimate eciaplaon (Equation # 28) a) Fro: covaiance exrapolabon (Equation 4.26) P= PML + Oe Stale estate cbsenationl uate (Equation &21) hat) Er covariance update (Equation #24) Ps) — Uh Fatal) Kalman gain max (Eavation 4.185 R BO) Riles HAI) PHT PRAT + I122 Lmiens GpriaaL RLTERS A0IO PREDICTORS Discrete gystem Newsurerent Digest Kalran fer Fig. 4.1. Stock chagram of system, measurement moda. arcd discrete-time Kalman itr Step 4 of the Kalian filer implementation [computation of %(4)] can be implemented only for state vector propagation where simulator or real data sets are available, An example of this is given in Seetion 4.12. In the design trade-olf tne covariances matrix update (steps 1 snd 3) should be checked lar syrametry and positive definiveress. Failure to attsin eilaer vanditioa is a sign thai something is wrong either a program “bug” or an illeconditioned problem. In order £9 avercome ill-conditioning, another equivalent expression for PCH is called the “Joseph form,”* as shown in Equation 4.23) PUG) = — RPA — Rtg + RRR Note thatthe right-hand side of this equation is the summation of two symmetric matrices, The fist of these is positive defini nd the second is aonnegtive definite, thereby making P+) a postive definite marx. ‘There sre many other forns® for K; and P,(-+) that might not be as usefial for robust computation. It can be sow that ate vector update, Kalman gi, and eror covariance equitions represent an asymptotically stable system, and tore, the cxtimate of sate & becomes independent ofthe inital ostimate 4. Py as F is inercased Figure 4.2 shows a typical Lime sequence of values assumed by the ih component of the esiimoted state vector iplolied with solid cies) and ss corresponding variance of estimion ameerniny éoted vith open ereles). The amows show the successive values assumed by the variables, with the annotation (in parentheses) on the nuows ancicating which input variables define te indicated trasiions. Note at act variable sesuries two gistnct values at cach giseree Hits prior value “ofr ues and Jeo [3 Some af the miemene Farms fer soma sad Soca [461 yond 7-4 sim be fo in Fri (23, Kuh 2)42 Raussen ruven 123 sain 4 ‘State _ {) sunt tek) wt Beh, i“ OR Feel Go) i INS (ea Qe_sb covatanes mc Hp dD tei) Pid LZ Pri lOhy O77 re near? Discrstetme f+! Fig. 42 Recresertatve sequence of values of itr variables mn discrete tine, corresponding tw the value defore the information in the measurement is used, and the @ posterior value coresponding to the value a/fvr lhe information is used. EXAMPLE 4.1 Let the system dynamics and observations be given by the following equations: Exo) = Bw, = 0, Etta.) = 20h — Hh), Elona.) = Alb ~ &), arb he Ps uO) Eat) — So(0} — The objective is wo lind fs and ihe steady-state covariance mals Pc. One can use tne equations in Table 4.3 with ®124 Lmiens GpriaaL RLTERS A0IO PREDICTORS for which PAS Py H+ PA) _ Path Paai+2 Ps3 Peso t Pet) =| P| 41), r= [RS] OP 2P HFN) Py RCS HR, Lat PAA) = PLE) =P (steady-state covarisnes), e+ P+3 0.

Você também pode gostar

• Shoe Dog: A Memoir by the Creator of Nike

  

  No Everand

  Shoe Dog: A Memoir by the Creator of Nike

  Phil Knight

  Nota: 4.5 de 5 estrelas

  4.5/5 (537)
• Grit: The Power of Passion and Perseverance

  

  No Everand

  Grit: The Power of Passion and Perseverance

  Angela Duckworth

  Nota: 4 de 5 estrelas

  4/5 (587)
• Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  

  No Everand

  Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  Margot Lee Shetterly

  Nota: 4 de 5 estrelas

  4/5 (890)
• • Revistas
  • Podcasts
  • Partituras
• The Yellow House: A Memoir (2019 National Book Award Winner)

  

  No Everand

  The Yellow House: A Memoir (2019 National Book Award Winner)

  Sarah M. Broom

  Nota: 4 de 5 estrelas

  4/5 (98)
• The Little Book of Hygge: Danish Secrets to Happy Living

  

  No Everand

  The Little Book of Hygge: Danish Secrets to Happy Living

  Meik Wiking

  Nota: 3.5 de 5 estrelas

  3.5/5 (399)
• On Fire: The (Burning) Case for a Green New Deal

  

  No Everand

  On Fire: The (Burning) Case for a Green New Deal

  Naomi Klein

  Nota: 4 de 5 estrelas

  4/5 (73)
• The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  

  No Everand

  The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  Mark Manson

  Nota: 4 de 5 estrelas

  4/5 (5794)
• Never Split the Difference: Negotiating As If Your Life Depended On It

  

  No Everand

  Never Split the Difference: Negotiating As If Your Life Depended On It

  Chris Voss

  Nota: 4.5 de 5 estrelas

  4.5/5 (838)
• Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  

  No Everand

  Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  Ashlee Vance

  Nota: 4.5 de 5 estrelas

  4.5/5 (474)
• Yes Please

  

  No Everand

  Yes Please

  Amy Poehler

  Nota: 4 de 5 estrelas

  4/5 (1891)
• A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  

  No Everand

  A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  Dave Eggers

  Nota: 3.5 de 5 estrelas

  3.5/5 (231)
• Principles: Life and Work

  

  No Everand

  Principles: Life and Work

  Ray Dalio

  Nota: 4 de 5 estrelas

  4/5 (599)
• The Emperor of All Maladies: A Biography of Cancer

  

  No Everand

  The Emperor of All Maladies: A Biography of Cancer

  Siddhartha Mukherjee

  Nota: 4.5 de 5 estrelas

  4.5/5 (271)
• The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  

  No Everand

  The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  Brene Brown

  Nota: 4 de 5 estrelas

  4/5 (1090)
• The World Is Flat 3.0: A Brief History of the Twenty-first Century

  

  No Everand

  The World Is Flat 3.0: A Brief History of the Twenty-first Century

  Thomas L. Friedman

  Nota: 3.5 de 5 estrelas

  3.5/5 (2219)
• Team of Rivals: The Political Genius of Abraham Lincoln

  

  No Everand

  Team of Rivals: The Political Genius of Abraham Lincoln

  Doris Kearns Goodwin

  Nota: 4.5 de 5 estrelas

  4.5/5 (234)
• The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  

  No Everand

  The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  Ben Horowitz

  Nota: 4.5 de 5 estrelas

  4.5/5 (344)
• Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  

  No Everand

  Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  Gilbert King

  Nota: 4.5 de 5 estrelas

  4.5/5 (265)
• Fear: Trump in the White House

  

  No Everand

  Fear: Trump in the White House

  Bob Woodward

  Nota: 3.5 de 5 estrelas

  3.5/5 (738)
• Angela's Ashes: A Memoir

  

  No Everand

  Angela's Ashes: A Memoir

  Frank McCourt

  Nota: 4.5 de 5 estrelas

  4.5/5 (440)
• Rise of ISIS: A Threat We Can't Ignore

  

  No Everand

  Rise of ISIS: A Threat We Can't Ignore

  Jay Sekulow

  Nota: 3.5 de 5 estrelas

  3.5/5 (137)
• Steve Jobs

  

  No Everand

  Steve Jobs

  Walter Isaacson

  Nota: 4.5 de 5 estrelas

  4.5/5 (806)
• John Adams

  

  No Everand

  John Adams

  David McCullough

  Nota: 4.5 de 5 estrelas

  4.5/5 (2409)
• The Unwinding: An Inner History of the New America

  

  No Everand

  The Unwinding: An Inner History of the New America

  George Packer

  Nota: 4 de 5 estrelas

  4/5 (45)
• Bad Feminist: Essays

  

  No Everand

  Bad Feminist: Essays

  Roxane Gay

  Nota: 4 de 5 estrelas

  4/5 (1015)
• The Glass Castle: A Memoir

  

  No Everand

  The Glass Castle: A Memoir

  Jeannette Walls

  Nota: 4.5 de 5 estrelas

  4.5/5 (1711)
• Wolf Hall: A Novel

  

  No Everand

  Wolf Hall: A Novel

  Hilary Mantel

  Nota: 4 de 5 estrelas

  4/5 (3811)
• The Outsider: A Novel

  

  No Everand

  The Outsider: A Novel

  Stephen King

  Nota: 4 de 5 estrelas

  4/5 (1839)
• The Perks of Being a Wallflower

  

  No Everand

  The Perks of Being a Wallflower

  Stephen Chbosky

  Nota: 4.5 de 5 estrelas

  4.5/5 (2099)
• The Woman in Cabin 10

  

  No Everand

  The Woman in Cabin 10

  Ruth Ware

  Nota: 3.5 de 5 estrelas

  3.5/5 (2322)
• The Light Between Oceans: A Novel

  

  No Everand

  The Light Between Oceans: A Novel

  M.L. Stedman

  Nota: 4.5 de 5 estrelas

  4.5/5 (789)
• The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  

  No Everand

  The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  Viet Thanh Nguyen

  Nota: 4.5 de 5 estrelas

  4.5/5 (119)
• Little Women

  

  No Everand

  Little Women

  Louisa May Alcott

  Nota: 4 de 5 estrelas

  4/5 (104)
• Brooklyn: A Novel

  

  No Everand

  Brooklyn: A Novel

  Colm Toibin

  Nota: 3.5 de 5 estrelas

  3.5/5 (1937)
• A Man Called Ove: A Novel

  

  No Everand

  A Man Called Ove: A Novel

  Fredrik Backman

  Nota: 4.5 de 5 estrelas

  4.5/5 (4609)
• The Art of Racing in the Rain: A Novel

  

  No Everand

  The Art of Racing in the Rain: A Novel

  Garth Stein

  Nota: 4 de 5 estrelas

  4/5 (4200)
• Manhattan Beach: A Novel

  

  No Everand

  Manhattan Beach: A Novel

  Jennifer Egan

  Nota: 3.5 de 5 estrelas

  3.5/5 (792)
• A Tree Grows in Brooklyn

  

  No Everand

  A Tree Grows in Brooklyn

  Betty Smith

  Nota: 4.5 de 5 estrelas

  4.5/5 (1929)
• Sing, Unburied, Sing: A Novel

  

  No Everand

  Sing, Unburied, Sing: A Novel

  Jesmyn Ward

  Nota: 4 de 5 estrelas

  4/5 (1103)
• The Constant Gardener: A Novel

  

  No Everand

  The Constant Gardener: A Novel

  John le Carre

  Nota: 3.5 de 5 estrelas

  3.5/5 (104)
• Her Body and Other Parties: Stories

  

  No Everand

  Her Body and Other Parties: Stories

  Carmen Maria Machado

  Nota: 4 de 5 estrelas

  4/5 (821)