OF NUMERICA’, ANALYSIS B Ox0)s4 (+) x (-) = 1. Faiz Ahmad A. Sen d Afzal Ranil 1.2 1.3 15 16 17 18 £2. 1.10 Lu 1.12 CONTENTS CHAPTER 1 ERRORS IN COMPUTATION Introduction Errors and their. sources Approximate numbers and significant figures Rounding of numbers and around off errors Truncation error Absolute, relative, and percentage errors The error of a sum The error of a difference The error of a product The error of a quotient The error of a power The error of a root CHAPTER 2 LINEAR OPERATORS ‘atol fe operators MATRICES RBS“my Tntroduction Tagrange’s formula The error of the in Divided differences Divided-difference table Newton's interpolation polynomial CHAPTER 5 INTERPOLATION WITH EQUALLY SPACED DATA The difference table 32 Newton's forward and backward difference formulae 5.3 The Lozenge diagram 5.4 Gauss formulae 55 Stirling’s interpolation formula terpolation polynomial 5.6 _Bessel’s interpolation formula 5.7 Everett's interpolation formula CHAPTER 6 THE SOLUTION OF NONLINEAR EQUATIONS 6.1 Introduction 6.2 Bisection method 6.3 The method of false position or regula falsi method 6.4 Secant method Newton-Raphson method Fixed point iteration Order of convergence erence equations ifference equations 8 Se 9 DM Ow ypweveyy~oComey, cONTEN CHAPTER 4 TION WITH UNEQUALLY SPACED Data ‘ange’s formula ‘The error of the interpol vided differences ivided-dif ference table Newton's interpolation polynomial CHAPTER 5 INTERPOLATION WITH EQUALLY SPACED DATA Jation polynomial waanne BLoONe ioe te difference table vion’s forward and backward difference formulae ne Lozenge diagram : formulae ‘ 's interpolation formula a 's interpolation formula a *s interpolation formula x ds CHAPTER 6 SOLUTION OF NONLINEAR EQUATIONS wows. Sepia rence equations ference equations‘CONTENTS CHAPTER 8 SOLUTION OF SYSTEMS OF LINEAR EQUATIONS 50 50 (DIRECT METHODS) = : 8.1 Introduction 15 56 8.2 Gauss’s elimination method 7 7. 8.3. LU decomposition 122 8.3.1 Dolittle’s method 123 8.3.2 Crout’s method 126 8.3.3 Cholesky’s method 128 62 CHAPTER 9 66 SOLUTION OF SYSTEMS OF LINEAR EQUATIONS (INDIRECT METHODS) 73 74 9.1 Iterative methods 133 i 9.1.1 Jacobi iterative method 134 9.1.2 Gauss Seidel iterative method 138 iL 9.1.3 Successive over-relaxation 145 i 9.2 Residuals 147 9.3 Convergence of iterative methods 150 9.4 Ill-conditioned system 153 CHAPTER 10 NUMERICAL DIFFERENTIATION 1S7 tiation formulae based on equally 188 forward 1S8 backward 162 formula 164 formula 168 "s formuia 170 172 174CHAPTER 11 NUMERICAL INTEGRATION integration oidal rule with error term rule with error term 3/8 rule with error term nation merical integration pezoidal rule son’s rule rapolation d quadrature formulae termined co-efficients single-step methods ultistep methods ons value problemsCONTENTS APPENDIX A Some important theorems APPENDIX B Vector and matrix norms exercises VII 268 272 277 278 290CHAPTER 1 ERRORS IN COMPUTATION lations with numsers, whether peformed by hand sare subject to aymerical errors. This form of ble in numerical computation and should not be oss errors’ arising from programming mistakes or ign or misuse of algorithms. For example, if we e can obtain the exact answer '8’, but if we exact answer is a decimal fraction with igures. We have to be content with an 0.143. which is in error by about red this error because, in practice, inite number of significant figures: in ting devices, from pencil-and-paper puter, have this restriction, Typ’ e limited to eight or ten signifi- omputers can handle a few more. formed there are many possible ecurate data and inaccurate formu- affect the accuracy of the solu- it is therefore evident that the due to one or both of two sour- ors of calculation. Errors of the t those of the second type can lease. Thus, when such a number ate value in a computation, we pproximation by taking e to as in some way approximate, to obtain results consis- and with better app- apter is to identity the effects on our computa~ and methods relating to4 to give methods a estimating euewen a jons ant te calculations © approxi tne results obtain | act c R SOURCES 8 1.2 ERRORS AND THE! | involved in mathematical problems may be diy, ; The errors v } i roups: : . ogee following ta. the statement of the problem. More I Eee acs are almost always idealizations made in a ety. earsemncicat model peor Sih Pe itation cot up a mi ven the model and the physical syst, The discrepancy betwe is a source of error problem. (Mathematica world are investigated ineering, etc.) ‘Ash i coring: crimes happens that it is either difficult or q impossible to solve a given problem when formulated preciy ly. If that ‘is the case, it is replaced by an approxi problem yielding almost the same results. This is the sow of an error regarded as the error of the method. 2. Errors occur due to the presence of infinite processes } mathematical analysis. Many numerical inethods express { In « desired result as the limit of an infinite sequence or and o sum of an infinite series. Since, generally speaking, |analys infinite process cannot be completed in a finite aN cua) cn forced to stop at some term of the sequen ie heat to be an approximation to the required si!-9 AP Soames ae a termination of the process git Borer a Se ch, Mion ae truncation error. In 1 sociated with the system of A to ma Al computation” there “are 2 of numeration. In numé ich he hte Wace. BA ‘undamental limitations imposed a le computing medium at our disPSppryn tation is subject to the numb, a8lic e time! ‘Finite space tilis use "to truncation & slight) "8 the exact ansWe The umber of figures ‘there an approximate mAlthou *-0,003333.... they ¢ we can When ‘which can be termed as error of ¢ 1 models and their relation to the y : in applied mathematics, physics, a 1evens OF NUMERICAL ANALYSIS 3 the original data into the final | ors of operation are inherent | 5, Errors due to numerical results. In this way, err- Parameters (in formulae) whose val- ues can es ie determined approximately. Such, for instan- divi ce, Physical constants and mathematical constants ide such as w and e. Although Teely. in mathenate Ue Such constants are defined pre- es eiuite coerce! terms, © thei representation requires an infinite sequence of digits, for example, m = 3.1415926 © = 2.7182818 .... these in numerical computation we must approximate lues to a finite number of digits, for instance, 142 (rounded to 3 decimal places or 4 significant figures), 71828 (rounded to 5 decimal places or 6 significant figures). sequence is a data error which may be called the problem, quite naturally, some errors are absent a negligible effect. But, generally, a complete with all types of errors. JERS AND SIGNIFICANT FIGURES approximate computation, it is convenient tween numbers which are exact and those ite values. ‘ate Numbers): An approximate value of a e number, is defined as a number which ion to an exact number and differs only ber for which it stands 1/3, 200 etc. are exact number's because or uncertainty associated with them. se, m, 12, etc. are exact numbers but actly by a finite number of digits. they must be written as 2.7183, 5 are therefore only approxima~ lied approximate numbers. t number x and its approximate Approximate number = x ~ XERRORS IN COMPUTATIoy Figures): Significant figures | ELEMENTS oF sty, menrh umber is; any one of the digt Foti representation, or any zero ING-OFF 5 point or to fill the place? Significant Tins “q the nth place, 4 0.00475 the significant figure, : as place holder, e used merely to TIX the po,” ee Sicate the place Values ofthe oy! ifthe significant. In the number gy «) { ee r including. zero, are gienitican Figur, Pan however, (9006070 the first four Bee ne ERticy s.oCcdetts the fy, In the pe nlag, digits, including the other two digits. All 5 ety non-zero igit of 0.0006070 cs if the last digi at eee’, are significant Tigre er must be written as 0.000607, (a) ir they significant, then eee numbers 0.0006070 and 0.000607 ary \ aah this point of view, has four significant figures ang digit is Ie the same, because the former later only three. When writing large numbel it is even ¢ the zeros on the Tight can se i i i ‘AMPLE 1.1: R ignificant figures and to fix the ou oe paces eee digits, This can lead to misunderstal, Ur. and three sig values ct numbers are written in the ordinary way. EELS TION, oaca example, the number 675,000. It is not clear how many signi, soe” promt ‘igures ‘there are, although we can say we have at least te3.442, 3.14. Ih ae “ambiguity can be removed by using powers-of-ten nota s ei EXAMPLE 1.2: jentific notation) as 6.75x10° if the number has three s;EXAI a figures, or as 6.750x10° if it has four significant fig 7.4721, 7 as 6.7500x10° if the number has five significant figv ing generally, this notation is convenient for nun! a large number of non significant zeros, suc oy = 2.30x10, and the like. SOLUTION: We obta Let us now introd ig _ DEFINITION 1.3 (R OF NUMBERS AND ROUND-OFF ERROR: requires more signi the resulting error 7, we obtain must be rounded of hie {n other words, the epresented to its f If we attempt to ‘raction with an in ent with an appro wound-off error of terminates. In order to use SY * sase, we do not kno We must cut it down to # "|. iome calculating devi 357143, °° It 1.857, or 1.85714, or ae 3s #@ that the resul re = 1.8571428571...,ELEMENTS OF NUMERICAL ANALYSIS Ss es z digite ROUNDING-OFF RULE. To round off or simply round a number to n ; eXcepy significant figures or digits, drop all digits to the right of laces gp) the nth place, or replace. th em by zeros if the zeros are needed as place holder. Im this‘rounding-off rule, note the following: (a) if the first of the the remaining digits (>) if the first: discard retained digit; (c) if the first discarded digit non-zero digits among those retained digit; (d) if the first discarded digit, however, is exactly 5 and Bi other’ discardededigive Sater zeros, the last retained digit is left unchanged or is increased by 1 according as it is even or odd (the even-digit rule), discarded digits is less than’ 5, leave unchanged; led digit exceeds 5, add 1 to the last is exactly 5 and there are discarded, add 1 to the last EXAMPLE 1.1: ff the number n = 3.14159265... to six, five, four, and three significant figures. SOLUTION: Rounding the number to given significant figures we obtain respectively the approximate numbers 3.14159, 3.1416, 3.142, 3.14. inl .zcitse arin 7 /EXAMPLE 1.2: Round the following number to four significant fig- syort vie ail ¥ 7.4727, 76346, 15.235 and 15.245 10 298Ke mB SOLUTION: We obtain respectively the approximate numbers , 15.24, 15.24. concept of the round-off error.ORS” IN COMPUTATION’ oe ELEMENTS be anything ; t, the answer may Gehows2ss0. voceni gang. In other words, Rern tinal ue to round-off by + 0.0005, pipearestnet 7 sition after the last significant figure w! EXAMPLE 1.3: Find the round-off errors introduced in Example |) SOLUTION: The round-off error introduced in each case of tj example is as follows: 1.6 ABSOL 7.4727 - 7.473 = -0.0003 76346 = 76350 = -4 _ Amber 15.235 - 15.24 = -0.005 15.245 ~ 15.24 = 0.005 DEFINITIO n decimal is the larg 1.5 TRUNCATION ERROR Many numerical methods express the desired results as the lini f an infinite sequence or the sum of an infinite series. | Practice’ only a finite number of terms'can be computed; the co For w = 31 Sequence of neglecting the remaining terms is the error known é the truncation error. A truncation error, therefore, leads to a approximate formula being used instead of an exact one. Consid a convergent infinite series and we see DEFINITION Eee ater ey mee exeaeer en 2 e 8 16 2 absolute err Thus, if | number of terms number x wt sterms are contrib 5 to S, we conclu hite number of ter! we might denote! ogi imSbs From this it 2 range possible to red the series. T# ding error for s can see that ‘ « For brevity, $.converging “ELEMENTS OF NUMERICAL ANALYSIS g Table 1.1 Truncation Error be ee n Se Error in S a 3 | 0.875” ~tortas S | 0.96875 | 0.03125 e Le, 10 | 0.99902 | 0.00098 a 15 | 0.99997 | 0.00003 4 1.6 ABSOLUTE, RELATIVE, AND PERCENTAGE ERRORS a DEFINITION 1.5: An approximation X to x is said to be correct to n decimal places\or positions (or agree to n decimal places) if n is the ‘largest ‘non-negative integer for which ~ swwioeds gaisimif, Sd? — 4 sivioeds gaisin i eee a sei} odd agdt aggrsd zi ~som to Yosws7e silt ai For m = 3.14159... and m = 3.142, the absolute error is and we see that 7 is ‘correct to 3 decimal places. DEFINITION 1.6 (Limiting Absolute or): The limiting absolute error of an is Be number not less than the absolute error of ‘Thus; if E, is the | number X which takes the Pe absolute error of an approximateERRORS IN COMPUTATION the limiting absolute error, of the number 1.4: Determine “ead of the number ™. 4 which is used instea UTION: It follows from the inequality 314 approximation %,tae, is said, to be correct to esietiieert figures “(ondigits}-arn is the Margest non- nega- tive integer for which (1.6) where m, an integer, is the highest power of ten when ¥ is’ repre- sented in the form of decimal. For example, the number x = 36.00 is an approximation to the exact number x =35.97 correct to three figures, since 10 = 5X01 san z DEFINITION 1.9: The limiting relative error E. of a gi appro- ximate number x is any number not less than the relative error of ‘that number. Thus, by definition we have eowgit tr ititgie ; PPE 3 BE SPORES: ia Be eres,ee °c is given as mine the limi- i ter at i ghdm’, of wa! or are Mpram(force)). Deven the weigh’ + (weighing the water: .001 8! mclopethe result TION: We have 4 = p = 999.847 ef- Sal ee o.corer, and) * ‘ie ta 0.001 x 10% ‘Thus E,. = 999.847 DEFINITION 1.10: The percentage errch is 100 times. the relative error. Thus percentage error of an approximate quantity = 100E, . ative and percentage errors are It is to be noted that rel: whereas absolute errors independent of the unit of measurement, are expressed in terms of the unit used. a. EXERCISE 1.1 1. Round-off the following number to four significant figures: 11.64498, 81.9772, 48.365, 67.495, 4.4995002, 39.63244. to three, four and 2. Round-off the number nm = 3.1415926535. five significant figures anc also find the absolute error in each case. 78340, 15.235, 15.245, and “digit? Tr Fi % A th ei 1.7 THI THEOR approx! ors of PROOF: sider th Then and, hen COROLL equal to Thus Followi numbers. RULE Fo! D Fi de ii) Rou oneELEMENTS OF NUMERICAL ANALYs?5 u 8. Round-off the following numbers: a) 8632574 to 4 significant figures 5) 3.1415926 to 5 decimal places ©) “8.5250 to 2 decimal’ places @). 1.6750 to 2 decimal. places. Find round-off error in each case, 9. A machinist measures a 0.5 inch bolt to the nearest thousand- th of an inch, and a carpenter a 10 fost beam to the nearest Cighth of an inch. Which measurement is more accurate? 1.7 THE ERROR OF A SUM THEOREM 1.1: The absolute error of an algebraic sum of several approximate numbers does not exceed the sum of the absolute err- ors of the numbers, PROOF: Let Xe X mn x, be the given approximate numbers. Con- sider their algebraic sum (1.9) + [ax] + + (ax, olute error of an algebraic sum is of limiting absolute errros of the termsERRORS IN COMPUTATION number of decimals. Hid the numbers, taking into account all retained deci- mals. ‘ v) Round off the result, reducing it by one decimal. EXAMPLE 1.7: Find the sum of the approximate numbers 0.348, 0.1834, 345.4, 235.2, 11.75, 9.27, 0.0849, 0.0214, 0.000354, cach ct to the indicated significant figures. c SOLUTION: We find the least accurate numbers 345.4 and 235.2 whose absolute error may attain 0.1. Rounding the remaining num- bers to two decimal places, we obtain 345.4 235.2 ) m1e7s 9.27 0.08 “909 0.02 “ad 0.35 0.18 % 0.00 602.25 Round the result to one decimal place by the cven-digit rule. We get the approximate value of the sum; 602.2. The total absolute error E, of the result is made up of three (0) { mi i) the sum of limiting errors of the original data: eet SAF E 210° + 10%+10'+ 10'+ 10° + 107 + 10° + 10° + 10° rounding errors 0 0.008054, < 0.009: ELEMEN] Hence and thi EXAMP 86939, nifican SOLUTI hundred give the Round sum; 154 1.8 THE | We cons Xp xX From e difference The limi where x is erence betw When one they must t tion,ELEMENTS OF NUMERICAL ANALYSIS 13 Hence EF. = E, +E, +E, = 0.222 + 0.009: 0.050 = 0.281'< 013 a a a and thus the desired sum is 602.2 + 0.3. EXAMPLE 1.8: Find the sum of 36490, 994, 557.32, 295000, and 86999, assuming that the number 29500 is known to only three sig- nificant figures. SOLUTION: Since the number 29500 is known only to the nearest hundred, we round off the others to the nearest ten, and add to give the sum as shown below: 29500 86940 36490 990 560 154480 Round the sum to hundred, we get the approximate value of the sum; 154500 or 1.545x10°. i XS. 1.8 THE ERROR OF A DIFFERENCE We consider the difference u = x,-x, of two approximate numbers = xa ee ; Th - From equation (1.10), the limiting absolute error E,_ of the is as: Ait, Bde ateibeonags.. 37%.eRRORS IN, COMPUTATION 9.8, assuming that each) num- ast figure place, therefore, place to obtain 1.9: Subtract 46.365 from a7 ‘approximate and correct only t? its I UTION: Since 779.8 is correct to one decimal e round off the number 46.365 to on® decimal 46.4. Thus 719.8 - 46.4 = 733.4 1.9 THE ERROR OF A PRODUCT THEOREM 1.2: The relative error of a product of several appr eds-the sum of the relative imate nonzero numbers does not exc errors of the numbers. PROOF: Let i aap » we obtain ELE an co to EX gui all so He No’ be son mu renELEMENTS OF NUMERICAL ANALYSIS and E, is the relative error of the product. + Consequently ESE EB 40. +E: (ua) “is clear that (1.11) also” holds true if’ the factors 1, 2, ..., n) have different signs. COROLLARY 1.2: The limiting relative error of a product is equal to the sum of’ the limiting relative errors of the factors: ree As EB (1.12) EXAMPLE 1.10: Find the product 349.1x863.4 and state how marv fi- gures of the result are trustworthy if the factors are corru.t to all written figures. SOLUTION: We have 0 -05 0.05 .000143 + .0000579 = 0.00020. ieee 4gmteeeees.4 = -° Now, u = 349.1x863.4 = 301412.94 ~ 301413. a, 2 ares is E. =u 113x0.0 020 = 60.2826 ~ 60. ~3 the result should sites Lol ow erERRORS IN COMPUTATION ELEMENTS OF 3 ff two or m ind the product of tw ay ee pig it is sufficient to: EXAMPLE 1. “a ntains one or two significan,S6.3/ ee factor; gure but no y significant figures ay accurate factor (op : When it is desi jate numbers of dif ‘ound them off so that each col figures more than the least accurate 2. in the final result, retain as man there are correct figures in the leas| keep one extra digit). SOLUTION: Since EXAMPLE 1.11: Find the product of the approximate numbers and x, = 72.397 correct to the number of digits written. tient SOLUTION: Rounding X, to one decimal place, we have As this err we take 25.2 = 72.4. Thus REMARKS: | = 181 = 1.8x10%, accurate tha: ded off so less accurate 1.10 THE ERROR OF A QUOTIENT cant figures Leta, b be two approximate numbers. If LU THE ERR a cee ots Here u =a = ¢ and then Inu=iIna-inb po Thus, we hav (1.13) Hence, the li is m times th EXAMPLE 1. (0.3862)*, ass e , ed thet asetiguré SOLUTION: Lé number a iste ereMENTS OF NUMERICAL ANALYSIS 0 EXAMPLE 1.12: Find the number of correct ‘figures in the quotient 56.3/ 15 , assuming that the numerator is correct to its last fi- gure but no further. es a a IN: We take {5 = 2.23 T fon SOLUTION: We take 45° = 2.236, Thus 0.05 | 0.0005 EB. = 36-3 * Za96- = -000888 + .000224 = 0.001 es6.3 . Us 55g = 25-18 = 25.2, ui Ei, = uxE,, = 25.18x.001 = 0.025. js this error does not affect the third figure of the quotient, we take 25.2 as the correct result REMARKS: If one of the numbers (divisor or dividend) is more accurate than the other, the more accurate number should be roun- ded off so as to contain one more significant figure than the less accurate one. The result should be given to as many signifi- cant figures as the less accurate number( or one it more). 1.11 THE ERROR OF A POWER Here u = a(m any natural number). Then In u = min a, aids as to goo1 tm 3 to 7 rore,,2xiteles ginal iE w aS Som a ha a i eyrtt to". aeAlso, Therefore, E 1 relative error of the re = 4xE = 4Xx0.0G01- = 9,00052. ERRORS IN COMPUTATION sult is Ir (0.3862)* = 0.022246 E 1 = 0.022246x0.00052 = 9.0000116. “Since this error affects the fourth significant figure oft! result, the best we can do is to write. (0.3862)* = 0,02225 1.12 THE ERROR OF A ROOT Suppose u= fa, then Inu = A IG + ina and, hence mn That is, the limiting relative error of the mth root of an appn ximate number is only (1/m)th of the limiting relative error .. the given number (i.e. radicand). 1.14: Find the number of trustworthy figures tt the number in parentheses is corre ‘arther. 5 (1.15; ELEMENTS rey Find 3.94 no f Com figu Deter 73.5€ facto Evalu round which to as a); Find 68.243 gits o1 Detern es int The nu only t how mi Find tl 876.3/4 true on Find th assumin digit gin Evaluate first pu maximum efficient: maximumTATION OF NUMERICAL ANALYSIS EXERCISE 1.2 L Find the sum of the approximate numbers 561.32, 86.954, Giger. 204 491-6, each being correct to its last figure but no farther. 2, Compute the difference {2703 _ AZ correct to five significant figures. Determine the product of the approximate numbers 12.2 and 73.56 and the number of Correct digits in the product if the factors are correct to all written digits. Evaluate the following expressions. round-off error in each figure, estimate the limits between, which your calculated value lies. Finally quote each result to as many significant figures as are reliable. Then, assuming maximum 2.362x1.76 a) P72x1 76 yy 7.62 + 54 1.32-0.463 S 15-3 ° i6.Ix2.17- Find the sum of the approximate numbers 136.421, 28.3, 321, 68.243, 17.482 if the numbers are correct to their last di- gits only. Determine the relative error and the number of correct figur- es in the product of the numbers 93.87 and 9.236. Sees AMMEPMIEE are. both approximate and true / their last digits. Find their difference and stave how aes y figures in the result are trustworthy. rs . figures in the quotient of jumb are approximate and the quotient u=25.7:3.6 ire correct to the lastCHAPTER 2 LINEAR OPERATORS 2.1 INTRODUCTION y Let F be a linear real function space, that is, F is a set of t functions f,g, ... with the Property that f ¢ F and ge F implies af + Bg € F. Here a and B are arbitrary constants. Let us define the following operators which map F into itself: E f(x) = r(x +h) (2.) — B f(x) = f(x + h) ~ f(x) (2.2) WV f(x) = f(x) - f(x - h) (2.3) 6 f(x) = f(x + h/2) - f(x - h2) (2.4) B f(x) = (172)[f(x + h/2) + f(x - hv2)] (2.5) The operators) £,A,V,5 and j are respectively called. the shifting operator, the forward difference operator, the backward ‘rence ope ntral difference operator and the meanLINEAR OPERATORS ELEMENT “72, . E*¢(x) = f(x - h/2) or we | Two operators iy and i are said to be equal if Pf= Pf for all f ¢ F where ¢ : Similar] EXAMPLE 2.1: Find Ae™. ae SOLUTION: Dek eet ick = (e" - eX. sy ete. An « EXAMPLE 2,2: Find A E™(sin x). ingful on tion f € | SOLUTION: AE 'sin x = A sin (x ~ h) = sin (x = h + h) - sin (x ~-h) = sin x ~ sin (x ~h) = 2 in Baber Cia Fs - ful Corn i (d + (2% = b) ineful. Hoy EXAMPLE 2.3: Find f(x). SOLUTION: Pex) = put (x)) = wC1/2)it(x + h72) + f(x = 2) pas gua ace = Ca)lult(x + h72) + tx ~ n/2)1) bom 272) (172) Fee) 4FGOH(72KF 4 DD 2.3 OPERA TT SUA) x + h) + 2-100 4 F(x - WI: ae a Any of th other operate in # by D, we can write (2.6) 3” tottow from th 4 S .. f(x) ae 3 be nanTORS : ELEMENTS OF NUMERICAL ANALYSIS or we write E=e a. peer sitief ined by means of the power séiies in (0.7). Similarly one can define, for any operator P, 8 yy, a Pp = P sinP = p- 5 al i 2 3 fie (14P) = pra Ps) fy P's z Pa et etc. An operator defined by means of an infinite series is) mean- ingful only if the series obtained when it operates on any func- tion f € F is convergent. For example, let Q be defined by D? 2 < D’ If F is the space of polynomials, the above definition is mean- ingful. Howeve if e“ € F then Qe™ does not have any meaning, since # =eU+1+1/2+1/3+...), Qe and the series on the right side does not converge. ‘s+ (2.8) (2.9) (2.10) (2.1)LINEAR OPERATORS af ELEMEN ya aire and # = (2B + ie 1, (2.14) 24D gives Ww? Lies, +E + 2). . ho 1 Thus 8742 = ay? This gives the useful relation we hai 2 Zi 8 (2.14) \ we ele = EXAMPLE 2.4: Prove that 8° = 0 (1 + 4)" SOLUTION: sae? op Later of Therefore oa Ey EM But an +t 2 1 hence S =A+1+ mh 2 pee 2:3 sagt A + 2) Ce 2 = 070 +1). EXAMPLE 2.5: Show that yi” = (1 + a/2)*(1 + a)? SOLUTION: — Since ji = (1/2)[EV? + EY), therefore w= (1/4)E + E* 4 2) But E=1+4, 1 Evalua there a/4il + a + +2] are hol TA 2. Prove | teers oe ae (13 ay? + VF 200 4a) tray == St 2th tO) ( (1 « 32 ig ( a (e (1) 2 que (8) é a> [i=] X (h)ATORS fLEMENTS OF NUMERICAL ANALYSIS 25 2.4 DIFFERNCE OPERATORS AND THE DERIVATIVE OPERATORS (2.13) From the relation Sanity we have A=E o> =\} (2.14) Vel- 1>e™ Ba Epa. or wr Sounn tho), = (v2)? + EY} = cosh (hD). Later on, in chapter 10, we shall derive the following formulae: 4 hp =a - 4 (2.15) n?p? = a? - 0° (2.16) Pav vv! gh og a (2.1 oH ae (2.18) iis a ee -ar etPRATORS, 27 CHAPTER 3 EIGENVALUES AND EIGENVECTORS OF MATRICES 3.1 INTRODUCTION Eigenvalues play a very important role in obtaining deep under- standing of many physical system. For example, the stability of craft in flight and the modes of vibration of a bridge are overned by eigenvalues. They are also important in the study of growth of certain types of population. Suppose X denotes the pumber of females of a certain age in a given population. If the population is in the state of age stability, the vector X satis- ‘ies an equation of the form AX = AX ere A is a square matrix and A is a positive real number which whether the population increases or not. If A < 1 the lation decreases, If A > 1 it imcreases, and if A = 1 it re- the same.EIGENVALUES AND EIGENVECTORS OF MATRICES (A - ADX = 0 fr. If xX (x, x, in n unknowns: x), then we have a system of n equations x), ibs + Gate Of (a, - Alx, + ax, + a,x, ee $ dee 0; ae tee Adx, a ‘Thu aK 0. are ax, + ax, tak tl Ha = Ax, Tt is well-known that a non-trivial solution exists if and only if the determinant of the system vanishes i.e., First Bij are —. “3 Boy ge San se. =0 (3.2) So x a ¥ nl 52 mn di | is an -+(-D>k =0, (3.3) is an . r that if we put A = 0 j t with det A. Hence k z Ua (3.4) wo? oe matrix A and of A. TMS a mari of them may ,.+ pire,LEMENTS OF NUMERICAL ANALYSIS 29 Find eigenvalues and corresponding eigenvectors of i] LE 3.1: the matrix aoa aie ‘ ines siete SA 6 A 2 s-6 Seoyan the characteristic equation of A and A = 6,-1 two eigenvalues. To find eigenvectors we have to solve (2 - adx, Pax 0, 4x, + (3 - ale, 0. o> A = 6, then the system becomes ax, JX, = 0, Inst ax, ox, be chosen -, Let x, = 1, then x, = 3/4. So s fe" % opi: ~ ee ES be Similarly we find thatepPEIGENVECTORS?OF MATRICES EIGENVALUES Al oe ante Sat =O, . d 7. To find A has eigenvalues “12 orn nL we 12,7. Hence as eee Corresponding tions X = (x.,x,.x,). Thus we have to solve the system of equa a ie 2. ~2x, + 6x, ~ 24x, = -X), + 3x, + 10x, = x, ~ 4x, + 13x, - 6x, + 24x, = ORs 3 2 x, - 4x, + 14x, =0, 2x, - 10x, = 0, 0. Since the equations are not linearly independent, we can choose 6 a 2, x, = 6. Thus ; is an Similarly we x, arbitrarily. Let x, eigenvector, corresponding to the eigenvalue 2, = -3 2 find that | 2] and | 1 respectively are the eigenvectors corres- 1 a Ponding to the eigenvalues A, = 2 and a, We now give some simple results about eigenvalues of a matrix. Suppose AY ay i a, are the (not necessarily distinct) eigen” values of a square matn is called the trace THEOREM 3,2: If eigenvalue A then s PROOF: PR ge or deeid oF NUMERICAL ANALYSIS 31 aeuenTs peFINITION 3.1: Two column vectors X, and X, are said to be orth- agonal if XX, = 0- A 1 ee 1, 7 e a Here x, = (x) , for example if x = [ then x = (-i,2+i,3). DEFINITION 3.2: A matrix is said to be hermitian if A = A. ? rics rasa adel EXAMPLE 3.3: If A ls Bl then A’ = hs s] oe fl 1-i] _ and w= x= (1, a] ea Thus A is a hermitian matrix. THEOREM 3.3: A hermitian matrix has only real eigenvalues. PROOF: Let A be hermitian and A be an eigenvalue and X the corresponding eigenvector i.e. By definition X is not the zero vector. Taking the transpose of both sides we have eavisvasgis is: epit x ots gouley (ax = ax’, tT 4 BA nee. Liat ctEIGENVALUES AND EIGENVECTORS oF MATRICES 32 genvalues nvectors corresponding to distinct ei THEOREM 3.4: Eige! deren of a hermitian matrix are or AX, = AX PROOF: Let ax, = 2%," and ad , . oa tian matrix, Proceeding as abov' and A is a hermi where au Ay have i . The above equati where we used the fact that A, is real quation implies a, xX =0. ) e x, Since x # AY we have which proves the theorem. An important corollary of the above theorem is as follows. COROLLARY: A symmetric real matrix has only real eigenvalues ani the eigenvectors corresponding to distinct eigenvalues are ortho gonal. It is easily shown [8], (page 84) that if x xX X ar 2 eigenvectors of A corresponding to distinct eigenvalues then th set X, +» X} is linearly i endent. Hence every nx1 column vector X i combination of these ve~ tors. Hoy can be shown that [1] even ! ge is orthogonal set of linear) lways be found. 2041 3 has an eigenvalue "°° x If Henents OF NUMERICAL ANALYSIS 33 Ix, =X, Tat 2, EXAMPLE 3.5: The matrix Peino) A=]011 oo1 has the eigenvalue unity repeated three times. However there is only one linearly independent eigenvector 3.3 GERSHGORIN’S THEOREM Now we consider some estimates for the absolute value |A| where A is any eigenvalue of a square matrix A = (a). If X = (xx, 7 is an eigenvector corresponding to A, then we have a eh (3.5) m= fax, esi “K":sa0 yo" 4 It [xy] = "2% |x], then we have a Ph |B seed =O beet PbSkSqaHbat okt mort &© MATRICES ENVALUES AND EIGENVECTORS 01 EIG - é (3.6) max J Ee o lal 1 i es, henc t ix has the same set of eigenvalu e Also the transposed matrix 3 , | (3.7) max al, EN crpee Veal ret From (3.5) we have E QQ - ax, = q aX: rey Thus JA-a lal ie Ther or Simil As before We replace the above result by a weaker one lies in one of the discs with centre at a, and radius From Finally In (3.13 (3.13) | EXAMPL. 4 Here we }ELEMENTS OF NUMERICAL ANALYSIS = 3.6) THEOREM 3.5: If A is an eigenvalue of A = {a,) then it satisfies the estimates (3.6) - (3.9), ce EXAMPLE 3.6: Consider ~_ -2 6 -24 A=1]0-3 10 1-4 13 Here 2) la | =2+6 +24 = 32 3 . Ll la, | =0+3+ 10219 e } E la] =0+4+ 3618 / — Therefore (3.6) gives JA] = 32. (3.10) Similarly (3.7) gives [Al = 47 (3.1) From (3.8) we get < 30 -13),55. (3.12) Ja +2[ $30 or [A + 3| = 10 or la 13] o abtyiaatp Finally (3.9) givesEIGENVALUES AND EIGENVECTORS OF MATRICES Ja-1) 4 or [a- 3] #4 or fa es] ss Ja - 1) = 7 or [a- 3] $50 or aA -s] 34 Thus all eigenvalues lie in the union of Ja-t] = 7 and Ja-8| = 4. The matrix has eigenvalues 3, 4 and 5. 3.4 THE POWER METHOD nique to find the dominant that a matrix has eigen- . = [a | and The power method is an iterative tech eigenvalue of a given matrix A. Assume values which are so arranged that |A,| > lal =. If u is I5 24 us, U, tod any arbitrary n x 1 vector then it can be expressed as a linear combination of the eigenvectors and we write u are n linearly independent eigenvectors. eee uf... tau. v1 2 nn 2 Then Au=aAu +aAu +... +a Au, ieee, 2 nn SB Auleaau+..+aru. rid 222 a nn It is clear oe k, k, k Aus aA du, + au, + au, ‘ ay k AL yk = — Ena, § fem ba(eye + +a(2)e] ow 1 pdt ze au as k is taken large 2 Hi oF If ScpieweNTS OF NUMERICAL ANALYSIS 37 ub A eo cece Cae nace Bao ee ates (3.15) (A wv) From (3.15) we have for large k AMY ayatu Thus ACA) © ACA‘) Zz : 5 er A GHEE {PProxinate elgeavector ‘corresponding 40 thé Rk erethod based on the above analysis fer tind. ing, the dominant (in absolute value) eigenvalue of a matrix, is called the power method. We proceed as follows: a) Choose a suitable arbitrary column vector u, Usually u is chosen as (1, 1, ..., 0)", b) Form As \ ) Multiply voye umber k, so as to make the largest ele | ment L to 1 If one is working with an automatic P ne is e to multiply, at every step, with % x, (i * Il? ay? ee - Be We do this because if v is an eigenvector then so is kv. 4 the numbers from getting too visable to a th ig phd’, date