{S6.leraive Ealinnle for Egenvalics 939 tea 4b leew eisinay 2 “The general real solution has the falling form. with reat Salas. “peaf sto} 3. Any salon sith e2 = 6, =0 4 atrcted wo the engin hecaune oF the negate exponential faci Other solaons hive components that grow witout hou. ad tne tryectries spa onan Be care oto maketh problem foronein Section 5.0 There theconaiion for atactin toma @ was that an egenvae he le than Vin maguitade to make if 0, Hee the eal put ofthe egenvale must be negative, to make 2" =» 0 sccond sm in) 5.8 ITERATIVE ESTIMATES FOR EIGENVALUES 1 scicaie applications of tinea algebra, eigenvalues ate sek knox press Forunately. lose umercl approximaies usally quit sasfactay In ck sone applications quite only 2 rough appisimaion to the gest eigewalue. The is Alport describebekow can work well fr this case. Albeit proves a fouaation for amore powerful meted tat can give fastesinues for ae eigenvalues 2s Wel. ‘The Power Method Te power metho applies wo ann» marx A wh a Sty demminant gen ae 2, Whichmeane that ust ie ages mabohie vale hana the ater eigen ies, In tis case. the power metnod proaices 2 salar sequence that appreaches 2 and 2 ‘ect seqence tat appreachesa corresponding eigenvect. The hackgremnd or te metho estson the eigenvector decane tion sed atthe beginning oF Sectann $6 ‘Assme fr silica Ais ditgonaliable and "ks a ass oF eigenvectors Yio aranged so thei coresponding eigenvaliess,... 2 decrewsein sie, with Ihe ated dominant eiervalue tt Thats Pal > [al [2a] 2 oo = Bl AL sic tere o ‘Ace sa in uation of Seton 4 ifin Be writen tm bemoan tadsMeetbalate dawn cy 0, Then ding by (98 La @ Fromincuaty (1). thefistion 4s A... 2o/As areal estan 1 in magni and sotheirpoviers goto er. Hence Bu MAecath a8 0 o320 HAPTER 5. Figenalues ae Eigenvectors ‘Thus. forge k 2 sults multiple of 4°x detsmines ames the ame direcion asthe eigenrecir cy, Sinceposiive aa riples do nocchuage the dvetion ofa vec ‘Af isel points anon inthe ste diction as vv, provided cy #0. exampLea tat =['$ ,$}0=[f]oma=[$]. mien 4 mw signe, Dad 1, an epson 2, #7 thee tg ed Fe Fenn moti #esad cont etm es Pe When rows SOLUTION Thetis te clans us Analogous calculations complete Table 1 TABLES RRerates ofa Vector f 2 7s eT CA) 10) Be 8] Fl 53] (3) The vetors 8... A ae shown in Fig The other voters se pring too lag ocispla. However tine segmentsare drawn showing te directions of ose sectors Infact, the dincconso he vectors are what we rally watt wee, notthe ver ters thonsclves The ns som toe approeching the lite rpesiting the egenspace spud Mane cy haste the ne (asps) deans and te Line (eigenspace) determined by ¥, goes 2549 at bs FIGURES Dicciossetsmined by xk PEA “The vectors 24)" Ain @) ate sealed to ake them converge tev, proved 174 0, Wecarnat aces Af inthis nay became we donot know dy. But we can scale (Sich AAxtomake ite legovtentry 21 ties at thatthe resating sequence [xy] will Cconvergetoa mkiplectv, where anges eatry 1 Figure 2showa the sealed quence‘88 Rerutive Estimates for Eigenvalues 924 for Example 1. The eigeavahie A, canbe estimated frm thesequsnce (3), t00. When X, isclese fo aneigensecor fr Jy, the vector AM, Seles 10 Dyke wilh eh entry AX, approximately 2; mest cosrespondin eny in x. Because the arpesteaty i xfs Ty the lags emry i isclose 10 A}. (Caf poots ofthese satenents ate ‘nsited) "THE POWER METHOD FOR ESTIMATING A STRICTLY DOMINANT EKGENVALUE, 4 Selestan niu Veco Xo Whose lgestemty 1 2 Fork =0.4 Compute Ax. bh Letts bean enty in Ae; whose abate vas isa large a posible 1 Compate x41 = (1/40) 3. For almost all choies of x, the sequence {11} approaches the dominant cigenvalie,and the sequence (3) approaches a cone sponding eigenvector EXAMPLE? ype pws 4=[4 {Joon [2] see when k= 5 ad estate the daminasteigenvaie Sid a comespanding sigenvector oa [SOLUTION Calcultions in this example and the net were male with MATLAB, hich computer with Iealigt accuracy, alhugh we show only afewaigniicant figures hte, To been. compte AN ad erty the lest tty ji ft ANE weft EIB) = eee[t aL]-[a] = ‘Seale AX) by 1/)y to $e», compute AX, an deny the ange entry Ax: ne =als)=[25] a)Los]= [tio] = 748322 CHAPTER 5. Hienvaluesand Higenvecors Seale AX by 1/0 ge, and on. The esas af MATLAR ealeulation forthe fi five eration ae aranged ivTable 2 TABLE 2 _The Power Method for Example 2 epe 412 3 7 z TET CT 2) Gy Ge] Tose] (2) GN) (a) (a) (ia) [in] 22s a0rs tums Tame a ‘The evidence from Table trngly auggeststhat (1) approaches (1,2) and (ee) approaches 7 Is, thea (1,2) sa eigenvoctorand7isthedowsnaateigenshie, This ically verified by computing “b)-E Ib )-L4]-12] . ‘Tre sequence {1} nExample 2 converged quicly wo Ay = Tbecause He vecond igenvalucot 4 wasmuch male. lata Jo = 1) In gnera. the raefconverpence {depen on the tt y/Ay),boausethe vectore(dy/Ay Fs ineyaation 2) sea source oferar when uring a scaled verdan of Ax as an estimate of cy. The other Fractions 2; /y ate Ukely 19 be sale) IF s/f les 1, then fy} aa) can gamverze vety show, and ober approsination methods may be peered ‘With he poner metod, dee isa ight chance ht de chosen ial vector x Will have ne component inthe vs directen (wbea cy = 0). But comper oanding ‘stots diag the eakulaions ofthe ate ilely to ctate a vector vi at Teast esl componcat inde daetonct¥.Irhateccastheay will Sato comers tamale ore ‘The Inverse Power Method This method provider an approcimition for any eigenvahi, pride a go initial estimate #01 teeigenvave) we kaown,Inthis ease, we et B= (=a) ap) Ine poser metind to. Itesn Bestown tha theeigensIneso ARE... Ate Ine elgenvatuesor Bare Rew Dew hee ud the comesponding cigenvectorsace he same as hove for A. (Seo Exereias 15 and 16 Suppose, for exarpl, tht ois clower to 1; than tothe eter eigenvalues of A Thee |/(s 8) will hea sity dominant eigeavalbe of B. I is wally cbse w 2 tHe 1/(2_ a) mac lazer thas the lhe eigen\lves of Band the inverse power meth! produces 3 very api appuorimation te for alot all aces of ty. The follwing slgonthm gives the deta58 lenstive Htimatestor Hvaes 323 THE INVERSE POWER METHOD FOR ESTIMATING AN EIGENVALUE A OF A 1. Selo intial ent a suicictly se A. 2 Select ania yector xg whose ngs en is 3. Fork =0.1 4, Solve (aly, = 4 fry 1, Let yy bean erty in y, whose absolue valu sas lageas possible © Compute ry = a+ (1/1) 4. Compute y= 1/0 )%. 4. For limest all choices of, the eouence {) sppreaces the eigenvalue (oF A. and the sequence [X apprches a comresponding eigenNector Notice that B,orrather (4a) *, does nol appear a he algorithm. Insead of computing U4~ a)~1xy to ge the nest vero in Be seqense it beter wo slbe Uhecquaton (4 af}yy = %; ry, (and then scale 3, to podace sy). Since is eqiaton fry, rt Resolved fo each Fan factonaton of A~ a wil peed up the process EXAMPLE 3 isno!unvommon in some sppictionsto need tok he smallest eigenvalue of 2 matric A unl to nae a hand rough esimaves oF the eigenalus. Suppose 21,33, and 1.9 arcestinates forthe eigenvalues of the masix 4 below. Fi the salle eigenvalue, accurate tosin devel places, wo 3-4 as|-3 4 “45 4 SOLUTION The two smallest eigenvalues seem cose Lope, so We Ne he inverse ower methal fa A~ 1.97, Resulsor a MATLAB calculation akshowan in Table 3. Here xs was chosen arbitrary, yy = (A~ 19d). je is th largest enty ia, se = 10% Upusandsers = (/pa)yy. Asitturas out theintleigemae estimate Ws lah god, andthe inverse power Secon converge quickly. The smallest eigenvalue exact 2. 7 TABLE2 The inverse Power Method Le 1 2 2 TF ac | 20820008 2aKDs —_2eenno 2H? anestirite for thea et sigenvali of amttin ic nt ailable,one can simply lake a = inte imese powermettod. Thisetnice of a works reasnably Wel fhe shallest izenvalue s much coer 200 hanto heer elgemalies