Srtexnpotation | Gntmoducdion :- im in Sntenpolation accnding te Theile ay as" tha oaxt of recoling bei ths. dines of ths doable ”. “J also means Insestion om Filling up inte nmedate ia [demms of ths Series. Jt wm ts dechniqne of obbaining- ths vale of a dunction don omy intenmeclats values of the — Sndlependént ~voniable ie. of ange nt When cths values of. dbs Funcdion Kesisespon dingy tr number | of values of the axgument ae | given. The prucess oF ramputirg ths value of dhs Funchon outside ths Monge Of given values oF ahs Variable uy called — eatorap. JaHen . AN itestpedation Fostnulae am based on eths Frypothesis abet ths HY Funcdim Ud polynomial on dk Con be ebsles enteol by A “pokymemial nith a geod degree oF abpotortime ~ dion , &3- Bhs population oF oO down in dhs decennial Census Wee as Unden ? T Yeon 1334 door | 4932 1921 population qe &e vi | 33 | 40| [race no. Dare: The process OF Finds ny ‘Apo pulahon dos Bhs Year Soy 1925 Widh dhs [hep oF above deta us called | Interpolation While LAs psvkess te Bind ths population Fon dhe Yeo 1925 will be known Os Ratniapolation . Finite Dishenences 2- Assume dhat we howe a dable | oF Values Qaiee) sc ds 0,4 wyat.n m> of | ony Suncdion ys G0, che values of |] 2 being » OY, por POD cated Fast Fomwanc AM efenences . the WShesences of the Fisch Foawand Asfeqences one called Sewnd Foswad A Sf$estences arc] one lenotedl by Arges Ato Ah Simidaouly One can define Shind Fenpand ou Benences, Svunth Foawond Aisfenences, ete. DB Yo = OF: — 440 = Yo - Ya ~ Cds -¥.) nD eae et one Avge = O45 ~ O7ge= Uae 249 +93 — CY.- 233 +32) = Ys~ 33, +345 - do b"So 2 S331 - D793. Du- 2d2t 390 -Bs = Odo 733 + 3y4 - Yeo) = Yu- 492 + S32 - 43, tdo -PAGE NO. Aine Cae Stain, Bhenegone , clear cthat any ies oxnden MAtgfenence so thar Nidal lyon oo Oe Vaso ees Uhes2 J w Called ths Backward WSF enence ©“Peaatin . gn a Similar Woy one Kan dedine backward Aiffenences of dughen oudens, Thus, We ~ Obtain, Von NVv.- Vy, Q Ye- Ds ~ Cds - Ye) = Ye - 293 + 90.ac WEY: Seco 395 + 499 ~ we [Sgmbodic Rela t= eae | Diffenence donmulae can easily be aaa @stablished by symbolic methed » using (ith shd+ Openodonr EF and the: a averaging © os) ths mean ‘opesaton Ny) |, In addition t chs operators A, | Ne xe s alswady — Aedined - | The avesiaging Openathes MK ob achi- [ned by he eqyeeation fF ae | M Ys ie CYvbays te Besa) | The Shift cpenaha EF wh dedrned | by Ay 2g” > \ BEG. 2 Daxaa 3 | Ushich = Shows that ths edhect of E [1b te Shg hs Funcsional Yalus dv ce ths nest highen Valks Yy 45 - A second equation with E gives a ECFSOr Ey os Niet aes | { and in genenal , eo tT —n Eola yee. 3+ BD new easy tw desive oa netahon ~ | ship ES Aue nee os ean ome , | 4y.= 323-9, = Ey. -Y. = CB-2) ye [and hence as een 7>@ We KON new ercpness any Wyhen- osnden Fxtwand vsgestence 17 terms |G the given purction voles. Pox | example | 4%y. = Ce-9)*y. = (B7-3E*43E-2)Y au SMe) NRC Hh gee) 3 Fuium ahs dedinitions , Ds Following refatims Kan easly be established * Weajpsrs & . | nS { C2 Cet! ere 3 0 I] LU ee a b acs lee|, As) an example) ae paeve, pee Dt CRS The hoe i by detinition, MYs = = Coe4en + Fray) Se ae pitty ) z is an Lots Fle, eal ops 4 Festi) ee | Hence, ae | (a Cela EROS) {| ard pe= 2 Ce? y go) | 24, (ey £442) | 2 = fete) y a | - eco | T We. thenetone hove He faed stAma. PAGE NO, DATE rea Finally» We dedine ths openofon D Such that Dy Cdoe < Yo). ao Sienhatel Gps tee Bl iivetus Poo Be | | ths Teyfon’s Semies Yor) = 9G thy G+ © g "GF bo y"Go 4+ This Kan be wWoitten in ds SYmbol*e Foam 2 Ey(™) = (2 +hpo+ eee eo. n ince the sesies in chs brackets wb ‘dhe expansion of 24® 5 We obtain ths intenesding me Sut RSet 2 Gano Using dha ~zietatien QD of numven F UseSul identidies Can be derived. This se Jem wD Used — Sepora- be One dAcBned | in teams of wenen vided MEfe nen — Les ee dhs Second WM vided Ai fhestences Ff Foy). Fox as drsiee amtgtiments wb given by PCa otg) = F (rte) 2H) and by 1 A, = | denoted by F Geta, Me) oa by Lat su). Similarly Bhs nth diviole || > te Ai ffenenxce Bb Sven Ly SSS So Mae a HF CHM Ha My He This WW denoted by F Cesta poe And OT boy x # (ow) Ad a+ ~ Wn Prop estties of clivided Ujfenenres i- Jt a>, Divided ditfenences ane sSymmepsu'c FuncHomns of khein ANJZuments » Fe- dhe. welus. ony divided Efe sven ce ae independent oF aks omden of ths amgument -zy The nih divided Agpenendes a a oa polynomial of ha nth degsice on Konstan . > | The nth hivided Adfenernce Kon be | Sepaiessed as ths wohent of 2 dete mminants each of onden +a D The nth divided aAidfenence «Kon be expotessed as ths parecluct of | multiple invegnn als fe. aE a dy + [F Oa 2m) 2 js J dds [ dds. | a [ihr 3 Cun) olf5 i [Whesce Un = Ch da)ota + Cd - 02) of + (dre nh) Apis + dyiy%r yp da, da, ds p dy Ow independent voowables, and FH means ths Ch-3)f, dlawva- dive of 2. i __ "Newton “S General in} e spo Latren Foomul Rt FCO) , FO) yi. FC) be ahs values ef (mn) cosmespending ths. arguments He, Ap pve ey Mn 7 OF necessastily -equelly spaced + Fam Bhs cedinitim oF divided olde men ££ —pate: 7 7 Fv %0) = SOOT FON 5 FO)= $F Ou) + me Wore m6) 2m Me) —@) | etry ttre ayy Again EM a, As) = AMER TT ra nA Hy fon Fires Fe, 11) + (m-m)F Cote a) samme) Simidorly — F Ot, Hey 214) =F Cw, Ma y2n) + Crea) | Boy me ay Pa) eee) & (estas Ma, Minne} = Flee ea ii 1nd Cuemrn) Flas, Ha MD. err ees eyucton @) by Co-ed Ge by. C= 50) Cre 22) and. So Mm an \ = He) (r-ry) equation WW by rr) Fine - Fo) and addias. te equation (a7 »we howe FO) = Fre) + =o) F Gro oa) + (ae) Corny) £ Cre Aa me) to + (ms Nn or Ma oe Ad Cn- M2) «=: (a= Atroa Fm t+ Ry, , Ff by Whe ste ths Slemaineen Rn (dace eee - =cee) om ba) o : 1h ae Ew Kato Zt SPO) f SF oun Sunctim FO) bo polynor: tak of nth ofeg nee, Athen F Orv 210 %y a | an) tnanishes Se thot = EG) = FG) 4 Crome) P (re, Ha) t Gore) Orn (GE) Ge» Cr= Ano) £ Cofet tty 1) yAw) WS This donmula » called Newton’s genesal inte rUpo Lodi Foamula With vid eof USF een es 2use Fisting »B-Splines ond Amar. App sori mation ee Least S4rames cue Sidting ‘ ~pr9ce clumess At dhs sed °F data. points be Cri, yi) » Jea52 0 ---,m, ond det ths uswe given by Y= FQ be. Fitted ode ahis data . At a= 21, ths given ondincte i Yu and th Comnes ponding vous ™ ths Fitting curve wb gO). 3G 8) ub the —ensten oF app now mation oF m7, . then ne howe es 37 - FCW => & a SF we write _ S= [¥y -F Ou) 14+ [a -F I]t ot Tam - FOm)dI* Benepe Cat Op te ne of ene > & Chen ths medhod of - deast squares onSiS45 in Winimising Ss. pe.) dhs Sum. of . th Sqtuanes of — the _ estos jn dhs Following Sections , we Shol/ Study Ths Wrneor and, nun Line or |dees+ Ss qu ares Sting ae Biven cata Cm, Jide Pst, km ‘ts Fitting a | Zot Ys do + agi be th St s1cisht uine te be. Fitted te ths given tidodacy vig Cn e dae, fe Bp 2 po cky mm, Then. , CHINES PHM HZ oh eqn @D We | hove Het S= Dyas - Cae + arm) ]* + [31 (204 aryl bo Da = Con + On tm)? — Fox S & be minimum ,we hove | OS =o sc -2[y, -Ca+oaro,)J-2) - T Sa. I Ca tual -2 lan (@e+ xn] i — wa and Os =o =~ 25 [yy ~ Cast Soy) - 27 | 4 P'g,- (e+ 92%) [= +++ = 21m [4, = Coo + 2) Gs The abowe “ Gs) on qx (© ors called dhs musa eyuctions Po ong ane iet meso ec mmeras am Go and Qa, Since Ap Grol Yj ord (demons, garntdes . : | We Kon obtain easily ~ mn m ~ a sty; 24} } | ae ean Tach get aa | ia 7 = 2G) | 1h mn Sor, = ori | a2 de. ) i and dhen ant Fag x >. &) | Smee “OFS. arte SES) ane both ‘positive Dics oO ay af ths = points Ao Gnd ay, ah jiFoliqus | that these values provide oa Pint mum oF 2&. ty cen (, a and ZF ore hs means of mend y, despectvely - Fmim 247% 0) we have Zz Ue t+ On, avhich. Shows thok dhe ditted Straight. Une . passes dhaough ths.~ enksoid og the data points 7 Somedimes , a geedness oF Ft "adopted > The Co mah SF fictend Cec) sg edined as cece f S-5 3 V Sy hese . he Sge Se Gi- gp? ey | and os dedined by eqn @. SF cc | close tH 2, then He Ft wh Karsi- dened fis be god » althougst, — dhis dy not — always Frtue - No - Linear kume Fitting | - Powen -duncdion: + - The gyven data may rot always — Follows a Kinear sielahmship . This <9” be ascertained drm a plod of the given data. JF a nomdinear model te be Fitted 51 af Kon be convenient i Yoonsfosmed or a dinean ielabonshit- Joma nentnesr faws and — dthsin EeSeeS creelPAGE NO: (bares 77 ATE dnonsfosimatons ans given os dellons « oa & eeoZki Ts This Can be Woudtten as ay 2 axwrth 7 & put ey = VY, ats x. With dhese Jdoancdes- mations . at becomes ao Linear mode, ag" sb \ Toking eg axrithns of beth Sides, We get deg + adegng = bgaab . dn this ease . We put Cue cana bog =x + bo,,b = Ay ard -% =Aa, Zo Shot “Ya Ae Aa lk 5 \ (SD) ie y= ab Taking Ligarudhms oF beth sles. we ob}nnWe Stowe Logs = hogy a + beg > Y2Aet Asx, whens Ya Lugae$, Ae = legs, 0 and * 2 begs. y - ae 2 Aa b dn this Coase, we pole dn y = Ina +t bx > Ys At Ary Wheste ’ Ye Ing 2Aez Ina ,A ae and x= Ht® Podiyno micl oF nth Reyne = oo {| Pa | Xet the paly nomi ol | Cr ees | degnee , V2 Got G2%4 Grtt ves + Oyo? > (h be Fitted GG dhs data prints iryi), d=39,2 ne we the, howe z as Sisley Case ates e ouen tae oe ats om + Ts, —~Caet As + GrmMtaiure + an") | ie 1) ey oe oe lyn, = (Oct O2%m + Oa PA te tOn re] — Ww Equating & Zesto ths Hash port ol desivetives and Simplidying. . We obtain | ths nemmal wi then gives pial of thsSe rere 4 a PP axXimahion (bare “I s Amore genenal abporeach vx oe minimize ths Weyghted sum of ths Squares of dhe esnons — daken oven all pal lata. ; is a Yesivad | d spoints . SR dhis Sum ub Aenshed ss ee doppene ef We howe iS= ifs ~ FO T+. [y,-F Guy Sem see [beet Cote) aN = = Se We)? + Wa 7 4st Wm en 9m gn th, above eq" Us Wi ae pres<- | suibed positive, numbens ard ors - Colled Woyts. A weight wv prescoubed Acco ndin G& te netlative accun- cy 2 a data -boint . dF au ths | date points ane Occunal , We Set Wj=2 Fn cel... We rconsident lazain ths Wnear. ard pondineos | Kases , ee Linear Weighted feast Spuores Appromaton ‘ = | Zt Y2a.t+ an be the Spuaight ] dine. te be ditted Le be given’ ata [pots 5 MZ (ato) Doe, Com 1 Sm). Ten2 bh _ __Ym eee ee ——s — S Ca ,a) 2, 2 Wj ry = Coot ayn} 132 \ : = Fost Masma 031 minima » We howe > i a Sr I mos PeiE(os aA => hy Oa Ba, \ | Ushich give a: fi | as 9 S 1, [yj (ota 1) =0 i Qac (22 az Wy ie : Par z pa (acta Hi) IH = 0 i aA ind 1 \ a SE Sim plidication Yields Abs Syshem of eq™ Fon Av aad at Si ee cy aa shew a2 ini = 23) 3 122 2a t= = eS = » = 2 WinhFo S im = = Aimy; > ph 2 Bhs nosmal €yuahons.in case Gnd arse solved du obhbain o and a, * ae |4.4.1. Orthogonal Polynomials In the previous section, we have seen that the method of determining a least square approximation to a continuous function gives satisfactory results. However, this method possesses the disadvantage of solving a large linear system of equations. Besides, such a system may exhibit a peculiar tendency called ill-conditioning, which means that small change in any of its parameters introduces large errors in the solution—the degree of ill-conditioning increasing with the order of the system. Hence, alternative methods of solving the aforesaid least-squares problem have gained importance, and of these the method that employs ‘orthogonal polynomials’ is currently in use. This method possesses the great advantage that it does not require a linear system to be solved and is described below. We choose the approximation in the form: ¥(x) = ag fy(x) + Mf (X) +++ Oy fy), (4.42) where f(x) is a polynomial in x of degree j. Then, we write b Sago. )= J WOR){ 92) Le fol) +a, f(a) +--bay fy DIP. (4.43) a For S to be minimum, we must have b 2 J WO (y6)=Lao fol) +a fi (2) +--+. (DD S00) de a b 2 J WE) [YO)MLao fol) +a fC) +--+. FAD AC A | 04 4y b =f WA) (y@)—L dg fo) + afi (0) 400+ On fn OM Sn) bt ‘a144 Charter 4: Least Squares and Fourier Transforms The normal equations are now given by b b b ao] Wo fFaddera | WarfolMh@rdre-+ a, [| WOO fola)fy(adar b = J WG) fo ax a b b 6 a] WENA) foladderay] WOIfPEdde+-+a, | WO) AG) fade 0 a a W(x) yx) fi @) dx ieee b b b a J WO Sy (2) foe de a, J WC fy 00 AC) dbo ay | WO Fe dd J a b = J We yf, de 7 (4.45) The above system can be written more simply as 5 5 f WO) fo) fyedarea | WO) fi@) fart 157/463 b b +4, ) Werf df de= JW) YO FEA F014 Qaoam a a (4.46) In Eq. (4.45), we find products of the type f(x) f(x) in the integrands, and if we assume that 0, b b J Wes, cof,oo de= fwenszinds pag a pds, p=a, p#q (4.47) then the system (4.45) reduces to b b ay J W(x) FG) dx = f WOR) ya) f(a) de < ‘° b b ay | Ox Fe 3) de= J Wa) 960) fy(2) deFrom the preceding equations we obtain b J ¥@ yo Fax B20 Qa 448) b Jwespoa Substitution ofa, a}, ..., d in Eq. (4.42) then yields the required least squares approximation, but the functions f(x), fix), --.. f(x) are still not known. The f(x), which are polynomials in x satisfying the conditions (4.47), are called orthogonal polynomials and are said to be orthogonal with respect to the weight function W(x). They play an important role in numerical analysis and a few of them are listed below in Table 4.1. Table 4.1 Orthogonal Polynomials* Name §) Interval wo) Jacobi PLO AYixy 14,1] (1-)%(14 x) (a, B >-1) Chebyshev Tal) 441 (1x22 (first kind) Chebyshev Up(x) i (= x22 (second kind) Legendre Palx) (4) 1 Laguerre Lyx) 10, ) ral Hermite Hin(x) (vs) oF A brief discussion of some important properties of the Chebyshev polynomials 7;,(x) and their usefulness in the approximation of functions will be given in a later section of this chapter. We now return to our discussion of the problem of determining the least squares approximation. As we noted earlier, the functions (x) are yet to be determined. These are obtained by using the ‘Gram-Schmidt orthogonalization process,’ which has important applications in numerical analysis. This process is described in the next section. 4.4.2 Gram-Schmidt Orthogonalization Process Suppose that the orthogonal polynomial f(x), valid on the interval [a, 6], has the leading term x’, Then, starting with Ae) =1 (4.49)We find that the linear polynomial f,(x), with leading term x, can be written as AG) = x + kyo fo), (4.50) where ;,o is a constant to be determined. Since f(x) and f(x) are orthogonal, we have 4 6 6 J Woof fie dv=0= f HC) fo(x) de+h,o f WOOF) de using Eqs. (4.47) and (4.49). From the above, we obtain b J Wet ar ko =--————_ (4.51) f woeosoe ae and Eq. (4.50) gives ® b J xWeofow ae A@=x- ——. J WO fe(a) de a Now, the polynomial f(x), of degree 2 in x and with leading term x?, may be written as Arla)=3? +hyo fol) +h iG, (4.52) where the constants ky and kp, are to be determined by using the orthogonality conditions in Eq. (4.47). Since f(x) is orthogonal to f(x), we have 6 J Wea fo0L2? + ke ofola) + ke falas =0. Since [> W(x) fo(a) f(x) dx =0, the above equation gives b b J PW (x) fo(adde J W(x) de a = _. (4.53) | a! fwoseoma [waar @ ‘ koAgain, since f(x) is orthogonal to fi(x), we have 6 J W(X) ACL? theo fola) thoy dx =0. Using the condition that {? W(x) fo(x) f,(x) dr =0, the above yields bd J PHA de ky, =-4$ (4.54) J WeosPoas Since ko and ky, are known, Eq. (4.52) determines f(x). Proceeding in this way, the method can be generalized and we write Fax thy oft h@te thy fae (4.58) where the constants k); are so chosen that f(x) is orthogonal to Ls ACOs +4 F102). These conditions yield b J PWeOf@ ae kj, =-4+___—_.. (4.56) Jwe@sproa J Since the a, and f(x) in Eq. (4.42) are known, the approximation ¥(x) can now be determined. The following example illustrates the method of procedure.4.5 APPROXIMATION OF FUNCTIONS The problem of approximating a function is a central problem in numerical analysis due to its importance in the development of software for digital computers. Function evaluation through interpolation techniques over stored Secrion 4.5: Approximation of Functions 149 table of values has been found to be quite costlier when compared to the use of efficient function approximations. Let fi, fay «+++ y be the values of the given function and 91, 2 «+s On be the corresponding values of the approximating function. Then the error vector is e, where the components of e are given by e, =f; — ¢,. The approximation may be chosen in a number of ways. For example, we may find the approximation such that the quantity /(e? +e} +--+e,) is minimum. This leads us to the least squares approximation which we have already studied. On the other hand, we may choose the approximation such that the maximum component of e is minimized. This leads us to the ‘celebrated Chebyshev polynomials’ which have found important application in the approximation of functions in digital computers. In this section, we shall give a brief outline of Chebyshev polynomials and their applications in the economization of power series.*4.5.1 Chebyshev Polynomials The Chebyshev polynomial of degree n over the interval [-1, 1] is defined by the relation T, (x) = cos (n cos™'x), (4.57) from which follows immediately the relation T, (4) =T_y(2)- (4.58) Let cos“!x=6 so that x=cos@ and (4.57) gives T, (x) = cos n6. Hence Ty(x)=1 and T(a)=x. Using the trigonometric identity cos (n—1)8 +c0s (n+1)8 = 2cos nd cos @, we obtain easily Tyas) +Ty4i (8) = 247 (2), which is the same as Tua (2) = 22Ty (2) — Tyas (2)- (4.59) This is the recurrence relation which can be used to successively compute all 7,(x), since we know Zy(x) and 7,(x). The first seven Chebyshev polynomials are:Tp(x)=1 T@=x T(Q=27 -1 T(x) =4x° —3x (4.60) Ty(x) =8x4 - 8x7 +1 Ts(x) =16x° - 20x +5x T(x) = 32x° — 48x* +18x? -1. The graph of the first four Chebyshev polynomials are shown in Fig. 4.1 Figure 4.1. Chebyshev polynomials Tp(x), n=1 2, 3, 4. It is easy to see that the coefficient of x" in T,(x) is always 2", Further, if we set y = T,(x) = cos n@, then we get dy _nsin nO dx sind and ? cos n8 +nsin nO cot@ _ —n?y + x(dy/dx) sin? @ 1-x* so that 2. ay dy _ Dye xe eG m=O (4.61) which is the differential equation satisfied by T,(x).Secrion 4.5: Approximation of Functions 154 Itis also possible to express powers of x in terms of Chebyshev polynomials. We find 1=Tay x= (x) IT) +70) (BT) +7300) (4.62) [3% (+474) +T,0] (107, (x) +575(x) +750] 1 yqllOT@) +1572) + 6TH) +To(0]- and so on. These expressions will be useful in the economization of power series ‘to be discussed later. ‘An important property of 7,(x) is given by 1 0, men f BOBQ | 1/2, mane0 (4.63) a fiex © A that is, the polynomials 7,(x) are orthogonal with the function 1/Y(1—2?). This property is easily proved since by putting x = cos®, the above integral becomes (cos 6) T,(cos 6) d0 = [cos m9 cos n6 48 ° in(m+n)@ , sin(m—n) 0)" m+n)" 2m—n) |," from which follow the values given on the right side of Eq. (4.63). We have seen above that 7,(x) is a polynomial of degree 7 in x and that the coefficient of x" in T,(x) is 2”. In approximation theory, one uses monic polynomials, i.e. Chebyshev polynomials in which the coefficient of x” is unity. If,(1) is a monic polynomial, then we can write P(x=2"T (x), (21). (4.64) A remarkable property of Chebyshev polynomials is that of all monic polynomials, P,(x), of degree n whose leading coefficient equals unity, the ome 3polynomial 2'"T,(x), has the smallest least upper bound for its absolute value in the range (1, 1). Since |T,(x)| < 1, the upper bound referred to above is 2'~”, Thus, in Chebyshev approximation, the maximum error is kept down to a minimum. This is often referred to as minimax principle and the polynomial in Eq. (4.64) is called the minimax polynomial. By this process we can obtain the best lower-order approximation, called the minimax approximation, to a given polynomial. This is illustrated in the following example. Example 4.13 Find the best lower-order approximation to the cubic 2x3+3x%, Using the relations given in Eq. (4.62), we write 2a? 430? = 2a) +3h 01430 230? tire dro a3 dedi), since 7; (x) = x. The polynomial 3x? + (3/2)x is the required lower-order approximation to the given cubic with a maximum error +1/2 in the range (-I, 1). A similar application of Chebyshev series in the economization of power series is discussed next.