ON ~lOLTIPLE POINT TAYLOR SERIES EXPANSIONS by CHESTER WAYNE DYCHE A THESIS submitted to OREGON STATE COLLEGE in partial fulfillment of the requirements tor the degree of MASTER OF ARTS June 1956 APPAOVED Redacted for Privacy Araoclatg Profeasor of hdathematlss In Chargc of MaJor Redacted for Privacy &ot1ng Shalrman of Departmant of Mathsmatloa Redacted for Privacy Chalrnan of Sohool" Srailuata 0onmlttea Redacted for Privacy Dean of Orad,uatc Sohoo1 prcscntcd Data thosl.s tr fypcdt by lY. M. $ltone s ACKNOWLEDGMENT The author is indebted to Professor w. M. Stone whose encouragement and assistance led to the completion of this thesis. TABLE OF CONTENTS PAGE CHAPTER I. II • III. IV. INTRODUCTION 1 THE TWO POINT THEOREM 3 SPECIAL CASES 10 NUMERICAL EXAMPLES 13 BIBLIOGRAPHY 17 ON MULTIPLE POINT TAYLOR SERIES EXPANSIONS INTRODUCTION The rapid advance in the electronic digital computer and its application has brought forth many advances in numerical and approximation methods. The literature pub­ lished in this field has increased and work is being done in many areas. One such area is the evaluation of a function in terms of its derivatives at one or more points. Such formulas have been developed and used by many authors al­ though no general approach has been made. Boola (2, 10?­ 109) used a semblance of this idea to solve a special problem. Milne (10, 53?-542} appears to have priority with a method for the numerical integration of different­ ial equations which depends on the values of the first three derivatives. Salzer (12, 103-106) has developed integration formulas using both the function and its first derivative. Luke (9, 298-30?), Hammer and Hollingsworth (4, 92-96) and Fettis (3, 85-91) obtained various modified forms of the trapezoidal formula. Householder (5, 232-240) discussed several types of quadrature formulas. Hummel and Seebeck (6, 243-24?) developed a treatment for the evalua­ tion of a function without regard to quadrature methods. 2 Lotkin (7, 171-179) developed a specialized version of the first integral of the function. Andress (1, 394­ 396) developed essentially the same method on the complex plane. The bibliography here only indicates the numerous approaches that have been made to this problem. The objective of this paper is to generalize Taylor's theorem of the expansion of a function about a single point to an expansion about two points . multiple point expansion are: The uses of the to obtain generalized formulas f or mechanical quadrature; generalization of the Milne-Lotkin approach to the numerical integration of dif­ ferential equations; a representation of functions by ra­ tional fractions which is of interest in the construction of tables. 3 THE TWO POINT THEOREM The following assumptions will be made: tinuous in an interval which contains derivatives up to and including the tinuous and the (m+.n+P+l)th The letters negative integers. noted by ( ~) are con­ n must be of exponential are defined to be non­ The binomial coefficients will be de­ q < 0 t or q > p. will mean to sum over n. (m+n+p)th x; all the with the understanding that the expression is zero for and and m and 0 derivative is sectionally continuous in that interval; F(x) order. F(x) is con­ Also A function from k t G(x,p} 0 o if = (I<) The sy.mhol . L f(m,n,k) k=O to the larger of p < m o. is defined by the definite inte­ gral X (2.1) G(x,p) = f . (mill-tp-f:l.) F m n (t) (x-t) t dt. 0 Application of the standard one-sided Laplace transforma­ tion to both sides yields (2.2) g(s,p) = d n ds n 4 m+n+p L m+n+p-k {k) s F (O)J. k=O Because of the Leibniz formula (2.2) becomes n n (-1) g(s,p} = m! L k:O n k P+k (m+n +p+l)! s r (k} (s) (m+p+l+k) ! (2.3) (k) m+P - .t!l! ~ (m+n+p-k)! ( 0) F sk+1-p (m+p-k)! - p -1 = mi L k =O n k (m+n +p+l) ! n (m+n+p+l) ! (k) p+k t s (s ) (m+p+l+k) n + m! L k=-P m+p - m! ~ k=O k (k) p+k s (s) f (m+p+l+k) ! (k) (m+n+p-k)! F (m+p-k): s (0) k+l-p • The inversion formulas (2.4) 1 - 1{ sp+k (k) t k (s)} = (-1) k+p _!. d~+p k X F(x) p < 0, 5 and = (2.5) k X lead to the form (m+n +p+l)! (m+p+l +k )! ~ -1 p+k (k) {s f (s)} The formula for interchange of order of summation, (2.7) n k k=O j=O L .L n f(k,j) = n LL j =0 f(k,j) ' 1:1 p+k (k) f ( S) }+ k =j yields (m;n +p +l)! { m+p+~+k} ! (2.8) {S 5 n ) (m+n +p+l)! j-p (j) ( m! j ::O x j-p F j: + ~ k-p c n-j +p) (-1} . k! (x) k= j k-j (m+k +l): -p-1 (k) (m-tn +P+l) ! = m! L k=O f ( s)} (m+P+l +k)! ~ k =O To eva luate the second s~~ation (with respect to r) it is necessary to itera te t he integra tion of both sides ot the identity n~ + ( -1) (2.9) r =o m+l r c n-j +P) r X r +j j = x (1-x) times, the last integration being between That is, n-j +p 0 and 1. 7 > n-j +p £__ ( -1) r c n-j +p ) r=O 1 -------­ (r+j+l}····(m+r + j +1) r (2.10} 1 = 1 m! J o j m+n-j +p x (1-x) dx j I (m+n-j +p }! = m! (m+n +p+1)! and (2.8) takes the form n (-1} G(x,p) > (~) -p-1 = .£_ A (m+p+1+k)! k=O (m-m +P+l}! m% l l - · {s p-+k (k) r (s)} The theorem of the mean for integrals gives as an estima te of G(;x,p), m-tn+l G(x,p) = x (m+n +p+1) F 1 r mn (X) J (1-u) u du 0 (2.12 ) m+n+l = ml n! x (m+n+p+l) F (X} • (m-tn +l)! where 0 < X < x. The final form of the two point Taylor 8 expansion is, therefore, ~ (~) --1'-1 i_ ml k=o p+k {s t (k) (s)J (m+p+l+k)! (2.13) ~( k)' Gm ) F (k) (0)- (-1) k-pIn =L _ m+n+p- • { k=o {m+n+p+l)! k-p ~ k-p )F(k) (x)J x k-p + R(x,p) , where n (2.14) R(x,p) = (-1) m! n! x m+n+l (m+n+p+l) F (X) • (m+n+l)!(m+n+p+l)! The Boole (2, 107-109) development is quickly dupli­ cated in modern notation by using the two-sided Laplace An obvious identity, -sh -sh 1 - e (1 - e )f(s) = ----sh (1 + e -sh )f(s) transformation. 1 + e (2.15) = {~ 2 00 ~ = ~ (sh) 24 - 3 •••• J -sh (1+ e )f(s) n 2n+l -sh (-1) En(sh) (1 + e )f(s) , n=O where the En are the Bernoulli numbers, is written down. 9 The inversion yields (2.16} F(x) - F(x-h) 00 ~ = L _ ( -1 ) n=O n 2n+1 En h { F (2n+1} (2n+1) ( x) + F (x-h)} , which is equivalent to the original form of Boole. 10 SPECIAL CASES The re placement of p by 0 in the general formula of (2.13) yields (3.1) + R(x,O) or F(x) - F(O) (3.2) = ~ (m-+n-k)! L_ k =l (m-tn)! () { mk F (k) k {0)- (-1) ( nk )F (k) +(m+n+l) R(x,O) , where ( 3 . 3) R(x,O) = n m+n+l (-1} m! n! x {(m+n+l)!} 2 F (m+n+l) (X) This formula is that of Hummel and Seebeck (6, 243-247), who state that many special cases arise from variations of m and n. If n =0 e quation (3.2) becomes the Tay­ lor series with remainder. Since (3.4) J F (t) dt X f -1{ f(s) } ~ s = 0 k (x) }x 11 the choice of p =-1 reduces (2.13) to X I +L F( t) dt 0 = R(x,-1) (3 .5) (m+n-1-k)! (m n)! ----{ k=O m)F (k) (0)- (-1) k+lGn J F (k) (x)}xk+l• 0 k+l k+l where (3.6) R(x,-1) = n m+n+l ( ) F m+n (I) (-1} m! n! x (m+n+1}! (m+n) ! Here again varied formulas are obtained by adjusting m and n. For x Jo m = n, n-1 F(t) dt ~ = L_ k=o ) t (Zn-l-k ·( n \p!' (2n)! \ k+~ (k) (0} - (-1) k+l (k) k+l F (x)Jx (3.7) + R(x,-1) , where (3.8) R(x,-1) = n 2 2n+l (-1) (d!) (2n) F (X) • ( 2n+l ) I ( 2n ) ! Thus (3.5) is a generalization of Lotkin's (8, 29-34) formula since he limited his study to the case m = n. In their papers on the numerical solution of differ­ ential equations Milne (10, 537-542) and Lotkin (8, 29-34) have used the two point expansion formula, n = m = 3, as 12 a "corrector", with the Taylor formula used as a "predict­ orn. Lotkin's proof of convergence for such a method may be generalized without difficulty. The transform pairs X J (x-t)F(t) dt 0 and -1 L {f (3.10) reduce (2.13) for X (1) ( s) s p = -2 to X J F(t) dt - (m-+n) 0 (3.11) ~( 2 k) ' +L _ m+n- - · k=o (m+n-1)! = -ftF(t) dt 0 X mx } J t F(t) dt 0 {G m \ F(k) (0)k+2 ) = R(x,-2) k ( k) k+2 (-1) / n ) F (x)}x • \k+2 Probably a better approach to finding a formula for X I -p+l t F ( t) dt 0 in (3. 5). would be to replace F(x) by -p+l F(x) X 13 NUMERICAL EXAMPLES To show the value of the two point formula it is first applied to the Bessel differential equation, (2) xy (x) + y (1) (x) + xy(x) 0, = (4.1) = 1, y(O) and y(x) =Jo(x}. y Taking (1) y(x) y(x,l) = 2 12 - X 2 2(6 +X ) (4.2) (6 - 2 n =m (1) and = the solution z(x) in terms 1, 2 and 3 are , 2 y(x,2) = - X ) 2 4 2(240+ 17x + x ) y(:x,3) = 201, 600 - 41, 280x + 1134x - 3:x X = 0, ( 0) = z(x) (x) of the two equations defining of their values at zero for y )( 80 2 2 4 4 6 6 8( 25 .zoo + 1140x + 33x + x ) with appropriate remainder terms. is less than By taking 2•10 m -6 = The error in • n ;= 2 and setting up the equations for a stepwise procedure the representations of z(x) y(l,3) in terms of values at an arbitrary x0 y(x) and become 14 (4.3) y =Yo + {x-xQ) 2 2 (x-x ) (y~ + y' ) + 0 ( y t t _ y I t) 0 12 and 2 {4.4) z = Zo + (x-xQ) (z'+z') + (x-xQ) (z''-z'') • 0 0 12 2 To reduce these to solvable equations in terms of and z(x) the values of differentiating (4,1). z' and z'' y(x) are obtained by Finally X (4.5) and y (x-x0 ) (7x-x 0 ) + z{l 2 + 2 6(x-x ) + (x-x 0 ) (2-x ) 2 ~ X X J (4.6) = 12y6 + 6(x•x 0 )y~' + (x-x 0 and successive appliea tiona for to the table x0 = 0, 2 ) y t t t 0 1, 2; • · · •• leads 15 J~(x) X y(x) .To(x) z(x) 1 .7655 .7652 -.4399 -.4401 2 . 2246 . 2239 -.5770 -.5767 3 -. 2596 -.2601 -.3400 -.3391 4 -.39?8 -.3971 .06503 .06604 5 -.1791 -.1776 .32?7 . 3276 6 .1496 .1506 . 2781 .2?67 7 .3007 . 3001 .00643 .00468 8 .1613 .1?1? -.2535 -.2346 9 -.1109 -.0903 -.2461 -. 2453 10 -.2467 -.2459 -.02758 -.04347 Table 1. That is, J (x) has been rather quickly estima ted from 0 0 to 10 with sufficient accuracy for drawing curves. The values of J 0 (x) .T~(x) and have been taken from Watson (13, 666-685). A form of the error function serves as a second numerical example. For 2 (4.7) y(x) = e X f 2 00 e -t dt X the two point formul a for m = n = 4 and x 0 = 0 reads 17 BIBLIOGRAPHY 1. Andress, w. R. The expansion of a function in terms of its values and derivatives at several points. American mathematical monthly 60: 394-396. 1953. 2. Boole, George. A treatise on differential equations. 4th edition. New York, G. E. Stechert, 1877. 235p. 3. Fettis, Henry E. Numerical calculation of certain definite integrals by Poisson's summation formula. Mathematical tables and other aids to computation, 9. 85-91. 1955. 4. Hammer, Preston c., and Jack w. Hollingsworth. Trape­ zoidal methods of approximating solutions of dif­ ferential equations. Mathematical tables and other aids to computation 9: 92-96. 1955. 5. Householder, Alton s. Principles of numerical analy­ sis. New York, McGraw-Hill, 1953. 268p. 6. Hummel, P.M., and c. L. Seebeck, Jr. A generalization of Taylor's expansion. American mathematical monthly 56: 243-247. 1949. 7. Lotkin, Mark. A new integration procedure. Journal of mathematics and physics 32: 171-179. 1953. 8. Lotkin, Mark. A new integrating procedure of high ac­ curacy. Journal of mathematics and physics 31: 29-34. 1952. 9. Luke, Yudell L. Simple formulas for the evaluation of some higher transcendental functions. Journal of mathematics and physics 34: 298-307. 1956. 10. Mi lne, William Edmund. A note on the numerical inte­ gration of differential equations. NBS journal of research 43: 537-542. 1949. 18 11. Rosser, J. Barkley. z J -x2 e 0 dx and Theory and application of Jz 0 - p 2Y2 e JY dy e -x2 dx. Brook­ 0 1yn, N. Y,, Mapleton House , 1948. l92p . 12. Salzer, Herbert E. Osculatory quadrature formulas . Journal of mathematics and physics 34: 103-106. 19!55. 13. Watson, G. N. A treatise on the theory of Bessel functions. 2nd edition. Cambridge, England, Cambridge University Press, 1944 . 804p .