Univ. u Novom Sadu Review of Research Zb. Rad. Prirod.—Mat. Fak. Faculty of Science Ser. Mat. 22, 1 (1992), 159-166 Mathematics Series NEWTON-RAPHSON’S METHOD AND CONVEXITY Miguel A. Hernandez Veron Dpto de Matematica Aplicada de la Universidad de Zaragoza Colegio Universitario de La Rioja C/ Obispo Bustamante 3 26001 Logroiio (La Rioja), Spain Abstract The influence that the convexity of a real function f has in the Newton-Raphson’s method is studied, in order to get the solution of f(z) = 0, and a method for accelerating this iterative process is ob- tained. A new iterative process of third order is obtained, as well. AMS Mathematics Subject Classification (1991): 65H05 Key words and phrases: Newton-Raphson’s method, nonlinear equa- tions, iterative method, convexity. 1. Introduction The sequence {z,,} obtained by the Newton-Raphson’s method for the solu- tion of the equation (1) f(x) = 0, consists in applying the iterative process F(z) f(z) Q) ty = F(tn-1) with F(x) = We will consider f € C™([a,6]), m > 2, (a,6] CR, a convex and strictly increasing function in [a,6], with f(a) <0 < f(b), then (1) has one and only one root s in {a,t], [4]. Besides {2p} converges to s when zo € (a,)) and 159160 M. A. Hernéndez +this convergence is quadratic, [1]. Moreover, if we take zo € [a,b] such that Ff (zo) > 0, then {zn} is a decreasing sequence. In general, the Newton-Raphson’s method converges for any function f provided that f’ # 0 and f” does not change sign on [a,b], [4]. Therefore, changing f(z) for f(—z), ~ f(z) or ~ f(-z) we obtain our conditions. The aim in this paper is to study the influence of the convexity in the Newton-Raphson’s method. In order to study the convexity of f, we are going to use the log-degree of convexity [3], which will provide us with a measure of the convexity of f at each point. The log-degree of convexity of f at uo is defined to be the positive real number given by (3) DL Al(t0) = F"(uo){F'(u0)]-?. If up is a minimum of f we set U[f](uo) = +00. ‘We will study the influence of the log-degree of convexity of f on the sequence {z,}. Then we will prove that if the log-degree of convexity of f decreases, in some conditions, the sequence {z,,} converges faster to s. So, we obtain a procedure for accelerating the Newton-Raphson’s method. By applying this procedure we obtain an acceleration in the form yn = G(a,-1)- Besides, this type of acceleration will provide us with a new iterative pro- cess of the third order, and also we will prove sufficient conditions for it’s convergence, 2. Influence of the convexity on Newton-Raphson’s method Let g bea function, in the same conditions that f in [a,6], such that g(s) = 0. We consider the sequence ta = G(Yn-1) being G(2) =2— a te Then, by means of the log-degree of convexity, we are going to compare the sequences {2q} and {yn}. We will denote La(z) = h(z)U{h|(2) for each h function. Theorem 1. if Lj{z) > La(z) for f(z) > 0, then the sequence {yn} con- verges faster to s than {zn}.Newton-Raphson’s method and convexity 161 Proof. In our conditions {zn} and {yo} are decreasing sequences to s. There- fore, we are going to prove by using induction that yn < zp for alln € N such that z, # . Besides yn = 6 if t,-1 # 3 and z, = 5. ‘Taking into account that F(s) = G(s) = # and zo # s, we have that 21-1 = (F — G)(z0) - (F - G)(s) and therefore there exists & € (3,0) such that a — 1 = (F — GY'(Go)(20 - 8). On the other hand as (F ~ G)'(2) = Ly(z) - 1,(2) and f(£) > 0 we obtain that 2 — y1 > 0. Now we assume that 2, > yg for k = 1, Zn-1 # 8 and Zn — Yn = F(2n-1) — G(Yn-1) 2 (F — G)(Zn-1) since G is an increasing function in [s,5] and {yn} C [s,é]. So, in a way similar to the case k = 1, we obtain that zn — Yn > 0. 2-1, if z, # 5 then By applying the above arguments it is easy to prove that if tai # s and z,=stheny,=s. O Below we are going to translate this result by means of the log-degree of convexity. Corollary 1. If U[f](z) > U[g\(z) and g'(s) < f'(s) then the sequence {yn} converges faster to s than {zn}. Proof. Since U[f|(z) > Ulg|(z), for all x € (s,6), taking into account (3), it is easy to prove that 7 2 a 7 1 v< [WU - Une = 5 - Fe5- GH FH) and therefore f“(z) > g'(z) in (s,6). So, f(z) > g(z) for f(z) > 0, ie. in (3,6), and then Ly(z) > L,(x) for f(z) > 0. Therefore, by applying Theorem 1 we obtain the thesis. O Note that if U[f)(z) > Ulg](z) for f(z) > 0 the previous result is valid too.162 : M. A. Hernéndez 3. An acceleration of Newtor-F.aphson’s method As a result of Corollary 1 we obtain that if the log-degree of convexity of f decreases, the sequence {zn} converges faster to s. Besides, if we consider Yn = G(2n-1) then {yq} is an acceleration of {z,}. Now, we are going to define a function that verifies the conditions of Corollary 1, and thus we will obtain an acceleration of Newton-Raphson’s method. Since straight lines have the minimum log-degree of convexity if we con- sider ” f'(s)(x—s)”, ie. the tangent line of f at s, this function verifies the conditions of Corollary 1. Then, we take 92) = f(z) - [6"(8)/21( - 8? that is the Taylor aproximation to the tangent line of f at s. But, as s is an unknown point, taking into account that we will define y, = G(z,-,) and that {cq} is a decreasing sequence to s, we can consider for k = 1,2: $"(8)(zn-a— 8) ~ f"(2n-s)(@n-t — 2a) = f"(tn-a Uf (@n—a))*f(Zn—1)* Since tim 2" )len- ~ 9)* Fut \@na — 2m) peg : we have alent) ~ F(Ena1) — LED (oy a) *f(ena? 9 (@n—1) ~ f\(@na1) — f"(@n—a)LF (Ana) F(2n—1)- ‘Therefore we obtain the sequence fens) 2 £)(an-1) “ Y= Baa Fan a) T= Lylema)” Corollary 2. Sequence (4) is an acceleration of Newton-Raphson’s method.Newton-Raphson’s method and convexity 163 Proof. It suffices to prove that limallyn — 8|/lzn — sl] = 0, whore.{zn} is the sequence given by (2). O Note that as lim, z, = s and L;(s) = 0, we can take zo such that Ly(2n) <1 for all n. 4. A new iterative process of third order Acceleration (4) will provide us with a new iterative process, for the solution of (1), given by the expression f(Zn-1) 7 1 F'Gn-1) 2° 1-2) (5) mm = Zn — H(Ls(zn-1)) with H(z) = Now, we obtain the result of global convergence for this iterative process by means of the value of L », so we will consider f a function in the previous conditions with m > 3, besides 2 € {a,6] with f(zo) > 0. Theorem 2. If a,b € R and M = max{U[f(z)/z € [a,6]} such that M < 1/f(b) and Ly(z) < 0 in(a,}] then the sequence {z,}, given by (5), decreases to s. Proof. As zo > s and L;(z) < f()ULf\(z) < 1 in [3,6], then (5) is a decreasing sequence. To finish, we are going to prove that 2, > s for all n € N. It is clear that ®) an = UF Gn-1) + Ge-1)] (where F(z) = 2 — #3 and G(z) = 2 - 3B with g = 4). Besides, F"(z) = Lj(z) and G'(z) = L,(z). Then, taking into account that ____ F@)L(2) 1al@) = — Fay = Ly(a)P and Lp(a) = FMM ~ Lyfe) 2 - Lye)164 M. A. Hernandez +it follows that L4(x) + L(x) = Ly(a)[l — Ly(a)[Ly(@) - Ly(z)] and therefore F’(z) + G’(z) > 0 in [s,6]. On the other hand, a —s=[(F + G)/2\(2) — ((F + G)/2\(s), so there exists 9 € (s, 20) such that a— = [(Ly + L,)/2\(0)(20 — 5) and then z ~s > 0. By induction, it is easy to prove that z,— 9 > 0 for all neéeN. As z, > s for all n € N and {z9} is a decreasing sequence then there exists limaz, = u > 3, and since Ls(u) < 1, one can conclude that u = 5 taking limits in (5). Now we are going to study the case where Ly(z) > 0. Lemma. Let f be as before with m > 3. If 3/f'(b) > ULFz) > 2/f'(@) @ (ii) in [a,b], U[f() < 1/K with K = max{|f(a)|, f(d)}, then |L,(z)| <1 in [a, 8]. Moreover, if Ly(z0) < 1/4 then |Lg(z)| <1 in (a,z0], with g = f/f’. Proof. It is clear that U[f] is an increasing function in [a,6] if and only if ULF"\(z) > 2/F'(z) in [a,}], then for (ii) we obtain that U[f] is an increasing function in [4,6]. Since U{f](z) > 0 in [a,8] and U[f](b) < 1/K, as Ly = FULf], it follows that [Ls(z)| <1. Now, taking into account that {L,(z)| < J, (i), (ii) and y(2) = Fa ~ 19(2)2 - Loa) it follows that Ly is an increasing function in (a, 6]. On the other hand, |L,(z)| < 1 if and only if (+) IF(2(-Ly@)| < f@)0. - Ly(2))?Newton-Raphson’s method and convexity 165 and this is always true in (a, s]. If x € (s,8), (+) is equivalent to Lj{z){3(1 — Ly(z)) + Ly(z)Ly(z)] <1. Hence, as Ly is in an increasing function in [a,b] and 3(1—L,(x)) < 3 turns out to be suficient, taking into account that Ly(zo) < 1/4 and Ly(z) < 3 for (i). 2 Theorem 3. Let f be as before. If %— 4 > 7f(%)/6f'(%), then the iterative process given by (5) is convergent to s. Proof. As A tn = 5LF (20-1) + G(2n-1)] by (6), from (iii) and the Lemma we have that 6 > z > 2 >a, |L;(z)| <1 and |L9(z)| < 1 in (a, 20]. Hence, F(z0) ~ F(s) = Ly(Bo)(z0 - ) and G20) - G(s) = Lg(u0)(20 — 5) with Bo,#0 € (s, 2), and therefore |z; — s| < Clz — s{, with C < 1. Re- iterating, we obtain that z, € [2,6] and {|z, — s{} is a strictly decreasing sequence to zero. O It is known [2], that if we have an iterative process in the form given by (5), with H(0) = 1, H’(0) = 1/2 and |H"(0)| < +00, then this iterative process has cubic convergence. Therefore, if m > 4, it is easy to prove that the iterative process given by (5) is of the third order. Note that as Ly(s) = 0, it is easy to take a,b and z as in the previous results. 5. Practical remarks In applying the new iterative process given by (5), it is necessary to compute f,f' and f" for each step, besides the specific calculations. But, the required166 M. A. Hernandez calculations do not have the negative effect on practice since, for the solutior. of equations, each of the usual iterative processes of the third order [2], needs a similar number of calculations. References [1] Ehrmann, H.: Konstruktion und durchfihrung von iteratonsverfahren hoherer ordnung, Arch. Rational Mech. Anal. 4 (1959), 65-88. [2] Gander, W.: On Halley’s iteration method, Monthly Vol. 92 No 2 (1985), 131-134. [3] Hernandez, M.A.: An acceleration procedure of Whittaker method by means of the convexity, Univ. u Novom Sadu Zb. Rad. prirod. - Mat. Fak, Ser. Mat. 21,1 (1991), 27-38. [4] Ostrowski, A.M.: Solution of equations ond systems of equations, Aca- demic Press, New York, 1973. REZIME NEWTON - RAPHSON -OV METOD I KONVEKSNOST Prouéava se uticaj konveksnosti realne funkcije f na Newton - Raphson - ov metod za nalazenje regenja jednatine f(z) = 0. Dobijen je metod za ubrzanje ovog iterativaog postupka, koji ima red konvergencije tri. Received by the editors September 26, 1990