Numer. Math.

59, 273-276 (1991)

Numerische
Mathematik
9 Springer-Verlag1991

A note on Halley's method

Miguel A. Hern~ndez Ver6n
Departamento de Matematica Aplicada, Universidad de Zaragoza,
Colegio Universitario de La Rioja, C/Obispo Bustamante n ~ 3, E-26001 Logrofio, Spain

Summary. We introduce the degree of logarithmic convexity which provides

a measure of the convexity of a function at each point. Making use of this
concept we obtain a new theorem of global convergence for Halley's method.
Subject classifications. AMS(MOS): 65H05; CR: G 1.5

Introduction

In this paper, we study the influence that the convexity of a real function f
has in Halley's method [-5, 7], in order to get the solution o f f ( x ) = 0. References
[-2] and [-3] give global convergence theorems of this method.
For each strictly convex function f i.e. f with f " > 0, we introduce an index of
convexity. This is the number of times that we need to compose the function with
the logarithmic function in order to get a concave one. This index, called degree
of logarithmic convexity, provides a measure of the convexity o f f at each point.

Degree of logarithmic convexity

Let V be an interval, a neighbourhood of x0, and C~2)+(V) the family of non-

negative functions in C<2)(V).
Definition 1 f ~ C<2)+(V) is a log-convex function at x0 if log f is strictly convex
at Xo [6].
If f is a log-convex function at x0, then there exists a neighbourhood W
of x0, such that h = log f is strictly convex in W:

h"(x) f ' ( x ) f ( x ) - [ f ( x ) ] 2 > O for all Xo~ W.

-~ [f(x)]Z
Therefore, feC~2~+(V) is a log-convex function at x0 if and only if
f (xo) f'(xo)> [ f (xo) ] z. I f f ( x o ) + 0 , we have
f (xo) f ' (xo) > 1
[f,(xo)]~
274 M.A.H. Ver6n

Iff'(xo) =0, i.e. ifxo is a minimum o f f , then x o is a minimum ofh.

Let us suppose that f(xo)= 1. (Otherwise, replace f(x) by f * (x)=f(xfff(xo).)
We define
E1 (x) = log f ( x )
F.+I(x)=log[I+F.(x)] (n~N).

N o t e that, by induction, F.6C~Z)(V) and F.(xo)=O(n~N).

Definition 2 We say that f is n-logarithmically convex at Xo if F. is strictly convex
at Xo. If f is n-logarithmically convex at Xo for all n ~ N , we say that it is
infinitely-logarithmically convex at Xo [4].
Theorem 1 (i) For all n~N we have
(1) F'(xo)=f'(xo)
(2) F" (Xo) = f " (Xo) -- n I f ' (Xo)] 2.
(ii) Xo is a minimum o f f if and only if f is infinitely-logarithmically convex
at Xo.
(iii) I f Xo is not a minimum o f f then f is n-logarithmically convex at Xo
if and only iff"(Xo) > n f ' (xo)z.
Proof. (i) It is obvious that (1) and (2) are satisfied for n = 1. Taking into account
that/7, + l ( x ) = log [1 + F,(x)] and F, (Xo)= 0, we obtain (1) by using induction.
On the other hand,
F~'(x) [1 + F.(x)] -- IF; (x)] z
F'+I (x) -
[i + F.(x)] z

We apply induction again and see that (2) holds for all n.
(ii) If Xo is a minimum of f, then (2) shows that F"(Xo)>0 for all heN, and
therefore f is infinitely-logarithmically convex at x o. Conversely, the condition
f ' ( x o ) - n [ f ( x o ) ] 2 > O for all n E N implies f ( x o ) = 0 . Since f is strictly convex
at Xo, this is a minimum of f
Part (iii) follows easily from (i). []
In order to generalize this measure of convexity, we introduce the degree
of logarithmic convexity.
Definition 3. Let f~C~2)+(V) be strictly convex, and f ( x o ) = 1. If x o is not a
minimum of f, the degree of logarithmic convexity of f is defined to be the
real positive n u m b e r given by
f"(Xo)
Lr = [f,(xo)]2 9

If Xo is a minimum o f f we let Li(xo)= + 0o.

Consider fr a strictly convex function with an arbitrary value at
Xo, and define Ly(xo)= Lf,(Xo). Some elementary calculations yield

f(xo) f"(Xo)
Ly(Xo) = Ly,(Xo) =" [f,(xo)]2
A note on Halley's method 275

When f is a locally convex function at Xo, i.e. f"(Xo)>O, all the previous
results turn out to be true with slight modifications, but the same definition
for the degree of logarithmic convexity [4]. Besides, we can extend this definition
to every function fsC(2)(V).

A theorem of global convergence for Halley's method

In terms of the degree of logarithmic convexity, Halley's method consists in

applying the iterative process given by

f(x) 2
(3) x,=F(x,_l) with F(x)=x
f'(x) 2--Ls(x ) "

It is known [-1] that Halley's method can be derived by applying Newton's

method to the function
f(x)
h(x)= [f,(x)]l/z.

In what follows, we assume that f satisfies the following conditions:

(4) feC(3)([a,b]), f(a)<O< f(b), f'(x)>0 and f"(x)>=O

for x e [a, b].

These conditions imply that there exists one and only one root st(a, b)
of the equation f(x)= 0. Suppose that the starting value Xo satisfies s =<x o =<b.
Studying the convergence of Newton's method for the function h by means
of the degree of logarithmic convexity we obtain a new theorem of global conver-
gence for Halley's method.
Theorem2 (i) If Ly,(x)<3/2 in [a, b] then (x.) is a decreasing sequence that
converges to s.
(ii) If Ly,(x)e(3/2, 2) and L y ( x ) < l in [a, b], for xo>a+2f(b)/f'(a), then the
sequence (x,) converges to s.
Proof (i) It is immediate that h has a point of inflexion at the root s since

f'(x)
h"(x) = 4 I f ' (x)] 1/2 Ly (x) [3 - 2 L f, (x)]

If LI,(x)<3/2, it follows that h is a concave function in (a, s) and a convex

function in (s, b). On the other hand, as h" is a positive function in (s, b), the
function h'(x)=f'(x)l/z(2-Ls(x))/2 is increasing in (s, b). Besides, h'(s)>0, and
therefore h is also an increasing function in (s, b). By applying Newton's method
to h, we obtain that (x,) is a decreasing sequence whose limit is s.
(ii) Since F(x)= x-h(x)/h'(x), it follows from (3) that

d h(x) d f(x) 2
with C(x)=
2 - Ly(x)"
276 M.A.H. Ver6n

As L h is a negative function in [a, b], we have

2 f(x) 2 L's(x ) > 0 .

L h(x) > - - 1 . ~ 2 - (1 - Ly (x)) 2 - L s (x) f ' (x) [2 - Ly (x)] 2

Then, [Lh (x)[ < 1 .~.f' (x)(2 -- L I ( x ) ) - f (x) L'y(x) > 0. On the other hand,

f"(x~
L'f (x) = ~ [ 1 + L f (x) (L I, (x) -- 2)3.

Therefore, we obtain
that [ L h ( x ) l < l if and only if 2 - L i ( x ) >
Li(x ) [l+Li(x)(Li,(x)--2)]. Taking into account that Li,(x)e(3/2, 2) and
L I ( x ) < 1 in I-a, b], it follows that JLh(x)l < 1 in [a, b]. Thus, there exists Me(0, 1)
such that ILh(x)l < M in [a, b].
h(x)
If we let H ( x ) = x - - h w ~ and x . = H ( x , _ 1), it is immediate that H(xo)e(a, s).
Since x l - s = H ( x o ) - - H ( s ) = H ' ( ~ o ) ( X o - S ) for r xo) and H'(x)=Lh(x), we
get I x l - s l < M b x o - S l . By induction, we obtain that H(x2.)e(s,b) and
H (x2. + l)e(a, s) for n > 0 . Besides, xn-- s = H (x,_ I) - H (s) = H' (~,_ O(xn_ 1 - s) for
~,_~ e(min {x._~, s}, max {X,_l, s}). Then I x , - s I < M Ix,-~ - s I. Reiterating this
procedure, we obtain I x . - s l < M " l x o - s l and therefore l i m x , = s . []
n

If the function f is decreasing, all the previous results can be seen to be

still valid by changing slightly the arguments used. The condition f ( x o ) > 0 does
not affect the results. W h e n f ( x o ) < 0 and Ls,(x)<3/2, (x,) turns out to be an
increasing sequence i f f is concave, we consider the respective degree of exponen-
tial concavity, and obtain analogous results.

References
1. Alefeld, G.: On the convergence of Halley's method. Monthly 88, 530-536 (1981)
2. Davis, M., Dawson, B.: On global convergence of Halley's iteration formula. Numer. Math.
24, 133-135 (1975)
3. Gander, W.: On Halley's iteration method. Monthly 92, 131 134 (1985)
4. Hernandez, M.A.: Funciones Cardanianas, Grado de Convexidad. Tesis Doctoral, Universi-
dad de Zaragoza (1985)
5. Ostrowski, A.M.: Solution of equations and systems of equations. New York, London:
Academic Press 1973
6. Roberts, A.W., Varberg, D.E.: Convex functions. New York, London: Academic Press 1973
7. Traub, J.F.: Iterative methods for solution of equations. Englewood-Cliffs, New Jersey: Pren-
tice Hall 1964