Journal of Computational and Applied Mathematics 177 (2005) 225 – 229

www.elsevier.com/locate/cam

Letter to the editor

Accelerated convergence in Newton’s method for
approximating square roots夡
M.A. Hernández∗ , N. Romero
Department of Mathematics and Computation, University of La Rioja, C/Luis de Ulloa s/n, 26004 Logroño, Spain
Received 20 April 2004; received in revised form 24 August 2004

Abstract
In this paper, we construct a modification of Newton’s method to accelerate the convergence of this method to
the approximation of the positive square root of a positive real number. From this modification, we can define a new
iterative process with prefixed order q ∈ N, q
2.
© 2004 Elsevier B.V. All rights reserved.
MSC: 65H05; 65B99

Keywords: Iterative processes; Newton’s method; Square root

1. Introduction

The problem of approximating the positive square root of a positive real number R has been widely
studied by different authors [2,7]. This problem is equivalent to solve the equation
f (t) = 0,
where
f (t) = t 2 − R. (1)

夡
Supported in part by a grant from the Ministry of Science and Technology (Ref. BFM 2002-00222) and a grant from the
University of La Rioja (Ref. API-02/18).
∗ Corresponding author.
E-mail addresses: [email protected] (M.A. Hernández), [email protected] (N. Romero).

0377-0427/$ - see front matter © 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2004.09.025
226 M.A. Hernández, N. Romero / Journal of Computational and Applied Mathematics 177 (2005) 225 – 229

To solve this problem, the famous Heron’s formula (75 b.C approx.), tn+1 = (tn + R/tn )/2, n
0, is well
known.
Modern calculus texts treat this algorithm as a special case of Newton’s method, tn+1 =tn −f (tn )/f  (tn ),
n
0. It is an established fact that the
√ convergence of Newton’s method is quadratic, at least for t0
sufficiently close to the solution
= R.
There are numerous studies to accelerate the convergence of Newton’s method [1,5].A modern approach
is given in [3], who applies Newton’s method to the function F (t) = t −2 f (t), where  ∈ R is looked
for in order to get cubic convergence. In fact, the value  = 23 is obtained as the appropriate one for
this purpose. The natural extension of this idea is to modify f (t) by the function F (t) = g(t)f (t), and
determine g(t) such that the order of convergence of Newton’s method
applied to F (t) is increased. So,
Gerlach [5], for a general real function f, obtains that, for g(t) = C/ f  (t), C ∈ R, the Newton method
has cubic convergence. Notice that, for C = 1, this method is a well-known third-order iteration: Halley’s
method or the method of tangent hyperbolas [4].
In this paper, we modify the algorithm of Newton’s process instead of the function. So, we consider
the modified algorithm
f (tn )
tn+1 = G(tn ) = tn − h(tn ) , n
0, (2)
f  (tn )
and we determine the function h such that we obtain an iterative process of qth order of convergence,
q ∈ N, q
2, for approximating
. This construction is realized in Section 2.
In Section 3, we analyze the monotonous convergence of these new iterative processes to the solution.
Finally, in Section 4, some of these new iterative processes are compared numerically.

2. A new qth-order method for approximating square roots, q
2

It is known, that the iterative process given in (2) is of qth order of convergence if G(
) =
, and
G(k) (
) = 0, for k = 1, 2, . . . , q − 1, and G(q) (
)  = 0. Then, if we consider the situation where G(t) =
,
for all t ∈ R, we obtain an iterative process of “infinite” order of convergence. Is this possible? To do
that, we consider
f (t)
G(t) = t − h(t) =
. (3)
f  (t)
Therefore, we can write
f  (t)(t −
) = h(t)f (t). (4)
Then, by Taylor’s approximation, we obtain f  (t)(t −
) = f (t) + f  (t)(
− t)2 /2!. Moreover, using
(4), it follows that: f  (t)f (t)2 h(t)2 /2! − f (t)f  (t)2 h(t) + f (t)f  (t)2 = 0, which is a second degree
polynomial in h(t). Therefore, the solutions are


1 − 1 − 2Lf (t)
h− (Lf (t)) = (5)
Lf (t)


and h+ (Lf (t))=(1+ 1 − 2Lf (t))/Lf (t), where Lf (t)=f (t)f  (t)/f  (t)2 is the degree of logarithmic
convexity [8]. Notice that, for f given in (1), it follows that Lf (t)  21 for all t ∈ R.
 M.A. Hernández, N. Romero / Journal of Computational and Applied Mathematics 177 (2005) 225 – 229 227

In consequence, the answer of the previous question is affirmative, but we cannot apply algorithm (2),
with h+ (Lf (t)), to approximate the positive square root
because the function has a singularity in the
root
. Then, we consider the Taylor’s approximation of the function h− (Lf (t)) given in (5), to define
algorithm (2), and it is easy to obtain that
 1 
h(Lf (t)) = h− (Lf (t)) = 2 (−1)j 2j +1 Lf (t)j .
j +1
j
0

In these conditions, we can define algorithm (2) in the following form:

 
q−2  1 
 f (tn )
tn+1 = G(tn ) = tn −  2 (−1)j 2j +1 Lf (tn )j   , n
0, (6)
j +1 f (tn )
j =0

for q
2. The question is then: Does this algorithm have q-order? The answer of this question is also
affirmative. Notice that, from (6), it is easy to prove that
q−2


G (t) = 1 − A0 + ((2j − 1)Aj −1 − (j + 1)Aj )Lf (t)j + (2q − 3)Aq−2 Lf (t)q−1 ,
j =1

j j +1
where Aj = ( j1/2 +1 )(−1) 2 , 0  j  q − 2. Then, as A0 = 1 and (2j − 1)Aj −1 = (j + 1)Aj , for
j = 1, 2, . . . , q − 2, it follows that G (t) = (2q − 3)Aq−2 Lf (t)q−1 . Therefore, since that Lf (
) = 0,
we obtain
G(
) =
, G (
) = · · · = G(q−1) (
) = 0, and G(q) (
) = (2q − 3)Aq−2 (q − 1)!
1−q  = 0.
So, the algorithm given by (6), is of qth order, for q
2. Thus, we have constructed a new algorithm
with any prefixed order of convergence q
2. Notice that, for q = 2, we obtain the Newton method and,
for q = 3, we obtain a famous iterative process of third order, the Chebyshev method [6].

Remark 1. Note that if we want to increase the order of convergence in one unity, as consequence of the
new sum appearing in the summatory, we only need to add the following term, Aj +1 Lf (tn )j +1 . Moreover,
the power Lf (tn )j has already been evaluated and Aj +1 = Aj (2j + 1)/(j + 2) is easily calculated.
Therefore, the operational cost increases by two multiplications and one addition. The procedure is
effective as far as concerns to the relation between the operational cost and the order of convergence.

3. Monotonous convergence

In this section, we study the monotonous convergence of the new algorithm given in (6). As other
authors [7], we consider t0 >
. Then, it is easy to check that
 
q−2
f (t0 )
t1 − t0 = −  Aj Lf (t0 )j    0,
f (t0 )
j =0

since Aj , Lf (t0 ), f (t0 ) and f  (t0 ) are positive numbers, j = 0, 1, . . . , q − 2.

228 M.A. Hernández, N. Romero / Journal of Computational and Applied Mathematics 177 (2005) 225 – 229

Table 1 √
Computation of values |tn − 35|

n Dubeau’s method, q = 3 Method (6), q = 3 Dubeau’s method, q = 4 Method (6), q = 4

1 0.169 · · · · ×10−4 0.817 · · · · ×10−5 0.454 · · · · ×10−6 0.142 · · · · ×10−6

2 0.140 · · · · ×10−15 0.781 · · · · ×10−17 0.385 · · · · ×10−27 0.123 · · · · ×10−29
3 0.782 · · · · ×10−49 0.682 · · · · ×10−53 0.199 · · · · ×10−111 0.709 · · · · ×10−122
4 0.137 · · · · ×10−148 0.454 · · · · ×10−161 0.144 · · · · ×10−448 0.763 · · · · ×10−491
5 0.732 · · · · ×10−448 0.134 · · · · ×10−485 0.397 · · · · ×10−1797 0.102 · · · · ×10−1967

0.01
kq (√ 35)
0.008

0.006

0.004 K q (√ 35)

0.002

3 4 5 6 7 8 9 10

Fig. 1. Asymptotic constants.

Besides, as G (t) > 0 in (
, t0 ), it follows that t1 −
= G (0 )(t0 −
) > 0, 0 ∈ (
, t0 ), and then t1 >
.
By induction, if tn−1 > tn and tn >
, by an analogous reasoning as before, we obtain that tn+1 < tn and
tn+1 >
. Then there exists l = limn tn , and, by letting n → +∞ in (6), it follows f (l) = 0, since that
h− (Lf (t)) > 0 for all t ∈ R, and therefore l =
.
In consequence, from the previous reasoning, the following result is obtained.

Theorem 1. If t0 >
, then the sequence {tn }, given in (6), is monotonically decreasing and converges to
the positive solution
of f (t) = 0, with f given from (1).

4. Numerical tests

When we want to apply the algorithm of qth order given in (6), it is interesting to note that, since
   
1 R f (t) t R
Lf (t) = 1− 2 , = 1− 2 ,
2 t f  (t) 2 t
 M.A. Hernández, N. Romero / Journal of Computational and Applied Mathematics 177 (2005) 225 – 229 229

we can write the algorithm given by (6) in the following simple form:
q−1  1 
  j
j R
tn+1 = G(tn ) = tn 2 (−1) 1− 2 , n
0,
j tn
j =0

with q
2. Next, we consider the application of algorithm (6) to the function f (t) = t 2 − 35 and we
observe the improvement with√regard to Dubeau’s method [2]. In Table 1, the values |tn − 351/2 | are
shown for the computation of 35 with t0 = 6.
Fig. 1, shows that the asymptotic constant of method (6): Kq (
) = (2q − 3)Aq−2
1−q /q is smaller
than the one of Dubeau’s method: kq (
) = −(q − 1)2q ( 1/2
q )

1−q where q is on the horizontal axis.

Moreover, if the two asymptotic constants are compared, we observe that

1
|Kq (
)| < |kq (
)| ⇐⇒ < 1 ⇐⇒ q > 2.
q −1
Therefore, from q = 3 (Chebyshev’s method), processes (6) are faster than the other one given in [2].

References

[1] I.K. Argyros, On a new Newton–Mysowskii-type theorem with applications to inexact Newton-like methods and their
discretizations, IMA J. Numer. Anal. 18 (1998) 37–56.
[2] F. Dubeau, Algorithms for n-th root approximation, Computing 57 (1996) 365–369.
[3] F. Dubeau, n-th Root extraction: double iteration process and Newton’s method, J. Comput. Appl. Math. 91 (1998)
191–198.
[4] W. Gander, On Halley’s iteration method, Amer. Math. Monthly 92 (1985) 131–134.
[5] J. Gerlach, Accelerated convergence in Newton’s method, SIAM Rev. 36 (1994) 272–276.
[6] J.M. Gutiérrez, M.A. Hernández, M.A. Salanova, A family of Chebyshev–Halley type methods in Banach spaces, Bull.
Austral. Math. Soc. 55 (1997) 113–130.
[7] J.M. Gutiérrez, M.A. Hernández, M.A. Salanova, Calculus of n-th roots and third order iterative methods, Nonlinear Anal.
47 (2001) 2875–2880.
[8] M.A. Hernández, M.A. Salanova, Index of convexity and concavity. Application to Halley method, Appl. Math. Comput.
103 (1999) 27–49.