Annales Mathematicae Silesianae 33 (2019), 21–40

DOI: 10.2478/amsil-2018-0008

EXTENDING THE APPLICABILITY

OF THE SUPER-HALLEY-LIKE METHOD
USING ω-CONTINUOUS DERIVATIVES
AND RESTRICTED CONVERGENCE DOMAINS

Ioannis K. Argyros, Santhosh George

Abstract. We present a local convergence analysis of the super-Halley-like

method in order to approximate a locally unique solution of an equation in
a Banach space setting. The convergence analysis in earlier studies was based
on hypotheses reaching up to the third derivative of the operator. In the present
study we expand the applicability of the super-Halley-like method by using
hypotheses up to the second derivative. We also provide: a computable error
on the distances involved and a uniqueness result based on Lipschitz constants.
Numerical examples are also presented in this study.

1. Introduction

In this study we consider the super-Halley-like method for approximating

a locally unique solution x∗ of equation

F (x) = 0,

where F is a twice Fréchet differentiable operator defined on a subset Ω of

a Banach space B1 with values in a Banach space B2 . In particular, we study
Received: 09.10.2017. Accepted: 25.08.2018. Published online: 27.10.2018.
(2010) Mathematics Subject Classification: 65D10, 65D99.
Key words and phrases: super-Halley-like method, Banach space, local convergence,
Fréchet derivative.
22 Ioannis K. Argyros, Santhosh George

the local convergence analysis of the super-Halley method defined for each
n = 0, 1, 2, . . . by

yn = xn − F 0 (xn )−1 F (xn ),

1
(1.1) xn+1 = yn + Ln (I − Ln )−1 (yn − xn ),
2

where x0 is an initial point and Ln = F 0 (xn )−1 F 00 (xn )F 0 (xn )−1 F (xn ). The ef-
ficiency and importance of method (1.1) were discussed in [4–9,14]. The study
of convergence of iterative algorithms is usually centered into two categories:
semilocal and local convergence analysis. The semilocal convergence is based
on the information around an initial point, to obtain conditions ensuring the
convergence of these algorithms, while the local convergence is based on the
information around a solution to find estimates of the computed radii of the
convergence balls. Local results are important since they provide the degree
of difficulty in choosing initial points. The local convergence of method (1.1)
was shown using hypotheses given in non-affine invariant form by
(C1 ) F : Ω ⊂ B1 → B2 is a thrice continuously differentiable operator.
(C2 ) There exists x0 ∈ Ω such that F 0 (x0 )−1 ∈ L(B2 , B1 ) and kF −1 (x0 )k ≤ β.
There exist η ≥ 0, β1 ≥ 0, β2 ≥ 0 and β3 ≥ 0 such that
(C3 ) kF 0 (x0 )−1 F 0 (x0 )k ≤ η,
(C4 ) kF 00 (x)k ≤ β1 , for each x ∈ D,
(C5 ) kF 000 (x)k ≤ β2 , for each x ∈ D,
(C6 ) kF 000 (x) − F 000 (y)k ≤ β3 kx − yk for each x, y ∈ D.
The hypotheses for the local convergence analysis of these methods are
the same but x0 is replaced by x∗ . Notice however that hypotheses (C5 ) and
(C6 ) limit the applicability of these methods. As a motivational example, let
us define function f on Ω = [− 21 , 52 ] by

x3 ln x2 + x5 − x4 , x 6= 0,

f (x) =
0, x = 0.

Choose x∗ = 1. We have that

f 0 (x) = 3x2 ln x2 + 5x4 − 4x3 + 2x2 , f 0 (1) = 3,

f 00 (x) = 6x ln x2 + 20x3 − 12x2 + 10x,
f 000 (x) = 6 ln x2 + 60x2 − 24x + 22.

Then, obviously, function f 000 is unbounded on Ω. Hence, the earlier results

using the (C) conditions cannot be used to solve equation f (x) = 0. Notice
that, in-particular there is a plethora of iterative methods for approximating
Extending the applicability of the super-Halley-like method... 23

solutions of nonlinear equations defined on B1 (see [1–16]). These results show

that if the initial point x0 is sufficiently close to the solution x∗ , then the
sequence {xn } converges to x∗ . But how close to the solution x∗ the initial
point x0 should be? These local results give no information on the radius of
the convergence ball for the corresponding method. We address this question
for method (1.1) in Section 2. We show in Remark 2.2 and Remark 3.3 as well
as in the numerical examples that the radius of convergence of method (1.1) is
smaller than the radius of convergence for Newton’s method. However method
(1.1) is faster than Newton’s method. Moreover, the radius of convergence is
larger and the error bounds on the distances kxn − x∗ k are tighter under our
new approach. Furthermore, we show that the equation f (x) = 0 mentioned
above can be solved under our new technique. Finally, our conditions do not
require (C5 ) and (C6 ). Hence, we expand the applicability of method (1.1) in
cases not covered before. The same technique can be used to other methods.
We study these methods in the more general setting of a Banach space
under hypotheses only on the first and second Fréchet derivatives of F and
under the same set of conditions (see conditions (H) that follow). This way
we expand the applicability of these methods.
The rest of the paper is organized as follows: Section 2 contains the local
convergence analysis and Section 3 contains the semi-local convergence anal-
ysis of method (1.1). The numerical examples are presented in the concluding
Section 4.

2. Local convergence analysis

Local convergence analysis of method (1.1) is based on some scalar func-

tions and parameters. Let w0 : [0, +∞) → [0, +∞) be a continuous and in-
creasing function with w0 (0) = 0. Suppose the equation w0 (t) = 1 has at least
one positive root. Denote by r0 the smallest such a root, i.e.

(2.1) r0 = min{t > 0 : w0 (t) = 1}.

Let also w : [0, r0 ) → [0, +∞), v : [0, r0 ) → [0, +∞) and v1 : [0, r0 ) → [0, +∞)
be continuous and increasing functions with w(0) = 0. Moreover, define func-
tions g1 and h1 on the interval [0, r0 ) by
R1
0
w((1 − θ)t)dθ
g1 (t) =
1 − w0 (t)
24 Ioannis K. Argyros, Santhosh George

and

h1 (t) = g1 (t) − 1.

We have h1 (0) = −1 < 0 and h1 (t) → +∞ as t −→ r0− . The intermediate

value theorem guarantees the existence of zeros for function h1 (t) in (0, r0 ).
Denote by r1 the smallest zero. Furthermore, define functions p and hp on the
interval [0, r0 ) by
R1
v1 (t) 0 v(θt)dθt
p(t) =
(1 − w0 (t))2

and

hp (t) = p(t) − 1.

We get that hp (0) = −1 < 0 and hp (t) → +∞ as t −→ r0− . Denote by rp the

smallest zero of function hp in the interval (0, r0 ). Finally, define functions g2
and h2 on the interval [0, rp ) by
R1
1 p(t) 0 v(θt)dθ
g2 (t) = g1 (t) +
2 (1 − p(t))(1 − w0 (t))

and

h2 (t) = g2 (t) − 1.

We obtain h2 (0) = −1 < 0 and h2 (t) → +∞ as t −→ rp− . Denote by r2

the smallest zero of function h2 on the interval (0, rp ). Define the radius of
convergence r by

(2.2) r = min{r1 , r2 }.

Then, for each t ∈ [0, r)

(2.3) 0 ≤ gi (t) < 1

and

(2.4) 0 ≤ p(t) < 1.

Let U (v, ρ), Ū (v, ρ) stand, respectively for the open and closed balls in B1
with center v ∈ B1 and of radius ρ > 0.
Extending the applicability of the super-Halley-like method... 25

Next, we present the local convergence analysis of method (1.1) using the
preceding notation.

Theorem 2.1. Let F : Ω ⊂ B1 → B2 be a twice continuously Fréchet-

differentiable operator. Suppose there exist x∗ ∈ Ω and a continuous and in-
creasing function w0 : [0, +∞) → [0, +∞) with w0 (0) = 0 such that for each
x∈Ω

(2.5) F (x∗ ) = 0, F 0 (x∗ )−1 ∈ L(B2 , B1 ),

and

(2.6) kF 0 (x∗ )−1 (F 0 (x) − F 0 (x∗ ))k ≤ w0 (kx − x∗ k).

Let Ω0 = Ω ∩ U (x∗ , r0 ). Suppose there exist functions w : [0, r0 ) → [0, +∞),

v : [0, r0 ) → [0, +∞) and v1 : [0, r0 ) → [0, +∞) such that for each x, y ∈ Ω0

(2.7) kF 0 (x∗ )−1 (F 0 (x) − F 0 (y))k ≤ w(kx − yk),

(2.8) kF 0 (x∗ )−1 F 0 (x)k ≤ v(kx − x∗ k),

(2.9) kF 0 (x∗ )−1 F 00 (x)k ≤ v1 (kx − x∗ k),

and Ū (x∗ , r) ⊆ Ω, where the radii r0 and r are defined by (2.1) and (2.2),
respectively. Then, sequence {xn } starting from x0 ∈ U (x∗ , r)\{x∗ } and given
by (1.1) is well defined, remains in U (x∗ , r) and converges to x∗ . Moreover,
the following estimates hold

(2.10) kyn − x∗ k ≤ g1 (kxn − x∗ k)kxn − x∗ k

and

(2.11) kxn+1 − x∗ k ≤ g2 (kxn − x∗ k)kxn − x∗ k,

where the functions gi , i = 1, 2, are defined previously. Furthermore, if there

exists r∗ ≥ r such that
Z 1
(2.12) w0 (θr∗ )dθ < 1,
0

then, the limit point x∗ is the only solution of equation F (x) = 0 in Ω1 =

Ω ∩ U (x∗ , r∗ ).
26 Ioannis K. Argyros, Santhosh George

Proof. We shall show estimates (2.10) and (2.11) using induction on the
integer k. Let x ∈ U (x∗ , r). Using (2.1), (2.2) and (2.5), we get that

(2.13) kF 0 (x∗ )−1 (F 0 (x) − F 0 (x∗ ))k ≤ w0 (kx − x∗ k) ≤ w0 (r) < 1.

In view of (2.13) and the Banach lemma on invertible operators ([1, 4, 15, 16])
we have that F 0 (x)−1 ∈ L(B2 , B1 ) and

1
(2.14) kF 0 (x)−1 F 0 (x∗ )k ≤ .
1 − w0 (kx − x∗ k)

In particular, y0 is well defined by the first substep of method (1.1) for n = 0.

We can write

(2.15) y0 − x∗ = x0 − x∗ − F 0 (x0 )−1 F (x0 ).

Using (2.2), (2.3) (for i = 1), (2.7), (2.14) and (2.15) we obtain in turn that

ky0 − x∗ k ≤ kF 0 (x0 )−1 F 0 (x∗ )k

Z 1
× F 0 (x∗ )−1 [F 0 (x∗ + θ(x0 − x∗ )) − F 0 (x0 )](x0 − x∗ )dθ
0
R1
0
w((1 − θ)kx0 − x∗ k)dθkx0 − x∗ k
≤
1 − w0 (kx0 − x∗ k)
(2.16) = g1 (kx0 − x∗ k)kx0 − x∗ k < kx0 − x∗ k < r,

which shows (2.10) for n = 0 and y0 ∈ U (x∗ , r). Next, we must show (I −
L0 )−1 ∈ L(B2 , B1 ). By (2.2), (2.8), (2.9) and (2.14) we have in turn that

kL0 k ≤ kF 0 (x0 )−1 F 0 (x∗ )kkF 0 (x∗ )−1 F 00 (x0 )k

× kF 0 (x0 )−1 F 0 (x0 )kkF 0 (x∗ )−1 F (x0 )k

R1
v1 (kx0 − x∗ k) 0 v(θkx0 − x∗ k)dθkx0 − x∗ k
≤
(1 − w0 (kx0 − x∗ k))2
(2.17) = p(kx0 − x∗ k) ≤ p(r) < 1,

so
1
(2.18) k(I − L0 )−1 k ≤ ,
1 − p(kx0 − x∗ k)
Extending the applicability of the super-Halley-like method... 27

where we also used the estimate

Z 1
kF 0 (x∗ )−1 F (x0 )k = F 0 (x∗ )−1 F 0 (x∗ + θ(x0 − x∗ ))(x0 − x∗ )dθ
0
Z 1
(2.19) ≤ v(θkx0 − x∗ k)dθkx0 − x∗ k.
0

We also have that x1 is well defined by the second substep of method (1.1)
and (2.18). Moreover, by (2.2), (2.3) (for i = 2), (2.14), (2.16)-(2.19), we get
in turn that
1
kx1 − x∗ k = kx0 − x∗ − F 0 (x0 )−1 F (x0 ) − L0 (I − L0 )−1 F 0 (x0 )−1 F (x0 )k
2
≤ kx0 − x∗ − F 0 (x0 )−1 F (x0 )k
1
+ kL0 kk(I − L0 )−1 kkF 0 (x0 )−1 F 0 (x∗ )kkF 0 (x∗ )−1 F (x0 )k
2
≤ g1 (kx0 − x∗ k)kx0 − x∗ k
R1
p(kx0 − x∗ k) 0 v(θkx0 − x∗ k)dθkx0 − x∗ k
+
2(1 − p(kx0 − x∗ k))(1 − w0 (t))
= g2 (kx0 − x∗ k)kx0 − x∗ k < kx0 − x∗ k < r,

which shows (2.11) for n = 0 and x1 ∈ U (x∗ , r). The induction is finished,
if we simply replace x0 , y0 , x1 by xk , yk , xk+1 , respectively in the preceding
estimates. Furthermore, from the estimate

kxk+1 − x∗ k ≤ c0 kxk − x∗ k < r,

where c0 = g2 (kx0 − x∗ k) ∈ [0, 1), we deduce that limk→∞ xk = x∗ and

xk+1 ∈ U (x∗ , r).
R1
Finally, to show the uniqueness part, let Q = 0 F 0 (x∗ + θ(y ∗ − x∗ ))dθ,
where y ∗ ∈ Ω1 with F (y ∗ ) = 0. Using (2.5) and (2.12), we obtain that
Z 1
kF 0 (x∗ )−1 (Q − F 0 (x∗ ))k ≤ w0 (θkx∗ − y ∗ k)dθ
0
Z 1
≤ w0 (θr∗ )dθ < 1,
0

so Q−1 ∈ L(B2 , B1 ). Then, from the identity 0 = F (x∗ ) − F (y ∗ ) = Q(x∗ − y ∗ ),

we conclude that x∗ = y ∗ . 
28 Ioannis K. Argyros, Santhosh George

Remark 2.2.
(a) Let w0 (t) = L0 t, w(t) = Lt and w∗ (t) = L∗ t (w∗ replacing w in (2.7)). In
[3], Argyros and Ren used instead of (2.7) the condition

(2.20) kF 0 (x0 )−1 (F 0 (x) − F 0 (y))k ≤ L∗ kx − yk for each x, y ∈ Ω.

But using (2.7) and (2.20) we get that

L ≤ L∗

holds, since Ω0 ⊆ Ω. In case L < L∗ , the new convergence analysis is

better than the old one. Notice also that we have by (2.6) and (2.20)

L0 ≤ L∗ .

The advantages are obtained under the same computational cost as before,
since in practice the computation of constant L∗ requires the computation
of L0 and L as special cases. In the literature (with the exception of our
works) (2.20) is only used for the computation of the upper bounds of the
inverses of the operators involved.
(b) The radius rA was obtained by Argyros in [4] as the convergence radius
for Newton’s method under conditions (2.3)-(2.7). Notice that the con-
vergence radius for Newton’s method given independently by Rheinboldt
([15]) and Traub ([16]) is given by

2 2
ρ= ∗
< rA = .
3L 2L0 + L

As an example, let us consider the function f (x) = ex − 1. Then x∗ = 0.

1
Set Ω = U (0, 1). Then, we have that L0 = e − 1 < L∗ = e, L = e e−1 , so
ρ = 0.24252961 < rA = 0.3827.
Moreover, the new error bounds ([4–7]) are:

L
kxn+1 − x∗ k ≤ kxn − x∗ k2 ,
1 − L0 kxn − x∗ k

whereas the old ones ([8, 9])

L
kxn+1 − x∗ k ≤ kxn − x∗ k2 .
1 − Lkxn − x∗ k

Clearly, the new error bounds are more precise, if L0 < L. Moreover,
the radius of convergence of method (1.1) given by r is smaller than rA
(see (2.2)).
Extending the applicability of the super-Halley-like method... 29

(c) The local results can be used for projection methods such as Arnoldi’s
method, the generalized minimum residual method (GMREM), the gen-
eralized conjugate method (GCM) for combined Newton/finite projection
methods and in connection to the mesh independence principle in order to
develop the cheapest and most efficient mesh refinement strategy ([4–7]).
(d) The results can be also used to solve equations where the operator F 0
satisfies the autonomous differential equation ([5, 7]):

F 0 (x) = p(F (x)),

where p is a known continuous operator. Since F 0 (x∗ ) = p(F (x∗ )) = p(0),

we can apply the results without actually knowing the solution x∗ . As an
example, let us consider F (x) = ex − 1. Then, we can choose p(x) = x + 1
and x∗ = 0.
(e) It is worth noticing that convergence conditions for method (1.1) are not
changing if we use the new instead of the old conditions. Moreover, for
the error bounds in practice we can use the computational order of con-
vergence (COC)

ln kxkxn+2 −xn+1 k
n+1 −xn k
ξ= kxn+1 −xn k
, for each n = 1, 2, . . .
ln kx n −xn−1 k

or the approximate computational order of convergence (ACOC)

∗
ln kx n+2 −x k
kxn+1 −x∗ k
∗
ξ = ∗ , for each n = 0, 1, 2, . . .
ln kxkxn+1 −x k
∗
n −x k

instead of the error bounds obtained in Theorem 2.1.

(f) In view of (2.6) and the estimate

kF 0 (x∗ )−1 F 0 (x)k = kF 0 (x∗ )−1 (F 0 (x) − F 0 (x∗ )) + Ik

≤ 1 + kF 0 (x∗ )−1 (F 0 (x) − F 0 (x∗ ))k ≤ 1 + L0 kx − x∗ k

condition (2.8) can be dropped and can be replaced by

v(t) = 1 + L0 t
or
v(t) = M = 2,

since t ∈ [0, L10 ).

30 Ioannis K. Argyros, Santhosh George

3. Semi-local convergence analysis

The following conditions (H) in a non affine invariant form have been used
to show the semi-local convergence analysis of method (1.1) ([5, 14]):
(H1 ) kF 0 (x0 )−1 k ≤ β.
(H2 ) kF 0 (x0 )−1 F (x0 )k ≤ η.
(H3 ) kF 00 (x)k ≤ M for each x ∈ Ω.
(H4 ) There exists a continuous and increasing function ϕ : [0, +∞) → [0, +∞)
with ϕ(0) = 0 such that for each x, y ∈ Ω kF 00 (x)−F 00 (y)k ≤ ϕ(kx−yk).
(H5 ) There exists a continuous and increasing function ψ : [0, 1] → [0, +∞)
such that ϕ(ts) ≤ ψ(t)ϕ(s) for each t ∈ [0, 1] and s ∈ [0, +∞).
(H6 ) Ū (x0 , r̄) ⊆ Ω, where r̄ is a positive zero of some scalar equation.
In many applications the iterates {xn } remain in a neighborhood of Ω0 . If
we locate Ω0 before we find M, q and ψ, then the new semi-local convergence
analysis will be weaker. Consequently, the new convergence domain will be
at least as large as if we were using Ω. To achieve this goal we consider the
weaker conditions (A) in an affine invariant form:
(A1 ) kF 0 (x0 )−1 F (x0 )k ≤ η.
(A2 ) There exists a continuous and increasing function w0 : [0, +∞) → [0, +∞)
with w0 (0) = 0 such that for each x ∈ Ω

kF 0 (x0 )−1 (F 0 (x) − F 0 (x0 ))k ≤ w0 (kx − x0 k).

Define r0 the smallest positive solution of equation w0 (t) = 1. Set Ω0 :=

Ω ∩ U (x0 , r0 ).
(A3 ) There exist continuous and increasing functions w, w1 , w2 : [0, r0 ) →
[0, +∞) with w1 (0) = w2 (0) = 0 such that for each x, y ∈ Ω0

kF 0 (x0 )−1 F 00 (x)k ≤ w(kx − x0 k),

kF 0 (x0 )−1 (F 00 (x) − F 00 (x0 ))k ≤ w1 (kx − x0 k)

and

kF 0 (x0 )−1 (F 00 (x) − F 00 (y))k ≤ w2 (kx − yk).

Define functions q, d, c on the interval [0, r0 ) by

w(t)η
q(t) = ,
(1 − w0 (t))2
Extending the applicability of the super-Halley-like method... 31
"R 1
1 0
w2 (θη)(1 − θ)dθ
d(t) =
1 − w0 (t) 1 − q(t)
R1
w(t + θη)(1 − θ)dθq(t)
0
+
1 − q(t)

1 1
Z
w(t)
+ w(1 + θη)dθ
2 0 (1 − w0 (t))2 (1 − q(t))
#
1 1
Z  
1 q(t) w(t)
+ w t+ θ (1 − θ)dθ
2 0 2 1 − q(t) (1 − w0 (t))2 (1 − q(t))

and
1 q(t)
c(t) = (1 + )d(t)t.
2 1 − q(t)
η
(A4 ) Equation 1−c(t) − t = 0 has zeros in the interval (0, r0 ). Denote by r
the smallest such a zero. Moreover, the zero r satisfies q(r) < 1 and
c(r) < 1.
(A5 ) Ū (x0 , r) ⊆ Ω.
R1
(A6 ) There exists r∗ ≥ r such that 0 w0 ((1 − θ)r + θr∗ )dθ < 1.
It is convenient for the semi-local convergence analysis of method (1.1)
that follows to define the scalar sequences {pn }, {qn }, {sn }, {tn } for each
n = 0, 1, 2, . . . by

w(kxn − x0 k)kF 0 (x0 )−1 F (xn )k

(3.1) pn = ,
(1 − w0 (kxn − x0 k))2

t0 = 0, s0 = η,

w(tn − t0 )(sn − tn )
(3.2) qn = ,
(1 − w0 (tn − t0 ))2
1 qn
(3.3) tn+1 = sn + (sn − tn ),
2 1 − qn
"R 1
1 0
w̄(θ(sn − tn ))(1 − θ)dθ(sn − tn )2
sn+1 = tn+1 +
1 − w0 (tn+1 − t0 ) 1 − qn
R1
w(tn − t0 + θ(sn − tn ))(1 − θ)dθqn (sn − tn )2
(3.4) + 0
1 − qn
32 Ioannis K. Argyros, Santhosh George

Z 1
1 qn
+ w(tn − t0 + θ(sn − tn ))dθ
(sn − tn )
2 0 1 − qn
#
1 1
 
qn (sn − tn )
Z
qn
+ w sn − t0 t + θ (1 − θ)dθ (sn − tn ) ,
2 0 2(1 − qn ) 1 − qn

where

w1 , n = 0,
w̄ =
w2 , n > 0.

We need an Ostrowski-type representation of method (1.1).

Assumption 3.1. Suppose that method (1.1) is well defined for each n =
0, 1, 2, . . . Then the following equality holds
Z 1
F (xn+1 ) = [F 00 (xn + θ(yn − xn )) − F 00 (xn )](1 − θ)(I − Ln )−1 (yn − xn )2 dθ
0
Z 1
− F 00 (xn + θ(yn − xn ))(1 − Ln )−1 Ln (xn )(1 − θ)dθ(yn − xn )2
0
Z 1
+ F 00 (xn + θ(yn − xn ))dθ(yn − xn )(xn+1 − yn )
0
Z 1
(3.5) + F 00 (yn + θ(xn+1 − yn ))(1 − θ)dθ(xn+1 − yn )2 .
0

Proof. Using method (1.1) and the Taylor series expansion about yn ∈ Ω,
we can write
Z xn+1
0
F (xn+1 ) = F (yn ) + F (yn )(xn+1 − yn ) + F 00 (x)(xn+1 − x)dx,
yn

which leads to (3.5). 

Next, we present the semi-local convergence analysis of method (1.1) using

the preceding notation and conditions (A).

Theorem 3.2. Let F : Ω ⊂ B1 → B2 be a twice continuously Fréchet-

differentiable operator. Suppose that the conditions (A) are satisfied. Then,
sequence {xn } generated for x0 ∈ Ω by method (1.1) is well defined in U (x0 , r),
remains in U (x0 , r) for each n = 0, 1, 2, . . . and converges to solution x∗ ∈
Ū (x0 , r) of equation F (x) = 0. Moreover, the following estimates hold

kxn − x∗ k ≤ r − tn .
Extending the applicability of the super-Halley-like method... 33

Furthermore, the limit point x∗ is the only solution of equation F (x) = 0 in

Ω1 = Ω ∩ Ū (x0 , r).

Proof. Let x ∈ U (x0 , r0 ), where r0 is defined in condition (A2 ) from

which we have that

(3.6) kF 0 (x0 )−1 (F 0 (x) − F 0 (x0 ))k ≤ w0 (kx − x0 k) ≤ w0 (r) < 1.

It follows from (3.6) and the Banach lemma on invertible operators that
F 0 (x)−1 ∈ L(B2 , B1 ) and

1
(3.7) kF 0 (x)−1 F 0 (x0 )k ≤ .
1 − w0 (kx − x0 k)

We also have that y0 is well defined by the first substep of method (1.1) and
L0 exists for n = 0. In view of the definition of L0 , (A1 ), (A3 ), (A4 ), (3.1),
(3.2) and (3.7), we get that

kL0 k = kF 0 (x0 )−1 F 00 (x0 )F 0 (x0 )−1 F (x0 )k

≤ kF 0 (x0 )−1 F 00 (x0 )kkF 0 (x0 )−1 F (x0 )k

≤ w(kx0 − x0 k)kF 0 (x0 )−1 F (x0 )k

w(kx0 − x0 k)kF 0 (x0 )−1 F (x0 )k

(3.8) ≤ = p0 ≤ q0 ≤ q(r) < 1,
(1 − w0 (kx0 − x0 k))2

so (I − L0 )−1 ∈ L(B2 , B1 ) and

1
(3.9) k(I − L0 )−1 k ≤ .
1 − q0

The point x1 is also well defined by the second substep of method (1.1) for
n = 0 and

kx1 − x0 k ≤ kx1 − y0 k + ky0 − x0 k

1
≤ kL0 kk(I − L0 )−1 kky0 − x0 k + ky0 − x0 k
2
 
1 q0
≤ 1+ ky0 − x0 k
2 1 − q0
 
1 q0
≤ 1+ η < r,
2 1 − q0
34 Ioannis K. Argyros, Santhosh George

so x1 ∈ U (x0 , r), ky0 − x0 k ≤ s0 − t0 and kx1 − y0 k ≤ t1 − s0 . Suppose

(3.10) kyk − xk k ≤ sk − tk

and

(3.11) kxk+1 − yk k ≤ tk+1 − sk .

Using (3.5), (A2 ), (3.1)–(3.4), (3.7)–(3.9), we get in turn that

kF 0 (x0 )−1 F (xk+1 )k

R1
w̄(kxk + θ(yk − xk ) − xk k)(1 − θ)dθkyk − xk k2
≤ 0
1 − pk
R1
w(kxk + θ(yk − xk ) − x0 k)(1 − θ)dθpk kyk − xk k2
+ 0
1 − pk
Z 1
+ w(kxk + θ(yk − xk ) − x0 k)dθkxk+1 − yk k
0
Z 1
+ w(kyk + θ(xk+1 − xk ) − x0 k)(1 − θ)dθkxk+1 − yk k
0
R1
0
w̄(θ(sk − tk ))(1 − θ)dθ(sk − tk )2
≤
1 − pk
R1
w((tk − t0 + θ(sk − tk ))(1 − θ)dθqk (sk − tk )2
+ 0
1 − qk
Z 1
+ w(tk − t0 + θ(sk − tk ))dθ(tk+1 − sk )
0
Z 1
+ w((sk − t0 + θ(tk+1 − sk ) − x0 k)(1 − θ)dθ(tk+1 − sk )
0

(3.12) := αk+1 ,

so

kyk+2 − xk+2 k ≤ kF 0 (xk+1 )−1 F 0 (x0 )kkF 0 (x0 )−1 F (xk+1 )k

αk+1
≤ = sk+2 − tk+2
1 − w0 (kxk+1 − x0 k)
Extending the applicability of the super-Halley-like method... 35

and
1
kxk+2 − yk+1 k ≤ kLk+1 kk(I − Lk+1 )−1 kkyk+1 − xk+1 k
2
1 qk+1
≤ (sk+1 − tk+1 ) = tk+2 − sk+1 ,
2 1 − qk+1

and the induction for (3.10) and (3.11) is completed. Moreover, we have that

kxk+2 − xk+1 k ≤ kxk+2 − yk+1 k + kyk+1 − xk+1 k

1 qk+1
≤ (sk+1 − tk+1 ) + (sk+1 − tk+1 )
2 1 − qk+1
 
1 qk+1
= 1+ (sk+1 − tk+1 )
2 1 − qk+1
 
1 q(r)
≤ 1+ d(r)r(sk − tk )
2 1 − q(r)

= c(sk − tk ) ≤ . . . ≤ ck+1 (s0 − t0 )

= ck+1 η, c = c(r),

so

kxk+2 − x0 k ≤ kxk+2 − xk+1 k + . . . + kx1 − x0 k

1 − ck+2 η
(3.13) ≤ ck+1 η + . . . + η = η< = r,
1−c 1−c

hence xk+2 ∈ U (x0 , r) and

kyk+1 − x0 k ≤ kyk+1 − xk+1 k + kxk+1 − x0 k,

(3.14) ≤ sk+1 − tk+1 + tk+1 − t0 = sk+1 ≤ r.

Scalar sequences {tk }, {sk } are increasing and bounded from above by r (see
(3.10), (3.11), (3.13) and (3.14)), so they converge to their unique least upper
bound r1 ≤ r. In view of (3.10) and (3.11) sequences {xk }, {yk } are Cauchy in
a complete space B1 and as such they converge to some x∗ ∈ Ū (x0 , r) (since
Ū (x0 , r) is a closed set). By letting k −→ ∞ in (3.12) we get F (x∗ ) = 0.
36 Ioannis K. Argyros, Santhosh George

To prove the uniqueness part, suppose Rthat there exists y ∗ ∈ Ω1 with

1
F (y ∗ ) = 0 and define linear operator T = 0 F 0 (x∗ + θ(y ∗ − x∗ ))dθ. Using
(A2 ) and (A3 ), we obtain in turn that
Z 1
0 −1 0
kF (x0 ) (T − F (x0 ))k ≤ w0 (kx∗ + θ(y ∗ − x∗ ) − x0 k)dθ
0
Z 1
≤ w0 ((1 − θ)kx∗ − x0 k + θky ∗ − x0 k)dθ
0
Z 1
≤ w0 ((1 − θ)r + θr∗ )dθ < 1,
0

so T −1 ∈ L(B2 , B1 ). Then, from the identity

0 = F (y ∗ ) − F (x∗ ) = T (y ∗ − x∗ ),

we conclude that x∗ = y ∗ . 

Remark 3.3.
(a) In view of the estimate

kF 0 (x0 )−1 F 00 (x)k = kF 0 (x0 )−1 [F 00 (x) − F 00 (x0 ) + F 00 (x0 )]k

≤ kF 0 (x0 )−1 (F 00 (x) − F 00 (x0 ))k + kF 0 (x0 )−1 F 00 (x0 )k

≤ w1 (kx − x0 k) + kF 0 (x0 )−1 F 00 (x0 )k,

the first condition in (A3 ) can be dropped, if we choose

w(t) = w1 (t) + kF 0 (x0 )−1 F 00 (x0 )k, f or each t ∈ [0, r0 ).

Moreover, the third condition in (A3 ) implies the second condition but

(3.15) w1 (t) ≤ w2 (t), for each t ∈ [0, r0 ).

Hence, it is important to introduce the second condition, in case strict

inequality holds in (3.15).
(b) Clearly conditions (A) are weaker than conditions (H) even, if the former
are specialized. Indeed, let us rewrite conditions (H3 )–(H5 ):
(H30 ) kF 0 (x0 )−1 F 00 (x)k ≤ M1 for each x ∈ Ω.
Extending the applicability of the super-Halley-like method... 37

(H40 ) kF 0 (x0 )−1 (F 00 (x) − F 00 (y))k ≤ ϕ1 (kx − yk) for each x, y ∈ Ω. Choose
w(t) = M0 . Then, we have that

M0 ≤ M1

and

w2 (t) ≤ ϕ1 (t),

since Ω0 ⊆ Ω.
(H50 ) ϕ1 (ts) ≤ ψ1 (t)ϕ1 (s), for each t ∈ [0, 1] and s ∈ [0, +∞).
Then, there exists a continuous and increasing function ψ0 : [0, 1] →
[0, +∞) such that

w2 (ts) ≤ ψ0 (t)w2 (s), for each t ∈ [0, 1] and s ∈ [0, +∞)

and

ψ0 (t) ≤ ψ(t)

leading to a tight convergence analysis (see also the numerical ex-

amples).

4. Numerical Examples

The numerical examples are presented in this section.

Example 4.1. Returning back to the motivational example at the in-

troduction of this study, we have w0 (t) = w(t) = v(t) = 146.6629073t and
v1 (t) = 2. The parameters for method (1.1) are

r1 = 0.6667, r2 = 0.4384 = r.

As already noted in the introduction earlier results using (C) conditions cannot
be used.
38 Ioannis K. Argyros, Santhosh George

Example 4.2. Let B1 = B2 = R3 , Ω = Ū (0, 1), x∗ = (0, 0, 0)T . Define

function F on Ω for w = (x, y, z)T by

e−1 2
F (w) = (ex − 1, y + y, z)T .
2
The Fréchet-derivative is defined by
 x 
e 0 0
F 0 (v) =  0 (e − 1)y + 1 0  .
0 0 1

Notice that using the conditions (2.5)–(2.9), we get w0 (t) = (e − 1)t, w(t) =
1 1
e e−1 t, v1 (t) = v(t) = e e−1 . The parameters for method (1.1) are

r1 = 0.3827, r2 = 0.1262 = r.

Under the old approach ([2, 3, 5–8, 12, 15, 16]) we have w̃0 (t) = w̃(t) = et,
ṽ1 (t) = ṽ(t) = e. The old parameters for method (1.1) are

r̃1 = 0.2453, r̃2 = 0.0540 = r̃.

Here g̃1 , g̃2 are g1 , g2 with w̃0 (t) = w̃(t) = et, ṽ1 (t) = ṽ(t) = e. Table 1
gives the comparison of g1 (kxn − x∗ k), g2 (kxn − x∗ k), g̃1 (kxn − x∗ k) and
g̃2 (kxn − x∗ k).
Table 1. Comparison table
n g1 (kxn − x k) g2 (kxn − x∗ k) g̃1 (kxn − x∗ k) g̃2 (kxn − x∗ k)
∗

2 0.0032 0.0138 0.0049 0.0213

3 9.4542e-08 3.9732e-07 1.4361e-07 6.0351e-07
4 4.7256e-17 1.9860e-16 7.1779e-17 3.0166e-16

Example 4.3. Let B1 = B2 = C[0, 1], the space of continuous functions

defined on [0, 1] equipped with the max norm. Let Ω = U (0, 1). Define function
F on Ω by
Z 1
F (ϕ)(x) = ϕ(x) − 5 xθϕ(θ)3 dθ.
0

We have that
Z 1
F 0 (ϕ(ξ))(x) = ξ(x) − 15 xθϕ(θ)2 ξ(θ)dθ, for each ξ ∈ Ω.
0
Extending the applicability of the super-Halley-like method... 39

Then, we get that x∗ = 0, w0 (t) = 7.5t, w(t) = 15t, v(t) = 15, v1 (t) = 30.
The parameters for method (1.1) are

r1 = 0.0667, r2 = 0.0021 = r.

Under the old approach w̃0 (t) = w̃(t) = 15t, ṽ(t) = 15, ṽ1 (t) = 30. The old
parameters for method (1.1) are

r̃1 = 0.0444, r̃2 = 0.0003 = r̃.

References

[1] Argyros I.K., On the Newton–Kantorovich hypothesis for solving equations, J. Comput.
Appl. Math. 169 (2004), 315–332.
[2] Argyros I.K., Ezquerro J.A., Gutiérrez J.M., Hernández M.A., Hilout S., On the semilo-
cal convergence of efficient Chebyshev-Secant-type methods, J. Comput. Appl. Math.
235 (2011), 3195–3206.
[3] Argyros I.K., Ren H., Efficient Steffensen-type algorithms for solving nonlinear equa-
tions, Int. J. Comput. Math. 90 (2013), 691–704.
[4] Argyros I.K., Computational Theory of Iterative Methods, Studies in Computational
Mathematics, 15, Elsevier B.V., New York, 2007.
[5] Ezquerro J.A., Hernández M.A., An optimization of Chebyshev’s method, J. Complex-
ity 25 (2009), 343–361.
[6] Ezquerro J.A., Grau A., Grau-Sánchez M., Hernández M.A., Construction of deri-
vative-free iterative methods from Chebyshev’s method, Anal. Appl. (Singap.) 11
(2013), 1350009, 16 pp.
[7] Ezquerro J.A., Gutiérrez J.M., Hernández M.A., Salanova M.A., Chebyshev-like meth-
ods and quadratic equations, Rev. Anal. Numér. Théor. Approx. 28 (1999), 23–35.
[8] Grau M., Díaz-Barrero J.L., An improvement of the Euler–Chebyshev iterative method,
J. Math. Anal. Appl. 315 (2006), 1–7.
[9] Grau-Sánchez M., Gutiérrez J.M., Some variants of the Chebyshev–Halley family of
methods with fifth order of convergence, Int. J. Comput. Math. 87 (2010), 818–833.
[10] Hueso J.L., Martinez E., Teruel C., Convergence, efficiency and dynamics of new
fourth and sixth order families of iterative methods for nonlinear systems, J. Comput.
Appl. Math. 275 (2015), 412–420.
[11] Magreñán Á.A., Estudio de la dinámica del método de Newton amortiguado, PhD
Thesis, Universidad de La Rioja, Servicio de Publicaciones, Logroño, 2013. Available
at http://dialnet.unirioja.es/servlet/tesis?codigo=38821
[12] Magreñán Á.A., Different anomalies in a Jarratt family of iterative root-finding meth-
ods, Appl. Math. Comput. 233 (2014), 29–38.
[13] Magreñán Á.A., A new tool to study real dynamics: the convergence plane, Appl. Math.
Comput. 248 (2014), 215–224.
[14] Prashanth M., Mosta S.S., Gupta D.K., Semi-local convergence of the Supper-Halley’s
method under w-continuous second derivative in Banach space. Submitted.
[15] Rheinboldt W.C., An adaptive continuation process for solving systems of nonlinear
equations, in: Tikhonov A.N., et al. (eds.), Mathematical Models and Numerical Meth-
ods, Banach Center Publ., 3, PWN, Warsaw, 1978, pp. 129–142.
40 Ioannis K. Argyros, Santhosh George

[16] Traub J.F., Iterative Methods for the Solution of Equations, Prentice-Hall Series in
Automatic Computation, Prentice Hall, Inc., Englewood Cliffs, New Jersey, 1964.

Department of Mathematical Sciences

Cameron University
Lawton
OK 73505
USA
e-mail: [email protected]

Department of Mathematical and Computational Sciences

National Institute of Technology Karnataka
575 025 Karnataka
India
e-mail: [email protected]