Comment

Comment on: On the Kung-Traub Conjecture for

Iterative Methods for Solving Quadratic Equations.
Algorithms 2016, 9, 1
Fayyaz Ahmad 1,2
1 Dipartimento di Scienza e Alta Tecnologia, Universita dell’Insubria, Via Valleggio 11, Como 22100, Italy
2 Departament de Física i Enginyeria Nuclear, Universitat Politècnica de Catalunya, Comte d’Urgell 187,
08036 Barcelona, Spain; [email protected]; Tel.: +34-632-066-627

Academic Editors: Alicia Cordero and Juan R. Torregrosa

Received: 21 January 2016; Accepted: 13 April 2016; Published: 26 April 2016

Abstract: Kung-Traub conjecture states that an iterative method without memory for finding the
simple zero of a scalar equation could achieve convergence order 2d−1 , and d is the total number
of function evaluations. In an article “Babajee, D.K.R. On the Kung-Traub Conjecture for Iterative
Methods for Solving Quadratic Equations, Algorithms 2016, 9, 1, doi:10.3390/a9010001”, the author
has shown that Kung-Traub conjecture is not valid for the quadratic equation and proposed an
iterative method for the scalar and vector quadratic equations. In this comment, we have shown
that we first reported the aforementioned iterative method.

Keywords: Kung-Traub conjecture; System of quadratic equations; Iterative methods

1. Iterative Methods for Solving Quadratic Equations Presented in [1]

According to Kung-Traub conjecture (KTC) [2], an iterative method without memory for solving
nonlinear equations in the case of simple zeros, could achieve a maximum convergence order of
2d−1 , where d is the number of function evaluations. Recently, an article was published for solving
quadratic equations [1] with arbitrary order of convergence by using one function and two derivatives
evaluations. The details of the proposed formulation in [1] is given as follows. Let f ( x ) = κ2 x2 +
κ1 x + κ0 be a quadratic function where κ2 6= 0 and κ1 , κ0 are constants. The proposed iteration
function ψ(r+2)th BQIM ( x ) in [1] is

 f (x)

 u( x ) = f 0 (x)
f 0 ( x −2/3u( x ))


τ=


 f 0 (x)
r (1)

 H (τ, r ) = 1 + ∑ ai (τ − 1)i
i =1




 ψ(r+2)th BQIM ( x ) = x − u( x ) H (τ, r )

   r +2
The error equation of Equation (1) is given as ψ(r+2)th BQIM ( x ) − x ∗ = Cr c2r+1 e(k) +
  r +3 
O e(k) where Cr is asymptotic error constant, x ∗ is the simple root of quadratic equation

Algorithms 2016, 9, 30; doi:10.3390/a9020030 www.mdpi.com/journal/algorithms

Algorithms 2016, 9, 30 2 of 11

and c2 = 1/2 f 0 ( x ∗ )−1 f 00 ( x ∗ ). The error equation clearly shows that KTC is not valid in the case of
quadratic equations. By using binomial expansion, the weight function H (τ, r ) can be written as

r i  
i
H (τ, r ) = 1 + ∑ ∑ ai (−1)(i− j) τ j
i =1 j =0
i−j
(2)
r
H (τ, r ) = 1 + ∑ bi τ i
i =0

where bi is constant and can be computed by comparing two expressions of H (τ, r ) in Equation (2).
The powers of τ can be computed recursively and hence the iteration function Equation (1) is
written as
f (x)



 u( x ) = f 0 (x)

 y = x − 2/3 u( x )



f 0 (y)


τ = f 0 (x)






 φ0 = u( x )

(3)

 for i = 1, r





 φi = τ φi−1

end





 r
 ψ(r+2)th BQIM ( x ) = x − φ0 − ∑ bi φi .


i =0

The computationally efficient vector version of iteration function Equation (3) is



 F0 (x) φ 0 = F(x)


y = x − 2/3 φ 0





for i = 1, r



(4)

 F 0 ( x ) φ i = F 0 ( y ) φ i −1

end





 r
 ψ (r+2)th BQIM (x) = x − φ 0 − ∑ bi φ i .


i =0

2. Iterative Methods for Solving Matrix-Vector Quadratic Equations Presented in [3]

In this direction, a manuscript [3] was posted on 4 May 2015 on Researchgate in which the author
provided models of three iterative methods, with their respective convergence orders, for computing
the solution of matrix-vector quadratic equations. As the proposed iterative methods are valid for
Algorithms 2016, 9, 30 3 of 11

solving systems of nonlinear equations with quadratic nonlinearity, they are also valid for scalar
quadratic equations. In the article [3] the model of iterative Method II can be written as:



 x0 = initial guess
F 0 ( x0 ) φ 1 = F ( x0 )










 x1 = x0 + α1,1 φ 1

for i = 1, m


F 0 ( x 0 ) φ i +1 = F 0 ( x 1 ) φ i (5)






 end

 m
x2 = x0 + ∑ α2,j φ j





 j =1

 x2 = x0

where F(x) = 0 is a quadratic equation i.e., F00 (x) = B is a bilinear form and F(s) (x) = O (zero tensor)
for s ≥ 3 and the convergence order is m + 2. The scalar version of iterative method Equation (5) can
be written as



 x0 = initial guess
f 0 ( x0 ) φ1 = f ( x0 )










 x1 = x0 + α1,1 φ1

for i = 1, m


f 0 ( x0 ) φi+1 = f 0 ( x1 ) φi (6)






 end

 m
x2 = x0 + ∑ α2,j φj






 j =1

x = x
0 2

where f ( x ) = a x2 + b x + c with a 6= 0 and convergence order is m + 2. The convergence proofs of

different iterative methods are established in Figures 1–4. We can see that the error equations in all
cases are the same. The Figure 4 shows that the iterative method (1) is a particular case of iterative
method Equation (6) for α1,1 = −2/3. Finally we provide the convergence proof of the iterative
method Equation (5).

Theorem 2.1. Let F : D ⊆ Rn −→ Rn be a function with all continuous Fréchet derivatives, F00 (x) = B
and F( j) (x) = O for j ≥ 3, where B is a bilinear form and D is convex open subset of Rn . If we take x0 in the
vicinity of a simple root x∗ of F(x) = 0 then the sequence of successive approximation generated by iterative
method Equation (5) for m = 6 and α1,1 = −2/3 converges to x∗ with convergence order eight.

Proof. We denote ek = xk − x∗ , C1 = F0 (x∗ ) and C2 = F0 (x∗ )−1 F00 (x∗ )/2. By expanding F0 (x0 )
around x∗ we get
 
F(x0 ) = C1 e0 + C2 e20 (7)

The Fréchet derivative of Equation (7) with respective to e0 is

F 0 ( x0 ) = C1 ( I + 2 C2 e0 ) (8)
Algorithms 2016, 9, 30 4 of 11

The inverse of Equation (8) is


F0 (x0 )−1 = I − 2 C2 e0 + 4 C22 e20 − 8 C32 e30 + 16 C42 e40 − 32 C52 e50 + 64 C62 e60 − 128 C72 e70
   (9)
+ 256 C82 e80 C1−1 + O e90

We compute φ 1 using Equations (9) and (7)

 
φ 1 = e0 − C2 e20 + 2 C22 e30 − 4 C32 e40 + 8 C42 e50 − 16 C52 e60 + 32 C62 e70 − 64 C72 e80 + O e90 . (10)

The expression for e1 is

 
e1 = 1/3 e0 + 2 C2 e20 − 4 C22 e30 + 8 C32 e40 − 16 C42 e50 + 32 C52 e60 + 64 C62 e70 + 128 C72 e80
  (11)
+ O e90

First order Fréchet derivative of F(x) at x1 is

 
F0 (x1 ) = C1 I + 1/3 2 C2 e0 + 4 C22 e20 − 8 C32 e30 + 16 C42 e40 − 32 C52 e50 + 64 C62 e60
   (12)
− 128 C72 e70 + 256 C82 e80 + O e90

Next we compute M = F0 (x0 )−1 F0 (x1 )

M = I − 4/3 C2 e0 + 4 C22 e20 − 32/3 C32 e30 + 80/3 C42 e40 − 64 C52 e50 + 448/3 C62 e60
  (13)
− 1024/3 C72 e70 + 768 C82 e80 + O e90

with help of Equation (13), we obtain the expressions for φ i for i ∈ {2, 3, 4, 5, 6, 7}

φ 2 = e0 − 7/3 C2 e20 + 22/3 C22 e30 − 64/3 C32 e40 + 176/3 C42 e50 − 464/3 C52 e60 + 1184/3 C62 e70
 
− 2944/3 C72 e80 + O e90
φ 3 = e0 − 11/3 C2 e20 + 130/9 C22 e30 − 460/9 C32 e40 + 168 C42 e50 − 1568/3 C52 e60
 
+ 4672/3 C62 e70 − 4480 C72 e80 + O e90
φ 4 = e0 − 5 C2 e20 + 70/3 C22 e30 − 2584/27 C32 e40 + 9712/27 C42 e50 − 34208/27 C52 e60
 
+ 114496/27 C62 e70 − 367616/27 C72 e80 + O e90
(14)
φ 5 = e0 − 19/3 C2 e20 + 34 C22 e30 − 4252/27 C32 e40 + 53512/81 C42 e50 − 208624/81 C52 e60
 
+ 767968/81 C62 e70 − 2697920/81 C72 e80 + O e90
φ 6 = e0 − 23/3 C2 e20 + 418/9 C22 e30 − 6448/27 C32 e40 + 89168/81 C42 e50
 
− 1137712/243 C52 e60 + 4543840/243 C62 e70 − 5747840/81 C72 e80 + O e90
φ 7 = e0 − 9 C2 e20 + 182/3 C22 e30 − 9236/27 C32 e40 + 46264/27 C42 e50 − 637376/81 C52 e60
 
+ 24618880/729 C62 e70 − 100011520/729 C72 e80 + O e90 .
Algorithms 2016, 9, 30 5 of 11

We obtain expression for e2 by using Equation (14)

 
e2 = 1 + α2,1 + α2,2 + α2,3 + α2,4 + α2,5 + α2,6 + α2,7 e0
 
+ − α2,1 − 7/3 α2,2 − 11/3 α2,3 − 5 α2,4 − 19/3 α2,5 − 23/3 α2,6 − 9 α2,7 C2 e20

+ 2 α2,1 + 22/3 α2,2 + 130/9 α2,3 + 70/3 α2,4 + 34 α2,5 C22 + 418/9 α2,6
 
+ 182/3 α2,7 C22 e30 + − 4 α2,1 − 64/3 α2,2 − 460/9 α2,3 − 2584/27 α2,4 − 4252/27 α2,5
 
− 6448/27 α2,6 − 9236/27 α2,7 C32 e40 + 8 α2,1 + 176/3 α2,2 + 168 α2,3 + 9712/27 α2,4
(15)
 
+ 53512/81 α2,5 + 89168/81 α2,6 + 46264/27 α2,7 C42 e50 + − 16 α2,1 − 464/3 α2,2
− 1568/3 α2,3 − 34208/27 α2,4 − 208624/81 α2,5 − 1137712/243 α2,6
 
− 637376/81 α2,7 C52 e60 + 32 α2,1 + 1184/3 α2,2 + 4672/3 α2,3 + 114496/27 α2,4
 
+ 767968/81 α2,5 + 4543840/243 α2,6 + 24618880/729 α2,7 C62 e70 + − 64 α2,1
− 2944/3 α2,2 − 4480 α2,3 − 367616/27 α2,4 − 2697920/81 α2,5 − 5747840/81 α2,6
  
− 100011520/729 α2,7 C72 e80 + O e90 .

By equating the coefficients of powers of e0 in Equation (15), we get system of seven equations

eq1 :=1 + α2,1 + α2,2 + α2,3 + α2,4 + α2,5 + α2,6 + α2,7 = 0

eq2 := − α2,1 − 7/3 α2,2 − 11/3 α2,3 − 5 α2,4 − 19/3 α2,5 − 23/3 α2,6 − 9 α2,7 = 0
eq3 :=2 α2,1 + 22/3 α2,2 + 130/9 α2,3 + 70/3 α2,4 + 34 α2,5 + 418/9 α2,6 + 182/3 α2,7 = 0
eq4 := − 4 α2,1 − 64/3 α2,2 − 460/9 α2,3 − 2584/27 α2,4 − 4252/27 α2,5 − 6448/27 α2,6
− 9236/27 α2,7 = 0
eq5 :=8 α2,1 + 176/3 α2,2 + 168 α2,3 + 9712/27 α2,4 + 53512/81 α2,5 + 89168/81 α2,6 (16)
+ 46264/27 α2,7 = 0
eq6 := − 16 α2,1 − 464/3 α2,2 − 1568/3 α2,3 − 34208/27 α2,4 − 208624/81 α2,5
− 1137712/243 α2,6 − 637376/81 α2,7 = 0
eq7 :=32 α2,1 + 1184/3 α2,2 + 4672/3 α2,3 + 114496/27 α2,4 + 767968/81 α2,5
+ 4543840/243 α2,6 + 24618880/729 α2,7 = 0 .

The solution set of seven equations is

(
43903 55767 497763 37719 416421
sol = α2,1 = − , α2,2 = , α2,3 = − , α2,4 = , α2,5 = − ,
1024 256 1024 64 1024
) (17)
38637 24057
α2,6 = , α2,7 = − .
256 1024

After simplification the error equation we get

 
e2 = 429 C72 e80 + O e90 . (18)
Algorithms 2016, 9, 30 6 of 11

(1)

(2)

(3)

6 (4)

(5)

Figure 1. Iterative method Equation (1).

Algorithms 2016, 9, 30 7 of 11

(1)

(2)

Figure 2. Iterative method Equation (6) for arbitrary finite α1,1 .

Algorithms 2016, 9, 30 8 of 11

(1)

(2)

Figure 3. Iterative method (6) for α1,1 = −1.

(1)

(2)

Figure 4. Iterative method (6) for α1,1 = −2/3.

Algorithms 2016, 9, 30 9 of 11

As we have seen, the convergence order of the iterative method Equation (5) is eight for m = 6
and this confirms our claimed order of convergence which is s + 2 for m = s. Now we provide the
proof of convergence order via mathematical induction.
Theorem 2.2. Let F : D ⊆ Rn −→ Rn be a function with all continuous Fréchet derivatives, F00 (x) = B
and F( j) (x) = O for j ≥ 3, where B is a bilinear form and D is a convex open subset of Rn . If we take x0
in the vicinity of a simple root x∗ of F(x) = 0 then the sequence of successive approximation generated by the
iterative method Equation (5) for m = s and α1,1 = −2/3 converges to x∗ with convergence order s + 2.
Proof. We suppose that our claim about the convergence order of the iterative method Equation (5)
is true for m = s, which means we have
 
(s)
e2 = Hs C2s−1 e0s+2 + O e0s+3 (19)

where Hs is asymptotic error constant and superscript “(s)” means the value of e2 when m = s. We
(s)
can write e2
 
(s)
e2 = e0 − 2/3 φ 1 + (α2,1 + α2,2 M + · · · + α2,s−1 Ms ) φ 1 + Hs C2s−1 e0s+2 + O e0s+3 (20)

It is convenient to express the combination

α2,1 + α2,2 M + α2,3 M2 + · · · + α2,s+1 Ms

in the powers of I − M [1], we establish the following identity

α2,1 + α2,2 M + · · · + α2,s−1 Ms ≡ β 2,1 + β 2,2 (I − M) + · · · + β 2,s−1 (I − M)s (21)

By comparing the same powers of M on both sides we can easily compute the value of β i ’s. By
using Equation (21), error Equation (20) can be written as

(s) ( s +1) ( s +2)

e2 =e0 − 2/3 φ 1 + β 2,1 + β 2,2 (I − M) + · · · + β 2,s−1 (I − M)s φ 1 + Hs C2


e0

( s +3)
 (22)
+ O e0

However, according to our assumption we can find the value of unknowns to make the following
expression equal to zero

e0 − 2/3 φ 1 + β 2,1 + β 2,2 (I − M) + · · · + β 2,s−1 (I − M)s φ 1 = 0

(23)

3 s
 
( s +1)
We can notice from Equation (13) that (I − M)s = C2s e0s + O e0


4 and φ 1 = O(e0 ).
Now we consider
( s +1) (s)
e2 =e2 + β 2,s (I − M)s+1 φ 1
  s +1
s +1 s +2 3 ( s +2)

( s +3)

= Hs C2 e0 + β 2,s C2s+1 e0 + O e0
4 (24)
  s +1 !
3 ( s +2)

( s +3)

= Hs + β 2,s C2s+1 e0 + O e0
4

( s +2)
As we know the value of Hs we can find β 2,s to make the coefficient of e0 equals to zero.
Hence we get  
( s +1) ( s +3)
e2 = O e0

which completes the proof.

Algorithms 2016, 9, 30 10 of 11

3. Numerical Testing
We adopt the following definition of computational convergence order (COC)


log ||F(xk+1 )||∞ /||F(xk )||∞
COC =
(25)
log ||F(xk )||∞ /||F(xk−1 )||∞

To verify the claimed convergence order of our proposed iterative method Equation (5), we study
the following system of quadratic equations



 x2 x3 + x4 ( x2 + x3 ) = 0


 x1 x3 + x4 ( x4 + x3 ) = 0
F(x) = (26)


 x1 x2 + x4 ( x1 + x4 ) = 0

x1 x2 + x3 ( x1 + x2 ) = 1



In Table 1, we listed the norm of the residue of F(x) and COC against the sequence of iterations
for different values of parameters α11 . The Table 1 confirms the claimed convergence order. For two
different values of the parameter α11 , we obtained the record of the norm of the residue of F(x) equal
to the system of quadratic Equations (26). The possible reason for his could be the same error equation
of iterative method for different values of parameter α11 in the iterative method Equation (5).

Table 1. Computational convergence order of iterative method Equation (5), initial guess =
[0.5, 0.5, 0.5, −0.25].

iter ||F(x)||∞ COC ||F(x)||∞ COC

α11 = −1 α11 = −2/3

0 2.50e1 2.50e1
1 2.71e–6 2.71e–6
2 7.56e–47 8.17 7.56e–47 8.17
3 2.80e–371 8.00 2.80e–371 8.00
4 9.79e–2967 8.00 9.79e–2967 8.00

4. Conclusions
We conclude that the iterative structure of iterative method Equation (1) was first reported in
article [3] as a particular case and our proposed iterative method [3] is general because α1,1 is a free
parameter.

Conflicts of Interest: The author declares no conflict of interest.

Algorithms 2016, 9, 30 11 of 11

References
1. Babajee, D.K.R. On the Kung-Traub Conjecture for Iterative Methods for Solving Quadratic Equations.
Algorithms 2016, 9, 1, doi:10.3390/a9010001.
2. Traub, J.F. Iterative Methods for the Solution of Equations; Prentice-Hall: Englewood Cliffs, NJ, USA, 1964.
3. Ahmad, F. Higher Order Iterative Methods for Solving Matrix Vector Equations. Researchgate 2015,
doi:10.13140/RG.2.1.1519.2487.

c 2016 by the author; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).