A family of derivative-free methods of seventh-order
convergence for solving nonlinear equations is suggested. In the proposed
methods, several linear combinations of divided differences are used in
order to get a good estimation of the derivative of the given function at
the different steps of the iteration. The efficiency indices of the members of
this family are equal to 1.6266. Also, numerical examples are used to show
the performance of the presented methods, on smooth and nonsmooth
equations, and to compare with other derivative-free methods, including
some optimal fourth-order ones, in the sense of Kung-Traub’s conjecture.
1. Introduction
Finding iterative methods for solving nonlinear equations is an important area of research in numerical analysis, and it has interesting applications in various branches of Science and Engineering. In this study, we describe new iterative methods to find a simple root α of a nonlinear equation f(x)=0, where f:I⊂ℝ→ℝ is a scalar function on an open interval I. The known Newton’s method for finding α uses the iterative expressionxk+1=xk-f(xk)f′(xk),k=0,1,…,
which converges quadratically in some neighborhood of α. If the derivative f′(xk) is replaced by the forward-difference approximationf′(xk)≈f(xk+f(xk))-f(xk)f(xk),
the Newton’s method becomesxk+1=xk-(f(xk))2f(xk+f(xk))-f(xk),
which is the known Steffensen’s method (SM), (see [1]). Newton and Steffensen’s methods are of second order, both require two functional evaluations per step, but in contrast to Newton’s method, Steffensen’s method is derivative-free.
The procedure of removing the derivatives usually increases the number of functional evaluations per iteration. Commonly in the literature the efficiency of an iterative method is measured by the efficiency index defined as I=p1/d (see [2]), where p is the order of convergence and d is the total number of functional evaluations per step. Kung and Traub conjectured in [3] that the order of convergence of any multipoint method cannot exceed the bound 2d-1, (called the optimal order). Thus, the optimal order for methods with 3 or 4 functional evaluations per step would be 4 or 8, respectively.
To improve the convergence properties, many variants of Steffensen’s method have been proposed in the last years. Some of these methods use forward or central divided differences for approximating the derivatives. For example, by composing Steffensen and Newton’s methods and using a particular approximation of the first derivative, Liu et al. derive in [4] an optimal fourth-order method, that we denote (LZM), with three functional evaluations per step. The iterative expression isxk+1=yk-f[xk,yk]-f[yk,zk]+f[xk,zk]f[xk,yk]2f(yk),k=0,1,…,
where yk is the approximation of the Steffensen’s method, zk=xk+f(xk) and f[x,y]=(f(x)-f(y))/(x-y) is the divided difference of order one.
Dehghan and Hajarian [5] proposed a variant of Steffensen’s method (DHM), which is written asxk+1=xk-2f(xk)[f(zk+1)-f(xk)]f(xk+f(xk))-f(xk-f(xk)),
where zk+1=xk+2(f(xk))2/(f(xk+f(xk))-f(xk-f(xk))). The method is obtained by replacing the forward-difference approximation in Steffensen’s method by the central-difference approximation. However, it is still a method of third order and requires four functional evaluations per iteration.
The authors have also presented in [6] a one-parameter family of optimal fourth-order derivative-free methods, denoted by (CTM), which will be used in this paper as a base in order to achieve higher orders of convergence. The iterative expression of this family isyk=xk-f(xk)2f(zk)-f(xk),xk+1=yk-f(yk)(f(yk)-βf(zk))/(yk-zk)+(f(yk)-δf(xk))/(yk-xk),
where parameters β and δ must verify β+δ=1. In the numerical section, we will work with the element of this family obtained by taking β=1 and δ=0.
In this paper, the technique used to improve the local order of convergence consists of the composition of two iterative methods of order p and q, respectively, to obtain a method of order pq (see [1]). Specifically, we compose Newton’s method and the CTM family (1.6). In addition, some particular approximations of the derivative will be made in order to obtain a Steffensen-type method. As we will show, the obtained family of methods is of seventh-order of convergence and requires four evaluations of the function f(x); therefore, this class of methods has efficiency index 71/4≈1.6266, which is higher than 21/2≈1.4142 of the Steffensen’s method, 31/4≈1.3161 of the DHM method (1.5), 41/3≈1.5874 of the LZM (1.4) and CTM (1.6) methods. Therefore, although the methods of the new family are not optimal in the sense of Kung-Traub’s conjecture, they are competitive from the point of view of the efficiency index.
Recently, some seventh-order methods have appeared in the literature: for example, Hu and Fang in [7] design a Jarratt-type scheme of order of convergence seven. Its iterative expression iszk=yk-2f′(xk)-f′(yk)f′(xk)f(yk)f′(xk),xk+1=zk+f′(xk)+f′(yk)f′(xk)-3f′(yk)f(zk)f′(xk),
where yk is the Newton’s iteration. We will denote this scheme by HFM. Let us note that this method is not derivative-free, and it uses five functional evaluations. So, its efficiency index is 71/5≈1.4758.
Noor et al. in [8, Algorithm 2.5] show an iterative method free from second derivative of order seven, with five functional evaluations. Its efficiency index is 71/5≈1.4758 and its iterative expression iszk=yk-2f(yk)f′(yk)2(f′(yk))2-f(yk)Pf(xk,yk),xk+1=zk-f′(xk)+3f′(yk)6f′(yk)-2f′(xk)f(zk)f′(xk),
where Pf(xk,yk)=(2/(yk-xk))(2f′(yk)+f′(xk)-3f[yk,xk]). We denote this scheme by NM.
Soleymani and Khattri in [9] (Theorem 1), design a derivative-free seventh-order method with four functional evaluations. Its iterative expression isyk=xk-f(xk)f[xk,wk],zk=yk-f(yk)f[xk,wk](1+f(yk)f(xk)+f(yk)f(wk)),xk+1=zk-f(zk)f[xk,wk](1+(2-f[xk,wk])f(yk)f(wk)+(11-f[xk,wk])(f(yk)f(xk))2+f(zk)f(yk)),
where wk=xk-f(xk). We denote this method by SKM. Its efficiency index is 71/4≈1.6266.
The rest of the paper is organized as follows: in Section 2, we describe our family of methods and analyze its convergence order for smooth equations. In Section 3, different numerical tests confirm the theoretical results and allow us to compare this family with other known methods mentioned in this section. We also analyze in this numerical section the behavior of the new family on nonsmooth equations.
2. The Methods and Analysis of Convergence
By direct composition of the CTM family (1.6) and Newton’s method, it is easy to see that the schemeyk=xk-f(xk)2f(zk)-f(xk),uk=yk-f(yk)(f(yk)-βf(zk))/(yk-zk)+(f(yk)-δf(xk))/(yk-xk),xk+1=uk-f(uk)f′(uk),
where zk=xk+f(xk), is of eighth-order convergence. In order to avoid the evaluation of the first derivative in the last step, we extend the estimation used in (1.6), by replacing f′(uk) by a linear combination of several divided differences:f′(uk)≈D(xk)=b1f(uk)-b2f(zk)uk-zk+b3f(uk)-b4f(xk)uk-xk+b5f(uk)-b6f(yk)uk-yk+b7f(yk)-b8f(zk)yk-zk+b9f(yk)-b10f(xk)yk-xk+b11f(zk)-b12f(xk)zk-xk,
where b1,b2,…,b12∈ℝ are parameters. So, we are going to prove that for some values of the parameters the family of methods described byyk=xk-f(xk)(f(zk)-f(xk))/f(xk),uk=yk-f(yk)(f(yk)-βf(zk))/(yk-zk)+(f(yk)-δf(xk))/(yk-xk),xk+1=uk-f(uk)D(xk),
is of seventh-order of convergence.
Theorem 2.1.
Let α∈I be a simple zero of a sufficiently differentiable function f:I⊆ℝ→ℝ in an open interval I. If x0 is sufficiently close to α, then the iterative method defined by (2.3) has seventh-order of convergence for β=1-δ, b2=1-b8-b4-b10-b6, b11=b12=0, b7=b8+b4+b10, b6=1-b9-b4, b5=1, and b8=-1+b9+b4-b10 and satisfies the error equation
xk+1-α=-(1+f′(α))3c22(c22-c3)[(-1+b1+b3+f′(α)(b3-b4))c22-(-2+b1+b3+f′(α)(b3-1))c3c22]ek7+O(ek8),
where ek=xk-α and ck=(1/k!)(f(k)(α)/f′(α)), k=2,3,… and δ,β,b1,b2,…,b12∈ℝ.
Proof.
By using Taylor’s expansion around x=α, it is easy to observe (see [6]) that if β+δ=1 then,
uk-α=(1+f′(α))2c2(c22-c3)ek4-(1+f′(α))(2(2+2f′(α)+f′(α)2)c24-2(4+5f′(α)+2f′(α)2)c22c3+(2+3f′(α)+f′(α)2)c32+(2+3f′(α)+f′(α)2)c2c4)ek5+O(ek6)
and then,
f(uk)=f′(α)(1+f′(α))2c2(c22-c3)ek4-f′(α)(1+f′(α))(2(2+2f′(α)+f′(α)2)c24-2(4+5f′(α)+2f′(α)2)c22c3+(2+3f′(α)+f′(α)2)c32+(2+3f′(α)+f′(α)2)c2c4)ek5+O(ek6).
Now, the approximation of f′(uk), D(xk), can be written as
D(xk)=[b11(1+f′(α))-b12+(b2+b8+b4+b10+b6)+f′(α)(b4-b7+(1+f′(α))(b2-b9)+(2+f′(α))(b10+b8+b11))]c2ek+f′(α)[(b7+b9+f′(α)(b2+b7+b8+b10+b11)-(1+f′(α))(2b6-b5))c22+((3+3f′(α)+f′(α)2)b2+b4+b5-b6-2b7+3b8+2b9+f′(α)(2b2+b5-b6-b7+3b8-3b9)+f′(α)2(b2+b8-b9)+(3+3f′(α)+f′(α)2)(b10+b11))c3]ek2+O(ek3),
and calculating the last step of the iterative process (2.3), we have
xk+1-α=(1+f′(α))2c2(c22-c3)×(1-f′(α)b11(1+f′(α))+f′(α)(b2+b8+b4+b10+b6)-b12)ek4+O(ek5).
It is necessary to assign the following values to the parameters in order to assure the sixth-order of convergence: b2=1-b8-b4-b10-b6, b11=b12=0, b7=b8+b4+b10, and b6=1-b9-b4. So, the error equation can be expressed as
xk+1-α=(1+f′(α))3c2(c22-c3)((2-b4-b5+b8-b9+b10)c22+(-1+b5)c3c22)ek6+O(ek7).
Finally, if b5=1 and b8=-1+b9+b4-b10, we have
xk+1-α=-(1+f′(α))3c22(c22-c3)[(-1+b1+b3+f′(α)(b3-b4))c22-(-2+b1+b3+f′(α)(b3-1))c3c22]ek7+O(ek8).
In terms of computational cost, the methods of this family require only four functional evaluations per step. So, they have efficiency indices 71/4=1.6266. If we denote by M7 any element of this family, we can establishIM7=ISKM>ICTM=ILZM>INM=IHFM>ISM>IDHM.
In the next section, we use the element of family (2.3) obtained by choosing (for simplicity) δ=b1=b3=b4=b9=b10=b11=b12=0 and, therefore, β=1, b2=b6=1, and b7=b8=-1. So, the resulting iterative expression of the method isyk=xk-f(xk)2f(zk)-f(xk),uk=yk-f(yk)(f(yk)-f(zk))/(yk-zk)+f(yk)/(yk-xk),xk+1=uk-f(uk)D(xk),
where D(xk)=(f(uk)-f(yk))/(uk-yk)-f(zk)/(uk-zk)-(f(yk)-f(zk))/(yk-zk) and zk=xk+f(xk).
It is well known that if α is a multiple zero of f(x), then α is a simple zero of f(x)/f′(x). In a similar way, for Steffensen-type methods, it is easy to prove that if we use f0(x)=(f(x))2/(f(x+f(x))-f(x)), we transform the problem of solving multiple roots of f(x)=0 into a simple roots one, f0(x)=0. Theoretically, this idea improves the order of convergence but in practice, results are not satisfactory.
As we have seen, the method (2.12) has seventh-order convergence for smooth equations but, what is its behavior for nonsmooth equations? As we will see in the next section, for this class of equations our method, in general, loses the seventh-order convergence and stability problems appear. For nonsmooth functions, Amat and Busquier in [10] presented an strategy to control the approximation of the derivative and the stability of the iteration. They applied this idea to Steffensen’s method, obtaining a new scheme (STM):xk+1=xk-f(xk)[f(xk+αk|f(xk)|f(xk))-f(xk)]/αk|f(xk)|f(xk),
where the parameters αk∈ℝ allow to control the approximation of the derivative. This procedure can be applied to any other derivative-free scheme. The authors showed the second-order convergence of (2.13) for nonsmooth functions and mentioned that, in order to control the stability in practice, the parameters αk should verifytolc≪|αk|f(xk)|f(xk)|≤tolu,
where tolc is related to the computer precision and tolu is a free parameter.
In the following section, we will apply this strategy to our proposed method, M7, obtaining a modified scheme that will be denoted as M7mod. Then, we will analyze how this new scheme improves in nonsmooth cases, although the order of convergence at singular points decreases to four.
3. Numerical Results
This numerical section is divided into two parts: one devoted to compare the different methods on smooth equations and other in which we analyze the behavior of our method on nonsmooth test functions.
In the first part of this section, we use some test functions in order to check the effectiveness of the new high-order method (2.12), we compare it with the classical Steffensen’s method, SM, the method DHM, and the optimal fourth-order methods, LZM and CTM with β=1 and δ=0. These methods are employed to find the zeros of some nonlinear functions, specifically,
f1(x)=sin2x-x2+1,α≈1.404492,
f2(x)=x2-ex-3x+2,α≈0.257530,
f3(x)=cosx-x,α≈0.739085,
f4(x)=(x-1)3-1,α=2,
f5(x)=x3-10,α≈2.154435,
f6(x)=cos(x)-xex+x2,α≈0.639154,
f7(x)=ex-1.5-arctan(x),α≈0.767653,
f8(x)=x3+4x2-10,α≈1.365230,
f9(x)=8x-cos(x)-2x2,α≈0.128077,
f10(x)=arctan(x),α=0.
The complexity of the iterative expressions plays an important role in the computational efficiency of the different methods. So, some authors use another index in order to compare the iterative methods, taking also into account the number of products and quotients involved in each step of the iterative process. The computational efficiency index is defined as CI=p1/(d+op), (see [11]), where p is the order of convergence, d is the number of functional evaluations, and op is the number of products and quotients per iteration. Under the point of view of this index, the relationship between the schemes that we use in this section isCICTM(41/8)=CISM(21/4)>CIM7(71/12)=CIHFM>CIDHM(31/7)>CILZM(41/10)>CINM(71/16)>CISKM(71/18).
Nowadays, high-order methods are important because numerical applications use high precision in their computations; for this reason, numerical computations have been carried out using variable precision arithmetic in Matlab 7.12 (R2011a) with 500 significant digits. The computer specifications are Intel(R) Core(TM) i5-2500 CPU @ 3.30 GHz with 16.00 GB of RAM. The stopping criterion used is |xk+1-xk|<10-150 or |f(xk)|<10-150. The information shown in Tables 1 and 2 is, for every method, the number of iterations needed to reach the required tolerance (if the method does not converge, it will be denoted by "nc"), the last value of |xk+1-xk| and |f(xk+1)|, and the approximated computational order of convergence (ACOC) ρ, defined by the authors in [12]:ρ=ln(|xk+1-xk|/|xk-xk-1|)ln(|xk-xk-1|/|xk-1-xk-2|).
By means of (3.2), a vector is obtained by using the different iterations calculated in the process. The value of ρ that appears in Tables 1 to 4 is the last coordinate of this vector when the variation between its components is small. Let us note that when the approximated convergence order is not stable (if the difference between two consecutive values is bigger than one unit), we will denote it by ‘—’.
Numerical results from f1 to f10.
SM
DHM
LZM
CTM
M7
f1,x0=0.9
|xk+1-xk|
3.9289e-112
1.7589e-055
3.7228e-122
1.5049e-124
1.9456e-023
|f(xk+1)|
4.4514e-223
2.8819e-164
0
0
1.8101e-159
ACOC
2.0000
3.0000
4.0000
4.0000
6.6629
iter
9
7
5
5
3
e-time
0.1217
0.1891
0.1139
0.1033
0.0927
f2,x0=1.2
|xk+1-xk|
1.4587e-149
7.6358e-103
3.4035e-138
2.6499e-141
3.1050e-029
|f(xk+1)|
2.0878e-298
2.5787e-306
0
0
1.0495e-202
ACOC
2.0000
3.0000
4.0000
4.0000
6.8723
iter
9
11
5
5
3
e-time
0.1282
0.2791
0.1161
0.1092
0.1075
f3,x0=2.1
|xk+1-xk|
8.3630e-085
1.9786e-059
1.0746e-143
1.4483e-112
5.6495e-024
|f(xk+1)|
1.7410e-169
1.8682e-177
0
0
3.7489e-167
ACOC
2.0000
3.0000
4.0000
4.0000
7.0731
iter
8
6
5
5
3
e-time
0.0746
0.0894
0.0874
0.0867
0.0468
f4,x0=2.2
|xk+1-xk|
1.9109e-116
1.8976e-072
3.3922e-110
1.0118e-116
3.4709e-027
|f(xk+1)|
4.3820e-231
1.5033e-214
0
0
5.1781e-184
ACOC
2.0000
3.0000
4.0000
4.0000
6.8325
iter
10
6
5
5
3
e-time
0.1176
0.1443
0.1246
0.0914
0.0930
f5,x0=2.3
|xk+1-xk|
7.8747e-085
1.2204e-123
9.1432e-142
8.5347e-144
1.2638e-030
|f(xk+1)|
5.9818e-167
0
0
0
6.8463e-207
ACOC
2.0000
3.0000
4.0000
4.0000
6.8181
iter
10
7
5
5
3
e-time
0.1168
0.1353
0.0945
0.1059
0.0596
f6,x0=2
|xk+1-xk|
1.4558e-087
—
2.1767e-109
5.9067e-112
5.4741e-023
|f(xk+1)|
5.7398e-174
—
0
0
9.2491e-157
ACOC
2.0000
—
4.0000
4.0000
5.9331
iter
8
nc
5
5
3
e-time
0.1245
—
0.1576
0.1418
0.1226
f7,x0=0.5
|xk+1-xk|
5.1639e-127
3.2959e-079
1.5312e-050
3.3808e-073
4.7872e-034
|f(xk+1)|
9.3020e-253
1.1763e-235
4.6052e-199
7.2079e-290
9.9787e-234
ACOC
2.0000
3.0000
3.9999
4.0000
6.8055
iter
11
6
5
5
3
e-time
0.1671
0.1696
0.1426
0.1349
0.0732
f8,x0=1.5
|xk+1-xk|
1.0817e-142
8.4333e-122
6.9628e-136
2.1376e-137
1.1249e-030
|f(xk+1)|
1.6591e-282
0
0
0
7.6946e-207
ACOC
2.0000
3.0000
4.0000
4.0000
6.7788
iter
11
7
5
5
3
e-time
0.1886
0.2356
0.1797
0.1463
0.1170
f9,x0=0.8
|xk+1-xk|
2.2055e-129
4.6810e-128
2.9693e-139
7.1679e-140
6.1073e-028
|f(xk+1)|
9.0498e-257
0
0
0
1.6582e-191
ACOC
2.0000
3.0000
4.0000
4.0000
6.7613
iter
15
6
7
8
4
e-time
0.2368
0.1680
0.2684
0.1902
0.1232
f10,x0=0.6
|xk+1-xk|
2.4132e-081
1.3419e-130
6.2415e-141
1.0766e-031
2.7207e-019
|f(xk+1)|
2.8106e-242
0
0
1.9282e-155
2.1785e-167
ACOC
3.0000
5.0000
5.0000
4.9922
8.7406
iter
7
5
5
4
3
e-time
0.0955
0.1471
0.0746
0.0733
0.0749
Numerical results from f1 to f10 and seventh-order schemes.
HFM
SKM
NM
M7
f1,x0=0.9
|xk+1-xk|
9.2403e-050
7.8995e-104
1.9456e-023
|f(xk+1)|
0
0
1.8101e-159
ACOC
—
7.0023
6.6629
iter
6
≥100
4
3
e-time
0.1126
0.1404
0.0927
f2,x0=1.2
|xk+1-xk|
1.1602e-045
1.0203e-036
6.3747e-058
3.1050e-029
|f(xk+1)|
0
0
0
1.0495e-202
ACOC
6.7965
7.2403
6.7108
6.8723
iter
3
4
3
3
e-time
0.0604
0.1445
0.1100
0.1075
f3,x0=2.1
|xk+1-xk|
1.8810e-040
9.8612e-037
1.4468e-052
5.6495e-024
|f(xk+1)|
0
0
0
3.7489e-167
ACOC
6.3228
6.2401
6.5225
7.0731
iter
3
4
3
3
e-time
0.0315
0.0969
0.0674
0.0468
f4,x0=2.2
|xk+1-xk|
5.1064e-034
—
9.9985e-042
3.4709e-027
|f(xk+1)|
0
—
0
5.1781e-184
ACOC
6.8073
—
6.9177
6.8325
iter
3
n.c.
3
3
e-time
0.0532
—
0.0842
0.0930
f5,x0=2.3
|xk+1-xk|
2.6845e-053
3.2107e-025
1.2399e-062
1.2638e-030
|f(xk+1)|
3.1147e-207
9.9146e-167
3.1147e-207
6.8463e-207
ACOC
6.9527
7.1651
6.9798
6.8181
iter
3
5
3
3
e-time
0.0406
0.0894
0.0961
0.0596
f6,x0=2
|xk+1-xk|
8.6332e-124
1.0547e-075
5.4741e-023
|f(xk+1)|
0
0
9.2491e-157
ACOC
6.9990
6.9948
5.9331
iter
4
≥100
4
3
e-time
0.0911
0.1409
0.1226
f7,x0=0.5
|xk+1-xk|
8.9879e-076
6.4335e-045
1.6607e-028
4.7872e-034
|f(xk+1)|
3.8934e-208
1.9467e-208
1.1707e-195
3.8434e-208
ACOC
7.0115
6.8734
7.2394
6.8055
iter
4
3
3
3
e-time
0.0747
0.0757
0.1214
0.0732
f8,x0=1.5
|xk+1-xk|
2.1424e-053
8.0693e-063
1.1249e-030
|f(xk+1)|
0
0
9.3442e-207
ACOC
6.9528
6.9783
6.7788
iter
3
≥100
3
3
e-time
0.0730
0.0967
0.1170
f9, x0=0.8
|xk+1-xk|
1.6408e-027
1.3528e-128
4.3711e-039
6.1073e-028
|f(xk+1)|
1.5509e-190
2.4334e-208
1.2167e-208
1.6582e-191
ACOC
7.3243
6.9999
7.1316
6.7613
iter
3
4
3
4
e-time
0.0552
0.1158
0.1112
0.1232
f10,x0=0.6
|xk+1-xk|
2.0089e-035
—
1.1046e-051
2.7207e-019
|f(xk+1)|
0
—
0
2.1785e-167
ACOC
10.9352
—
10.9217
8.7406
iter
3
n.c.
3
3
e-time
0.0367
—
0.0922
0.0749
On the other hand, in Tables 1 and 2, the mean elapsed time, calculated by means of the command "cputime" of Matlab (e-time), after 100 performances of the program, appears. It can be observed that, in most cases, the elapsed time taken by M7 to obtain the solution is lower than the corresponding ones of the other methods involved. In terms of computational effort, the efficiency of the proposed method is not lower than that of the optimal fourth-order methods. These elapsed times are in concordance with the computational efficiency index of each method.
Numerical results in Table 1 confirm the theoretical statements developed in this paper, showing that the estimated order of convergence coincides with the theoretical one, except in case f10: the second derivative of this nonlinear function at the solution is zero, so the order of convergence increases: from second to third in SM, from third (and fourth) to fifth in DHM (and LZM or CTM), and from seventh to ninth in M7. Nevertheless, in this case the e-time from the new method M7 does not improve the other ones. In fact, in case f10 the best time is obtained by Steffensen’s method, followed by CTM and DHM. In general, there can be stated that the new high-order scheme improves the results obtained by other known methods, even optimal fourth-order methods such as LZM and CTM.
In Table 2, we compare the new method M7 with other known seventh-order schemes described in the introduction, HFM, SKM, and NM. It can be observed that M7 performs better than SKM and NM, but HFM is more precise than the rest of methods.
Now, we are going to make some numerical tests in order to check how the methods SM and M7 behave in nonsmooth cases. Moreover, we apply the αk-procedure to both methods to avoid some stability problems. In these cases, numerical computations have been carried out using simple precision arithmetic, so tolc=10-16, and the stopping criterion used has been |xk+1-xk|<tolu=10-11 or |f(xk+1)|<tolu=10-11. From a sufficiently small α0, we use the following algorithm to compute the different αk:αk+1={αk2if|αk2|f(xk)|f(xk)|≥tolc,tolc∥f(xk)|f(xk)|,elsewhere.
The first test has been made on the function:f11(x)={x(x+1)ifx<0,-2x(x-1)ifx≥0,
that can be found in [13]. We use three initial estimations in order to approximate the three different roots of the equation, {0,1,-1}. In Table 3, we show for each initial estimations and every method, the exact absolute error at first and last iterations, the absolute difference between the two last iterations |xk+1-xk|, the value of f in the last iteration |f(xk+1)|, and the ACOC. From Table 3 it can be inferred that the order of convergence of M7 method decreases to five and stability problems appear when it is applied to nonsmooth equations. Nevertheless, it usually performs better than or equal to Steffensen’s method and its modifications by the αk procedure. Indeed, when this strategy is applied on the seventh-order method (M7mod), the stability of the method is improved and it results in more precise estimations with less iterations. In this example, the ACOC is not stable in some cases.
Numerical results for function f11(x).
x0=0.1
SM
STM (tolu=10-11)
M7
M7mod
iter
error
iter
error
iter
error
iter
error
α=0
1
4.52e-2
1
1.50e-2
1
1.50e-3
1
2.28e-4
2
4.60e-3
2
2.32e-4
2
3.49e-11
2
0
⋮
3
5.38e-8
3
0
4
8.32e-9
4
0
5
0
|xk+1-xk|
8.32e-9
5.38e-8
3.49e-11
2.28e-4
|f(xk+1)|
2.08e-16
5.80e-15
3.76e-37
4.08e-15
ACOC
2.0919
2.0000
4.2042
—
x0=3
iter
error
iter
error
iter
error
iter
error
α=1
1
0.29
1
0.80
1
3.45e-2
1
6.22e-2
2
0.15
2
0.25
2
6.91e-11
2
2.35e-9
⋮
⋮
3
NaN
3
0
5
7.98e-8
6
5.36e-12
6
1.0e-14
7
0
xk+1-xk|
7.98e-8
5.36e-12
NaN
2.35e-9
|f(xk+1)|
1.27e-14
0
NaN
0
ACOC
1.9997
2.0000
—
4.9686
x0=-20
iter
error
iter
error
iter
error
iter
error
α=-1
1
18.44
1
9.26
1
7.17
1
2.31
2
17.88
2
4.39
2
2.38
2
8.82e-2
⋮
⋮
⋮
3
2.29e-10
30
5.65e-5
8
1.98e-6
4
2.72e-8
4
0
31
1.80e-13
9
3.93e-12
5
NaN
|xk+1-xk|
5.65e-5
1.98e-6
NaN
2.29e-8
|f(xk+1)|
1.80e-13
3.93e-12
NaN
0
ACOC
—
1.9991
—
—
Numerical results for function f12(x).
x0=2.8
SM
STM (tolu=10-11)
M7
M7mod
iter
error
iter
error
iter
error
iter
error
α=3
1
0.44
1
7.14e-3
1
0.17
1
7.54e-5
2
0.89
2
8.39e-6
2
1.69e-11
2
1.0e-14
⋮
3
1.17e-11
3
0
24
6.0e-10
4
0
25
0
|xk+1-xk|
6.0e-10
1.17e-11
1.69e-8
7.54e-5
|f(xk+1)|
0
0
0
6.93e-14
ACOC
2.0009
1.9984
—
—
x0=-2.8
iter
error
iter
error
iter
error
iter
error
α=-3
>104
1
7.14e-3
1
5.66e-3
1
7.54e-5
2
8.39e-6
2
4.41e-3
2
1.0e-14
3
1.17e-11
⋮
4
0
4
2.74e-10
5
NaN
|xk+1-xk|
1.17e-11
NaN
7.54e-5
|f(xk+1)|
0
NaN
6.93e-14
ACOC
1.9984
—
—
x0=-10
iter
error
iter
error
iter
error
iter
error
α=-3
>104
1
2.45
1
0.70
1
7.98e-2
2
0.55
2
0.73
2
3.2e-13
⋮
⋮
5
2.05e-8
7
2.23e-9
6
0
8
NaN
|xk+1-xk|
2.05e-8
NaN
7.98e-2
|f(xk+1)|
0
NaN
1.95e-12
ACOC
1.9368
—
—
Finally, we consider the nonsmooth functionf12(x)=|x2-9|.
The numerical experiments made on this function are summarized in Table 4. In this case, the advantages of the modified methods over original Steffensen’s method and M7 method are more evident when the initial estimation is far from the zero of the function. When the initial estimation is good enough, it is clear that the behavior of M7mod improves lower-order methods, in terms of precision and number of iterations.
4. Conclusions
A new seventh-order family of derivative-free iterative methods for solving nonlinear equations has been presented. As only four functional evaluations are required per iteration, the efficiency index of each member of this family is equal to 71/4=1.6266. In addition, these methods use a small amount of products and quotients and are derivative-free, which allow us to apply them also to nonsmooth equations with positive and promising results.
The generalization of these methods to nonlinear systems F(x)=0 is similar to the classical Steffensen’s method (see [1]):x̃k=xk+αk‖F(xk)‖F(xk),xk+1=xk-[xk,x̃k;F]-1F(xk),
where [u,v;F]:ℝn→ℝn is a linear operator such that [u,v;F](u-v)=F(u)-F(v), which is called divided difference.
Acknowledgments
The authors would like to thank the referees for their valuable comments and for their suggestions to improve the readability of the paper. This research was supported by Ministerio de Ciencia y Tecnología MTM2011-28636-C02-02 and by Vicerrectorado de Investigación, Universitat Politècnica de València PAID-06-2010-2285.
OrtegaJ. M.RheinboldtW. C.1970New York, NY, USAAcademic Pressxx+5720273810ZBL0241.65046OstrowskiA. M.1966New York, NY, USAAcademic Pressxiv+338Pure and Applied Mathematics, Vol. 90216746ZBL0222.65070KungH. T.TraubJ. F.Optimal order of one-point and multipoint iteration197421643651035365710.1145/321850.321860ZBL0289.65023LiuZ.ZhengQ.ZhaoP.A variant of Steffensen's method of fourth-order convergence and its applications201021671978198310.1016/j.amc.2010.03.0282647066ZBL1208.65064DehghanM.HajarianM.Some derivative free quadratic and cubic convergence iterative formulas for solving nonlinear equations2010291193010.1590/S1807-030220100001000022608446ZBL1189.65091CorderoA.TorregrosaJ. R.A class of Steffensen type methods with optimal order of convergence2011217197653765910.1016/j.amc.2011.02.0672799778ZBL1216.65055HuY.FangL.A seventh-order convergence Newton-type method for solving nonlinear equations2nd International Conference on Computational Intelligence and Natural Computing2010NoorM. A.KhanW. A.NoorK. I.Al-saidE.High-order iterative method free from second derivative for solving nonlinear equations20116818871893SoleymaniF.KhattriS. K.Finding simple roots by seventh and eighthorder derivative-free methods2012164552AmatS.BusquierS.On a Steffensen's type method and its behavior for semismooth equations2006177281982310.1016/j.amc.2005.11.0322292007ZBL1096.65047CorderoA.HuesoJ. L.MartínezE.TorregrosaJ. R.A modified Newton-Jarratt's composition2010551879910.1007/s11075-009-9359-z2679752CorderoA.TorregrosaJ. R.Variants of Newton's method using fifth-order quadrature formulas2007190168669810.1016/j.amc.2007.01.0622338747ZBL1122.65350AmatS.BusquierS.On a higher order secant method20031412-332132910.1016/S0096-3003(02)00257-61972912ZBL1035.65057