An Efficient Sixth-Order Newton-Type Method for Solving Nonlinear Systems

In this paper, we present a new sixth-order iterative method for solving nonlinear systems and prove a local convergence result. The new method requires solving five linear systems per iteration. An important feature of the new method is that the LU (lower upper, also called LU factorization) decomposition of the Jacobian matrix is computed only once in each iteration. The computational efficiency index of the new method is compared to that of some known methods. Numerical results are given to show that the convergence behavior of the new method is similar to the existing methods. The new method can be applied to smalland medium-sized nonlinear systems.


Introduction
We consider the problem of finding a zero of a nonlinear function F : D ⊂ R m → R m , that is, a solution α of the nonlinear system F(x) = 0 with m equations and m unknowns.Newton's method [1,2] is the well-known method for solving nonlinear systems, which can be written as: where F (x (k) ) is the Jacobian matrix of the function F evaluated in the kth iteration and F (x (k) ) −1 is the inverse of F (x (k) ).Newton's method is denoted by NM.Newton's method converges quadratically if F (α) is nonsingular and F (x) is Lipschitz continuous on D. The method (1) can be written as: which requires (m 3 − m)/3 multiplications and divisions in the LU decomposition (lower upper, also called LU factorization) and m 2 multiplications and divisions for solving two triangular linear systems.So, the computational cost (multiplications and divisions) of the method (1) is (m 3 − m)/3 + m 2 .
In order to accelerate the convergence or to reduce the computational cost and function evaluation in each step of the iterative process, many efficient methods have been proposed for solving nonlinear systems, see [3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21] and the references therein.Cordero et al. [3,4] developed some variants of Newton's method.One of the methods is the following fourth-order method [3]: −1 F(x (k) ), where I is the identity matrix.Method (3) is denoted by CM4 and requires LU decomposition of the Jacobian matrix only once per full iteration.Based on method (3), Cordero et al. [4] presented the following sixth-order method: where I is the identity matrix.Method ( 4) is denoted by CHM and requires two LU decompositions of the Jacobian matrix, one for F (x (k) ) and one for F (y (k) ).Grau-Sánchez et al. [5] obtained the following sixth-order method: where F[•, •] denotes the first-order divided difference of D. Method ( 5) is denoted by SNAM and requires two LU decompositions for solving the linear systems involved.
It is well known that the computational cost of the iterative method greatly influences the efficiency of the iterative method.The number of LU decompositions that are used in the iterative method thus plays an important role when it comes to measuring the computational cost.So, we can reduce the computational cost of the iterative method by reducing the number of LU decompositions in each iteration.
The purpose of this paper is to construct a new sixth-order iterative method for solving small-and medium-sized systems.The theoretical advantages of the new method are based on the assumption that the Jacobian matrix is dense and that LU factorization is used to solve systems with the Jacobian.This assumption is not correct for sparse Jacobian matrices.An important feature of the new method is that the LU decomposition is computed only once per full iteration.This paper is organized as follows.In Section 2, based on the well-known fourth-order method (3), we present a sixth-order iterative method for solving nonlinear systems.The new method only increases one function evaluation, F. For a system of m equations, each iteration uses 3m + 2m 2 evaluations of scalar functions.The computational efficiency is compared to some well-known methods in Section 3. In Section 4, numerical examples are given to illustrate the convergence behavior of our method.Section 5 offers a short conclusion.

The New Method and Analysis of Convergence
Based on the method (3), we construct the following iterative scheme: where I is the identity matrix.We note that the first two steps of method (6) are the same as those of method (3), and the third step of method (6) increases one function evaluation F(z (k) ).Method (6) will be denoted by M6.For method (6), we have the following convergence analysis.
Theorem 1.Let α ∈ R m be a solution of the system F(x) = 0 and F : D ⊂ R m → R m be sufficiently differentiable in an open neighborhood D of α.Suppose that F (x) is nonsingular in D.Then, for an initial approximation sufficiently close to α, the iterations converge with order 6.
In order to simplify the calculation, the new method ( 6) can be written as: From ( 22), we can see that the LU decomposition of the Jacobian matrix F (x (k) ) would be computed only once per iteration.
The C i (µ, m) is given by [8]: where a i (m) represents the number of evaluations of the scalar functions used in the evaluations of F, F and [y, x; F].The p i (m) represents the computational cost per iteration.µ > 0 is the ratio between multiplications (and divisions) and evaluations of functions that are required to express C i (µ, m) in terms of multiplications (and divisions).The divided difference [y, x; F] is defined by [9]: where F(x) and F(y) are computed separately.When we compute a divided difference, we need m 2 quotients and m(m − 1) scalar function evaluations.We must add m multiplications for the scalar product, m 2 multiplications for the matrix-vector multiplication and m 2 evaluations of scalar functions for any new derivative F .In order to factorize a matrix, we require (m 3 − m)/3 multiplications and divisions in the LU decomposition.We require m 2 multiplications and divisions for solving two triangular linear systems.
Taking into account the previous considerations, we give the computational cost of each iterative method in Table 1.
From Table 1, we can see that these sixth-order iterative methods have the same number of function evaluations.Our method (M6) needs less LU decompositions than other methods with the same order.Therefore, the computational cost of our method is lower.
We use the following expressions [10] to compare the computational efficiency index of each iterative method: For R i,j > 1 the iterative method M i is more efficient than M j .We have the following theorem: Theorem 2. For all µ > 0 we have: 1. CEI 6 > CEI 3 for all m ≥ 5, 2.

1.
Based on the expression (26), the relation between SNAM and M6 can be given by: Subtracting the denominator from the numerator of (27), we have: Equation ( 28) is positive for m ≥ 4.541.Thus, we get CEI 6 > CEI 3 for all m ≥ 5 and µ > 0.

2.
The relation between CHM and M6 is given by: Subtracting the denominator from the numerator of (29), we have: Equation ( 30) is positive for m ≥ 6.606.Thus, we get CEI 6 > CEI 4 for all m ≥ 7 and µ > 0.

3.
The relation between CHM and M6 can be given by: Subtracting the denominator from the numerator of (31), we have: Equation ( 32) is positive for m ≥ 8.772.Thus, we obtain CEI 6 > CEI 5 for all m ≥ 9 and µ > 0. This completes the proof.
This completes the proof.
In Tables 2 and 3, we show the computational efficiency indices of NM, CM4, SNAM, CHM, CTVM and M6, for different values of the size of the nonlinear systems.CEI CEI > over it (see Figure 1).Boundary (36) cuts axes at points ( , ) (0, 19.2012) m μ = and (5.3984, 2) .Thus, we get  In Tables 2 and 3, we show the computational efficiency indices of NM, CM4, SNAM, CHM, CTVM and M6, for different values of the size of the nonlinear systems.The results shown in Tables 2 and 3 are in concordance with the Theorems 2 and 3. We can see that, for small-and medium-sized systems, the computational efficiency index of our method (M6) is similar to the other methods in this paper.

Numerical Examples
In this section, we compare the related methods by numerical experiments.The numerical experiments are performed using the MAPLE computer algebra system with 2048 digits.The method M6 is compared with NM, CM4, SNAM, CHM, CTVM by solving some nonlinear systems.
Table 4 presents the results showing the following information: the number of iterations k needed to converge to the solution, the value of the stopping factors at the last step and the computational order of convergence ρ.The computational order of convergence ρ is defined by [18]: The numerical results shown in Table 4 are in concordance with the theory developed in this paper.The order of convergence of our method (M6) is 6, which is higher than the methods NM and CM4.The iterative method SNAM is not convergent (nc) for F 3 with the corresponding initial value.The convergence behavior of our method is similar to the existing methods in this paper.

Conclusions
In this paper, we have proposed a new iterative method of order six for solving nonlinear systems.Although five linear systems are required to be solved in each iteration, the LU decomposition of the linear systems are computed only once per full iteration.Numerical results are given to show that our method has a similar convergence behavior as the existing methods in this paper.The new method is suitable for solving small-and medium-sized systems.In order to obtain an efficient iterative method for solving nonlinear systems, we should make the iterative method achieve an as high as possible convergence order consuming an as small as possible computational cost.In addition, the theoretical advantages of the new method are based on the assumption that the Jacobian matrix is dense and that LU factorization is used to solve the systems with the Jacobian.This assumption is not correct for sparse Jacobian matrices.

Table 1 .
Computational cost of the iterative methods.