- Research
- Open Access

# A superlinearly convergent hybrid algorithm for systems of nonlinear equations

- Lian Zheng
^{1}Email author

**2012**:180

https://doi.org/10.1186/1029-242X-2012-180

© Zheng; licensee Springer 2012

**Received:**28 March 2012**Accepted:**6 August 2012**Published:**24 August 2012

## Abstract

We propose a new algorithm for solving systems of nonlinear equations with convex constraints which combines elements of Newton, the proximal point, and the projection method. The convergence of the whole sequence is established under weaker conditions than the ones used in existing projection-type methods. We study the superlinear convergence rate of the new method if in addition a certain error bound condition holds. Preliminary numerical experiments show that our method is efficient.

**MSC:** 90C25, 90C30.

## Keywords

- nonlinear equations
- projection method
- global convergence
- superlinear convergence

## 1 Introduction

*S*denote the solution set of (1.1). Throughout this paper, we assume that

*S*is nonempty and

*F*has the property that

The property (1.2) holds if *F* is monotone or more generally pseudomonotone on *C* in the sense of Karamardian [1].

Nonlinear equations have wide applications in reality. For example, many problems arising from chemical technology, economy, and communications can be transformed into nonlinear equations; see [2–5]. In recent years, many numerical methods for problem (1.1) with smooth mapping *F* have been proposed. These methods include the Newton method, quasi-Newton method, Levenberg-Marquardt method, trust region method, and their variants; see [6–14].

Recently, the literature [15] proposed a hybrid method for solving problem (1.1), which combines the Newton, proximal point, and projection methodologies. The method possesses a very nice globally convergent property if *F* is monotone and continuous. Under the assumptions of differentiability and nonsingularity, locally superlinear convergence of the method is proved. However, the condition of nonsingularity is too strong. Relaxing the nonsingularity assumption, the literature [16] proposed a modified version for the method by changing the projection way, and showed that under the local error bound condition which is weaker than nonsingularity, the proposed method converges superlinearly to the solution of problem (1.1). The numerical performances given in [16] show that the method is really efficient. However, the literatures [15, 16] need the mapping *F* to be monotone, which seems too stringent a requirement for the purpose of ensuring global convergence property and locally superlinear convergence of the hybrid method.

To further relax the assumption of monotonicity of *F*, in this paper, we propose a new hybrid algorithm for problem (1.1) which covers one in [16]. The global convergence of our method needs only to assume that *F* satisfies the property (1.2), which is much weaker than monotone or more generally pseudomonotone. We also discuss the superlinear convergence of our method under mild conditions. Preliminary numerical experiments show that our method is efficient.

## 2 Preliminaries and algorithms

*x*onto Ω is defined as

We have the following property on the projection operator; see [17].

**Lemma 2.1**

*Let*$\mathrm{\Omega}\subset {\mathbb{R}}^{n}$

*be a closed convex set*.

*Then it holds that*

**Algorithm 2.1** Choose ${x}_{0}\in C$, parameters ${\kappa}_{0}\in [0,1)$, *λ*, $\beta \in (0,1)$, ${\gamma}_{1}$, ${\gamma}_{2}>0$, $a,b\ge 0$, $max\{a,b\}>0$, and set $k:=0$.

Stop if ${d}^{k}=0$. Otherwise,

*m*satisfying

Let $k=k+1$ and return to Step 1.

**Remark 2.1** When we take parameters $a=0$, $b=1$, and the search direction ${d}^{k}={\overline{x}}^{k}-{x}^{k}$, our algorithm degrades into one in [16]. At this step of getting the next iterate, our projection way and projection region are also different from the one in [15].

Now we analyze the feasibility of Algorithm 2.1. It is obvious that ${d}^{k}$ satisfying conditions (2.1) and (2.2) exists. In fact, when we take ${d}^{k}=-{({G}_{k}+{\mu}_{k}I)}^{-1}F({x}^{k})$, ${d}^{k}$ satisfies (2.1) and (2.2). Next, we need only to show the feasibility of (2.3).

**Lemma 2.2** *For all nonnegative integer* *k*, *there exists a nonnegative integer* ${m}_{k}$ *satisfying *(2.3).

*Proof* If ${d}^{k}=0$, then it follows from (2.1) and (2.2) that $F({x}^{k})=0$, which means Algorithm 2.1 terminates with ${x}^{k}$ being a solution of problem (1.1).

*k*. By the definition of ${r}^{k}$, the Cauchy-Schwarz inequality and the positive semidefiniteness of ${G}_{k}$, we have

*m*,

*i.e.*,

*F*, we have

Which, together with (2.5), ${d}^{{k}_{0}}\ne 0$, and ${\sigma}_{k}\le {\kappa}_{0}<1$, we conclude that $\lambda \ge 1$, which contradicts $\lambda \in (0,1)$. This completes the proof. □

## 3 Convergence analysis

In this section, we first prove two lemmas, and then analyze the global convergence of Algorithm 2.1.

**Lemma 3.1** *If the sequences* $\{{x}^{k}\}$ *and* $\{{y}^{k}\}$ *are generated by Algorithm * 2.1, $\{{x}^{k}\}$ *is bounded and* *F* *is continuous*, *then* $\{{y}^{k}\}$ *is also bounded*.

*Proof*Combining inequality (2.5) with the Cauchy-Schwarz inequality, we obtain

From the boundedness of $\{{x}_{k}\}$ and the continuity of *F*, we conclude that $\{{d}^{k}\}$ is bounded, and hence so is $\{{y}^{k}\}$. This completes the proof. □

**Lemma 3.2**

*Let*${x}^{\ast}$

*be a solution of problem*(1.1)

*and the function*${h}_{k}$

*be defined by*(2.4).

*If condition*(1.2)

*holds*,

*then*

*In particular*, *if* ${d}^{k}\ne 0$, *then* ${h}_{k}({x}^{k})>0$.

*Proof*

where the inequality follows from condition (1.2) and the definition of ${y}^{k}$.

If ${d}^{k}\ne 0$, then ${h}_{k}({x}^{k})>0$ because ${\sigma}_{k}\le {\kappa}_{0}<1$. The proof is completed. □

**Remark 3.1**Lemma 3.2 means that the hyperplane

strictly separates the current iterate from the solution set of problem (1.1).

where the first inequality follows from condition (1.2), the second one follows from (2.3), and the last one follows ${d}^{k}\ne 0$, which shows that $-(aF({x}^{k})+bF({y}^{k}))$ is a descent direction of the function $\frac{1}{2}{\parallel x-{x}^{\ast}\parallel}^{2}$ at the point ${x}^{k}$.

We next prove our main result. Certainly, if Algorithm 2.1 terminates at Step *k*, then ${x}^{k}$ is a solution of problem (1.1). Therefore, in the following analysis, we assume that Algorithm 2.1 always generates an infinite sequence.

**Theorem 3.1** *If* *F* *is continuous on* *C*, *condition* (1.2) *holds and* ${sup}_{k}\parallel {G}_{k}\parallel <\mathrm{\infty}$, *then the sequence* $\{{x}^{k}\}\subset {\mathbb{R}}^{n}$ *generated by Algorithm * 2.1 *globally converges to a solution of* (1.1).

*Proof*Let ${x}^{\ast}$ be a solution of problem (1.1). Since ${x}^{k+1}={\mathrm{\Pi}}_{{C}_{k}}({x}^{k}-{\alpha}_{k}F({y}^{k}))$, it follows from Lemma 2.1 that

*i.e.*,

*F*, we have that $\{F({y}^{k})\}$ is bounded; that is, there exists a positive constant

*M*such that

*λ*, we have

Now, we consider the following two possible cases:

*F*is continuous and $\{{x}^{k}\}$ is bounded, which implies that the sequence $\{{x}^{k}\}$ has some accumulation point $\stackrel{\u02c6}{x}$ such that

This shows that $\stackrel{\u02c6}{x}$ is a solution of problem (1.1). Replacing ${x}^{\ast}$ by $\stackrel{\u02c6}{x}$ in the preceding argument, we obtain that the sequence $\{\parallel {x}^{k}-\stackrel{\u02c6}{x}\parallel \}$ is nonincreasing, and hence converges. Since $\stackrel{\u02c6}{x}$ is an accumulation point of $\{{x}_{k}\}$, some subsequence of $\{\parallel {x}^{k}-\stackrel{\u02c6}{x}\parallel \}$ converges to zero, which implies that the whole sequence $\{\parallel {x}^{k}-\stackrel{\u02c6}{x}\parallel \}$ converges to zero, and hence ${lim}_{k\to \mathrm{\infty}}{x}^{k}=\stackrel{\u02c6}{x}$.

*F*is continuous, we obtain by letting $j\to \mathrm{\infty}$ that

From (2.5) and (3.5), we conclude that $\lambda \ge 1$, which contradicts $\lambda \in (0,1)$. Hence, the case of ${lim}_{k\to \mathrm{\infty}}{t}_{k}=0$ is not possible. This completes the proof. □

**Remark 3.2** Compared to the conditions of the global convergence used in literatures [15, 16], our conditions are weaker.

## 4 Convergence rate

In this section, we provide a result on the convergence rate of the iterative sequence generated by Algorithm 2.1. To establish this result, we need the following conditions (4.1) and (4.2).

*δ*, ${c}_{1}$, and ${c}_{2}$ such that

*x*to solution set

*S*, and

*F*is differentiable and $\mathrm{\nabla}F(\cdot )$ is locally Lipschitz continuous with modulus $\theta >0$, then there exists a constant ${L}_{1}>0$ such that

In 1998, the literature [15] showed that their proposed method converged superlinearly when the underlying function *F* is monotone, differentiable with $\mathrm{\nabla}F({x}^{\ast})$ being nonsingular, and ∇*F* is locally Lipschitz continuous. It is known that the local error bound condition given in (4.1) is weaker than the nonsingular. Recently, under conditions (4.1), (4.2), and the underlying function *F* being monotone and continuous, the literature [16] obtained the locally superlinear rate of convergence of the proposed method.

Next, we analyze the superlinear convergence rate of the iterative sequence under a weaker condition. In the rest of section, we assume that ${x}^{k}\to {x}^{\ast}$, $k\to \mathrm{\infty}$, where ${x}^{\ast}\in S$.

**Lemma 4.1** *Let* $G\in {R}^{n\times n}$ *be a positive semidefinite matrix and* $\mu >0$. *Then*

(1) $\parallel {(G+\mu I)}^{-1}\parallel \le \frac{1}{\mu}$;

(2) $\parallel {(G+\mu I)}^{-1}G\parallel \le 2$.

*Proof* See [18]. □

**Lemma 4.2** *Suppose that* *F* *is continuous and satisfies conditions* (1.2), (4.1), *and* (4.2). *If there exists a positive constant* *N* *such that* $\parallel {G}_{k}\parallel \le N$ *for all* *k*, *then for all* *k* *sufficiently large*,

(1) ${c}_{3}\parallel {d}^{k}\parallel \le \parallel F({x}^{k})\parallel \le {c}_{4}\parallel {d}^{k}\parallel $;

(2) $\parallel F({x}^{k})+{G}_{k}{d}^{k}\parallel \le {c}_{5}{\parallel {d}^{k}\parallel}^{3/2}$, *where* ${c}_{3}$, ${c}_{4}$ *and* ${c}_{5}$ *are all positive constants*.

*Proof*For (1), let ${x}^{k}\in N({x}^{\ast},\frac{1}{2}\delta )$ and ${\stackrel{\u02c6}{x}}^{k}\in S$ be the closest solution to ${x}^{k}$. We have

*i.e.*, ${\stackrel{\u02c6}{x}}^{k}\in N({x}^{\ast},\delta )$. Thus, by (2.1), (2.2), (4.2), and Lemma 4.1, we have

We obtain that the left-hand side of (1) by setting ${c}_{3}:=\frac{{c}_{1}^{2}{\gamma}_{1}(1-{\kappa}_{0})}{{c}_{2}{M}_{1}+2{\gamma}_{1}{c}_{1}}$.

We obtain the right-hand side part by setting ${c}_{4}:=N+{\gamma}_{1}{M}_{1}+{\kappa}_{0}{\gamma}_{1}{M}_{1}$.

By setting ${c}_{5}:=(1+{\kappa}_{0}){\gamma}_{1}{c}_{4}^{1/2}$, we obtain the desired result. □

**Lemma 4.3**

*Suppose that the assumptions in Lemma*4.2

*hold*.

*Then for all*

*k*

*sufficiently large*,

*it holds that*

*Proof*By ${lim}_{k\to \mathrm{\infty}}{x}^{k}={x}^{\ast}$ and the continuity of

*F*, we have

*k*sufficiently large. Hence, it follows from (4.2) that

*k*sufficiently large, we obtain

where the last inequality follows from ${lim}_{k\to \mathrm{\infty}}F({x}^{k})=0$.

which implies that (2.3) holds with ${t}_{k}=1$ for all *k* sufficiently large, *i.e.*, ${y}^{k}={x}^{k}+{d}^{k}$. This completes the proof. □

From now on, we assume that *k* is large enough so that ${y}^{k}={x}^{k}+{d}^{k}$.

**Lemma 4.4**

*Suppose that the assumptions in Lemma*4.2

*hold*.

*Set*${\tilde{x}}^{k}:={x}^{k}-{\alpha}_{k}F({y}^{k})$.

*Then for all*

*k*

*sufficiently large*,

*there exists a positive constant*${c}_{6}$

*such that*

*Proof*Set

*k*sufficiently large, we have

where ${c}_{6}={c}_{4}^{1/2}({\gamma}_{2}+\frac{{c}_{2}}{{c}_{3}{\gamma}_{1}})$. This completes the proof. □

Now, we turn our attention to local rate of convergence analysis.

**Theorem 4.1** *Suppose that the assumptions in Lemma * 4.2 *hold*. *Then the sequence* $\{dist({x}^{k},S)\}$ *Q*-*superlinearly converges to* 0.

*Proof*By the definition of ${\tilde{x}}^{k}$, Lemma 4.2(1) and (4.4), for sufficiently large

*k*, we have

*k*sufficiently large, which, together with (4.2), Lemma 4.2, Lemma 4.4, and the definition of ${\tilde{x}}^{k}$, we obtain

*S*is the solution set of problem (1.1). Since ${x}^{k+1}={\mathrm{\Pi}}_{C\cap {H}_{k}}({\tilde{x}}^{k})$, it follows from Lemma 2.1 that

which implies that the order of superlinear convergence is at least 1.5. This completes the proof. □

**Remark 4.1** Compared with the proof of the locally superlinear convergence in literatures [15, 16], our conditions are weaker.

## 5 Numerical experiments

In this section, we present some numerical experiments results to show the efficiency of our method. The MATLAB codes are run on a notebook computer with CPU2.10GHZ under MATLAB Version 7.0. Just as done in [16], we take ${G}_{k}={F}^{\prime}({x}^{k})$ and use the left division operation in MATLAB to solve the system of linear equations (2.1) at each iteration. We choose $b=1$, $\lambda =0.96$, ${\kappa}_{0}=0$, $\beta =0.7$, and ${\gamma}_{1}=1$. ‘Iter.’ denotes the number of iteration and ‘CPU’ denotes the CPU time in seconds. We choose $\parallel F({x}^{k})\parallel \le {10}^{-6}$ as the stop criterion. The example is tested in [16].

**Example**Let

*C*be taken as

*a*. When we take $a=0$, the operation results are not best, that is to say, the direction $F({y}^{k})$ is not an optimal one.

**Numerical results of Example with**
$\mathit{a}\mathbf{=}{\mathbf{10}}^{\mathbf{-}\mathbf{15}}$

Initial point | Iter. | CPU | $\mathbf{\parallel}\mathit{F}\mathbf{(}{\mathbf{x}}^{\ast}\mathbf{)}\mathbf{\parallel}$ |
---|---|---|---|

(3,0,0,0) | 11 | 0.10 | 1.07 × 10 |

(1,1,0,0) | 13 | 0.09 | 1.62 × 10 |

(0,1,0,1) | 15 | 0.04 | 2.46 × 10 |

(0,0,0,1) | 21 | 0.18 | 9.92 × 10 |

(1,0,0,2) | 16 | 0.54 | 5.66 × 10 |

**Numerical results of Example with**
$\mathit{a}\mathbf{=}\mathbf{0}$

Initial point | Iter. | CPU | $\mathbf{\parallel}\mathit{F}\mathbf{(}{\mathbf{x}}^{\ast}\mathbf{)}\mathbf{\parallel}$ |
---|---|---|---|

(3,0,0,0) | 11 | 0.10 | 1.07 × 10 |

(1,1,0,0) | 13 | 0.12 | 1.62 × 10 |

(0,1,0,1) | 19 | 0.14 | 1.17 × 10 |

(0,0,0,1) | 18 | 0.18 | 1.44 × 10 |

(1,0,0,2) | 15 | 0.21 | 7.88 × 10 |

## Declarations

### Acknowledgements

The author wish to thank the anonymous referees for their suggestions and comments. This work is also supported by the Educational Science Foundation of Chongqing, Chongqing of China (Grant No. KJ111309).

## Authors’ Affiliations

## References

- Karamardian S:
**Complementarity problems over cones with monotone and pseudomonotone maps.***J. Optim. Theory Appl.*1976, 18: 445–454. 10.1007/BF00932654MathSciNetView ArticleMATHGoogle Scholar - Dirkse SP, Ferris MC:
**MCPLIB: a collection of nonlinear mixed complementarity problems.***Optim. Methods Softw.*1995, 5: 319–345. 10.1080/10556789508805619View ArticleGoogle Scholar - El-Hawary ME:
*Optimal Power Flow: Solution Techniques, Requirement and Challenges*. IEEE Service Center, Piscataway; 1996.Google Scholar - Meintjes K, Morgan AP:
**A methodology for solving chemical equilibrium system.***Appl. Math. Comput.*1987, 22: 333–361. 10.1016/0096-3003(87)90076-2MathSciNetView ArticleMATHGoogle Scholar - Wood AJ, Wollenberg BF:
*Power Generations, Operations, and Control*. Wiley, New York; 1996.Google Scholar - Bertsekas DP:
*Nonlinear Programming*. Athena Scientific, Belmont; 1995.MATHGoogle Scholar - Dennis JE, Schnabel RB:
*Numerical Methods for Unconstrained Optimization and Nonlinear Equations*. Prentice Hall, Englewood Cliffs; 1983.MATHGoogle Scholar - Ortega JM, Rheinboldt WC:
*Iterative Solution of Nonlinear Equations in Several Variables*. Academic Press, San Diego; 1970.MATHGoogle Scholar - Polyak BT:
*Introduction to Optimization*. Optimization Software, Inc. Publications Division, New York; 1987.MATHGoogle Scholar - Tong XJ, Qi L:
**On the convergence of a trust-region method for solving constrained nonlinear equations with degenerate solution.***J. Optim. Theory Appl.*2004, 123: 187–211.MathSciNetView ArticleMATHGoogle Scholar - Zhang JL, Wang Y:
**A new trust region method for nonlinear equations.***Math. Methods Oper. Res.*2003, 58: 283–298. 10.1007/s001860300302MathSciNetView ArticleMATHGoogle Scholar - Fan JY, Yuan YX:
**On the quadratic convergence of the Levenberg-Marquardt method without nonsingularity assumption.***Computing*2005, 74: 23–39. 10.1007/s00607-004-0083-1MathSciNetView ArticleMATHGoogle Scholar - Fan JY:
**Convergence rate of the trust region method for nonlinear equations under local error bound condition.***Comput. Optim. Appl.*2006, 34: 215–227. 10.1007/s10589-005-3078-8MathSciNetView ArticleMATHGoogle Scholar - Fan JY, Pan JY:
**An improved trust region algorithm for nonlinear equations.***Comput. Optim. Appl.*2011, 48: 59–70. 10.1007/s10589-009-9236-7MathSciNetView ArticleMATHGoogle Scholar - Solodov MV, Svaiter BF:
**A globally convergent inexact Newton method for systems of monotone equations.**In*Reformulation: Piecewise Smooth, Semismooth and Smoothing Methods*. Edited by: Fukushima M, Qi L. Kluwer Academic, Dordrecht; 1998:355–369.View ArticleGoogle Scholar - Wang CW, Wang YJ:
**A superlinearly convergent projection method for constrained systems of nonlinear equations.***J. Glob. Optim.*2009, 44: 283–296. 10.1007/s10898-008-9324-8View ArticleMathSciNetMATHGoogle Scholar - Zarantonello EH:
*Projections on Convex Sets in Hilbert Spaces and Spectral Theory*. Academic Press, New York; 1971.MATHGoogle Scholar - Zhou GL, Toh KC:
**Superlinear convergence of a Newton-type algorithm for monotone equations.***J. Optim. Theory Appl.*2005, 125: 205–221. 10.1007/s10957-004-1721-7MathSciNetView ArticleMATHGoogle Scholar

## Copyright

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.