# Nonlinear Conjugate Gradient Conjugate Gradient ¢ are provided for the conjugate gradient...

date post

01-Oct-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Nonlinear Conjugate Gradient Conjugate Gradient ¢ are provided for the conjugate gradient...

Nonlinear Conjugate Gradient Methods∗

Yu-Hong Dai

State Key Laboratory of Scientific and Engineering Computing,

Institute of Computational Mathematics and Scientific/Engineering Computing,

Academy of Mathematics and Systems Science, Chinese Academy of Sciences,

Zhong Guan Cun Donglu 55, Beijing, 100190, P.R. China.

E-mail: dyh@lsec.cc.ac.cn

Abstract

Conjugate gradient methods are a class of important methods for solving linear equations and for solving nonlinear optimization. In this article, a review on conjugate gradient methods for unconstrained optimization is given. They are divided into early conjugate gradi- ent methods, descent conjugate gradient methods and sufficient de- scent conjugate gradient methods. Two general convergence theorems are provided for the conjugate gradient method assuming the descent property of each search direction. Some research issues on conjugate gradient methods are mentioned.

Key words. conjugate gradient method, line search, descent prop- erty, sufficient descent condition, global convergence.

Mathematics Subject Classification: 49M37, 65K05, 90C30.

∗This work was partially supported by the Chinese NSF grants 10831106 and the CAS grant kjcx-yw-s7-03.

1

1 Introduction

Conjugate gradient methods are a class of important methods for solving unconstrained optimization problem

min f(x), x ∈ Rn, (1.1)

especially if the dimension n is large. They are of the form

xk+1 = xk + αkdk, (1.2)

where αk is a stepsize obtained by a line search, and dk is the search direction defined by

dk =

{ −gk, for k = 1; −gk + βkdk−1, for k ≥ 2,

(1.3)

where βk is a parameter, and gk denotes ∇f(xk). It is known from (1.2) and (1.3) that only the stepsize αk and the pa-

rameter βk remain to be determined in the definition of conjugate gradient methods. In the case that f is a convex quadratic, the choice of βk should be such that the method (1.2)-(1.3) reduces to the linear conjugate gradient method if the line search is exact, namely,

αk = arg min{f(xk + αdk);α > 0}. (1.4)

For nonlinear functions, however, different formulae for the parameter βk result in different conjugate gradient methods and their properties can be significantly different. To differentiate the linear conjugate gradient method, sometimes we call the conjugate gradient method for unconstrained opti- mization by nonlinear conjugate gradient method. Meanwhile, the parame- ter βk is called conjugate gradient parameter.

The linear conjugate gradient method can be dated back to a seminal paper by Hestenes and Stiefel [46] in 1952 for solving a symmetric posi- tive definite linear system Ax = b, where A ∈ Rn×n and b ∈ Rn. An easy and geometrical interpretation of the linear conjugate gradient method can be founded in Shewchuk [77]. The equivalence of the linear system to the minimization problem of 12 x

T Ax− bT x motivated Fletcher and Reeves [37] to extend the linear conjugate gradient method for nonlinear optimiza- tion. This work of Fletcher and Reeves in 1964 not only opened the door of nonlinear conjugate gradient field but greatly stimulated the study of non- linear optimization. In general, the nonlinear conjugate gradient method without restarts is only linearly convergent (see Crowder and Wolfe [16]),

2

while n-step quadratic convergence rate can be established if the method is restarted along the negative gradient every n-step (see Cohen [15] and MicCormick and Ritter [54]). Some recent reviews on nonlinear conjugate gradient methods can be found in Hager and Zhang [44], Nazareth [60, 61], Nocedal [62, 63], etc. This paper aims to provide a perspective view on the methods from the angle of descent property and global convergence.

Since the exact line search is usually expensive and impractical, the strong Wolfe line search is often considered in the implementation of non- linear conjugate gradient methods. It aims to find a stepsize satisfying the strong Wolfe conditions

f(xk + αkdk)− f(xk) ≤ ραk gTk dk, (1.5) |g(xk + αkdk)T dk| ≤ −σ gTk dk, (1.6)

where 0 < ρ < σ < 1. The strong Wolfe line search is often regarded as a suitable extension of the exact line search since it reduces to the latter if σ is equal to zero. In practical computations, a typical choice for σ that controls the inexactness of the line search is σ = 0.1.

On the other hand, for a general nonlinear function, one may be satisfied with a stepsize satisfying the standard Wolfe conditions, namely, (1.5) and

g(xk + αkdk)T dk ≥ σ gTk dk, (1.7)

where again 0 < ρ < σ < 1. As is well known, the standard Wolfe line search is normly used in the implementation of quasi-Newton methods, another important class of methods for unconstrained optimization. The work of Dai and Yuan [30, 33] indicates that the use of standard Wolfe line searches is possible in the nonlinear conjugate gradient field. Besides this, there are quite a few references (for example, see [19, 41, 81, 93]) that deal with Armijo-type line searches.

A requirement for an optimization method to use the above line searches is that, the search direction dk must have the descent property, namely,

gTk dk < 0. (1.8)

For conjugate gradient methods, by multiplying (1.3) with gTk , we have

gTk dk = −‖gk‖2 + βk gTk dk−1. (1.9)

Thus if the line search is exact, we have gTk dk = −‖gk‖2 since gTk dk−1 = 0. Consequently, dk is descent provided gk 6= 0. However, this may not be

3

true in case of inexact line searches for early conjugate gradient methods. A simple restart with dk = −gk may remedy these bad situations, but will probably degrade the numerical performance since the second derivative information along the previous direction dk−1 is discarded (see [68]). Assume that no restarts are used. In this paper we say that, a conjugate gradient method is descent if (1.8) holds for all k, and is sufficient descent if the sufficient descent condition

gTk dk ≤ −c ‖gk‖2, (1.10)

holds for all k and some constant c > 0. However, we have to point out that the borderlines between these conjugate gradient methods are not strict (see the discussion at the beginning of § 5).

This survey is organized in the following way. In the next section, we will address two general convergence theorems for the method of the form (1.2)- (1.3) assuming the descent property of each search direction. Afterwards, we divide conjugate gradient methods into three categories: early conjugate gradient methods, descent conjugate gradient methods and sufficient descent conjugate gradient methods. They will be discussed in Sections 3 to 5, respectively, with the emphases on the Fletcher-Reeves method, the Polak- Ribière-Polyak method, the Hestenes-Stiefel method, the Dai-Yuan method and the CG DESCENT method by Hager and Zhang. Some research issues on conjugate gradient methods are mentioned in the last section.

2 General convergence theorems

In this section, we give two global convergence theorems for any method of the form (1.2)-(1.3) assuming the descent condition (1.8) for all k. The first one deals with the strong Wolfe line search, while the second treats the standard Wolfe line search.

At first, we give the following basic assumptions on the objective func- tion. Throughout this paper, the symbol ‖ · ‖ denotes the two norm. Assumption 2.1. (i) The level set L = {x ∈ Rn : f(x) ≤ f(x1)} is bounded, where x1 is the starting point; (ii) In some neighborhood N of L, f is continuously differentiable, and its gradient is Lipschitz continuous; namely, there exists a constant L > 0 such that

‖g(x)− g(y)‖ ≤ L ‖x− y‖, for all x, y ∈ N . (2.1)

4

Sometimes, the boundedness assumption for L in item (i) is unneces- sary and we only require that f is bounded below in L. However, we will just use Assumption 2.1 for the convergence results in this survey. Under Assumption 2.1 on f , we state a very useful result, which was obtained by Zoutendijk [94] and Wolfe [83, 84]. The relation (2.2) is usually called as the Zoutendijk condition.

Lemma 2.2. Suppose that Assumption 2.1 holds. Consider any iterative method of the form (1.2), where dk satisfies gTk dk < 0 and αk is obtained by the standard Wolfe line search. Then we have that

∞∑

k=1

(gTk dk) 2

||dk||2 < +∞. (2.2)

To simplify the statements of the following results, we assume that gk 6= 0 for all k for otherwise a stationary point has been found. Assume also that βk 6= 0 for all k. This is because if βk = 0, the direction in (1.3) reduces to the negative gradient direction. Thus either the method converges globally if βk = 0 for infinite number of k, or one can take some xk as the new initial point. In addition, we say that a method is globally convergent if

lim inf k→∞

‖gk‖ = 0, (2.3)

and is strongly convergent if

lim k→∞

‖gk‖ = 0, (2.4)

If the iterations {xk} stay in a bounded region, (2.3) means that there exists at least one cluster point which is a stationary point of f , while (2.4) indicates that every cluster point of {xk} will be a stationary point of f .

To analyze the method of the form (1.2)-(1.3), besides (1.9), we derive another basic relation. By (1.3), we have dk + gk = βkdk−1 for all k ≥ 2. Squaring both sides of this relation yields

‖dk‖2 = −2gTk dk − ‖gk‖2 + β2k‖dk−1‖2. (2.5)

The following theorem give

*View more*