A SEMISMOOTH EQUATION APPROACH TO THE SOLUTION OF NONLINEAR COMPLEMENTARITY PROBLEMS Tecla De Luca 1, Francisco Facchinei 1 and Christian Kanzow 2 1 Universita di Roma \La Sapienza" Dipartimento di Informatica e Sistemistica Via Buonarroti 12, 00185 Roma, Italy e-mail (De Luca): deluca@newton.dis.uniroma1.it e-mail (Facchinei): soler@peano.dis.uniroma1.it 2 University of Hamburg Institute of Applied Mathematics Bundesstrasse 55, D-20146 Hamburg, Germany e-mail: kanzow@math.uni-hamburg.de Abstract: In this paper we present a new algorithm for the solution of nonlinear complementarity problems. The algorithm is based on a semismooth equation reformulation of the complementarity problem. We exploit the recent extension of Newton's method to semismooth systems of equations and the fact that the natural merit function associated to the equation reformulation is continuously dierentiable to develop an algorithm whose global and quadratic convergence properties can be established under very mild assumptions. Other interesting features of the new algorithm are an extreme simplicity along with a low computational burden per iteration. We include numerical tests which show the viability of the approach. Key Words: Nonlinear complementarity problem, semismoothness, smooth merit function, global convergence, quadratic convergence.
1 Introduction We consider the nonlinear complementarity problem, NCP(F ) for short, which is to nd a vector in IR n satisfying the conditions x 0; F (x) 0; x T F (x) = 0; where F : IR n! IR n is a given function which we shall always assume to be continuously dierentiable. The fastest algorithms for the solution of NCP(F ) are Newton-type methods, which, however, are in general only locally convergent. In the last years much attention has been devoted to techniques for globalizing these local algorithms. To this end there have been several dierent proposals, but a common scheme can be seen to underlie most of these methods: (a) reformulate NCP(F ) as a system of (possibly nonsmooth) equations (x) = 0; (b) dene a local Newton-type method for the solution of the system of equations; (c) perform a linesearch to minimize a suitable merit function, usually k(x)k 2 or k(x)k, in order to globalize the local method. There are several possibilities to redene NCP(F ) as a system of equations. The rst proposal is probably due to Mangasarian, see [29], where a class of reformulations, mainly smooth ones, is described; the smooth reformulation approach has been further explored in [9, 15, 25, 26, 27, 30, 32, 49]. In the last years, however, nonsmooth reformulations have attracted much more attention [5, 7, 8, 10, 14, 16, 18, 20, 32, 34, 35, 36, 41, 46, 54, 55], since they enjoy some denite advantages: they allow to dene superlinearly convergent algorithms also for degenerate problems, and the subproblems to be solved at each iteration tend to be more numerically stable. However, there is a price to pay: the globalization (point (c) above) becomes more complex since the merit function k(x)k 2 is nonsmooth. Furthermore, the subproblems which have to be solved at each iteration, usually (mixed) linear complementarity problems, are more complex than in the smooth reformulation case. We note that a common analytical property of the recent nonsmooth reformulations is that, though not F-dierentiable, they are B-dierentiable, so that the machinery developed in [34, 43, 44, 45] can be usefully employed. In this paper we describe a new algorithm for the solution of NCP(F ) which is both globally and superlinearly convergent. We use a nonsmooth equation reformulation which is based on the simple function (a; b) := p a 2 + b 2? (a + b) which was rst introduced by Fischer [11] and further employed by several authors, see [7, 8, 12, 16, 25, 27, 39, 52]. The resulting system of equations is not F-dierentiable, but nevertheless it is semismooth [31, 40]. Semismoothness is a stronger analytical property than B-dierentiability so that, using the recent powerful theory for the solution of semismooth equations [37, 38, 40], it is possible to develop a fast local algorithm for the solution of NCP(F ) which only requires, at each iteration, the solution of one linear system. Furthermore, the natural merit function k(x)k 2 is, surprisingly, smooth, so that global convergence through a linesearch procedure can 2
easily be enforced. The assumptions under which global and superlinear convergence can be proved compare favourably with those of existing algorithms and the overall algorithm seems to convey the advantages of both smooth and nonsmooth reformulations of NCP(F ) in a very simple scheme. The use of the function mentioned above to reformulate nonlinear complementarity problems as systems of nonsmooth equations seems a very promising approach. Algorithms for linear complementarity problems based on this function were considered in [12, 27], while the nonlinear case was studied in [7, 8, 16, 25]. The results of this paper, however, are an improvement on previous results because the semismooth nature of the reformulation is here fully exploited for the rst time and the overall analysis carried out is much more rened and detailed than in previous works. While we were completing this paper, we became aware of a recent report of Jiang and Qi [23] which has some similarities with this work in that the same local approach is adopted. However, the global algorithm is quite dierent and the analysis is restricted to the case in which F is a uniform P -function. This paper is organized as follows, in the next section we collect some background material, in Section 3 we state the algorithm and its main properties, Sections 4 and 5 are devoted to some technical, but fundamental results, while in Section 6 we prove and further discuss the properties of the algorithm. In Section 7 we report some preliminary numerical results to show the viability of the approach adopted. Some concluding remarks are made in the last section. Throughout this paper, the index set f1; : : : ; ng is abbreviated by the upper case letter I: For a continuously dierentiable function F : IR n! IR n ; we denote the Jacobian of F at x 2 IR n by F 0 (x); whereas the transposed Jacobian is denoted by rf (x): k k denotes the Euclidean norm and S(x; ) is the closed Euclidean sphere of center x and radius, i.e. S(x; ) = fx 2 IR n : kx? xk g. If is a subset of IR n, distfxjg := inf y2 ky? xk denotes the (Euclidean) distance of x to. If M is an n n matrix with elements M jk, j; k = 1; : : : n, and J and K are index sets such that J; K f1; : : : ng, we denote by M J;K the jjj jkj submatrix of M consisting of elements M jk, j 2 J, k 2 K and, assuming that M J;J is nonsingular, by M=M J;J the Schur-complement of M J;J in M, i.e. M=M J;J = M K;K? M K;J M J;JM?1 J;K, where K = I n J. If w is an n vector, we denote by w J the subvector with components w j, j 2 J. 2 Background Material In this section we collect several denitions and results which will be used throughout the paper. 3
2.1 Basic Denitions A solution to the nonlinear complementarity problem NCP(F ) is a vector x 2 IR n such that F (x ) 0; x 0; F (x ) T x = 0: Associated to the solution x we dene three index sets: := fijx i > 0g; := fijx i = 0 = F i (x )g; := fijf i (x ) > 0g The solution x is said to be nondegenerate if = ;. Denition 2.1 We say that the solution x is R-regular if rf ; (x ) is nonsingular and the Schur-complement of rf ; (x ) in! is a P -matrix (see below). rf ; (x ) rf ; (x ) rf ; (x ) rf ; (x ) Note that R-regularity coincides with the notion of regularity introduced by Robinson in [47] (see also [42], where the same condition is called strong regularity) and is strictly related to similar conditions used, e.g., in [10, 32, 36]. We shall also employ several denitions concerning matrices and functions and need some related properties. Denition 2.2 A matrix M 2 IR nn is a - P 0 -matrix if every of its principal minors is non-negative; - P -matrix if every of its principal minors is positive; - S 0 -matrix if fx 2 IR n j x 0; x 6= 0; Mx 0g 6= ;: It is obvious that every P -matrix is also a P 0 -matrix and it is known [4] that every P 0 -matrix is an S 0 -matrix. We shall also need the following characterization of P 0 -matrices [4]. Proposition 2.1 A matrix M 2 IR nn is a P 0 -matrix i for every nonzero vector x there exists an index i such that x i 6= 0 and x i (Mx) i 0. Denition 2.3 A function F : IR n! IR n is a - P 0 -function if, for every x and y in IR n with x 6= y, there is an index i such that x i 6= y i ; (x i? y i )[F i (x)? F i (y)] 0: 4
- P -function if, for every x and y in IR n with x 6= y, there is an index i such that (x i? y i )[F i (x)? F i (y)] > 0: - uniform P -function if there exists a positive constant such that, for every x and y in IR n, there is an index i such that (x i? y i )[F i (x)? F i (y)] kx? yk 2 : It is obvious that every uniform P -function is a P -function and that every P -function is a P 0 -function. Furthermore, it is known that the Jacobian of every continuously dierentiable P 0 -function is a P 0 -matrix and that if the Jacobian of a continuously dierentiable function is a P -matrix for every x, then the function is a P -function. If F is ane, that is if F (x) = Mx + q, then F is a P 0 -function i M is a P 0 -matrix, while F is a (uniform) P -function i M is a P -matrix (note that in the ane case the concept of uniform P -function and P -function coincide). We nally note that every monotone function is a P 0 -function, every strictly monotone function is a P -function and that every strongly monotone function is a uniform P -function. 2.2 Dierentiability of Functions and Generalized Newton's Method Let G : IR n! IR n be locally Lipschitzian; by Rademacher's theorem G is dierentiable almost everywhere. If we indicate by D G the set where G is dierentiable, we can dene the B-subdierential of G at x [38] as @ B G(x) := and the Clarke subdierential of G at x as lim x k 2D G x k!x G 0 (x k ); @G(x) := co @ B G(x); where co denotes the convex hull of a set. Semismooth functions were introduced in [31] and immediately shown to be relevant to optimization algorithms. Recently the concept of semismoothness has been extended to vector valued functions [40]. Denition 2.4 Let G : IR n! IR n be locally Lipschitzian at x 2 IR n. We say that G is semismooth at x if lim Hv 0 (1) exists for any v 2 IR n. H2@G(x+tv 0 ) v 0!v;t#0 5
Semismooth functions lie between Lipschitz functions and C 1 functions. Note that this class is strictly contained in the class of B-dierentiable functions. Furthermore it is known that if G is semismooth at x then it is also directionally dierentiable there, and its directional derivative in the direction v is given by (1). A slightly stronger notion than semismoothness is strong semismoothness, de- ned below (see [38] and [40] where, however, dierent names are used). Denition 2.5 Suppose that G is semismooth at x. We say that G is strongly semismooth at x if for any H 2 @G(x + d), and for any d! 0, Hd? G 0 (x; d) = O(kdk 2 ): In the study of algorithms for the local solution of semismooth systems of equations, the following regularity condition plays a role similar to that of the nonsingularity of the Jacobian in the study of algorithms for smooth systems of equations. Denition 2.6 We say that a semismooth function G : IR n! IR n is BD-regular at x if all the elements in @ B G(x) are nonsingular. A generalized Newton method for the solution of a semismooth system of n equations G(x) = 0 can be dened as x k+1 = x k? (H k )?1 G(x k ); H k 2 @ B G(x k ); (2) (H k can be any element in @ B G(x k )). The following result holds [38]. Theorem 2.2 Suppose that x is a solution of the system G(x) = 0 and that G is semismooth and BD-regular at x. Then the iteration method (2) is well dened and convergent to x superlinearly in a neighborhood of x. If, in addition, G is directionally dierentiable in a neighborhood of x and strongly semismooth at x, then the convergence rate of (2) is quadratic. We nally give the denition of SC 1 function. Denition 2.7 A function f : IR n! IR is an SC 1 function if f is continuously dierentiable and its gradient is semismooth. SC 1 functions can be viewed as functions which lie between C 1 and C 2 functions. 2.3 Reformulation of NCP(F ) Our reformulation of NCP(F ) as a system of equations is based on the following two variables convex function: (a; b) := p a 2 + b 2? (a + b): The most interesting property of this function is that, as it is easily veried, (a; b) = 0 () a 0; b 0; ab = 0; (3) 6
note also that is continuously dierentiable everywhere but in the origin. The function was introduced by Fischer [11] in 1992, since then it has attracted the attention of many researchers and it has proved to be a valuable tool in nonlinear complementarity theory [7, 8, 12, 16, 22, 23, 25, 27, 39, 52]. An up-to-date review on the uses of the function can be found in [13]. Exploiting (3) it is readily seen that the nonlinear complementarity problem is equivalent to the following system of nonsmooth equations: (x) := 2 6 4 (x 1 ; F 1 (x)). (x i ; F i (x)). (x n ; F n (x)) It is then also obvious that the nonnegative function (x) := 1 2 k(x)k2 = 1 2 nx i=1 3 7 5 = 0: (x i ; F i (x)) 2 is zero at a point x if and only if x is a solution of NCP(F ), so that solving NCP(F ) is equivalent to nding the unconstrained global solutions of the problem fmin (x)g. The following results have been proven in [7] or easily follow from [39, Lemma 3.3]. Part (c) has also been shown in the recent paper [22]. Theorem 2.3 It holds that (a) is semismooth everywhere; furthermore, if every F i is twice continuously differentiable with Lipschitz continuous Hessian, then is strongly semismooth everywhere; (b) is continuously dierentiable; furthermore, if every F i is an SC 1 function, then also is an SC 1 function; (c) If F is a uniform P -function then the level sets of are bounded. 3 The Algorithm In this section we present the algorithm for the solution of NCP(F ). Roughly speaking this algorithm can be seen as an attempt to solve the semismooth system of equations (x) = 0 by using the generalized Newton's method described in the previous section (see (2)). To ensure global convergence, a linesearch is performed to minimize the smooth merit function ; if the search direction generated according to (2) is not a \good" descent direction, we resort to the negative gradient of. Global Algorithm 7
Data: x 0 2 IR n, > 0, p > 2, 2 (0; 1=2), " 0. Step 0: Set k = 0. Step 1: (stopping criterion) If kr (x k )k " stop. Step 2: (search direction calculation) Select an element H k in @ B (x k ). Find the solution d k of the system If system (4) is not solvable or if the condition is not satised, set d k =?r (x k ). H k d =?(x k ): (4) r (x k ) T d k?kd k k p (5) Step 3: (linesearch) Find the smallest i k = 0; 1; 2; : : : such that (x k + 2?ik d k ) (x k ) + 2?ik r (x k ) T d k : (6) Set x k+1 = x k + 2?ik d k, k k + 1 and go to Step 1. We note that the above algorithm is virtually indistinguishable from a global algorithm for the solution of an F-dierentiable system of equations. Formally, the only point which requires some care is the calculation of an element H belonging to @ B (x k ) in Step 2. This, however, turns out to be an easy and cheap task, as it will be discussed in Section 7. Another point which is worth of attention is that, if it exists, the direction obtained by solving (4) is always a descent direction for the function, unless (x k ) = 0. This is a standard property of the Newton direction for the solution of a smooth system of equations, but it is no longer true, in general, when the system of equations is nonsmooth. To see that this assertion is true, it is sucient to note that, as it is stated in Theorem 2.3 (b), is dierentiable at x k ; on the other hand, applying standard nonsmooth calculus rules (see [3]), we obtain @ (x k ) = fr (x k )g = V T (x k ) for every V 2 @(x k ); so that the expression in the rightest hand side is independent of the element V 2 @(x k ) chosen. Hence we can take V = H k and write, taking into account (4), r (x k ) T d k = (x k ) T H k d k =?k(x k )k 2 : This fact clearly shows, once again, the similarity of our algorithm with Newton's method for the solution of a system of equations; furthermore, in our opinion, it also indicates that the direction (4) is a \good" search direction also when far from a solution of the system (x) = 0, which is not true for a general semismooth system. 8
The stopping criterion at Step 1 can be substituted by any other criterion without changing the properties of the algorithm. In what follows, as usual in analyzing the behaviour of algorithms, we shall assume that " = 0 and that the algorithm produces an innite sequence of points. The following result, summarizing the main properties of the algorithm, will be proved in Section 6, where we shall also discuss the properties of the algorithm in greater detail. Theorem 3.1 It holds that (a) Each accumulation point of the sequence fx k g generated by the algorithm is a stationary point of. (b) If one of the limit points of the sequence fx k g, let us say x, is an isolated solution of NCP(F ), then fx k g! x. (c) If one of the limit points of the sequence fx k g, let us say x, is a BD-regular solution of the system (x) = 0, and if each F i is an SC 1 function, then fx k g! x and 1. Eventually d k is always given by the solution of system (4) (i.e., the negative gradient is never used eventually). 2. Eventually the stepsize of one is always accepted so that x k+1 = x k + d k. 3. The convergence rate is superlinear; furthermore if each F i is twice continuously dierentiable with Lipschitz continuous Hessian, then the convergence rate is quadratic. The results stated in this theorem are somewhat \crude" and need to be completed by answering the following two questions. Under which conditions is a stationary point x of the function a global solution and hence a solution of NCP(F )? In the next section sucient and necessary-and-sucient conditions will be established in order to ensure this key property. The second question is: under which conditions is a solution of NCP(F ) a BD-regular solution of the system (x) = 0? In fact, according to Theorem 3.1, a superlinear (at least) convergence rate of the algorithm can be ensured under the assumption of BD-regularity of solutions of the system (x) = 0. More in general, the possibility of using the (hopefully) good \second order" search direction (4) even far from a solution is closely related to this issue. Conditions guaranteeing the nonsingularity of all the elements in @(x) will be analyzed in Section 5. We note that, as discussed in [38], other strategies are possible to globalize the local Newton method (2) for the system (x) = 0. However the one we chose here, i.e., using the gradient of in \troublesome" situations, appears to be by far the simplest choice. The possibility of using the gradient, in turn, heavily depends of the surprising fact that kk 2 is continuously dierentiable, which is not true, in general, for the square of a semismooth system of equations. This peculiarity of the system (x) = 0 paves the way for a simple extension of practically any 9
standard globalization technique used in the solution of smooth systems of equations: Levenberg-Marquardt methods, trust region methods etc. In this paper we have chosen the simplest of these techniques since we were mainly interested in showing the potentialities of our approach and its numerical viability. 4 Regularity Conditions In this section we give some (necessary-and-) sucient conditions for stationary points of our merit function to be solutions of NCP(F ). We call these conditions regularity conditions. We rst recall a result of Facchinei and Soares [7, Proposition 3.1]. Lemma 4.1 For an arbitrary x 2 IR n ; we have @(x) T D a (x) + rf (x)d b (x); where D a (x) = diag(a 1 (x); : : : ; a n (x)); D b (x) = diag(b 1 (x); : : : ; b n (x)) 2 IR nn diagonal matrices whose i-th diagonal element is given by a i (x) = x i k(x i ; F i (x))k? 1; b i(x) = F i (x) k(x i ; F i (x))k? 1 are if (x i ; F i (x)) 6= 0; and by a i (x) = i? 1; b i (x) = i? 1 for every ( i ; i ) 2 IR 2 such that k( i ; i )k 1 if (x i ; F i (x)) = 0: In the following, for simplicity, we sometimes suppress the dependence on x in our notation. For example, the gradient of the dierentiable function can be written as follows: r (x) = D a (x)(x) + rf (x)d b (x)(x) = D a + rf D b : In the analysis of this section, the signs of the vectors D a 2 IR n and D b 2 IR n play an important role. We therefore introduce the index sets C := fi 2 Ij x i 0; F i (x) 0; x i F i (x) = 0g R := I n C and further partition the index set R as follows: (complementary indices), (residual indices), P := fi 2 Rj x i > 0; F i (x) > 0g N := R n P (positive indices), (negative indices). Note that these index sets depend on x, but that the notation does not reect this dependence. However, this should not cause any confusion because the given vector x 2 IR n will always be clear from the context. 10
The names P and N of the above index sets are motivated by the following simple relations, which can be easily veried: (D a ) i > 0 () (D b ) i > 0 () i 2 P; (D a ) i = 0 () (D b ) i = 0 () i 2 C; (D a ) i < 0 () (D b ) i < 0 () i 2 N ; (7) in particular, the signs of the corresponding elements of D a and D b are the same. The following denition of regular vector is motivated by some similar denitions in the recent papers of Pang and Gabriel [36, Denition 1], More [32, Denition 3.1] and Ferris and Ralph [10, Denition 2.4]. Denition 4.1 A point x 2 IR n is called regular if for every vector z 6= 0 such that z C = 0; z P > 0; z N < 0; (8) there exists a vector y 2 IR n such that y P 0; y N 0; y R 6= 0 and y T rf (x)z 0: (9) It is dicult to compare our denition of regular vector with the corresponding ones in [10, 32, 36] since we use dierent index sets (the denition of the index sets depends heavily on the merit function chosen). Nevertheless, we think that the following points should be remarked: (a) As pointed out by More [32], the denition of regularity given by Pang and Gabriel [36] depends on the scaling of the function F; i.e., a vector x 2 IR n could be regular for NCP(F ) in the sense of Pang and Gabriel [36], but not for the equivalent problem NCP(F s ), where F s (x) = sf (x) for some positive scaling parameter s: Obviously our denition is independent of the scaling of F: (b) In the related denition of regularity given by More [32] and Ferris and Ralph [10], a condition similar to (9) is employed. However, they have to assume y T rf (x)z > 0 instead of (9). In this respect, our denition of regularity seems to be weaker. For example, if rf (x) is a positive semidenite matrix, then we can choose y = z and directly obtain from Denition 4.1 that x is a regular vector. More in general, whenever directly comparable, our regularity conditions appear to be weaker than those introduced in [10, 32, 36]. The following result shows why the notion of regular vector is so important. 11
Theorem 4.2 A vector x 2 IR n is a solution of NCP(F ) if and only if x is a regular stationary point of : Proof. First assume that x 2 IR n is a solution of NCP(F ). Then x is a global minimum of and hence a stationary point of by the dierentiability of this function. Moreover, P = N = ; in this case, and therefore the regularity of x holds vacuously since z = z C, and there exists no nonzero vector z satisfying conditions (8). Suppose now that x is regular and that r (x) = 0: As mentioned at the beginning of this section, the stationary condition can be rewritten as Consequently, we have D a + rf (x)d b = 0: y T (D a ) + y T rf (x)(d b ) = 0 (10) for any y 2 IR n : Assume that x is not a solution of NCP(F ). Then R 6= ; and hence, by (7), z := D b is a nonzero vector with z C = 0; z P > 0; z N < 0: Recalling that the components of D a and z = D b have the same signs, and taking y 2 IR n from the denition of regular vector, we have (since y R 6= 0), and y T (D a ) = y T C (D a ) C + y T P(D a ) P + y T N (D a ) N > 0 (11) y T rf (x)(d b ) = y T rf (x)z 0: (12) The inequalities (11) and (12) together, however, contradict condition (10). Hence R = ;: This means that x is a solution of NCP(F ). 2 From Theorem 4.2 and the remark (b) after Denition 4.1, we directly obtain that a stationary point x is a solution of NCP(F ) if the Jacobian matrix F 0 (x) is positive semidenite. In particular, we have the result that all stationary points of are solutions of NCP(F ) for monotone functions F; i.e., we reobtain in this way a recent result of Geiger and Kanzow [16, Theorem 2.5]. However, in the following we present some weaker conditions which also ensure that stationary points are global minimizers of and hence solutions of NCP(F ). Let T 2 IR jrjjrj denote the diagonal matrix T = diag(t 1 ; : : : ; t jrj ) with entries ( +1 if i 2 P; t i :=?1 if i 2 N ; and note that T T = I: Using this notation, we can prove the following result. 12
Theorem 4.3 Let x 2 IR n be a stationary point of ; and assume that the matrix T F 0 (x) RR T is an S 0 -matrix. Then x is a solution of NCP(F ). Proof. We shall prove that x is regular. The result then follows from Theorem 4.2. By the denition of S 0 -matrix and the assumptions made, there exists a nonzero vector ~y R such that ~y R 0 and T F 0 (x) RR T ~y R 0: (13) Let y be the unique vector such that by the denition of T and (13) we have that y R 6= 0 and y C = 0; y R = T ~y R ; (14) y P 0; y N 0: Then it is easy to see that, for every z 2 IR n such that z 6= 0 and we can write, recalling (14), z C = 0; z P > 0; z N < 0; y T rf (x)z = yrrf T (x) RR z R = yr(t T T )rf (x) RR (T T )z R = (yr T T )(T rf (x) RRT )(T z R ) = ~y RT T (F 0 (x) RR ) T T (T z R ) 0; (15) where the last inequality follows by (13) and the fact that (T z R ) > 0. 2 In the following corollary, we summarize some simple, but noteworthy consequences of Theorem 4.3. First recall that a vector x 2 IR n is called feasible for NCP(F ) if x 0 and F (x) 0: Corollary 4.4 It holds that (a) If x 2 IR n is a stationary point of such that the submatrix F 0 (x) RR is a P 0 -matrix, then x is a solution of NCP(F ). (b) If F : IR n! IR n is a P 0 -function, then all the stationary points of are solutions of NCP(F ). (c) If x 2 IR n is a feasible stationary point of such that the submatrix F 0 (x) RR is an S 0 -matrix, then x is a solution of NCP(F ). Proof. (a) Since a square matrix M is a P 0 -matrix if and only if the matrix DMD is a P 0 -matrix for all nonsingular diagonal matrices D; the assertion is a direct consequence of Theorem 4.3 and the fact that every P 0 -matrix is also an S 0 -matrix. (b) Since F is a continuously dierentiable P 0 -function, its Jacobian F 0 (x) is a P 0 -matrix for all x 2 IR n ; see More and Rheinboldt [33, Theorem 5.8]. The 13
second assertion is therefore a direct consequence of part (a) since, by denition, all principal submatrices of a P 0 -matrix are also P 0 -matrices. (c) Since x is assumed to be a feasible vector, we have N = ;: Hence the diagonal matrix T introduced before Theorem 4.3 reduces to the identity matrix, and part (c) follows directly from Theorem 4.3. 2 Note that part (b) of Corollary 4.4 has recently been proved also by Facchinei and Soares [7, Theorem 4.1]. Before proving our next result, we introduce some notation. Without loss of generality, assume that the Jacobian F 0 (x) be partitioned in the following way: and dene F 0 (x) = 0 B @ F 0 (x) CC F 0 (x) CP F 0 (x) CN F 0 (x) PC F 0 (x) PP F 0 (x) PN F 0 (x) N C F 0 (x) N P F 0 (x) N N J(x) := T F 0 (x)t; where the transformation matrix T has the diagonal structure T := 0 B @ +I C 0 +I P 0?I N (if C = ;; this matrix coincides with the matrix T introduced before Theorem 4.3, so we use the same symbol here). Then we have J(x) = 0 B @ 1 C A F 0 (x) CC F 0 (x) CP?F 0 (x) CN F 0 (x) PC F 0 (x) PP?F 0 (x) PN?F 0 (x) N C?F 0 (x) N P F 0 (x) N N 1 C A ; 1 C A : (16) Transformed Jacobian matrices of this kind were also considered by Pang and Gabriel [36], More [32] and Ferris and Ralph [10]. Using this matrix, we can prove the following result. Theorem 4.5 Let x 2 IR n be a stationary point of ; and assume that the submatrix F 0 (x) CC is nonsingular and the Schur-complement of this matrix in J(x) is an S 0 - matrix. Then x is a solution of NCP(F ). Proof. We rst show that there exist ~y P ; ~y N 0 with ~y R 6= 0 and q P ; q N 0 such that J(x)~y = q; (17) where ~y = (~y C ; ~y P ; ~y N ) T ; q = (0 C ; q P ; q N ) T 2 IR n : In view of (16), system (17) can be rewritten as F 0 (x) CC ~y C + J(x) CR ~y R = 0 C ; (18) J(x) RC ~y C + J(x) RR ~y R = q R : (19) 14
Solving the rst equation for ~y C yields ~y C =?F 0 (x)?1 CC J(x) CR ~y R : (20) Substituting this into (19) and rearranging leads to J(x)RR? J(x) RC F 0 (x)?1 CC J(x) CR ~yr = q R : (21) Since, by assumption, the matrix of the linear system (21) is an S 0 -matrix, there exists an ~y R 0; ~y R 6= 0; and a q R 0 satisfying this system. Then dene ~y C by (20) and set y := T ~y: Then y C = ~y C ; y P = ~y P 0; y N =?~y N 0 and y R 6= 0; i.e., the vector y 2 IR n satises all the conditions required in Denition 4.1 for a regular vector x: Now let z 6= 0 be an arbitrary vector with z C = 0; z P > 0 and z N < 0: Then ~z := T z satises the conditions ~z C = 0 C ; ~z P > 0; ~z N > 0; and we therefore obtain from (17) and the fact that T T = I : y T rf (x)z = z T F 0 (x)y = z T T T F 0 (x)t T y = ~z T J(x)~y = ~z T q = ~z T Pq P + ~z T N q N 0; i.e., x is regular. 2 If C = ;; then Theorem 4.5 reduces to Theorem 4.3. We note that similar results were shown in [10, 32, 36], but in all these papers a certain Schur-complement is assumed to be an S-matrix, which, again, is a stronger assumption than the one used in our Theorem 4.5. In order to illustrate our theory, consider NCP(F ) with F : IR 4! IR 4 being dened by 0 3x 2 + 2x 1 1x 2 + 2x 2 + x 1 2 3 + 3x 4? 6 F (x) := 2x B 2 1 + x 1 + x 2 2 + 3x 3 + 2x 4? 2 @ C 3x 2 1 + x 1 x 2 + 2x 2 2 + 2x 3 + 3x 4? 1 A : x 2 1 + 3x 2 2 + 2x 3 + 3x 4? 3 This small example is due to Kojima [28] and is a standard test problem for nonlinear complementarity codes. Harker and Xiao [20] report a failure of their method for this example. Their method converges to the point ^x = (1:0551; 1:3347; 1:2681; 0) T ; where F (^x) = (4:9873; 6:8673; 9:8472; 5:9939) T ; 15
which is obviously not a solution of NCP(F ). The corresponding Jacobian matrix is 0 1 9:0001 7:4491 1 3 B F 0 5:2204 2:6695 3 2 C (^x) = B @ C 7:6653 6:3940 2 3 A : 2:1102 8:0084 2 3 Obviously, F 0 (^x) is an S 0 -matrix. Since ^x is also feasible, Corollary 4.4 (c) guarantees that our method will not terminate at ^x: 5 Nonsingularity Conditions In this section we study conditions which guarantee that all elements in @(x) are nonsingular. We begin with three lemmas that will be needed later in this section. Lemma 5.1 If M 2 IR nn is a P 0 -matrix, then every matrix of the form D a + D b M is nonsingular for all positive denite (negative denite) diagonal matrices D a ; D b 2 IR nn. Proof. We only consider the case in which the two matrices are positive definite, the other case is analogous. Assume that M is a P 0 -matrix and D a := diag(a 1 ; : : : ; a n ); D b := diag(b 1 ; : : : ; b n ) are positive denite diagonal matrices. Then (D a + D b M)q = 0 for some q 2 IR n implies D a q =?D b Mq and therefore q i =?(b i =a i )(Mq) i (i 2 I): Since M is a P 0 -matrix, we get from Proposition 2.1 that q = 0; i.e., the matrix D a + D b M is nonsingular. 2 We note that it is also possible to prove the converse direction in the above lemma, i.e., the matrix D a + D b M is nonsingular for all positive (negative) denite diagonal matrices D a ; D b 2 IR nn if and only if M is a P 0 -matrix. In the following, however, we only need the direction stated in Lemma 5.1. Lemma 5.2 Let M be any matrix and consider the following partition Assume that (a) M L;L is nonsingular, (b) M=M L;L is a P -matrix (P 0 -matrix),! M = M L;L M L;J : M J;L M J;J 16
then, for every index set K such that L K L [ J, M K;K =M L;L is a P -matrix (P 0 -matrix). Proof. The assertion easily follows by the fact that M K;K =M L;L is a principal submatrix of M=M L;L which, in turn, is a P -matrix (P 0 -matrix). 2 We now introduce some more index sets: := fi 2 Ij x i > 0; F i (x) = 0g; := fi 2 Ij x i = 0; F i (x) = 0g; := fi 2 Ij x i = 0; F i (x) > 0g; := I n f [ [ g (once again, we suppress in our notation the dependence on x of the above index sets). If the point x under consideration is a solution of NCP(F ) then = ; and the sets, and coincide with those already introduced in Section 2. Using the notation of the previous section, the index sets ; and form a partition of the set C; whereas corresponds to what has been called R in Section 4 (here we changed the name of this set just for uniformity of notation). Note that denotes the set of degenerate indices at which is not dierentiable. Further note that at an arbitrary point x 2 IR n ; the index sets ; and, in particular, are usually very small (and quite often empty). The last lemma is introduced in order to simplify the proof of the main Theorem (Theorem 5.4). It gives conditions which guarantee that all elements in @(x) are nonsingular, assuming = ;. This result will then be used to prove the general case 6= ;: Lemma 5.3 Let x 2 IR n be given and set M := F 0 (x). Assume that (a) (x) = ;; (b) M ; is nonsingular, (c) the Schur-complement M [;[ =M ; is a P 0 -matrix. Then the function is dierentiable at x and the Jacobian matrix 0 (x) is nonsingular. Proof. Since (x) = ;; the function is dierentiable at x with 0 (x) = D a +D b M where D a and D b are the diagonal matrices introduced in Lemma 4.1. Let q 2 IR n be an arbitrary vector with (D a + D b M)q = 0: Writing D a = 0 B @ D a; 17 D a; D a; 1 C A ;
D b = M = 0 B @ 0 B @ D b; D b; D b; M ; M ; M ; M ; M ; M ; M ; M ; M ; q = (q ; q ; q ) T ; where D a; ; D a; : : : are abbreviations for the diagonal matrices (D a ) ; ; (D a ) ; : : : ; the equation (D a + D b M)q = 0 can be rewritten as 1 C A ; 1 C A ; M ; q + M ; q + M ; q = 0 ; (22) D a; q = 0 ; (23) D a; q + D b; M ; q + D b; M ; q + D b; M ; q = 0 ; (24) where we have taken into account that, by Lemma 4.1, D a; = 0 ; ; D b; = 0 ; and that D b; is nonsingular. Since the diagonal matrix D a; is also nonsingular, we directly obtain q = 0 (25) from (23). Hence equations (22) and (24) reduce to M ; q + M ; q = 0 ; (26) D a; q + D b; M ; q + D b; M ; q = 0 : (27) Due to the nonsingularity of the submatrix M ; ; we directly obtain from (26) q =?M?1 ; M ;q : (28) Substituting this into (27) and rearranging terms yields Da; + D b; h M;? M ; M?1 ;M ; i q = 0 : (29) According to the negative deniteness of the diagonal matrices D a; and D b; ; we obtain from assumption (c), equation (29) and Lemma 5.1 that q = 0 : This, in turn, implies that q = 0 by (28). In view of (25), we therefore have q = 0; which proves the desired result. 2 Theorem 5.4 Let x 2 IR n be given, and set M := F 0 (x). Assume that (a) the submatrices M ~;~ are nonsingular for all ~ ( [ ); (b) the Schur-complement M [[;[[ =M ; is a P 0 -matrix. Then all elements G 2 @(x) are nonsingular. 18
Proof. We rst note that by Lemma 5.2 and the assumptions made, the Schurcomplement M [^;[^ =M ; is a P 0 -matrix for every ^ ( [ ). Let us now consider ~ ( [ ): We prove that the Schur-complement M ~[ ~ ;~[ ~ =M ~;~ is a P 0 -matrix for all ~ [ [( [ ) n ~]: By assumption (a) and the quotient formula for Schur-complements (see Cottle, Pang and Stone [4, Proposition 2.3.6]), we have that M ~;~ =M ; is nonsingular and M ~[ ~ ;~[ ~ =M ~;~ = M ~[ ~ ;~[ ~ =M ; = (M~;~ =M ; ) : (30) By the rst part of the proof, the matrix M ~[ ~ ;~[ ~ =M ; is a P 0 -matrix. Hence, by using Lemma 2.3 of Chen and Harker [2], the Schur-complement on the right-hand side of (30) is a P 0 -matrix. Therefore the matrix on the left-hand side of (30) is also a P 0 -matrix. Now, recalling that denotes the set of degenerate indices at which (x i ; F i (x)) = 0; and considering ( i ; i ) 2 IR 2 such that k( i ; i )k 1 as in Lemma 4.1, let us partition the index set into the following three subsets: 1 := fi 2 j i > 0; i = 0g; 2 := fi 2 j i = 0; i > 0g; 3 := n ( 1 [ 2 ); and dene ~ := [ 1 ; ~ := [ 2 and ~ := [ 3 : It is now very easy to see that the nonsingularity of any element of @(x) can be proved following exactly the lines of the proof of Lemma 5.3 by replacing the index sets ; and by ~; ~ and ~ ; respectively, and taking into account assumptions (a) and (b). 2 In view of Lemma 5.2 the assumptions (c) of Lemma 5.3 and (b) of Theorem 5.4 are satised, e.g., if the Schur-complement M=M ; is a P 0 -matrix. Due to Chen and Harker [2, Lemma 2.3], Schur-complements of nonsingular submatrices of a P 0 - matrix are again P 0 -matrices, hence assumption (b) of Theorem 5.4 is satised in particular if F 0 (x) is a P 0 -matrix. In turn it is known, see [33, Theorem 5.8], that F 0 (x) is a P 0 -matrix if F itself is a P 0 -function. In the next two corollaries we point out two simple situations in which the assumptions of the previous Theorem are satised. The rst of the two corollaries was already proved by Facchinei and Soares [8, Proposition 3.2]. Corollary 5.5 Let x 2 IR n be an R-regular solution of NCP(F ). Then all the elements of the generalized Jacobian @(x ) are nonsingular. Proof. Since x is a solution of NCP(F ), we have (x ) = ;: We prove that assumptions (a) and (b) of Theorem 5.4 are satised. Assumption (a) follows by R- regularity and by [35, Lemma 1], while assumption (b) is obvious by the R-regularity. 2 19
Due to the upper semi-continuity of the generalized Jacobian (see Clarke [3, Proposition 2.6.2 (c)]), Corollary 5.5 remains true for all x in a suciently small neighbourhood of an R-regular solution x of NCP(F ). Corollary 5.6 Suppose that F 0 (x) is a P -matrix. generalized Jacobian @(x) are nonsingular. Then all the elements of the Proof. Since every principal submatrix of a P -matrix is nonsingular and the Schurcomplement of every principal submatrix of a P -matrix is a P -matrix, we obviously have that the assumptions of Theorem 5.4 are satised. 2 Jiang and Qi [23, Proposition 1] recently proved that all elements of the generalized Jacobian @(x) are nonsingular if F is a uniform P -function. Since it is not dicult to see that the Jacobian matrices of a uniform P -function are P -matrices, we reobtain their result as a direct consequence of Corollary 5.6. 6 Convergence of the Algorithm The main aim of this section is to prove and discuss Theorem 3.1. Proof of point (a) of Theorem 3.1. The proof is by contradiction. We rst note that it can be assumed, without loss of generality, that the direction is always given by (4). In fact if, for an innite set of indices K, we have d k =?r (x k ) for all k 2 K, then any limit point is a stationary point of by well known results (see, e.g., Proposition 1.16 in [1]). Suppose now, renumbering if necessary, that fx k g! x and that r (x ) 6= 0; Since the direction d k always satises (4), we can write k(x k )k = kh k d k k kh k k kd k k (31) from which we get kd k k k(xk )k kh k k (32) (recall the kh k k cannot be 0, otherwise (31) would imply (x k ) = 0, so that x k would be a stationary point and the algorithm would have stopped). We now note that 0 < m kd k k M (33) for some positive m and M. In fact if, for some subsequence K, fkd k kg K! 0 we have from (32), that fk(x k )kg K! 0 because H k is bounded on the bounded sequence fx k g by known properties of the generalized Jacobian. But then, by continuity, (x ) = 0 so that x is a solution of the nonlinear complementarity problem, thus contradicting the assumption r (x ) 6= 0. On the other hand kd k k cannot be 20
unbounded because, taking into account that r (x k ) is bounded and p > 2, this would contradict (5). Then, since (6) holds at each iteration and is bounded from below on the bounded sequence fx k g we have that f (x k+1 )? (x k )g! 0 which implies, by the linesearch test, f2?ik r (x k ) T d k g! 0: (34) We want to show that 2?ik is bounded away from 0. Suppose the contrary. Then, subsequencing if necessary, we have that f2?ik g! 0 so that at each iteration the stepsize is reduced at least once and (6) gives (x k + 2?(ik?1) d k )? (x k ) 2?(ik?1) > r (x k ) T d k : (35) By (33) we can assume, subsequencing if necessary, that fd k g! d 6= 0, so that, passing to the limit in (35), we get r (x ) T d r (x ) T d: (36) On the other hand we also have, by (5), that r (x ) T d?kdkp < 0, which contradicts (36); hence 2?ik is bounded away from 0. But then (34) and (5) imply that fd k g! 0, thus contradicting (33), so that r (x ) = 0. 2 Proof of point (b) of Theorem 3.1. Let fx k g be the sequence of points generated by the algorithm and let be x the locally unique limit point in the statement of the theorem; then x is an isolated global minimum point of. Denote by the set of limit points of the sequence fx k g; we have that x belongs to which is therefore a nonempty set. Let be the distance of x to n x, if x is not the only limit point of fx k g, 1 otherwise, i.e. 8 < inf y2nx fky? x kg if n x 6= ; = : 1 otherwise; since x is an isolated solution, we have > 0. Let us now indicate by 1 and 2 the following sets, 1 = fy 2 IR n : dist(yjg =4g; 2 = fy 2 IR n : kyk kx k + g: We have that for k suciently large, let us say for k k, x k belongs at least to one of the two sets 1, 2. Let now K be the subsequence of all k for which kx k?x k =4 (this set is obviously nonempty because x is a limit point of the sequence). Since all points of the subsequence fx k g K are contained in the compact set S(x ; =4) and every limit point of this sequence is also a limit point of fx k g, we have that all the subsequence fx k g K converges to x, the unique limit point of fx k g in S(x ; =4). Furthermore, since x is a solution of NCP(F ) we have that fkr (x k )kg K! 0 21
which in turn, by (5), implies that fd k g tends to 0. So we can nd ~ k k such that kd k k =4 if k 2 K and k ~ k. Let now ^k be any xed k ~ k belonging to K; we can write: distfx^k+1 j n x g inf y2nx fky? x kg? (kx? x^kk + kx^k? x^k+1 k)? =4? =4 = =2: (37) This implies that x^k+1 cannot belong to 1 n S(x ; =4); on the other hand, since x^k+1 = x^k + ^kd^k for some ^k 2 (0; 1], we have kx^k+1 k kx^kk + k^kd^kk kx + (x^k? x )k + kd^kk kx k + kx^k? x k + kd^kk kx k + =4 + =4; so that x^k+1 does not belong to 2. Hence we get that x^k+1 belongs to S(x ; =4). But then, by denition, we have that ^k+1 2 K, so by induction (recall that ^k+1 > ~ k also, so that kd^k+1 k =4) we have that every k > ~ k belongs to K and the whole sequence converges to x. 2 Remark 6.1 We explicitly point out that in the proof of point (b) we have shown that if the sequence of points generated by the algorithm is converging to a locally unique solution of the nonlinear complementarity problem, then fkd k kg! 0. This fact will be used also in the proof of point (c). Proof of point (c) of Theorem 3.1. The fact that fx k g! x follows by part (b) noting that the BD-regularity assumption implies, by [37, Proposition 3], that x is an isolated solution of the system (x) = 0 and hence also of NCP(F ). Then, we rst prove that locally the direction is always the solution of system (4) and then that eventually the stepsize of one satises the linesearch test (6), so that the algorithm eventually reduces to the local algorithm (2) and the assertions on the convergence rate readily follow from Theorems 2.2 and 2.3. Since fx k g converges to a BD-regular solution of the system (x) = 0, we have, by Lemma 2.6 [38], that the determinant of H k is bounded away from 0 for k suciently large. Hence system (4) always admits a solution for k suciently large. We want to show that this solution satises, for some positive 1, the condition We rst note that, by (4), r (x k ) T d k? 1 kd k k 2 (38) kd k k k(h k )?1 k k(x k )k; so that, recalling that r (x k ) can be written as H k (x k ), r (x k ) T d k =?k(x k )k 2? kdk k 2 22 M 2 ; (39)
where M is an upper bound on k(h k )?1 k (note that this upper bound exists because the determinant of H k is bounded away from 0). Then (38) follows by (39) by taking 1 1=M 2. But now it is easy to see, noting that fkd k kg! 0, that (38) eventually implies (5) for any p > 2 and any positive. Now to complete the proof of the theorem it only remains to show that eventually the stepsize determined by the Armijo test (6) is 1, that is, that eventually i k = 0. But this immediately follows by Theorem 3.2 in [6], taking into account (38) and the fact that is SC 1 by Theorem 2.3. 2 We can now use Theorem 3.1 and the results of Sections 4 and 5 to easily obtain the following two theorems. Theorem 6.1 Let x be a limit point of the sequence generated by the algorithm. Then x is a solution of NCP(F ) i it is a regular point according to Denition 4.1. In particular x is a solution of NCP(F ) if it satises any of the conditions of Theorem 4.3, Corollary 4.4, Theorem 4.5, Theorem 5.4, Corollary 5.5, or Corollary 5.6. Proof. The proof is obvious except for the fact that the conditions of Theorem 5.4, Corollary 5.5, or Corollary 5.6 are sucient to guarantee that x is a solution of NCP(F ). But to prove this it suces to note that, by Theorem 3.1 (a), r (x ) = 0. But employing elementary rules on the calculation of subgradients, it can be shown that r (x ) = @(x ) T (x ) (see discussion in Section 3). If x is not a solution of NCP(F ), then (x ) 6= 0, so that, since Theorem 5.4, Corollary 5.5, and Corollary 5.6 guarantee the nonsingularity of all the elements of @(x ), it cannot be r (x ) = 0 (note that to prove the result it would be sucient to show that there exists at least one nonsingular element in @(x )). 2 The next result is an immediate consequence of the analysis carried out in Section 5. Theorem 6.2 Let x be a limit point of the sequence generated by the algorithm, and suppose that x satises any of the conditions of Theorem 5.4, Corollary 5.5, or Corollary 5.6. Then x is a BD-regular solution of system (x) = 0 so that all the assertions of point (c) of Theorem 3.1 hold true. It is interesting to note that Theorem 5.4, Corollary 5.5, or Corollary 5.6 guarantee the nonsingularity of all the elements of @(x ) while, according to Theorem 2.3, BD-regularity, i.e. nonsingularity of all the elements in @ B (x ), is sucient to have superlinear/quadratic convergence. Hence a fast convergence rate could take place also when the conditions of Theorem 5.4, Corollary 5.5, or Corollary 5.6 are not met. We illustrate this with a simple example. Consider the unidimensional complementarity problem with F (x) =?x; obviously the unique solution of this problem is x = 0. For this problem we have (x) = p x 2 + x 2? x + x = p 2jxj. Hence @ B (0) = f? p 2; p 2g, so that x = 0 is a BD-regular solution of (x) = 0. 23
On the other hand none of the conditions of Theorem 5.4, Corollary 5.5, or Corollary 5.6 can be satised since @(0) = [? p 2; p 2] contains the singular element 0. To complete the discussion of the properties of the algorithm we note that if the function F is a uniform P -function then the level sets of are bounded by Theorem 2.3 (c) so that at least one limit point of the sequence produced by the algorithm exists. By Theorem 3.1 (a), each limit point is a stationary point of ; and by Corollary 4.4 (b) every stationary point is a solution of NCP(F ). Since a complementarity problem with a uniform P -function has a unique solution, we obtain the result that any sequence fx k g generated by our algorithm converges to this unique solution. We note that the case of a uniform P -function is, basically, the case considered in [23]. 7 Numerical Results In this section we perform some numerical tests in order to show the viability of the approach proposed. In Section 3 we left unanswered the problem of calculating an element of @ B (x), which is needed in Step 2 of the algorithm. Our rst task is therefore to show how this can be accomplished. Procedure to evaluate an element H belonging to @ B (x) Step 1: Set = fi : x i = 0 = F i (x)g. Step 2: Choose z 2 IR n such that z i 6= 0 for all i belonging to. Step 3: For each i 62 set the ith row of H equal to!! x i kx i ; F i (x)k? 1 F e T i (x) i + kx i ; F i (x)k? 1 rf i (x) T : (40) Step 4: For each i 2 set the ith row of H equal to!! z i kz i ; rf i (x) T zk? 1 rf e T i (x) T z i + kz i ; rf i (x) T zk? 1 rf i (x) T : (41) Theorem 7.1 The element H calculated by the above procedure is an element of @ B (x). Proof. We shall build a sequence of points fy k g where (x) is dierentiable and such that r(y k ) T tends to H; the theorem will then follow by the very denition of B-subdierential. Let y k = x + " k z, where z is the vector of Step 2 and f" k g is a sequence of positive numbers converging to 0. Since, if i 62, either x i 6= 0 or F i (x) 6= 0, and 24
z i 6= 0 for all i 2, we can assume, by continuity, that " k is small enough so that, for each i, either y k i 6= 0 or F i (y k ) 6= 0, and is therefore dierentiable at y k. Now, if i does not belong to ; it is obvious, by continuity, that the ith row of r(y k ) T tends to the ith row of H; so the only case of concern is when i belongs to. We recall that, according to Lemma 4.1, the ith row of r(y k ) T is given by (a i (y k )? 1)e T i + (b i (y k )? 1)rF i (y k ) T : (42) where a i (y k ) = " k z i q(" k ) 2 z 2i + F i (y k ) 2 ; b i (y k ) = F i (y k ) q(" k ) 2 z 2i + F i (y k ) 2 : We note that by the Taylor-expansion we can write, for each i 2, F i (y k ) = F i (x) + " k rf ( k ) T z = " k rf ( k ) T z; with k! x: (43) Substituting (43) in (42) and passing to the limit, we obtain, taking into account the continuity of rf that also the rows of r(y k ) T relative to indices in tend to the corresponding rows of H dened in Step 4. 2 Changing z we will obtain a dierent element of @ B (x). In our code we chose to set z i = 0 if i 62 and z i = 1 if i 2. We note that the overhead to evaluate an element of @ B (x) when is not dierentiable is negligible with respect to the case in which is dierentiable. This, again, is a favourable characteristic of our reformulation, which is not true, in general, for semismooth systems of equations. The implemented version of the algorithm diers from the one described in Section 3 in the use of a nonmonotone linesearch, which can be viewed as an extension of (6). To motivate this variant we rst recall that it has been often observed in the eld of nonlinear complementarity algorithms that the linesearch test used to enforce global convergence can lead to very small stepsizes; in turn this can bring to very slow convergence and even to a numerical failure of the algorithm. To circumvent this problem many heuristics have been used, see, e.g., [20, 25, 36]. Here, instead, we propose to substitute the linesearch test (6) by the following nonmonotone linesearch: (x k + 2?i d k ) W + 2?i r (x k ) T d k ; (44) where W is any value satisfying (x k ) W max j=0;1;:::;m k (x k?j ) (45) and M k are nonnegative integers bounded above for any k (we suppose for simplicity that k is large enough so that no negative superscripts can occur in (45)). This kind of linesearch has been rst proposed in [17] and since then it has proved very useful in the unconstrained minimization of smooth functions. Adopting the same proof techniques used in [17] it is easy to see that all the results described in the previous 25