IBM Almaden Research Center,650 Harry Road Sun Jose, Calijornia and School of Mathematical Sciences Tel Aviv University Tel Aviv, Israel

Transcription

1 and Nimrod Megiddo IBM Almaden Research Center,650 Harry Road Sun Jose, Calijornia and School of Mathematical Sciences Tel Aviv University Tel Aviv, Israel Submitted by Richard Tapia ABSTRACT Smale proposed a framework for applying Newton's method to the linear p gramming problem. It is shown that his method is closely related to recent inte point methods, in the sense that it also traces the path of centers, even though tracing is done outside the affine hull of the feasible domain. Also, an equivalenc the fundamental theorems is pointed out. It is well known (see [2]) that the linear programming problem and its dual Minimize ctx subject to Ax 2 b, x 2 0 Maximize bty subject to ATy < c, y 2 0 LINEAR ALGEBRA AND ITS APPLICATIONS 152: (1991) OElsevier Science Publishing Co., Inc., Avenue of the Americas, New York, NY 10010

2 which can in turn be viewed as a system of piecewise linear equa where xf = max{xj,o), x,: = min{xj, 01, x ' = (xp,..., x,+ IT, SO th and w = xf. Note that here x E Rn is not the same as the x in programming problem above. Smale [lo] proposed a "regularization" of the piecewise = q as follows. For a 2 0, denote Approximate x * by is approximated (x) + Ma,- (x). This approximation is good in the sense tends on Rn as a -t 0. Incidentally, the other claim that "for as llxll +m" [lo, p. 178, line 81 is incorrect. For example,

3 but x, does not necessarily tend to infinity when llxll does. Smale proposes to solve the linear programming problem as follows. every sufficiently large a > llqll, the zero vector lies in the domain quadratic convergence of Newton's method for solving the Starting with such an a, a solution for the linear programm problem can be obtained by following the path of solutions as a is driven to zero. Although it does not seem to be an interior point method, it turns out Smale's method is very closely related to recent interior path follow algorithms. The "path of centers" for a general linear complementarity problem [ defined to be the set of solutions of the following system: In the special case of the linear programming problem, this system defin unique path which is obtained by combining the primal and dual logarith barrier trajectories [4, 1, 7, 9, 111. Given a > 0, let us associate w any x E Rn a pair of vectors 5 = g(x) = and q = q(x) = q, then obviously q = Mg + q and 5, q > 0. Surprisingly, Sjqj = a2/4, and hence the point (g, q) lies on the path of centers wh p = a2/4. Conversely, if (5, q) is on the path of centers for a certain value p, de x = q - 5 and a = 2 6. We = - = q, Thus, theorems concerning the path of centers correspond to theor concerning the set of solutions of the = q. The main theo talks about the existence and uniqueness of the path of centers. Thi discussed below..., n) and g(x), q(x) Interestingly, for any x, tj(~)qj(~) = p (j = 1,. However, if x is not exactly on the path of centers, then q(x) # M5(x) This means that although Smale's method traces the ~ ath of centers, it not do that within the interior of the feasible domain but rather as an exte point method, although the iterates stay in the positive orthant.

4 nonempty interior (i.e., there exist x,y > 0 such that y = Mx+ path of centers exists and is unique, i.e., for every p > 0, th uniquepairx,y>o such that y=mx+qand xjyj=p for j=l The equivalent form of the theorem was independently Kojima, Mizuno, and Yoshise [5], continuing the analysis of [7 uses arguments of convex optimization. The theorem also fol more general result on complementarity problems with maxim multifunctions given by McLinden [6]. See Theorems 2 and 3 o In order to trace the solutions = q, one needs t starting point, an approximate zero2 = q for some a > 0 a point is not always available, Smale proposes to start with an zero of +,(x)= q+ aa(0). In fact, he argues that the zero approximate zero of the latter for a sufficiently large. The proof is based on Smale's "a-theorem." Now, Thus, the choice of 5 = -q = fie gives an approximate solution f However, the question of what is a sufficiently large p can '~t least in one place (p. 189, line 11) the proof relies on the special struc derived from a linear programming problem, but it is claimed that it applies to th case of a positive semidefinite M. '~n approximate zero of differentiable map F: Rn + Rn is defined to be domain of quadratic convergence of Newton's method for the equation F(x) = 0

5 REFERENCES D. A. Bayer and J. C. Lagarias, The nonlinear geometry of linear programmin Affine and projective rescaling trajectories, Trans. Amer. Math. Soc., to appe R. W. Cottle and G. B. Dantzig, Complementary pivot theory of mathematic programming, Linear Algebra Appl. 1: (1968). B. C. Eaves and H. Scarf, The solution of systems of piecewise linear equatio Math. Oper. Res. 1:l-27 (1976). A. V. Fiacco and G. P. McCormick, Nonlinear Programming: Sequential Unc strained Minimization Techniques, Wiley, New York, M. Kojima, S. Mizuno, and A. Yoshise, A polynomial-time algorithm for a class linear complementarity problems, Math. Programming 44:l-26 (1989). L. McLinden, The complementarity ~roblem for maximal monotone multifun tion~, in Variational lnequalities and Complementarity Problems (R. W. Cottle, Giannessi, and J.-L. Lions, Eds.), Wiley, New York, 1980, pp N. Megiddo, Pathways to the optimal set in linear programming, in Progress Mathematical Programming: Interior-Point and Related Methods (N. Megid Ed.), Springer-Verlag, New York, 1988, pp N. Megiddo and M. Kojima, On the existence and uniqueness of solutions nonlinear complementarity theory, Math. Programming 12: (1977). J. Renegar, A polynomial-time algorithm, based on Newton's method, for lin programming, Math. Programming 40: (1988). S. Smale, Algorithms for solving equations, in Proceedings of the Internatio Congress of Mathematicians (Berkeley, 1986), Amer. Math. Soc., Providen 1987, pp Gy. Sonnevend, An "Analytic Center" for Polyhedrons and New Classes Global Algorithms for Linear (Smooth, Convex) Programming, Inst. of Mat matics, ~tv6s Univ., Hungary, Received 16 October 1989; final manuscript accepted 22 October 1990