In this paper, a modified simple penalty function is proposed for a constrained nonlinear programming problem by augmenting the dimension of the program with a variable that controls the weight of the penalty terms. This penalty function enjoys improved smoothness. Under mild conditions, it can be proved to be exact in the sense that local minimizers of the original constrained problem are precisely the local minimizers of the associated penalty problem.
MSC: 47H20, 35K55, 90C30.
Keywords:constrained minimization; exact penalty function; nonlinear programming problem
Merit function has always taken an important role in optimization problem. It is traditionally constructed to solve nonlinear programs by augmenting the objective function or a corresponding Lagrange function some penalty or barrier terms with respect of the constraints. Then it can be optimized by some unconstrained or bounded constrained optimization softwares or sequential quadratic programming (SQP) techniques. No matter what kind of techniques are involved, the merit function always depends on a small parameter ε or large parameter . As , the minimizer of a merit function such as a barrier function or the quadratic penalty function, converges to a minimizer of the original problem. By using some exact penalty function such as penalty function (see [1,9,20-22]), the minimizer of the corresponding penalty problem must be a minimizer of the original problem when ε is sufficiently small. There are some nonsmooth penalty functions for nonsmooth optimization problems, such as the exact penalty function using the distance function for the nonsmooth variational inequality problem in Hilbert spaces  and the one in .
The traditional exact penalty functions  are always nonsmooth. When it is used as a merit function to accept a new iterate in an SQP method, it may cause the Maratos effect . On the other hand, a traditional smooth penalty function like the quadratic penalty function cannot be an exact one. So we must compute a sequence of minimization subproblem as . At that time, ill-conditioning may occur when the penalty parameter is too large or small, which also brings difficulty of computation. In  and , some kinds of augmented Lagrangian penalty functions have been proposed with improved exactness under strong conditions. In , exact penalty functions via regularized gap function for variational inequalities have also been given. All these functions enjoy some smoothness, but at the very beginning, to use this smoothness we need second-order or third-order derivative information of the problem function that is difficult to estimate in practice. Besides, all the above kinds of penalty functions (see [2-4,7,16] for summary) may be unbounded below even when the constrained problem is bounded, which may make it difficult to locate a minimizer.
In the paper , a new penalty function is proposed for the constrained optimization problem. By augmenting the dimension of the program with an additional variable ε that controls the weight of the penalty terms, this new penalty function enjoys properties of smoothness and exactness, and remains bounded below under reasonable conditions. Its important new idea is that the penalty function is considered as a function of variable x and the additional variable ε simultaneously. Under proper assumptions, the minimizer of the merit function satisfies , and is a minimizer of the original problem. However, the penalty function given in  is not smooth in a small neighborhood of , where the minimizer of the original constrained problem lies. In this paper, we give a penalty function which enjoys the properties of the penalty function given in  and has improved smoothness.
The rest of this paper is organized as follows. In Section 2, a penalty function is introduced for a smooth nonlinear optimization problem with equality constraints and bounded constraints. The smoothness of this penalty function is discussed, as well as other properties, including being bounded below under mild assumptions. Section 3 shows the exactness of our penalty function in the sense that under certain conditions, local minimizer of our penalty function has the form with and is a local minimizer of the original problem, and a converse result holds.
2 New penalty function
We consider the smooth nonlinear optimization problem with equality constraints and bound constraints:
with the constraint violation measure
The main difference between (2.3) and the penalty function given in  is that in (2.3), , which does not have the property that , as .
2.1 Boundedness of the penalty function
Now we see a simple example:
is unbounded below for all penalty parameters since, e.g., for , . It is also the case for traditional penalty functions, including multiplier penalty functions that use an additional term . On the other hand, our new penalty function is bounded below. Set , it reads
where . Since , if , it is obvious that our penalty function is bounded below. See Figure 1 for the display of the contour of the penalty function on this example.
Figure 1. ,,, and.
3 Exactness of the penalty function
In this section, we show that our penalty function is exact in the sense that under certain conditions, local minimizer of our penalty function has the form in which and is a local minimizer of the original problem and a converse proposition holds.
Firstly, recall the Mangasarian-Fromovitz condition. We say that the Mangasarian-Fromovitz condition (see) for Problem (P) holds atifhas full rank and there is a vectorwithand
(3.4) is equivalent to
By (3.3), we obtain
Furthermore, by (3.2), it holds that
Proof Now let be a local minimizer of () with finite and is sufficiently large. If , then must be a Kuhn-Tucker point of (), which is a contradiction with Theorem 3.1. Therefore, , and since is finite, . It implies that , and by (2.5) there is a neighborhood of where for feasible x. Therefore, is a local minimizer of (P). □
We now show a converse result of Theorem 3.2, which will use the following lemmas.
Proof Since and has full rank, there exists a matrix such that the augmented matrix is nonsingular. By the continuity of at , there exists a neighborhood of such that is nonsingular, for any . Take for the closed convex hull of , then for all , the matrix is nonsingular. We now show that for any , there exists a matrix such that
where , so (3.8) holds. Set the mapping , for . By the proof in , Theorem 4.5], we have that there exists a neighborhood of such that for each , and each subset J of , there exists a vector with
Therefore, combining (3.8) with (3.9), we have
and this complete the proof. □
where the last inequality holds from (3.10).
which complete the proof. □
where the second inequality is by (3.11).
From Case 1 and Case 2, we obtain the conclusion. □
4 Conclusion remarks
In this paper, a modified exact penalty function for equality constrained nonlinear programming problem is constructed by augmenting a new variable that controls the constraint violence. This function enjoys smoothness, and with very mild conditions it is proved to be an exact penalty function.
Since in practice, a lot of applied problems are nonsmooth, it is a meaningful work to extend the results in this paper to the nonsmooth case. By using the limiting subgradients that is presented in two books written by Mordukhovich [14,15], as well as Clarke’s generalized gradients in , we can extend the penalty function with the mentioned good properties to nonsmooth optimization problems, just as that has been done in [17-19]. That will be our future research direction.
The authors declare that they have no competing interests.
All authors read and approved the final manuscript.
The authors wish to thank the anonymous referees for their endeavors and valuable comments. The authors would also like to thank Professor Zhang Liansheng for some very helpful comments on a preliminary version of this paper. This research was supported by the National Natural Science Foundation of China under Grants 10971118, 11101248, Natural Science Foundation of Shandong Province under Grants ZR2012AM016, and the foundation 4041-409012 of Shandong University of Technology.
Antczak, T: Exact penalty functions method for mathematical programming problems involving invex functions. Eur. J. Oper. Res.. 198, 29–36 (2009). Publisher Full Text
Boukari, D, Fiacco, AV: Survey of penalty, exact-penalty and multiplier methods from 1968 to 1993. Optimization. 32, 301–334 (1995). Publisher Full Text
Han, SP, Mangasarian, OL: Exact penalty functions in nonlinear programming. Math. Program.. 17, 251–269 (1979). Publisher Full Text
Hoheisel, T, Kanzow, C, Outrata, J: Exact penalty results for mathematical programs with vanishing constraints. Nonlinear Anal.. 72, 2514–2526 (2010). Publisher Full Text
Huyer, W, Neumaier, A: A new exact penalty function. SIAM J. Optim.. 13, 1141–1158 (2003). Publisher Full Text
Mangasarian, OL, Fromovitz, S: The Fritz John necessary optimality conditions in the presence of equality and inequality constraints. J. Math. Anal. Appl.. 17, 37–47 (1967). Publisher Full Text
Soleimani-damaneh, M: The gap function for optimization problems in Banach spaces. Nonlinear Anal.. 69, 716–723 (2008). Publisher Full Text
Soleimani-damaneh, M: Penalization for variational inequalities. Appl. Math. Lett.. 22, 347–350 (2009). Publisher Full Text
Soleimani-damaneh, M: Nonsmooth optimization using Mordukhovich’s subdifferential. SIAM J. Control Optim.. 48, 3403–3432 (2010). Publisher Full Text
Zangwill, WI: Nonlinear programming via penalty function. Manag. Sci.. 13, 344–358 (1967). Publisher Full Text
Zaslavski, AJ: A sufficient condition for exact penalty in constrained optimization. SIAM J. Optim.. 16, 250–262 (2005). Publisher Full Text