Skip to main content

Solving nonlinear programming problems with unbounded non-convex constraint sets via a globally convergent algorithm

Abstract

In this paper, we propose a set of unbounded conditions, under which we are able to solve nonlinear programming problems in a class of unbounded non-convex sets via the combined homotopy interior point algorithm. We also obtain the global convergence of the combined homotopy interior point algorithm and analyze the implementation of this algorithm in detail.

1 Introduction

Consider the constrained minimization problem:

min x R n f ( x ) s.t.  u i ( x ) 0 , i = 1 , , m , s.t.  v j ( x ) = 0 , j = 1 , , l ,
(1)

where f: R n R 1 , u i : R n R 1 , i=1,,m and v j : R n R 1 , j=1,,l are three times continuously differentiable.

We call a point x is a Karush-Kuhn-Tucker (K-K-T) point of (1) and y , z are the corresponding Lagrangian multiplier vectors if ( x , y , z ) satisfies

f ( x ) + u ( x ) y + v ( x ) z = 0 , v ( x ) = 0 , Y u ( x ) = 0 , u ( x ) 0 , y 0 ,
(2)

where y R m , z R l , f(x)= ( f ( x ) / x ) T R n , u(x)=( u 1 (x),, u m (x)) R n × m , v(x)=( v 1 (x),, v l (x)) R n × l , u(x)= ( u 1 ( x ) , , u m ( x ) ) T R m , v(x)= ( v 1 ( x ) , , v l ( x ) ) T R l and Y=diag(y) R m × m .

Since Kellogg et al. (see [1]) and Smale (see [2]) proposed the notable homotopy method, this method has become a powerful tool in dealing with various nonlinear problems, for example, zeros or fixed points of maps (see [35], etc. and the references therein). However, the homotopy method has seldom been touched upon in constrained optimization until 1988, Megiddo (see [6]) and Kojima et al. (see [7]) discovered the Karmarkar interior point method for linear programming was a kind of path-following method. From then on, the central path-following methods for mathematical programming have become an active research subject. Furthermore, it was extended to convex nonlinear programming problems recently (see [811], etc.). But all their convergence results were obtained under the assumptions that the logarithmic barrier function is strictly convex and the solution set is nonempty and bounded.

Recently, a combined homotopy interior point method (denoted as CHIP method for convenience) for nonlinear programming problems was presented in [12, 13] (detailed abstract of them was announced in [14]). In [13], compared with the central path-following methods, the authors removed the convexity of the logarithmic barrier function and the boundedness and nonemptiness of the solution set. This shows that the CHIP method can solve the problems that interior path-following methods cannot solve. In [15], by taking a piecewise technique, under the commonly used conditions, the polynomiality of the CHIP method was given, which shows that the efficiency of the CHIP method is also very well. In [16], we introduced C 2 functions α i (x) R n , i=1,,m and β j (x) R n , j=1,,l to extend the results in [12, 17] to more general non-convex sets. However, there are no results reported in the literature about the work in [16] extended to unbounded non-convex sets. In this paper, we attempt to complete this work. To this end, by using some inequality techniques and the ideas of infinite solutions which were introduced in [18, 19], we develop a set of new unboundedness conditions. Under these conditions, we obtain the global convergence results of the CHIP method and therefore extend the work in [16] to unbounded non-convex sets.

2 Main results

In this section, the nonnegative and positive orthants of R m are denoted as R + m and R + + m , respectively. We also denote by B(x)={i{1,,m}: u i (x)=0} the active set at x. In addition, let X={x R n : u i (x)0,i=1,,m, v j (x)=0,j=1,,l} be the feasible set, let X 0 ={x R n : u i (x)<0,i=1,,m, v j (x)=0,j=1,,l} be the strictly feasible set and let X=X X 0 be the boundary set of X.

In [16], to solve (2), the following homotopy was constructed:

H ( w , w ( 0 ) , λ ) = ( ( 1 λ ) [ f ( x ) + u ( x ) y + λ ξ ( x ) y 2 ] + [ v ( x ) + λ ( η ( x ) v ( x ) ) ] z + λ ( x x ( 0 ) ) v ( x ) Y u ( x ) λ Y ( 0 ) u ( x ( 0 ) ) ) = 0 ,
(3)

where w=(x,y,z) R n + m + l , w ( 0 ) =( x ( 0 ) , y ( 0 ) ,0) X 0 × R + + m × R l , and Y ( 0 ) =diag( y ( 0 ) ) R m × m .

In this section, we utilize the concept of infinite solutions [18, 19] and hence give a new set of unboundedness conditions. The nonlinear programming problem is said to have a solution at infinite, if there exists a sequence { x ( k ) } satisfying { x ( k ) }X, x ( k ) as k, and for any given ζX, there exist y ( k ) R + m and z ( k ) R l such that

lim k ( ζ x ( k ) ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) 0.

Then we assume that there exist smooth mappings ξ i (x) R 1 , i=1,,m and η j (x) R 1 , j=1,,l such that:

(A1) X 0 is nonempty; nonlinear programming problems have no infinite solutions.

(A2) xX, if

i B ( x ) ( y i g i ( x ) + α i ξ i ( x ) ) +v(x)z+η(x)β=0, y i 0, α i 0,z R l ,β R l ,

then y i =0, α i =0 and z=β=0, iB(x).

(A3) xX, we have

{ x + i B ( x ) y i ξ i ( x ) + z η ( x ) : y i 0 , i B ( x ) , z R l } X={x},

where η(x)=( η 1 (x),, η l (x)).

(A4) xX, v(x) is of full of column rank and v ( x ) T η(x) is nonsingular.

However, only applying the infinite solution technique to the items f(x), u(x), and v(x), we cannot extend the results in [16] to unbounded cases because of the existence of the items ξ(x) and η(x). So for the items ξ(x) and η(x), we need to use other techniques. In this paper, we still need the following assumption:

(A5) ζX,

u i ( ζ ) T u i ( x ) T ( ζ x ) T ξ i (x) y i ,i=1,,m,v ( ζ ) T v ( x ) T ( ζ x ) T η(x).

For a given w ( 0 ) , rewrite H(w, w ( 0 ) ,λ) as H w ( 0 ) (w,λ). The zero-point set of H w ( 0 ) is

H w ( 0 ) 1 (0)= { ( w , λ ) X × R + m × R l × ( 0 , 1 ] : H w ( 0 ) ( w , λ ) = 0 } .

The inverse image theorem tells us that, if 0 is a regular value of the map H w ( 0 ) , then H w ( 0 ) 1 (0) consists of some smooth curves. The regularity of H w ( 0 ) can be obtained by the following lemma.

Lemma 2.1 (Parameterized Sard theorem)

Let V R n , U R m be open sets, and Φ:V×U R ( k ) a C r map, where r>max{0,mk}. If 0 R ( k ) is a regular value of Φ, then for almost all aV, 0 is a regular value of Φ a Φ(a,).

Lemma 2.2 Let H be defined as in (3). In addition, let assumptions (A1)-(A5) hold, let u i (x), i=1,,m and v j (x), j=1,,l be C 3 functions, and let ξ i (x), i=1,,m and η j (x), j=1,,l be C 2 functions. Then, for almost all w ( 0 ) X 0 × R + + m × R l , 0 is a regular value of H w ( 0 ) , and H w ( 0 ) 1 (0) consists of some smooth curves, among which there exists a smooth curve Γ w ( 0 ) that starts from ( w ( 0 ) ,1).

Lemma 2.3 Let H be defined as in (3). In addition, let assumptions (A1)-(A5) hold, let u i (x), i=1,,m and v j (x), j=1,,l be C 3 functions, and let ξ i (x), i=1,,m and η j (x), j=1,,l be C 2 functions. Then, for almost all w ( 0 ) X 0 × R + + m × R l , the projection of the smooth curve Γ w ( 0 ) onto the x-plane is bounded.

Proof If not, then there exists a sequence of points { ( x ( k ) , y ( k ) , z ( k ) , λ k ) } k = 1 such that x ( k ) as k.

It is easy to show that the following inequality holds:

x ζ 2 x ( 0 ) ζ 2 2 ( x ζ ) T ( x x ( 0 ) ) .
(4)

By the homotopy equation (3), we have

H ( w ( k ) , w ( 0 ) , λ k ) = ( ( 1 λ k ) [ f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + λ k ξ ( x ( k ) ) ( y ( k ) ) 2 ] + λ k ( x ( k ) x ( 0 ) ) + [ v ( x ( k ) ) + λ k ( η ( x ( k ) ) v ( x ( k ) ) ) ] z ( k ) v ( x ( k ) ) Y ( k ) u ( x ( k ) ) λ k Y ( 0 ) u ( x ( 0 ) ) ) = 0 .
(5)

Multiplying the first equation in (5) by ( x ( k ) ζ ) T , we get

( 1 λ k ) ( x ( k ) ζ ) T f ( x ( k ) ) + ( 1 λ k ) ( x ( k ) ζ ) T u ( x ( k ) ) y ( k ) + ( 1 λ k ) λ k ( x ( k ) ζ ) T ξ ( x ( k ) ) ( y ( k ) ) 2 + ( x ( k ) ζ ) T v ( x ( k ) ) z ( k ) + ( x ( k ) ζ ) T λ k ( η ( x ( k ) ) v ( x ( k ) ) ) z ( k ) + λ k ( x ( k ) ζ ) T ( x ( k ) x ( 0 ) ) = 0 ,
(6)

i.e.,

λ k ( x ( k ) ζ ) T ( x ( k ) x ( 0 ) ) = ( 1 λ k ) ( x ( k ) ζ ) T f ( x ( k ) ) ( 1 λ k ) ( x ( k ) ζ ) T u ( x ( k ) ) y ( k ) ( 1 λ k ) λ k ( x ( k ) ζ ) T ξ ( x ( k ) ) ( y ( k ) ) 2 ( x ( k ) ζ ) T [ v ( x ( k ) ) + λ k ( η ( x ( k ) ) v ( x ( k ) ) ) ] z ( k ) = ( 1 λ k ) ( x ( k ) ζ ) T f ( x ( k ) ) ( 1 λ k ) ( x ( k ) ζ ) T u ( x ( k ) ) y ( k ) ( 1 λ k ) λ k ( x ( k ) ζ ) T ξ ( x ( k ) ) ( y ( k ) ) 2 ( 1 λ k ) ( x ( k ) ζ ) T v ( x ( k ) ) z ( k ) λ k ( x ( k ) ζ ) T η ( x ( k ) ) z ( k ) .
(7)

So

λ k ( x ( k ) ζ 2 x ( 0 ) ζ 2 ) 2 λ k ( x ( k ) ζ ) T ( x ( k ) x ( 0 ) ) = 2 ( 1 λ k ) ( x ( k ) ζ ) T f ( x ( k ) ) 2 ( 1 λ k ) ( x ( k ) ζ ) T u ( x ( k ) ) y ( k ) 2 ( 1 λ k ) λ k ( x ( k ) ζ ) T ξ ( x ( k ) ) ( y ( k ) ) 2 2 ( 1 λ k ) ( x ( k ) ζ ) T v ( x ( k ) ) z ( k ) 2 λ k ( x ( k ) ζ ) T η ( x ( k ) ) z ( k ) = 2 ( 1 λ k ) ( x ( k ) ζ ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) + 2 ( 1 λ k ) λ k ( ζ x ( k ) ) T ξ ( x ( k ) ) ( y ( k ) ) 2 + 2 λ k ( ζ x ( k ) ) T η ( x ( k ) ) z ( k ) 2 ( 1 λ k ) ( x ( k ) ζ ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) + 2 ( 1 λ k ) λ k ( u ( ζ ) T y ( k ) u ( x ( k ) ) T y ( k ) ) + 2 λ k ( v ( ζ ) T z ( k ) v ( x ( k ) ) T z ( k ) ) .
(8)

Since u ( ζ ) T 0 and y ( k ) 0, then u ( ζ ) T y ( k ) 0. Besides, by the third equation in (5), we have u ( x ( k ) ) T y ( k ) = λ k u ( x ( 0 ) ) T y ( 0 ) , thus the following inequality holds:

λ k ( x ( k ) ζ 2 x ( 0 ) ζ 2 ) 2 ( 1 λ k ) ( x ( k ) ζ ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) 2 ( 1 λ k ) λ k u ( x ( k ) ) T y ( k ) = 2 ( 1 λ k ) ( x ( k ) ζ ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) 2 ( 1 λ k ) λ k 2 u ( x ( 0 ) ) T y ( 0 ) .
(9)

Then by (9), we have

( ζ x ( k ) ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) λ k 2 ( 1 λ k ) ( x ( k ) ζ 2 x ( 0 ) ζ 2 ) + λ k 2 u ( x ( 0 ) ) T y ( 0 ) .
(10)

When x ( k ) , by (10), we have

lim k ( ζ x ( k ) ) T ( f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + v ( x ( k ) ) z ( k ) ) lim k λ k 2 ( 1 λ k ) ( x ( k ) ζ 2 x ( 0 ) ζ 2 ) + λ k 2 u ( x ( 0 ) ) T y ( 0 ) 0 ,
(11)

which contradicts assumption (A1). □

Theorem 2.1 Let H be defined as in (3), let f(x), u i (x), i=1,,m and v j (x), j=1,,l be three times continuously differentiable functions, let assumptions (A1)-(A5) hold, and let ξ i (x), i=1,,m and η j (x), j=1,,l be twice times continuously differentiable functions. Then for almost all w ( 0 ) X 0 × R + + m × R l , there exists a C 1 curve (w(s),λ(s)) of dimension 1 such that

H ( w ( s ) , w ( 0 ) , λ ( s ) ) =0, ( w ( 0 ) , λ ( 0 ) ) = ( w ( 0 ) , 1 ) .
(12)

When λ(s)0, w(s) tends to a point w =( x , y , z ). In particular, the component x of w is a K-K-T point of problem (1).

Proof By Lemma 2.2, there must be a C 1 curve (w(s),λ(s)) of dimension 1 (denoted by Γ w ( 0 ) ) such that

H ( w ( s ) , w ( 0 ) , λ ( s ) ) =0, ( w ( 0 ) , λ ( 0 ) ) = ( w ( 0 ) , 1 ) .

By the classification theorem of one-dimensional smooth manifolds, Γ w ( 0 ) is diffeomorphic either to a unit circle or to a unit interval. For any w ( 0 ) X 0 × R + + m × R l , H w ( 0 ) ( w ( 0 ) ,1)/w is nonsingular, so Γ w ( 0 ) cannot be diffeomorphic to a unit circle. That is, Γ w ( 0 ) is diffeomorphic to a unit interval.

Let ( w , λ ) be a limit point of Γ w ( 0 ) . If ( w , λ ) X 0 × R + + m × R l ×(0,1), because 0 is a regular value of H w ( 0 ) , ( w , λ ) H w ( 0 ) 1 (0), and the Jacobian matrix of H at ( w , λ ) is of full row rank, then by the implicit function theorem, Γ w ( 0 ) can be extended at ( w , λ ). This result contradicts the fact that ( w , λ ) is a limit point of Γ w ( 0 ) .

Let ( w , λ )=( x , y , z , λ ). Thus, ( w , λ )(X× R + m × R l ×(0,1]) and the following three cases are possible:

  1. (a)

    ( w , λ )=( x , y , z , λ )X× R + m × R l ×{0}.

  2. (b)

    ( w , λ )=( x , y , z , λ ) X 0 × R + + m × R l ×{1}.

  3. (c)

    ( w , λ )=( x , y , z , λ )(X× R + m × R l )×(0,1].

Since the equation H(w, w ( 0 ) ,1)=0 has a unique solution ( w ( 0 ) ,1) in X 0 × R + + m × R l ×{1}, case (b) is impossible.

By the homotopy equation (3), we have

( 1 λ k ) [ f ( x ( k ) ) + u ( x ( k ) ) y ( k ) + λ k ξ ( x ( k ) ) ( y ( k ) ) 2 ] + λ k ( x ( k ) x ( 0 ) ) + [ v ( x ( k ) ) + λ k ( η ( x ( k ) ) v ( x ( k ) ) ) ] z ( k ) = 0 ,
(13)
v ( x ( k ) ) =0,
(14)
Y ( k ) u ( x ( k ) ) λ k Y ( 0 ) u ( x ( 0 ) ) =0.
(15)

Let

I(x)= { i = 1 , , m : lim k y i ( k ) = + } ,J(x)= { j = 1 , , l : lim k z j ( k ) = + } .

If J(x), then

( 1 λ k ) [ f ( x ( k ) ) + i I ( x ) ( u i ( x ( k ) ) y i ( k ) + λ k ξ i ( x ( k ) ) ( y i ( k ) ) 2 ) ] + λ k ( x ( k ) x ( 0 ) ) + [ ( v ( x ( k ) ) + λ k ( η ( x ( k ) ) v ( x ( k ) ) ) ) z ( k ) + ( 1 λ k ) i I ( x ) ( u i ( x ( k ) ) y i ( k ) + λ k ξ i ( x ( k ) ) ( y i ( k ) ) 2 ) ] = 0 .
(16)

Because X and (0,1] are bounded, by assumption (A2), the third part on the left-hand side of (16) tends to infinity as k, but the other two parts are bounded, this is impossible. Hence we conclude that the projection of the smooth curve Γ w ( 0 ) onto the z-plane is also bounded.

In case (c), first, we prove that y R + m . If y R + m , then there exist i 0 {1,,m} and a sequence of points {( w ( k ) , λ k )} Γ w ( 0 ) such that y i 0 ( k ) y i 0 =0 as k+. From (15), we have

y i 0 ( k ) u i 0 ( x ( k ) ) = λ k y i 0 ( 0 ) u i 0 ( x ( 0 ) ) .
(17)

Because X and (0,1] are bounded, when k+, the left-hand side of (17) tends to 0. At the same time, the right-hand side of (17) tends to λ y i 0 ( 0 ) u i 0 ( x ( 0 ) ), which is strictly less than 0. This fact results in a contradiction.

Then we only need to prove that the remainder of case (c) is impossible. If not, then there exists a sequence of points {( w ( k ) , λ k )} Γ w ( 0 ) such that u i ( x ( k ) )0 for some i{1,,m} as k+. From (15), we obtain y ( k ) +. Because X and (0,1] are bounded, there exists a subsequence of points (denoted also by {( w ( k ) , λ k )}) such that x ( k ) x , y ( k ) +, z ( k ) z , and λ k λ as k+.

When λ >0, from (15), the active index set is B( x )=I( x ). When λ =0, the index set I( x )B( x ).

(1) When λ =1, from (16), by the fact that y i ( k ) is bounded for each iI( x ), assumptions (A1)-(A2), we conclude that lim k + (1 λ k ) y i ( k ) =0 and lim k + (1 λ k ) ( y i ( k ) ) 2 = y i . Therefore, when k+, (16) becomes

η ( x ) z + i I ( x ) ξ i ( x ( k ) ) y i + x = x ( 0 ) ,
(18)

which contradicts assumption (A3).

(2) When 0< λ <1, rewrite (16) as

i I ( x ) ( u i ( x ( k ) ) y i ( k ) + λ k ξ i ( x ( k ) ) ( y i ( k ) ) 2 ) + λ k 1 λ k ( x ( k ) x ( 0 ) ) = f ( x ( k ) ) i I ( x ) ( u i ( x ( k ) ) y i ( k ) + λ k ξ i ( x ( k ) ) ( y i ( k ) ) 2 ) 1 1 λ k ( v ( x ( k ) ) + λ k ( η ( x ( k ) ) v ( x ( k ) ) ) ) z ( k ) .
(19)

Let y I ( k ) be a vector with components ( y i ( k ) ) 2 , iI( x ). Then set

ρ i ( k ) = ( y i ( k ) ) 2 y I ( k ) ,iI ( x ) .
(20)

Note that 0 ρ i ( k ) 1; then there exists a subsequence of { ρ i ( k ) }, still denoted by { ρ i ( k ) }, such that ρ i ( k ) ρ i for each iI( x ) as k+. Furthermore, the vector with components ρ i , iI( x ) is denoted by ρ ; thus, ρ =1. Dividing both sides of (19) by y I ( k ) and letting k+, we have

i I ( x ) λ i ρ i ξ i ( x ) =0,

which contradicts assumption (A2).

(3) When λ =0, because the nonempty index set I( x )B( x ), the proof of (3) is similar to that of (2).

From the discussion above, we conclude that (a) is the only possible case. Hence, the x-component of w is a K-K-T point of (1). □

3 Algorithmic analysis

For almost all w ( 0 ) =( x ( 0 ) , y ( 0 ) ,0) X 0 × R + + m × R l , by Theorem 2.1, the homotopy generates a C 1 curve Γ w ( 0 ) , by differentiating the first equation in (12), we obtain the following theorem.

Theorem 3.1 The homotopy path Γ w ( 0 ) is determined by the following initial value problem to the ordinary differential equation:

D H w ( 0 ) ( w ( s ) , λ ( s ) ) ( w ˙ ( s ) λ ˙ ( s ) ) =0, ( w ( 0 ) , λ ( 0 ) ) = ( w ( 0 ) , 1 ) ,
(21)

where s is the arc length of the curve Γ w ( 0 ) .

As for how to trace numerically the homotopy path, there have been many predictor-corrector algorithms, see [20], etc. for reference. Hence we omit them in this paper. In the implementation of the algorithm, generally we need to be devoted to finding the positive direction of the tangent vector at a point on Γ w ( 0 ) which keeps the sign of the determinant | D H w ( 0 ) ( w , λ ) p T | invariant. On the first iteration, the sign is determined by the following lemma.

Lemma 3.1 If Γ w ( 0 ) is smooth, then the positive direction p ( 0 ) at the initial point ( w ( 0 ) ,1) satisfies sign | D H w ( 0 ) ( w ( 0 ) , 1 ) p ( 0 ) T | = ( 1 ) m + l + 1 .

Proof Let

a=f ( x ( 0 ) ) u ( x ( 0 ) ) +(12λ)ξ ( x ( 0 ) ) ( y ( 0 ) ) 2 ,b= Y ( 0 ) g ( x ( 0 ) ) ,

then

H w ( 0 ) ( w ( 0 ) , 1 ) ( w , λ ) = ( I 0 η ( x ( 0 ) ) a v ( x ( 0 ) ) T 0 0 0 Y ( 0 ) u ( x ( 0 ) ) T diag ( u ( x ( 0 ) ) ) 0 b ) =( M 1 , M 2 ),

where M 1 R ( n + m + l ) × ( n + m + l ) , M 2 R ( n + m + l ) × 1 . The tangent vector p ( 0 ) of Γ w ( 0 ) at ( w ( 0 ) ,1) satisfies

( M 1 , M 2 ) p ( 0 ) =( M 1 , M 2 ) ( p 1 ( 0 ) p 2 ( 0 ) ) =0,
(22)

where p 1 ( 0 ) R n + m + l , p 2 ( 0 ) R 1 .

It is easy to get p 1 ( 0 ) = M 1 1 M 2 p 2 ( 0 ) , then

| D H w ( 0 ) ( w ( 0 ) , 1 ) p ( 0 ) T | = | M 1 M 2 p 1 ( 0 ) T p 2 ( 0 ) T | = | M 1 M 2 M 2 T M 1 T 1 | p 2 ( 0 ) = | M 1 | p 2 ( 0 ) ( 1 + M 2 T M 1 T M 1 1 M 2 ) = | I 0 η ( x ( 0 ) ) v ( x ( 0 ) ) T 0 0 Y ( 0 ) u ( x ( 0 ) ) T diag ( u ( x ( 0 ) ) ) 0 | p 2 ( 0 ) ( 1 + M 2 T M 1 T M 1 1 M 2 ) = | diag ( u ( x ( 0 ) ) ) | | I η ( x ( 0 ) ) v ( x ( 0 ) ) T 0 | p 2 ( 0 ) ( 1 + M 2 T M 1 T M 1 1 M 2 ) = ( 1 ) l | diag ( u ( x ( 0 ) ) ) | | v ( x ( 0 ) ) T η ( x ( 0 ) ) | p 2 ( 0 ) ( 1 + M 2 T M 1 T M 1 1 M 2 ) .

Since u( x ( 0 ) )<0, (1+ M 2 T M 1 T M 1 1 M 2 )>0, and p 2 ( 0 ) <0, so the sign of

| D H w ( 0 ) ( w ( 0 ) , 1 ) p ( 0 ) T |

is ( 1 ) m + l + 1 . □

The following pseudocode describes the basic steps of a generic predictor-corrector method.

Algorithm 3.1 (Euler-Newton method)

Step 0. Provide an initial guess ( w ( 0 ) ,1), an initial step length h 0 >0, and three small positive numbers ϵ 1 >0, ϵ 2 >0, and ϵ 3 >0. Set k=0.

Step 1. Compute the direction θ ( k ) of the predictor step.

  1. (a)

    Compute a unit tangent vector p ( k ) .

  2. (b)

    Determine the direction θ ( k ) of the predictor step as follows:

If the sign of the determinant | D H w ( 0 ) ( w ( k ) , λ k ) p ( k ) T | is ( 1 ) m + l + 1 , then θ ( k ) = p ( k ) .

If the sign of the determinant | D H w ( 0 ) ( w ( k ) , λ k ) p ( k ) T | is ( 1 ) m + l , then θ ( k ) = p ( k ) .

Step 2. Compute a corrector point ( w ( k + 1 ) , λ k + 1 ).

( w ¯ ( k ) , λ ¯ k ) = ( w ( k ) , λ k ) + h k θ ( k ) , ( w ( k + 1 ) , λ k + 1 ) = ( w ¯ ( k ) , λ ¯ k ) D H w ( 0 ) ( w ¯ ( k ) , λ ¯ k ) + H w ( 0 ) ( w ¯ ( k ) , λ ¯ k ) .

If H w ( 0 ) ( w ( k + 1 ) , λ k + 1 ) ϵ 1 , then let h k + 1 =min{ h 0 ,2 h k }, and go to Step 3.

If H w ( 0 ) ( w ( k + 1 ) , λ k + 1 )( ϵ 1 , ϵ 2 ), then let h k + 1 = h k , and go to Step 3.

If H w ( 0 ) ( w ( k + 1 ) , λ k + 1 ) ϵ 2 , then let h k + 1 =max{ 2 25 h 0 ,( h k /2)}, k=k+1, and go to Step 2.

Step 3. If λ k + 1 ϵ 3 , then stop. Otherwise, k=k+1, and go to Step 1.

4 Conclusions

In this paper, we present a set of unboundedness conditions, under which, we are able to solve nonlinear programming problems on a class of unbounded non-convex sets via the combined homotopy interior point algorithm. The main advantage of the algorithm presented in this paper is that it is a globally convergent algorithm whose initial points can be chosen more easily than the locally convergent algorithms. Since nonlinear programming problems have wide applications in engineering, management, economics, and so on, our results may be useful to propose a powerful solution tool for dealing with these nonlinear problems. In future, we devote our efforts to proposing new techniques to solve nonlinear programming problems in a broader class of unbounded non-convex sets.

References

  1. Kellogg RB, Li TY, Yorke JA: A constructive proof of the Brouwer fixed-point theorem and computational results. SIAM J. Numer. Anal. 1976, 13: 473–483. 10.1137/0713041

    MathSciNet  Article  MATH  Google Scholar 

  2. Smale S: A convergent process of price adjustment and global Newton method. J. Math. Econ. 1976, 3: 1–14. 10.1016/0304-4068(76)90002-1

    MathSciNet  Article  MATH  Google Scholar 

  3. Chow SN, Mallet-Paret J, Yorke JA: Finding zeros of maps: homotopy methods that are constructive with probability one. Math. Comput. 1978, 32: 887–899. 10.1090/S0025-5718-1978-0492046-9

    MathSciNet  Article  MATH  Google Scholar 

  4. Garcia CB, Zangwill WI: An approach to homotopy and degree theory. Math. Oper. Res. 1979, 4: 390–405. 10.1287/moor.4.4.390

    MathSciNet  Article  MATH  Google Scholar 

  5. Li Y, Lin ZH: A constructive proof of the Poincaré-Birkhoff theorem. Trans. Am. Math. Soc. 1995, 347: 2111–2126.

    Article  MATH  Google Scholar 

  6. Megiddo N: Pathways to the optimal set in linear programming. In Progress in Mathematical Programming, Interior Point and Related Methods. Edited by: Megiddo N. Springer, New York; 1988:131–158.

    Google Scholar 

  7. Kojima M, Mizuno S, Yoshise A: A primal-dual interior point algorithm for linear programming. In Interior Point and Related Methods. Edited by: Megiddo N. Springer, New York; 1988:29–47.

    Google Scholar 

  8. Kortanek KO, Potra F, Ye Y: On some efficient interior point algorithms for nonlinear convex programming. Linear Algebra Appl. 1991, 152: 169–189.

    MathSciNet  Article  MATH  Google Scholar 

  9. McCormick GP: The projective SUMT method for convex programming. Math. Oper. Res. 1989, 14: 203–223. 10.1287/moor.14.2.203

    MathSciNet  Article  MATH  Google Scholar 

  10. Monteiro RDC, Adler I: An extension of Karmarkar type algorithm to a class of convex separable programming problems with global linear rate of convergence. Math. Oper. Res. 1990, 15: 408–422. 10.1287/moor.15.3.408

    MathSciNet  Article  MATH  Google Scholar 

  11. Zhu JA: A path following algorithm for a class of convex programming problems. ZOR-Methods Models Oper. Res. 1992, 36: 359–377. 10.1007/BF01416235

    Article  MathSciNet  MATH  Google Scholar 

  12. Feng GC, Lin ZH, Yu B: Existence of interior pathway to the Karush-Kuhn-Tucker point of a nonconvex programming problems. Nonlinear Anal. 1998,32(6):761–768. 10.1016/S0362-546X(97)00516-6

    MathSciNet  Article  MATH  Google Scholar 

  13. Lin ZH, Yu B, Feng GC: A combined homotopy interior point method for convex nonlinear programming. Appl. Math. Comput. 1997, 84: 193–211. 10.1016/S0096-3003(96)00086-0

    MathSciNet  Article  MATH  Google Scholar 

  14. Feng GC, Yu B: Combined homotopy interior point method for nonlinear programming problems. Lecture Notes in Numerical and Applied Analysis 14. In Advances in Numerical Mathematics: Proceedings of the Second Japan-China Seminar on Numerical Mathematics. Edited by: Fujita H, Yamaguti M. Kinokuniya, Tokyo; 1995:9–16.

    Google Scholar 

  15. Yu B, Xu Q, Feng GC: On the complexity of a combined homotopy interior method for convex programming. J. Comput. Appl. Math. 2007,200(1):32–46. 10.1016/j.cam.2005.12.018

    MathSciNet  Article  MATH  Google Scholar 

  16. Su ML, Lv XR: Solving a class of nonlinear programming problems via a homotopy continuation method. Northeast. Math. J. 2008,24(3):265–274.

    MathSciNet  Google Scholar 

  17. Liu QH, Yu B, Feng GC: A combined homotopy interior point method for nonconvex nonlinear programming problems under quasi normal cone conditions. Acta Math. Appl. Sin. 2003, 26: 372–377.

    Google Scholar 

  18. Xu Q, Yu B, Feng GC, Dang CY: Condition for global convergence of a homotopy method for variational inequality problems on unbounded sets. Optim. Methods Softw. 2007, 22: 587–599. 10.1080/10556780600887883

    MathSciNet  Article  MATH  Google Scholar 

  19. Xu Q, Lin ZH: The combined homotopy convergence in unbounded set. Acta Math. Appl. Sin. 2004, 27: 624–631.

    MathSciNet  Google Scholar 

  20. Allgower EL, Georg K: Introduction to Numerical Continuation Methods. SIAM, Philadelphia; 2003.

    Book  MATH  Google Scholar 

Download references

Acknowledgements

This work was supported by NSFC-Union Science Foundation of Henan (No. U1304103) and Natural Science Foundation of Henan Province (No. 122300410261).

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Menglong Su.

Additional information

Competing interests

The authors declare that they have no competing interests.

Authors’ contributions

MS carried out the main work and drafted the manuscript. JL participated in completing the proof of Lemma 2.3. ZX participated in completing the proof of Theorem 2.1. All authors read and approved the final manuscript.

Rights and permissions

Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0), which permits use, duplication, adaptation, distribution, and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Su, M., Li, J. & Xu, Z. Solving nonlinear programming problems with unbounded non-convex constraint sets via a globally convergent algorithm. J Inequal Appl 2014, 281 (2014). https://doi.org/10.1186/1029-242X-2014-281

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/1029-242X-2014-281

Keywords

  • unbounded conditions
  • nonlinear programming problems
  • non-convex sets