On the minimization of the general entropy function on a convex set

- G. Kéri

Computer and Automation Institute

Hungarian Academy of Sciences

Budapest, Hungary

- Abstract: The asymptotic behaviour of the general entropy function, i. e. the dual geometrical programming objective function along straight lines will be examined,
thereafter on this ground two transformations of different kinds will be introduced for the study of those optimization problems where the function indicated in the title is minimized on a closed convex set.
After recognizing and demonstrating with an example, that the boundedness of the objective function on an unbounded polyhedral convex set does not involve that the objective function takes on its minimum on the same set,
a necessary and sufficient condition will be developed for the objective function to be bounded on the condition set and another one for the existence of optimal solutions.
As side results, a small collection of lemmas and an algorithmic process are also given in connection with the basic aim of the paper.

- Keywords: geometric programming, general entropy function, dual geometric program

- INTRODUCTION
- NOTATIONS AND BASIC ASSUMPTIONS
- BASIC PROPERTIES OF THE FUNCTION
- A NECESSARY CONDITION OF THE BOUNDEDNESS FROM BELOW OF THE FUNCTION
- A SUFFICIENT CONDITION FOR THE ACHIEVEMENT OF THE MINIMUM OF THE FUNCTION
- SHORTENING THE NON-LINEAR PART OF THE FUNCTION
- CONTRACTION OF THE LINEAR PART OF THE FUNCTION
- A NECESSARY AND SUFFICIENT CONDITION FOR THE BOUNDEDNESS OF THE FUNCTION FROM BELOW ON A CONVEX POLYHEDRON
- SOME POSSIBLE GENERALIZATIONS

The dual geometric programming problem can be written as follows: The function

(1.1)

is to be minimized under the conditions

(1.2) .

After a long process of preparations, towards the end of this paper a necessary and sufficient condition is given for the dual geometrical programming objective function to be bounded on its condition set and another one for the existence of an optimal solution (Theorems 8.1 and 8.2). To demonstrate that the boundedness of the objective function from below does not imply the existence of an optimal solution, consider the function

(1.3) ,

and the condition set*K*defined by inequalities(1.4) .

Then , and we can see by the help of the features described in Section 3, that for all and , that is does not take its minimum on

*K*for this concrete example.

Denote by

*c*or_{k}*x*the vector obtained from those or that are indexed with the elements of set_{k}*I*; by the objective function under consideration, and by a nonlinear term of it. Two vectors written side by side denote the scalar product of these. In this notation, the dual geometric programming objective function (i.e. the general entropy function) is the following:_{k}(2.1) .

Denote by

*m*the greatest element of the index set_{k}*I*, i.e. let_{k}(2.2) .

Then we can write the function also in the following way

(2.3) .

The domain of the function is the set of those in space for which

(2.4) if

and

(2.5) if

hold. The set defined by (2.4)-(2.5) will be denoted by

*D*.Herebelow the following problem more general than problem (1.1)-(1.2) will be considered: The function is to be minimized under the constraint , where

*K*is a closed convex set. We require of the constraint set*K*right through that(2.6)

be satisfied and there exist a vector

*x*in*K*for which(2.7) if .

In the remaining part of this paper

*K*will always denote a closed convex set endowed with these properties, although the prescription of these requirements will not be repeated every time in the particular theorems.The stipulation of the existence of a vector satisfying (2.7) does not restrict the generality and serves only for the simplification of the discussion. The theorems and other statements that will be put forward can be simply in principle -- but in a complicated way in respect of notation technique -- transferred to the more general case.

The closed convex cone

(2.8)

is called the recession cone of set

*K*and will be denoted by*K*^{2}. Similarly, for the recession cone of any other set superscript 2 will be used. If*K*is a polyhedron, then according to the Motzkin theorem*K*can be produced in the form(2.9)

where

*K*^{1}is a convex polytope.

It is known in the literature of geometric programming that the functions have the following properties:

a) for all .

b) is fulfilled if and only if at most one of the variables takes a value different from zero.

c) if and .

d) if and . Equality is here fulfilled if and only if

(3.1)

holds for every where denotes the

*j*-th component of*s*._{k}Properties c) and d) of functions are inherited by function the following way:

(3.2) if and ,

(3.3) if and .

Equality in (3.3) is satisfied if and only if (3.1) holds for all and .

As a further property, the asymptotic behaviour of function will be written along straight lines. It can be proved by the means of elementary analysis that in case of arbitrary vectors ,

(3.4)

The formulae (3.2)-(3.4) enable us to state the following properties of function : The value of the expression does not depend on if and only if and (3.1) is fulfilled for all , . If, on the other hand, but (3.1) does not hold for every such index pair (

*k*,*j*), then is a strictly monotone decreasing function of .

**Theorem 4.1.**If the function is bounded from below on set*K*, then the following assertions hold:a) for every .

b) If

*s*is an arbitrary vector for which and , then either*s*= 0 or_{k}*s*>> 0 for every ._{k}*Proof.*Let*x*be such a point of set*K*for which*x*>> 0 for every . According to the general assumption taken in Section 2, such a point_{k}*x*exists. According to the definition of the recession cone (or in case of*s*= 0 trivially) it is fulfilled that for arbitrary and . Therefore it follows from the boundedness from below of the function on the set K that(4.1)

for arbitrary . According to formula (3.4) describing the asymptotic behaviour of the function this is possible only if and in case of , and imply for arbitrary , . On account of , this postulate can be formulated also in the form of conditions a) and b).

The sufficient condition is expressed more generally in the following statement.

**Lemma 5.1.**If*K*is a non-empty closed convex set,*f(x)*is a continuous homogeneous convex function on a closed convex cone containg*K*and*f(s )*> 0 is fulfilled for every , then the function*f(x)*takes on its minimum on set*K*.The proof will be given by the use of the subsequent Lemmas 5.2-5.5. These, with one exception, can be found among the theorems published in [9].

**Lemma 5.2.**(**Corollary 8.3.3**in [9]). If*K*and*L*are closed convex sets,*K*^{2}and*L*^{2}are their recession cones and is not empty, then the recession cone of is identical with the set .**Lemma 5.3.**(**Theorem 8.4**in [9]). A non-empty closed convex set*K*is bounded if and only if*K*^{2}contains only the null vector.**Lemma 5.4.**(**Theorem 9.6**in [9]). If*K*is a non-empty closed convex set, which does not contain the null vector, then(5.1) .

**Lemma 5.5.**If*K*is a closed convex set,*f(x)*a homogeneous convex function on*K*and

*f(y)*> 0 is fulfilled for every , then the function*f(x)*takes on its minimum on*K*.**Proof.**It obviously follows from the assumptions that*f(x)*is bounded from below on set*K*. Assume indirectly that*f(x)*does not take on its minimum on*K*. Then there is such a series of vectors for which and the series*f(x*is convergent. Denote by^{(i)})*y*the unit vector for which^{(i)}(5.2) .

Then

(5.3) ,

and the series is convergent. Because of , so , and therefore for any condensation point

*y*^{(o)}of the series*y*,^{(i)},

contrary to the assumptions.

**Proof of Lemma 5.1.**Let*C*be a closed convex cone containing*K*and assume that the function*f(x)*is continuous and homogeneous convex on*C*. Let.

Obviously L is also a closed convex cone.

Consider first the case when there exists such a vector for which . The same condition can be laid down also such that the set is not empty. As

*f(s)*> 0 is fulfilled for of every , therefore the set contains only the null vector. According to Lemma 5.2, the set is the recession cone of the set . (*L*is a non-empty closed convex cone, therefore obviously*L*^{2}=*L*.) Now it follows by the application of Lemma 5.3 that the set is bounded. The continuous function*f(x)*takes on its minimum in a point*x*^{(0)}of the non-empty bounded closed convex set , and this point is obviously a minimum place also for the whole set*K*.In the not yet discussed case,

*f(x)*> 0 for every , and consequently*f(y)*> 0 for every .

Considering also the assumption of Lemma 5.1 we obtain that*f(y)*> 0 is fulfilled for every .

Since for the homogeneous convex function*f(x)*we have*f*(0) = 0, consequently*K*cannot contain the null vector, thus Lemma 5.4 is applicable for the present case. Therefore the above assertion with*y*can be also said declaring that*f(y)*> 0 is satisfied for every i.e. the assumptions of Lemma 5.5 are satisfied. Then, according to this lemma,*f(x)*takes on its minimum on*K*.

In this section there will be derived from the function another function of the same type which generally contains less nonlinear terms than the original function.

**Theorem 6.1.**Assume that there exists a vector for which and for arbitrary either*s*= 0 or_{k}*s*>> 0. Then we state the following:_{k}a) is bounded from below on the set

*K*if and only if the function(6.1)

is bounded from below on K; further

(6.2)

holds.

b) takes on its minimum on

*K*if and only if takes on its minimum on K, on the one hand, and for at least one vector from the minimum places of it is fulfilled that the value of the expression does not depend on , on the other hand.**Proof.**According to the assumptions, equality (3.4) holds with the last (lowermost) right hand side for every , and then, according to the remark made at the end of Section 3, the value of either does not depend on , or it is a strictly monotone decreasing function of for every ; therefore inequality is satisfied anyway. Consequently(6.3) .

According to the definition of the recession cone it follows from that and consequently

(6.4)

holds for every , therefore (6.3) holds also with the inequality of the opposite direction.

Assertion b) follows simply from a) and from the fact that is a non-increasing function of for every .

**Theorem 6.2.**Assume that there exists a vector for which and*s*>> 0 for every , and let_{k}(6.5) .

In this case for to be bounded from below on

*K*it is necessary (and in case of a polyhedral set it is also sufficient), that for every .**Proof.**The function is now a linear function of*x*. According to a well known theorem of linear programming, a linear function*f(x)*is bounded from below on a polyhedron*K*if and only if on the recession cone*K*^{2}. The assertion of Theorem 6.2 for the polyhedral special case follows simply from this and from Theorem 6.1. The necessity declared for the more general case follows from Theorems 4.1 and 6.1.

In this section we show that if the theorems contained in Sections 4-6 cannot be applied to the function and the set

*K*, then can be reduced to a function(7.1)

of similar type, which contains less variables than does. Now first we give some lemmas that will be applied for this purpose.

**Lemma 7.1.**(**Theorem 8.1**in [9]). If*K*is a non-empty convex set, while*K*^{2}its recession cone, then(7.2) .

**Lemma 7.2.**Let be a linear mapping which is defined on a set and which assigns to the vector the vector . Let*K*be a non-empty closed convex set, for which .Denote by

*K*^{2}the recession cone of*K*, and by*M(K)*^{2}the recession cone of the set(7.3) .

Moreover, let

(7.4) ;

then we assert that

(7.5) ,

and in case of a polyhedral

*K*(7.6) .

**Proof.**It is known from the theory of linear mappings that the image of a closed convex set is also a closed convex set, the image of a closed convex cone is also a closed convex cone, and the image of a polyhedral closed convex set is also a polyhedral closed convex set.Assume that and . Then there exist such vectors and , for which and . According to Lemma 7.1, , therefore .

Now by applying Lemma 7.1 to the image sets we obtain that , therefore (7.5) holds indeed.

If

*K*is a polyhedral closed convex set, then according to the Motzkin theorem there exists such a convex polytope*K*^{1}for which(7.7)

*K*=*K*^{1}+*K*^{2}and therefore

(7.8)

*M*(*K*) =*M*(*K*^{1}) +*M*(*K*^{2}) .As the set

*M(K*^{1}*)*is bounded, therefore the recession cone of the set*M(K)*is identical with the recession cone of the set*M(K*^{2}*)*, which latter is nothing else than*M(K*^{2}*)*itself . Exactly this fact is expressed by equality (7.6).**Theorem 7.3.**Consider the linear mapping(7.9)

which assigns to the vector

(7.10)

the vector

(7.11) .

Denote by

*M(K)*^{2}the recession cone of*M(K)*and let(7.12) .

Then we assert the following:

a) is bounded from below on

*K*if and only if is bounded from below

on*M(K)*, moreover in this case(7.13) .

b) takes on its minimum on

*K*if and only if takes on its minimum on*M(K)*.c) There exists such a vector , for which is fulfilled for

*j*=*m*_{0}+1,*m*_{0}+2, . . . ,*m*._{p}d) If in the set there exists a vector differing from the zero vector, then for all such vectors we have for at least one .

**Proof.**According to the definition of function , we have for all , therefore the range of function on set*K*is identical to the range of function on set*M(K)*, from which assertions a) and b) of Theorem 7.3 follow obviously.The validity of assertion c) follows from our similar assumption concerning

*K*taken at (2.7).The validity of assertion d) can be realized by indirect reasoning. If for a vector , holds for all then and therefore either or .

**Theorem 7.4.**If*K*is a polyhedron, and for the function , holds for every , where(7.14)

then for the function defined in Theorem 7.3, is satisfied for every .

**Proof.**In case of , we have , and thus is fulfilled for every . From this and from Lemma 7.2 it follows that for every . If for some we have , then according to assertion d) of Theorem 7.3, either or holds for at least one .Assume now indirectly that is fulfilled for some . Consider a vector for which . For this vector , such that according to the assumption of the theorem, has to be valid. According to the definition of set Z, holds for

*k*= 1, 2, … ,*p*. This way; however, we have come into contradiction with assertion d) of Theorem 7.3.**Theorem 7.5.**If*K*is a polyhedron, and for the function , is fulfilled for every , where*Z*is the set defined at (7.14), then the function takes on its minimum on*K*.**Proof.**According to Theorem 7.4, the function defined in Theorem 7.3 and the set*M(K)*satisfy the conditions of Lemma 5.1. From the application of this lemma it follows that takes on its minimum on the set*M(K)*. The definition of the function and the mapping*M*makes it obvious that also the function takes on its minimum on*K*.Hereafter it will be shown that jointly the theorems proved in Sections 4-7. in every imaginable case give some sort of usable information concerning the minimization of function on set

*K*. The range of the applicability of those theorems together with the conclusions which can be drawn are summarized in Tables 1 and 2. About set*K*it is still assumed now and henceforward thata)

*K*is a convex closed set,b) is fulfilled for all vectors ,

c) A vector

*x*in*K*can be found such that for .Under these assumptions the possible cases can be classified according to Table 1. If we assume, in addition, that

*K*is a polyhedron, then the classification of the cases is shown in Table 2.The assumptions belonging to the cases of Tables 1 and 2 are formulated in such a way that the relations between each cases and the corresponding theorems are easily recognizable. We have to remark; however, that the numbered cases do not exclude each other, e.g. case IV in Table 2 can be substantially considered as a part of case III, while cases I-IV contained in Table 1 exhaust all possibilities, and the same can be said of cases I-III contained in Table 2.

In order to prove our previous assertion concerning Table 1, examine the situation when the possibility of cases I-III is excluded. Then all of the following must be true:

a) for every .

b) If and , then

*s*= 0 for every ._{k}c) There exists a vector for which .

Now let us first assume that ; then b) and c) cannot hold simultaneously, because in case of

*m*_{0}= 0 a logical contradiction comes about between them, while in case of*m*_{0}= 1, the functions and are identical, therefore in this case assertion d) of Theorem 7.3 just asserts that b) and c) cannot hold simultaneously. Accordingly, only is possible, i.e. the exclusion of the possibility of cases I-III entails the validity of case IV.

In this section we restrict ourselves to the case when

*K*is a closed convex polyhedron.It was shown in the previous section that one of the cases I-III of Table 2 must be true. Therefore, if a vector

*s*corresponding to case III cannot be found, then for having case II it is necessary and sufficient that be bounded from below in*K*. If the question of boundedness cannot be decided this way, i.e. case III is in existence, then denote by*s*^{(1)}(next by*s*^{(2)},*s*^{(3) }etc. in course of the possible iterations of the process) the vector named*s*in the assumption belonging to case III; by*S*_{1}(next*S*_{2},*S*_{3}etc. in course of the possible iterations) the set of the indices*k*for which*s*_{k}^{(1)}>>0 (*s*_{k}^{(2)}>>0 , etc.); by (next , etc. in course of the possible iterations) the function derived by the shortening of the nonlinear part according to Theorem 6.1. After repeating the iteration with the derived function of the process, in a finite number of steps, -- say in the*q*-th step -- only case I or case II may be in existence.What has been said in the above train of thought can be formulated in the following theorem.

**Theorem 8.1.**The function is bounded from below on the polyhedron*K*if and only if for some there exist such vectors*s*^{(1)},*s*^{(2)}, . . . ,*s*and such, pairwise disjoint, non empty index sets^{(q)}*S*_{1},*S*_{2}, . . . ,*S*for which_{q}the following holds:

a)

(8.1) .

b) Denoting

(8.2) ,

also holds

(8.3) .

c)

(8.4) ,

where

(8.5) .

In the case of the fulfilment of the assumptions of Theorem 8.1, we will say that '

*problem is q-regular*'. (One can notice that in this terminology, 0-regularity means the fulfilment of the assumptions of Theorem 7.5 that is the fulfilment of the*Slater-condition*.) Theorem 8.1 can be formulated now also in the way that the function is bounded from below on the polyhedron K if and only if the problem is*q*-regular for some non-negative integer*q*.According to Theorem 7.5 the assumptions of Theorem 8.1 also guarantee that the function takes on its minimum on

*K*. In the knowledge of this fact, the assertions of the following theorems can be shown by the use of Theorem 6.1.**Theorem 8.2.**In the case of the fulfilment of the assumptions of Theorem 8.1, the function takes on its minimum on*K*if and only if the following hold:a) takes on its minimum on the polyhedron

(8.6) .

b)

(8.7) .

**Theorem 8.3.**If the conditions of Theorem 8.1 hold and eq. (8.7) is fulfilled, then the set of all optimal solutions of problem(8.8)

is identical with the set of all optimal solutions of the original problem

(8.9) .

Theorems similar to the ones occuring in Section 8 can be formulated if we do not restrict ourselves to the case of a polyhedral constraint set. For this more general case, the algorithm otlined at the beginning of Section 8 should be modified such that for every function the transformation defined in Theorem 7.3 is applied, and after that cases I-III of Table 1 are considered. The train of thought; however, can be expressed in the more general case in a way only slightly different from the earlier one. Essentially, we exploit only the properties of functions , without explicit application of the transformation defined in Theorem 7.3. Then the following theorem -- parallel to Theorem 8.1 -- can be formulated.

**Theorem 9.1.**The function is bounded from below on the set*K*if and only if for some there exist such vectors*s*^{(1)},*s*^{(2)}, . . . ,*s*and such pairwise disjoint non-empty index sets^{(q)}*S*_{1},*S*_{2}, . . . ,*S*{1, 2, . . . , p} for which the following are true:_{q}a)

.

b) Denote by

*M*the linear mapping which assigns to the vector_{i}*x*the vector whose components are.

Let be formally the same as the function defined at (8.2), i.e.

.

With these notation our assumption is the following:

.

c)

.

For a similar generalization of Theorems 8.2 and 8.3 it is enough to replace '8.1' by '9.1' and 'polyhedron' by 'set' in the text of the assertions of these theorems.

A further generalization can be formulated by disregarding the assumption of canonical representation -- i.e. not requiring the existence of a vector

*x*in*K*satisfying (2.7). Now for*k*= 1, 2, … ,*p*, denote.

Then the reqirement

*s*>> 0 can be relaxed to and similarly_{k}*x*>> 0 to in all their occurances in the text of Sections 4 to 9 of this paper._{k}TABLES

REFERENCES

- Avriel, M. and Williams, A.C.,
*On the primal and dual constraint sets in geometric programming*, J. of Math. Anal. Appl. 32(1970)684-688. - Avriel, M. and Williams, A.C.,
*Complementary geometric programming*, SIAM J. Appl. Math. 19(1970)125-141. - Duffin, R.J. and Peterson, E.L.,
*Duality theory for geometric programming*, SIAM J. Appl. Math. 14(1966)1307-1349. - Duffin, R.J. and Peterson, E.L.,
*The proximity of (algebraic) geometric programming to linear programming*, Math. Progr. 3(1972)250-253. - Duffin, R.J., Peterson, E.L., and Zener. C.,
*Geometric Programming*, John Wiley, New York, 1966. - Gale, D.,
*The Theory of Linear Economic Models*, McGraw-Hill, New York, Toronto, and London, 1960. - Gochet, W. and Smeers, Y.,
*Constraint sets of geometric programs characterized by auxiliary problems*, SIAM J. Appl. Math. 29(1975)708-718. - Klafszky, E.,
*Geometric Programming*, Hungarian Committee for Systems Analysis , Seminary Notes, Mathematics, No. 11.,Budapest, 1976. - Rockafellar, R.T.,
*Convex Analysis*, Princeton University Press, 1970.