Multi-criteria analysis with partial information about the weighting coefficients

ELSEVIER European Journal of Operational Research 81 (1995) 291-301 EUROPEAN JOURNAL OF OPERATIONAL RESEARCH Theory and Methodology Multi-criteria analysis with partial information about the weighting coefficients E. Carrizosa, E. Conde, F.R. Fernandez, J. Puerto * Departamento de Estadfstica e Investigaci&n Operativa, Universidad de Sevilla, 41012 Sevilla, Spain Received October 1992; revised April 1993 Abstract In this paper we address the problem of ranking a set of alternatives with partial information about the weighting coefficients. We introduce a family of quasiorders that are easily interpretable and manageable, which includes among others, the natural quasiorder in R n and other well known preference structures in the literature. The enrichment of the preference structure with respect to the natural quasiorder is measured by means of an absolute measure we introduce. Keywords: Multiple criteria decision making; Partial information; Weights 1. Introduction Nearly all the real world decision problems involve more than one objective and can be formulated in a natural way using the multi-criteria approach: 'Max' f(x) - (f,(x),...,fn(X)) (1) where fi:X--* ~ is the evaluation of the i-th objective (i = 1 .... , n) and X is the set of alternatives to be ranked. In this formulation, and without additional information about the objectives, one can obtain a partial order R I in X: Given two alternatives x, y ~ X, we say that x is as preferred as y following R, (xRzy) iff * Corresponding author. fi(x)>fi(y) Vi=l,...,n, or equivalently, xRiy iff ~_, Will(X)> ~_, wifi(y ) l <i <_n l <_i <_n Vw>O, w~ ~. AS the partial order Rz is typically too vague (and the set of nondominated alternatives is too large) a number of procedures has been proposed in order to enrich the preference structure above (Promethee [5], Electre [19], interactive methods [11], the utility approach [9], etc.). The interested reader is referred to the seminal paper of Roy [20] for a synthesis of the main approaches to this problem which have been studied. In the utility approach, one assumes the existence of a function U: ~ ~ R, in such a way that an alternative x is considered as preferred to y iff U(fi(x) ..... fn(X)) >-- U(fl(Y) ..... fn(Y))" 0377-2217/95/$09.50 © 1995 Elsevier Science B.V. All rights reserved SSDI 0377-2217(93)E0270-8 292 E. Carrizosa et al. ~European Journal of Operational Research 81 (1995) 291-301 Furthermore, under mild regularity conditions [12], U is a linear function, i.e.: one assumes the existence of a vector w* = (w*,...,w*)_> 0 such that the Decision Maker (D-M) preferences are given by the order R: xRy iff ~,, wTk(x ) ~ ~_, w~k(y ). l~i~n l~i~n However, one of the main drawbacks of this approach is that the D-M would have to estimate numerically the weight (the vector w*) that each criterion brings to the final score of an alternative and he is not always willing to do so [6]. This choice is a critical step, and, although some methodologies have been developed to attain this goal, such as the Entropy method [11], the Saaty method [21], the Solymosi and Dombi technique [23,15], etc., they require, in our opinion, too much specialized information from the D-M. Instead of providing the exact w*, the D-M might have some knowledge about w* and this information can be used to obtain a (partial) ranking in X which enriches the original preference structure R I. In this paper we show that exact numerical weights are not always necessary. Instead of this exact estimation, the D-M gives only certain linear relations which express partial information about the marginal substitution rates between the criteria. This approach is not new and some work in this field can be found in the literature; Kirwood and Satin [13], Hazen [10], or Eiselt and Laporte [8], derive conditions to determine ordinal rankings between alternatives using partial information about weighting constants. For instance, in the qualitative approach [17] the D-M is requested to estimate only the rank order of the criteria. If the criteria are arranged in decreasing order of preference, with w 1 > • • • > wn, then xny iff w(f(x)-f(y))>O Vw>_O, WI~_ ... ~_Wn or, equivalently (see [17]) xRy iff gfj(x)> ~]fj(y) Vi=l,...,n. j<_i j<i In a more general setting, the D-M is requested to estimate a linear operator which mixes the weights, i.e. the D-M asserts that the vector w* belongs to a certain polyhedron. Thus, given a linear operator A, we define the preference between the alternatives by means of the binary relation R A as xRay iff w(f(x)-f(y))>_O Vw>O, Aw > O. (2) All these binary relations are quasi-orders (i.e. reflexive and transitive) [26] and can be used to partially rank the alternatives, although their use requires the solution of a linear problem to compare every pair of alternatives. However, if the extreme points wl,...,w m of n the polytype {w : Aw > O, w > O, F~i= lwi - 1} are known, the relation R a is xRAy iff wk(f(x)--f(y)) >O Vk = 1,...,m, (3) which simplifies its use. The following example iUustrates the comments. Example 1.1. The staff manager of a consulting firm must rank four different executives of a bank according to the budget they estimate for next year based on three financial criteria. These criteria are: loans given to clients (C1), clients' savings deposits (C 2) and redemptions achieved (C3). Each manager's budget must adhere to the financial policy of the bank. This policy is given in form of three constraints: 1. The bank wishes that the credits are greater than the sum of the redemptions plus 1.5 times the savings. 2. The redemptions must be less than 0.1 times the savings deposits. 3. The redemptions must be non-negative. Consider the following matrix where the row i represents the action proposed by the manager i, i= 1,...,4: C1 C2 C3 a 1 11 12.2 0 a 2 5 4 5 a3 11 11 13 a 4 11 12 2.3 E. Carrizosa et al. / European Journal of Operational Research 81 (1995) 291-301 293 Let F denote the above matrix. The constraints imposed by the bank can be written as Wl>I.5w2+w 3, W2>10W 3, W3>__0. Thus, the linear operator A and the extreme points matrix E, of the polytope {w : Aw > O, w > 0, Ei~__ lW~ = 1} are respectively (1.-1.5-1) 1527 A= 0 1 -10 and E= 2 10 5 27 1 0 0 1 0 and the qnasiorder given by the above constraints in the action's set is obtained from the matrix F × E b means of the natural quasiorder. Hence, -11 287 298 ] 25 27 5 2...~3 125 5 27 F×E= 11 11 22- ~ 11 zs~ 298.3. 25 27 generates a relation which can be represented by the graph of Fig. 1. However, with a different set of constraints imposed by the bank given by the operator A t , and whose matrix representation and extremes are 1 -2.5 -2.5) A1 = 0 1 -13 , 0 0 1 1 5 35 7 49 E 1 0 2 13 7 49 , 0 0 +9 it is easy to see that the quasiorder generated is actually an order whose graph is given in Fig. 2. Thus, it is possible in many cases to identify the Fig. 1. The graph of R A. [K]-- [?N-- [] Fig. 2. The graph of RAf preferred alternatives without knowing the numerical estimates of the criterion weights. However, the preference structure determination in its general formulation (the determination of the set of non-dominated alternatives) is equivalent to the vertex polyhedron enumeration, which is not a polynomially solvable problem in the size of the inequality system defining the polyhedron [7,22]. The paper is structured as follows: In Section 2, we introduce a class of matrices (Q-operators) and study some properties of the quasiorders they induce. In Section 3, we characterize some Q-operators that are easily understandable and manageable. In Section 4, we extend the results obtained in previous sections to a broader class. The paper finishes with some conclusions and possible extensions. 2. The class of Q-operators Every linear operator A belonging to the set of real matrices of dimension k×n, defines a quasi-order R n on the set X of alternatives: Given a matrix A ~ R kxn we define the polytopes C~ and CA: C~=(w:Aw>O,w>O, i=l~Wi=l} (4) and CA = w:Aw>_O, wi=l , (5) i= and the quasiorder R a on X: xRaY iff ~wifi(x)>__ ~wifi(y ) i=1 i=1 Vw ~ C + . (6) In this section we introduce a class of qua- 294 E. Carrizosa et aL / European Journal of Operational Research 81 (1995) 291-301 siorders induced by linear operators (Q-oper- ators) that have very interesting properties: the extreme points of the corresponding C~ can be very easily obtained. Definition 2.1. A linear operator A ~R.x. is said to be a Q-operator if det(A) ~ 0 and A -1 _> 0 (componentwise). For a Q-operator A, the inverse operator A-1 exists; we denote its elements by aii, and by /x the vector of sums of the columns of A-1: The matrix above is easily shown to be a Q-oper- ator, thus Paelinck's theorem appears as a direct consequence of our Theorem 2.1. Indeed, A -a can be readily obtained, and the extreme points of C~ are the columns of the following matrix: 1 1/n 1 ~ ... 1 1/n 0 -~ ... 0 0 ... 1/n 1/n 0 0 ... 1/n A -l= (aij) and /xj= E aij" (7) l <_i <_n Theorem 2.1. If A E ~Xn is a Q-operator, then C] is the convex hull of the columns of A- 1, each one normalized in order to add 1. Proof. As A-l>-0, it follows that /~i > 0, Vj = 1,...,n. Let D be the diagonal matrix such that dii = 1/Izi, Vi, and let e e R ~ denote the vector of ones. Given w e E', one has w~C~ iff3z such that w=A-aDz, Dz>_O, A-aDz >_ O, eA-1Dz = 1 i.e. (recall that D >_ 0, A -1 _> 0, eA-1D = e)" weC~ iff3z such that w=A-1Dz, z>O, ez=l. In other words, w ~ C~ iff w can be represented as a convex combination of the columns of A-aD, as asserted. [] Remark 2.1. The well-known Paelinck theorem [17] proven, among others, in [17,2,6,13], reduces to the calculus of the extreme points of the polyhedron {w e R ~ : w a > w 2 > • • • >_ w~ >_ O, ~i= aWi = 1}, which is of the form C~ for the following matrix A: (010 0 1 ~1 ... 0 0 0 ... 1 Remark 2.2. In the centroid method of Solymosi and Dombi [23,15], given the polyhedron C~, one proposes as w* the average vector of the extreme points of C]. By the theorem above, this w* is easy to obtain when A is a Q-operator: w*= (1/n)A-ae. Remark 2.3. It should be noted that these kinds of quasi-orders do not necessarily need n linear relations. If the D-M is only able to supply k < n, these operators can be transformed without incorporating any additional information. This is possible by adding the natural relations w i > 0 and removing the redundant ones. As an illustration, consider a problem with three objectives, where the D-M states that w a > w 2 but is unable to provide more information about the weights. Then he will have the following operator A and its inverse A-a: A= 1 , A -1= 1 0 , 0 0 1 and the quasi-order would be (fa(x) > fl(Y), XRAY iff ~fa(x) +f2(x) >fx(Y) +f2(Y), If3(x) >f3(Y)" Remark 2.4. Another important property of Q- operators is the fact that they induce interval weights, which are easy to obtain, allowing a certain degree of sensitivity analysis in the nu- E. Carrizosa et al. / European Journal of Operational Research 81 (1995) 291-301 295 merical estimation of weights. The exact interval of w i ~ C~, Vi = 1 ..... n, is given by WiG [Inj!n(olij//~j), max(otij/IXj)l, i= 1,...,n, where aij and /xy were defined in (7). Indeed, for i = 1,...,n, let z i be the optimal value of the linear program max{w i : w ~ C~ }. This z~ is attained at an extreme point of C~ and, thus, by Theorem 2.1, z~ = maxj(aiJixj). Similarly, one concludes with the minimum. For instance, for the quasiorder R A described in Remark 2.1, it is easily seen that wl~[1/n, 11, w2~[0,½1,..., w,~[O, 1/n]. As a final consequence, observe that one can also obtain the maximum and minimum value associated with each alternative x E X when the weight w varies in C~, which is the basis of some decision-making methods (see, e.g. [3]). Indeed, as finding the maximum (respect. the minimum) value of wf(x) when w varies in C~ reduces to solving the linear program max{wf(x):w ~ C~} (respectively min{wf(x) : w ~ C~}), Theorem 2.1 implies that min ~ ai---k fi(x) <_wf(x) < max ~ °tik fi(x ) k i=1 tZk k i=1 ~k Vw~C~, x~X. The family of operators proposed in the previous theorem is maximal in the sense that the unique set of weights in ~n with n extreme points whereby all the weights generated are non-nega- tive are those with A- ~ _> 0. This is stated in the following theorem: First, we show that D A = D~. Indeed, it is evident that D~ c D A. Now, let w ~ D A, and we will show that w ~ D~. Obviously, if w = 0, then w ~ D~, so we can assume that w v~ 0; in other words, we only have to consider the cases (ew ~ O) and (ew < O, w 4= 0). Case 1. ew>O. Let w l=(1/(ew))w~C A= C~. Hence, w 1 > 0, thus w >_ 0, which (recall that w E D A) implies that w ~ D~. Case 2. ew < O, w 4= O. As by assumption C A v~ ¢, there exists w°~ C A = CJ. Define w 1 as follows: 1 -- ew w 1 -- --w dw °. 1 - ew 1 - ew Such w 1 verifies that w 1 ~ D A and ew I = 0. Furthermore, at least one component of w 1 is negative. Indeed, if w 1 > 0, then, as ew 1 = O, it would follow that w I = 0, thus (ew)w ° = w; as w 4= 0, ew < O, one would obtain ew < 0, thus 0 <Aw = (ew)Aw°; as 0<Aw ° and ew <0, this would imply that Aw °= O, i.e.: (recall that A -x exists) w ° = 0, which is a contradiction. Hence, w a has at least a negative component, thus there exists some A, 0 < h < 1 such that the vector w 2 = hw* + (1 - h)w ° verifies that Aw 2 > O, ew 2 > 0, w 2 has at least a negative component. By case 1, w e ~D~, which is a contradiction. Hence, D A = D~, as we claimed. With this, we are in position to show that A-l> 0. Indeed, let v be a column of A-l; as AA -1 gives the identity matrix, it follows that Av >_ O, i.e., v ~ D A, thus v > 0. Then, we have shown that all the columns v of A-1 verify that v >_ 0, thus A-1 >_ 0, as asserted. [] Theorem 2.2. Let A ~ ~n×n be a linear operator such that det(A) ~ 0 and C A -4= ~J. Then C + = C A iff A-l >O. Proof. It is evident that, if A-l> 0 then C~ = C A. We now show the converse. Let A be an n × n matrix with det(A) ~ 0 such that C~ = C A. Define the sets D n and D~: D4={wE~n:Aw >>. O}, D~={w~n:Aw >O, w>O}. The process of supplying information to the initial multi-criteria problem transforms the preference scheme. Thus in the beginning, i.e. when no information is available, one alternative x ~ X is preferred to another y~X iff f(x)>_f(y) (component-wise), That is, with no information, the preference scheme coincides with the Pareto quasi-order. So, it seems natural that in the process of supplying information, the more precise information the D-M gives, the more accurate quasi-order will be generated. It is evident that 296 E. Carrizosa et al. / European Journal of Operational Research 81 (1995) 291-301 every quasi-order R A with A- 1 >_ 0, improves the no-information-relation because it reduces the set of weights. But given two relations R A and R B it is not clear how to determine which of them is the most accurate. However, this is very important because it allows us to know the degree of knowledge shown by the D-M about his own problem. The more accurate the quasi-order, the better the knowledge of the problem. Moreover, it seems natural that the accuracy of a relation is inversely proportional to the magnitude of its set of weights as is proposed by Rios [18], so we define the accuracy of R A in the following way. Definition 2.2. Given a Q-operator A, the accuracy of RA, AC(RA), is defined as IXn_l(C~) / /xn_l(C~), where /x,_ 1 represents the Lebesgue measure in Nn-~. The limit values for the accuracy are given in the following proposition• Proposition 2.1. If A is a Q-operator, then 1 < AC(RA) < + oo. Proof. First of all, ix~_l(C~)<lx,_~(Cf-), VA E N~xn. Second, as A -1 > 0, C~ is a simplex in the hyperplane El<_i<,wi = 1 and /X,_x(C~)>0. Then 1 _< AC(R A) < +m. [] The value 1 corresponds to the first no-infor- mation case and, hence, we can see the accuracy as a reduction measure of the set of weights. More precisely every operator that generates an order on R n has an accuracy +~. An order relation of this kind could be seen as the limit case of a convergent sequence of quasi-orders with increasing and more precise information. Finally we shall give the expression of the accuracy: Theorem 2.3. Let A be a Q-operator. Then, AC(R A) = [det(A)lFIl<j_<~/xj, where tz k was defined in (7). Proof. First, recall that the volume of the simplex generated by the points xX,..., x n in ~-1 is [1] 1 1 ... 1 1 x~ x 2 ... x~' (n- 1)~ det ". X 1 X 2 X n n--1 n--1 • " " n--1 Let e i ~ ~" be the vector whose i-th component is one and zero the rest. Consider the system of reference ~ in the hyperplane H = {w:ew = 1} that has e 1 as origin and the vectors e i-e 1, i = 2 ..... n, as generators. Then, any point w = (wa,. •., w n) ~ H has coordinates (w2,..., w n) in ~9~. By Theorem 2.1, CJ is generated by the vectors Olli// ]Z i I azz/tZi. ], i= l,...,n. I Olni/~i ] Hence, AC(RA) 0/21/],-£1 0/22///£ 2 ... 012n/1.£ n = det . . O/nl//l-L 1 ang///L~ 2 ... Olnn//l~n Thus (recall that /-~i = E]=laij, Vi), AC(RA) [ a11//]£ 1 O/12///'£2 ... Olln//~n --1. = det " -. O/nl//~L~ 1 Oln2/~ 2 ... ann/]£n In other words, 1 AC(RA) = [det(A-1)[FI]=l(1//xj) n = [det(A) I 1-I/xj, j=l as asserted. [] Example 2.1. For the quasiorder R A described in Remark 2.1, one has that det(A)= 1 and /xj =j, Vj', thus AC(R A) = n! E. Carrizosa et al. / European Journal of Operational Research 81 (1995) 291-301 297 For the quasiorders R A and RA~ introduced in Example 1.1, the improvement in the order relation from R A to RA1 can be measured by means of the accuracy: AC(R A) = 67.5 and AC(Rq)= 171.5. 3. Some families of Q-operators The widest class of operators we can deal with is characterized in Theorem 2.2. However, the condition shown in the previous paragraph is difficult to check beforehand. So in order to enable the D-M to apply these relations, we propose two sub-classes belonging to the original one with three important properties: 1. to know beforehand that they belong to the broad class; 2. to be easy for the D-M to understand and accept; 3. to be sure that the set of weights it generates is not empty. A procedure to obtain a class of these operators consists in offering the D-M the comparison of each criterion fi(-) with at most a coalition of the remaining criteria. Usually, the exact determination of the weights is made by the trade-off between a criterion and the remainders. But this methodology cannot be used if the D-M does not give its preferences so precisely. Alternatively, our approach proposes to replace this equivalence by inequalities which are quite acceptable for the D-M. In this process, the more precise information the D-M supplies, the more accurate the quasi-order it generates, and hence it is closer to the cardinal utility. Therefore, the information required from the D-M about the criterion fi(.) which he is willing to give, will have the following form: Wi ~--- E aijwj, aij ~ O, E aij ~ 1, j4=i j~i where aij represents the minimum marginal substitution rate of fi for fj. We should notice that when no information is available, this kind of relation will be wi > 0. But, even by supplying small values of ais one improves the accuracy of the quasi-order given by the D-M and avoids the problem of the exact estimation of the weights (cardinal utility). Example 3.1. We shall deal with the following example in which the D-M has three objectives, a set X of feasible alternatives, and he is able to give the required information in the form of inequalities. W 1 ~ 0.5W2, W 2 ~ 0.5W 1 q- 0.5W3, w 3 > 0.5w 2 . Then the operator is defined by the matrix A, whose extreme points are given by the columns of the second matrix and the interval weights are 1 -0.5 0 A = -0.5 1 -0.5 0 -0.5 1 1 1 1 2 4 ~ W1 ~ [61 ' 1] 1 1 1 E = 3 2 X W2 ~ [31-' 1]. : : : w.~[L½] 6 4 g Hence its accuracy AC(R A) = 18. This example suggests the possibility that these kinds of operators have the property of inverse positive. In order to clarify the terms used in the following results we introduce two classical concepts [16]. Definition 3.1. A linear operator A ~ R nxn is diagonally dominant if ~_, laijl < laiil, i= l,...,n, ]~i and strictly diagonally dominant if strict inequalities hold for all i = 1 ..... n. The result suggested by the example above is stated in the following theorem. Theorem 3.1. Let A ~ ~nXn be a diagonally dominant linear operator such that aij <_ 0, i v~ j, aii > 0, V i = 1 ..... n, and det(A) v~ 0. Then A -: > O. 298 E. Carrizosa et aL ~European Journal of Operational Research 81 (1995) 291-301 Proof. For simplicity and without loss of generality, we consider aii = 1, Vi. As A is diagonally dominant, the inequalities aii >~ --Y'.j~iaij, Vi = 1 .... ,n, hold. Then every cofactor Aij of the matrix A is nonnegative [4]. Let A -1 be the inverse of A, whose elements are a u =Aii/det(A). Then the sign of aij., Vi, j, coincides with the sign of det(A) and det(A) 1 a12 ... aln 0 1 -- a12a21 ... a2n -- alna21 = det 0 an2 -- al2anl ... 1 -- alnanl 1 a12 "'" aln ~ = det O A 1 J =dee(A1). Besides, the matrix A~ is diagonally dominant because for all i = 2 .... , n their elements verify 1 - aliail > 0 and 1+ ~ aij+ail (- ~au)>l+ Zaij>_O. j¢i,j~a j4=l / j#i Thus, we can repeat this reasoning n times and we obtain that det(A) > 0. [] Corollary 3.1. Let A ~ ~x~ be a strictly diagonally dominant linear operator such that aii ~ 0, i =~ j, aii > O, Vi = 1 ..... n. Then A- 1 ~ O. he can compare the importance of several criteria between them. So the expression obtained for the criterion f~ ranked in i-th position is wi>- ~,auw j, ai~>-O, j>i where air represents the minimum marginal substitution rate of fi for fp In this sub-class the three properties enumerated at the beginning of Section 3 also hold and they belong to the class characterized by Theorem 2.2. Theorem 3.2. Let A ~ W '×n be a diagonal positive triangular linear operator such that a u < O, Vj > 1. Then A- 1 > O. Proof. In this situation det(A)= ~[l<i<naii > 0 and the cofactors are Aij= ]'-I akk>0 Vi < j, k~i,k~j Aii = YI akk > O, i = l,...,n, k~i Aq=0 Vi>j. Hence, reasons analogous to the ones we used in Theorem 3.1 prove this theorem. [] 4. Non-homogeneous Q-operators This class of operators is closely related to a well known family of linear operators called M- operators [16,25], but the first one exhibits in its favor the easy interpretability and manipulation because the M-operators require properties of irreducibility or strict diagonal dominance, not needed in the proposed class. It must also be noted that the well known ordinal relation mentioned in Remark 2.1 is an example of such an operator. Another important sub-class of operators belonging to the class characterized by Theorem 2.2 are the triangular and nonpositive off-diagonal elements. This type of operator corresponds to the situation in which the D-M is able to do a certain rank order in the criteria and in addition, On many occasions, the D-M is willing to obtain at least certain levels in his weights [14], so the quasi-order is given by a linear operator A and a level vector A _> 0. In this case, the set of feasible weights is given by the polytope (w ~ ~ :w >__ O, Aw >_,t, C~,~) = w i = 1 / . (8) E i=l,...,n I Hence the pair (A, A) induces the relation R(A,a ) given by XR(A,;~)y iff w(f(x)--f(y)) >O Vw e C(5,, ). (9) E. Carrizosa et al. /European Journal of Operational Research 81 (1995) 291-301 299 An interesting type of these relations are those defined by a matrix A with non-negative inverse (componentwise) and eA-1A < 1. In what follows we say these relations are induced by a non-ho- mogeneous Q-operator (A, A). To this kind of relations we can extend all the previous results with minimum effort as can be seen from the following theorem. Theorem 4.1. Let (A, A) be a non-homogeneous Q-operator. Then w -A-1A w~C~,A) iff l_eA_lh ~CJ. Proof. w~CfA,h ) iff 3t>OIAw=h +t, w>0, ew = 1; i.e. (recall that A -1 >__ 0): w~C~,~) iff 3t>OIw-A-~h=A-lt, ew = 1. As e(w -A-1A) = 1 - eA-1h > 0, it follows that 1 w~C~,x) iff :lt>Ol l_eA_lh (W-A-1h) t =A -1 iff 3t>__0[W =A-if, 1 - eA-1h e~ = 1, w -A-1A whereff- 1-eA-1A iff ff~C~. Which concludes the proof. [] Corollary 4.1. If (A, A) is a non-homogeneous Q-operator, then the set C(~,x) is the convex hull of the columns of the following matrix: /x11(1 - Eke_ltZkAk) A2 A_I[ 1 - mogeneous case, we wish to pay some attention to the following problem. We know (recall Remark 2.3) that there exist many quasi-orders that produce the same intervals for their weights; but, are there any of them which must be emphasized? The answer is affirmative and among these operators one is particularly interesting; it is known in the literature as E-cone [24] which is one of our non-homogeneous Q-operators. This kind of operator (the E-cone) is based on relations of the form w i>_k i, i = 1,...,n (one per objective)• Their two main properties are that the set of weights it defines includes the interval weights considered beforehand and also is minimal (in the inclusion sense among the operators of this kind). We state and prove this in the following theorem• Let k = (k 1 .... , k n) >_ 0. Consider the set C+(l,k)---- ( ~n Wi__ Wi= } w ~ : >ki; ~ 1 l <_i <_n and for each a,/3 ~ ~n such that 0 _< ag _< fig _< 1, Vi, and F,~=la i < 1 < E~=I/3g, the set (see [2]) X(a,13)=(w~n:OtNWN/3; ~ wi=l}. l <_i <_n Theorem 4.2. The vector k = (kl,..., k n) with k i = max(a i, 1 - Ej~//3j), i = 1 ..... n, determines the minimum set CfLk) containing the set X(~,t~). A 1 1 -/xKl(1 - Ek,2/ZkAk) An where IX j, j = 1 ..... n, were defined in (7). ... A 1 ''• /~2 ... 1--/X21(1--Ek.n~k;tk) Although all the properties holding in the homogeneous case can be extended to the non-ho- Proof. For all (kl,..., k n) >_ O, El <_i<_nki __< 1, one can obtain, using Corollary 4,1, that C(~,k ) is the