CONSECUTIVE OPTIMIZORS FOR A PARTITIONING PROBLEM WITH APPLICATIONS TO OPTIMAL INVENTORY GROUPINGS FOR JOINT REPLENISHMENT* A. K. Chakravarty J. B. Orlin and U. G. Rothblum Sloan W.P. No. 1383-82 November 1982 *Support received from the National Science Foundation Grant ENG-78-25182 and Grant ECS-8205022. A. K. Chakravarty, Washington State University, Department of Management, Pullman, Washington 99163 J. B. Orlin, Massachusetts Institute of Technology, Sloan School of Management, Cambridge, Massachusetts 02139. U. G. Rothblum, Yale University, School of Organization and Management, New Haven, Connecticut 06520 CONSECUTIVE OPTIMIZORS FOR A PARTITIONING PROBLEM WITH APPLICATIONS TO OPTIMAL INVENTORY GROUPINGS FOR JOINT REPLENISHMENT by A. K. Chakravarty J. B. Orlin and U. G. Rothblum Abstract We consider several subclasses of the problem of grouping (indexed 1, ..., n) g(S1, ..., S ). into m items subsets so as to minimize the function In general, these problems are very difficult to solve to optimality, even for the case conditions on n g(-) m = 2 . Here we provide several sufficient which guarantee that there is an optimum partition in which each subset consists of consecutive integers (or else the partition satisfies a more general condition called "semi-consecutiveness"). S1 ... , S Moreover, by restricting attention to "consecutive" (or "semi-consecutive") partitions we can solve the partition problem efficiently for small values of g m . If in addition, is symmetric then the partition problem reduces to a shortest path problem, solvable in O(n2m) steps. We apply the above results to the problem of grouping inventory items into subgroups with a common order cycle per subgroup so as to minimize the resulting economic order quantity costs. Under relatively minor assumptions on the cost structure and on the set of feasible policies, we may reindex the items a priori so as to guarantee the optimality of a "consecutive" partition. - 1. 1 - Let al, ..., a some integer 0 and r Introduction bl, ..., n, bl, bn br ..., be real numbers ordered so that for are negative, br, ..., b r+l' n are non- negative and a1 / b1 where for a > 0 . bi = 0 If whose coordinates are let N = {, bS = sets, say ai / b i / br+l , to be +* a i / bi As usual, we let and . ar+ a •a or -- /b (1) according as is defined arbitrarily so and b denote the vectors b i , respectively. For each subset S of N let aS = iS bi .iS . We consider the problem of partitioning the set S, ... gm(S1s hm (-) function ai ..., n S real-valued function where and ai = bi = that inequality (1) holds. Let / b we consider ai < O . or a N and into m , including possibly some empty sets, so as to minimize a g (S1 '' Sm ..., S ), hm(as1' b 1 where g () ' a 1 is a real valued function of h (.) m ai 2m is assumed to have the form ' bs'' ) m (2) im variables. We allow that the be nonsymmetric, in which case the cost of a partition depends on the order in which the sets are selected. A subset S c N integers, e.g., secutive. is called consecutive if its elements are consecutive {4,5,6}. A partition consecutive for each P = {S1, ..., S} i . if there is a permutation S (1) u ... u S In particular, the empty set is considered to be con- A partition S3, S3 = P {S1, ..., S } of the set of integers are consecutive for {S1 ,S 2 ,S 3 } = {{3,7,8}-, {1,2,91 subsets is called consecutive if {4,5,61} u S1 , S3 u S 1 u S2 is is called semi-consecutive 1, ..., m j - 1, ..., m . Si such that the sets For example, the partition is semi-consecutive because the three are all consecutive. II 2- The purpose of this paper is to provide sufficient conditions that the optimization problem defined above has an optimal consecutive or semi-consecutive partition. Specifically, we show that if h m is concave in its then there is an optimal semi-consecutive partition. addition, b 0 or Also, we show that if a 2m variables Moreover, if in 0 , then there is an optimal consecutive partition. b. = 1 1 for each i and , in addition, h m is concave in the "a variables" for each fixed value of the "b variables" then a consecutive optimal partition exists. S c {l, ..., n} a set a given partition Of course, when bS P = equals S1 ..., ISI, S } b P ... b S ) (bS we have that for S . In this case, for is called the shape of m m of sets in a partition P we call P We observe that the number of ordered m-Dartitions that are con- O(m!n m secutive is i (e.g., Hwang [19811). When we wish to specify the number an m-partition. for each the cardinality of 1 the partition 1 -l) and the number of semi-consecutive partitions is Thus for a fixed value of m O(m!n2m-2) we can determine an optimal consecutive or semi- consecutive m-partition in polynomial time via complete enumeration. Although complete enumeration is quite impractical even for moderate sizes of m , the above fact contrasts sharply with the NP-completeness of the optimal partition problem for any fixed m partition problem for m is m .) changes in 2 , as the knapsack problem is a special case of the = 2 . (Of course, the total number of ordered m-partitions We note that the sensitivity of optimal consecutive partitions to m can be studied by using the results of Denardo, Huberman and Rothblum [1982]. We formally state our main results concerning the existence of consecutive and semi-consecutive optimal partitions in Section 2, where we also survey a number of places where special cases of our results are used. In Section 3, we show how our results apply to problems that involve optimal groupings for joint - 3- replenishment of inventories. where g (-) Next, in Section 4, we discuss the special case is separable and symmetric, in which case optimal consecutive partitions can be determined in 2 (mn ) steps and optimal semi-consecutive partitions can be determined in 4 O(mn ) steps. of NP-completeness which apply to our results. are given in Section 6. In Section 5, we discuss issues Finally, proofs of our results 111 - 4 - 2. The Main Results We next state our main results concerning optimal partitions that are either consecutive or semi-consecutive. Theorem 1: b 0 or a Theorem 2: Suppose 0 . h The proofp are deferred to Section 6. is concave in its m 2m variables and in addition Then there exists a consecutive optimal partition. Suppose that h m is concave in its 2m variables. Then there exists a semi-consecutive optimal partition. Theorem 3: Suppose b. = 1 for i = 1, ... , n and h (-) the "a variables" for each fixed value of the "b variables." is concave in Then there exists a consecutive optimal partition. Although the partition problems considered in the above Theorems are quite restricted, there have been several realms where special cases of these results have been used. We next give a brief survey of such applications. One classical partitioning problem occurs in Hypothesis Testing in Statistics. Outcomes of a random phenomenon are partitioned into two sets, one of which corresponds to the region where the null hypothesis is accepted and the other to the region where it is rejected. In this case the "a variables" correspond to proba- bility of type 1 error, the "b variables" correspond to probability of type 2 error and the corresponding objective function is linear. applies and an optimal consecutive partition exists. Neyman-Pearson Lemma. DeGroot [1975, p. 374]. It follows that Theorem 1 This fact is the well-known For more details on the applicaton to statical testing see -5- We next give a number of examples of studies where special cases of the results of Theorem 3 were obtained. where [1982]. h In particular, the special case of Theorem 3 is separable and symmetric is discussed in Chakravarty, Orlin and Rothblum They apply the corresponding result to a problem of inventory grouping with joint replenishment (see Section 4 for more complex such problems). Also, Hwang [1981] and Hwang, Sun and Yao [1982] consider restricted classes of functions for which a corresponding optimization problem have consecutive optimizers. Their conditions entail monotonicity and additivity requirements, and some of their results are special cases of Theorem 3. The above papers provide applications of their results to problems in storage and problems in group testing. Finally, Barnes and Hoffman [1982] considered a special instance of the partitioning problem studied in Theorem 3. The motivation for their work arose in connection with the partitioning of eigenvalues of a matrix so as to derive estimates on a graph theoretic partitioning problem. They used the theory of submodular set functions to establish the existence of consecutive optimal partitions. Their results were developed independently of those in Chakravarty, Orlin and Rothblum [1982], and their proof is an interesting application of the theory of submodular set functions. In Section 5, we show that the Barnes-Hoffman partitioning problem is NP-complete. III - 6 - 3. An Application to Inventory Grouping Consider an economic order quantity model involving per unit time and a fixed cost to partition the n items into Ki m items, where the Di , a unit inventory holding cost i-th item has (deterministic) demand rate hi n for placing an order. The problem is subgroups and choose order cycles for the groups out of a given set of allowable (joint) order cycles, so as to minimize the net average cost per unit time. Motivation and more detailed explanation of This model special cases of this model are given in Chakravarty [1982a, 1982b]. is a generalization of the joint replenishment models considered by Goyal [1974], who study the problem in which a unit of order Silver [1975], and Nocturne [1973] X is determined for the group of time X . item is an integer multiple of n items, and the order cycles for each They give heuristic solutions to this problem. A related model was studied by Chakravarty, Orlin and Rothblum [1982]. a. = 21h D i Let holds. i and bi = K where the items are labeled so that (1) As in the ordinary EOQ model (e.g., Wagner [1969, pp. 18-19]), if item has order cycle then its order quantity is T S per unit time of a group c(S,T) = I 2 of items having the same order cycle 1 tD.h ieS -1 K.t- + ..., t = Ta l + T b T . is: (3) iS So, if the items are partitioned into groups t, Di , and the average net cost S1, ..., S having order cycles , respectively, the total average cost per unit time is m C (S1 .Sm, ... tj) = (4) (S j=1 The m-vector of order cycles is assumed to be taken out of a set m-vectors of order cycles. T of allowable This formulation allows one to impose joint restrictions on the order cycles, e.g., the requirement that all order cycles are integer -7 multiples of the smallest one, or individual restrictions, e.g., that each order cycle be an integer in the set {1,7,30}. It follows that for a fixed partition of the items, the minimum average cost per unit time is: S1 , ..., S gm(S, ..., Sm ) c (Sl, inf S, t, ..., t) m- tja = inf teT Evidently, as c (S m l' .. S , t, m 1 (5) b j ..., t) m '( ) , where subsets so as to minimize g is linear in (for fixed and + t j > 0 , the infimum defining g is The problem is to find a partition of the items into finite for each partition. aS S j=l bS gm is given above. t ), we conclude that Now, as c(S,t) c (S1, ..., S , bS and therefore g(S, ..., S ) aS , bs1, ..., a 1 1 m m as the infimum of linear functions, is concave in aS , b, a bS i.e., 1 1 m m the assumptions of Theorem 1 are satisfied. Hence, our results apply and there tl, ... tm ) is linear in exists an optimal partition consisting of consecutive sets. We next consider a modified version of the model described above where the terms and Ta S and T lbS f( -lbs) , where in (3) are replaced, respectively, by expressions e (aS) S T e (.) and f () are real valued functions. This formulation allows one to introduce discounting on the holding cost as well as the setup cost, where the value of the discount depends on the monetary volume. Under the above modification (5) has to be replaced by: m gm(S1, ..., S) = inf Z etj(tja S ) + fS eT j=l i aj 3 It is easily seen that if the function tions of Theorem 1 hold and if Ki eT and f is independent of are concave then the assumptions of Theorem 3 hold. 3 (6) are concave then the assumi and the functions e In either case we con- clude that there exists an optimal grouping of the products where each set is I -8- consecutive. We emphasize that concavity of the functions e and f is reasonable as it represents higher discounting as the volume increases. A special case of the general model allows the decision maker to choose the order cycles of the groups independently out of a given set of the enumeration of the group), i.e., In this case gm T = U x U x ... x U (which is independent for some set U c R is symmetric and separable and the results of (the forthcoming) Section 4 apply. In particular, the problem of finding an optimal grouping can be reduced to a shortest path problem. Finally, we consider the case in which the cost of ordering in period P is concave in the total quantity ordered in that period, where we might have a simultaneous reorder from different groups. If we let P be the least common multiple of the order cycles, then the total cost of ordering is: P-1 I p=O where tjlp e( I tja ) (7) tjIP means that t. is an integral divisor of p . As before, the infimum of concave functions is concave, and thus there is a consecutive optimal partition. -9- 4. The Symmetric Separable Case We next consider the case where some function h : R 2 is symmetric and separable, i.e., for g R m g (S1, S) ... = h(as j=1 J ) .b j In this case the order of the subsets in a consecutive partition does not matter. In particular, without loss of generality, one can assume that in an (optimal) consecutive partition the indices in precede the indices in Si Sj for i < j Chakravarty, Orlin and Rothblum [1982] showed that the problem of determining an optimal consecutive partition reduces to the problem of finding the shortest path between two nodes in a graph where the number of edges in the path is at most m . 0(mn2 ) This problem can be solved in steps using a standard dynamic programming recursion. In the symmetric and separable case, one can determine an optimal semiconsecutive partition as follows. O < i of Then n . j Sij into 1 ij f Let t - ih'(S fi- = fij subsets. iJ) for Let Sij = {i+l, ..., j} for all for each value of with be the optimal value of a semi-consecutive partitioning Finally, let 0 < i h'(S) - h(as,bS) for each he S c N j < n. min h'(Sik U S ) + ft (S) Itis that easy tosee the above recursions can be computed in It is easy to see that t i,j a recursions bove can be computed in 0(n4) 4 (8) steps 0(n ) steps t , and thus the total number of computations is (mn4 ) . - 10 - The NP-Completeness of an Optimal Partition Problem 5. Barnes and Hoffman [19821 consider the following partition problem. of n m 1 di = N, ii for al, ... , integers partition i = 1, ..., m N into non-negative integers m and a m subsets S, ... Sm Given a set dl, ..., dm, where ISi = di i= such that and so as to maximize 2 m I d. j=l 3 1 (9) ( ai ) itS. Barnes and Hoffman demonstrated that there is always an optimal partition consisting of consecutive subsets. in which m Moreover, since they were interested in problems was quite small (e.g., m = 3) , implicit enumeration of all feasible consecutive partitions was quite practical. The Barnes-Hoffman partition problem satisfies the conditions of Theorem 3 (except for the fact that they maximize a convex function rather than minimize a concave function). Thus the optimality of consecutive partitions is a special case of Theorem 3. Although the Barnes-Hoffman problem is polynomially solvable for fixed m the corresponding recognition problem is NP-complete if jointly with n . m , is allowed to vary We next prove the NP-completeness via a transformation from 3-partition. 3-PARTITION INPUT: non-negative integers QUESTION: such that bl, ..., b3k Is it possible to partition such that {1, ..., 3k} b into k k 2 for subsets i = 1, _.., 3k S1 , ..., Sk - 11 - 3k bi = (l/k) I bi itS. i=l for j = 1, ...,k ? (10) We consider the version of 3-Partition in which the subsets are not specified to havw three elements each. Garey and Johnson [1973] showed that the above variant of the 3-Partition problem is NP-complete. Theorem 4. is NP-complete. The recognition version of the Barnes-Hoffman partition problem - 12 - 6. Proofs Proof of Theorem 1. We consider only the case where from analogous arguments. m = 2 . be ignored. some set 0 as the case where The case m = 1 is trivial. We first remark that the case where 0 follows and that for such ai = bi = 0 = S, aN\ S aN\SaN can write the optimal partitioning problem with and m = 2 m of We next consider the Noting that each 2-partition may be written as S c N a Our proof follows by induction on the number subsets in the partitions. case b for some {S, N \ S} bN\S i can for bN b sS we as the following optimization problem: min ScN g2(S, N\S) = min ScN = min ScN Since the set of ordered pairs which will be denoted by h2(as, bs, aN\S' bN\S) aN - aS' bN - b h2 (aS , b, {(as, b) (11) : S c N} S) is finite, its convex hull, C , is a convex polyhedron. It now follows from (11) that amin g 2 (S, N\S) = min ScN ScN > h 2 (as, b, min (x,y)2cC h2 (x, y, a- h (x, y, aN - x, bN - y) Since the function aN - as, b (12) -b x, bN - y) is concave in the variables the latter minimization problem attains a minimum at an extreme point of therefore suffices to show that for each extreme point (x , y ) exists a set and S c N such that a consecutive partition. (x , y ) (as, bs*) of C {S , N \ S (x, y) , C . there is It - 13 - Let (x , y ) be an extreme point of C . well known that there exists a linear function minimum over C f(x, y) = ax + at y (x , y ) . where as the minimizer of observe that as ax The set S + f a and over C Since B f on C is a polyhedron, it is that attains its unique is linear it has a representation The uniqueness of y = ax + By b i < 0} (as, b) = min ScN f aa+bS=min ScN C that (x , y ) - a = 0 then {i :b. > N \ S } if B O0 , and B < 0 . are consecutive. S Let P = b 0 * < S }-t m 3 . S} be an optimal m-partition of max Sj . ij We see this as follows. Observe that __ and For 1' 2i} > i min {M1, then / ) Finally, S and N S = S and m = 2 of N , let d(S) = max S - min S that minimizes We claim that for each d(S) j = 1, ..., m , Suppose that not all subsets are S2 , min S 1 < i < max S1 . 3 B > 0 For each subset for some = if This completes our proof of Theorem 1 when Then we may determine two sets of i .(13) N \ S a > 0 In either case we trivially have that both consecutive. M. s/sll is consecutive. If and then S * { : ai / bi > and N \ S are consecutive. with repsect to all such (optimal) partitions. Sj S 0 U{ : i E S} , min S = min {i : i {S1, ..., S } b i) as the minimizer of assures that both - {i : b i < We next consider the case where max S = max (aa.+ iES * (x , y ) We consider a number of cases separately. S = {i : a i / b i < - / and if In either case (1) assures that both if Next (aS*, bs*) We next show that the fact that are consecutive. 0 . is clearly optimal for the last optimization Hence, we conclude from the uniqueness of over B (x , y ) : S c N} , * problem. a assures that we do not have is the convex hull of = {i : aa i + f C are real numbers. C min (x,y)EC Since P , say j = 1, 2, and S1 let and S2 ,such mj = min S max {ml1, m2 } < i . that and Hence, II - 14 - {ml min {M1 , M2} -max Let {S, S} be an optimal parition of consecutive with respect to (14) O m2 } S1 u S2 S1 u S 2 where into two subsets which are S3, ..., S m = 2 possible by the established result of Theorem 1 when that h m is concave in the variables corresponding to remaining variables are fixed). d(Si) + d(S') remain fixed (as is and the observation S1 1 and S2 2 when the Then < max {M1 , M2 } - min {ml, m2 (15) - 1 and therefore, using (14), we conclude that d(S 1) + d(S2) = M 1 - m 1 + M 2 - (16) 2 = max {M1 , M2 } + min {M1 , M 2} - min {ml, m 2} - max {m1 , m 2} > d(S{) + d(S') + 1 . This contradicts the choice of L S1 , ..., Sm , completing our proof. Proof of Theorem 2. Our proof follows by induction on the number of subsets in the partitions. m = 2 . We next consider the case is trivial. m = 1 The case m The argument used in the proof of Theorem 1 shows that it suffices to prove that for every pair of real numbers a B , the set and S = {i : bi < 0} ai + N , or equivalently, that perty that {S , N \ S 3 is a semi-consecutive partition of N \ S either S or ately. If a > O then is a consecutive set. S = ai / bi < - We consider a number of cases separ- {i = 1, ..., r : a i / bi > -B / {i = r+l, ..., n : ai / b i < -B / a} and if has the pro- a < 0 / a} u {i = r+l, ..., n : ai / bi > -B a u then S = {i - 1, ..., r / a . In the former case - 15 - if $ > 0 and = {i : b S 8 then and = {i : bi < 0 S ..., r} = {1l, is not consecutive if N \ S {S , N \ S} and therefore the partition i, a i > 0 = b i) is semi- S In eitner case we have that <0 . if > 0) 0 X (though in the latter case is consecutive and for some a = 0 Finally, if consecutive. {S , N \ S } Hence, in either case the partition is consecutive. N \ S is consecutive and in the latter case (1) assures that S (1) assures that r > 1 is semi- consecutive. Next assume that the conclusions of Theorem 2 hold whenever the number of sets in the partitioning problem is less than partitioning problems where the number of sets is m ( m . 3), and consider We prove the existence of a semi-consecutive optimal partition for the corresponding partitioning problems by induction on the number of elements in the partitioned set The cases where the number of elements in forward. less than is N Next assume that each partitioning problem of a set consisting of n sets, where the assumptions of Theorem 2 are m elements into We next consider such = S n {1, ..., r S S c N, let For each subset where the number of elements in N partitioning problems of a set r S, Si,... ..., S m S 1 U ... U S+ 1 m N+ all consecutive. S+ ... , S+ S }, N fixed and repartition such that the sets in the new (optimal) partition of N are The above arguments establish the existence of an optimal r = 0 for which r = n or S1, ..., S , SS, .. 1m then the partition secutive and therefore semi-consecutive. of this proof that S+ = S n {r+l, ..., n}, Similarly, by applying an appropriate modification of {Sl' ..., S I m We remark that if n fixed and, by applying Theorem 1, repartition Theorem 1, one can hold the (new) sets = N is such that the sets in the new (optimal) partition of are all consecutive. S1 u ... u S and N Given any optimal partition, say {S1,... is defined through (1). one can hold partition are straight 1, 2 or 3 satisfied, has a semi-consecutive optimal partition. where N 1 ____P__IIII·_l__l__1__1__1111111____1 r < n. S. m are all consecutive. S {S , ..., 1 I m is con- Henceforth, we assume through the end II - 16 - We continue our proof by assuming that the partitioning problem of m sets has an optimal partition > 1 IS ..., S } for which the sets or > 1 combine all the elements in and a "b" value have that if bT T T Let T = S Si . >1 One can into a single element and attach an "a" value to this combined element. ai / bbi a ..., Sm We consider only the case where as the alternative case follows from similar arguments. aT S, into ~~mm 1' with either {S1, N for each i E T Evidently, as then ac T N aT / b T < B we It follows that the corresponding partitioning problem where the elements in T are combined into a single element satisfies (1) and the other assumptions of Theorem 2 . As the number of elements in the partitioned set of this modified problem is less than n , we conclude from the induction assumption that a semi-consecutive optimal partition exists. It immediately follows that the original problem too has a semi-consecutive optimal solution. It remains to consider the case where any optimal partition {S1 , ..., S } m for which Si, ... each of the sets {S1 , ..., S } and S are all consecutive has the property that S .. Si, ... ... S, S , has at most a single element. Let be such an optimal partition (whose existence follows from our earlier arguments). r S , S Let S and S be the sets in r+l , respectively (recall that 1 r < n ). {S1 , ..., S } which contain We establish existence of a semi-consecutive optimal partition by considering two cases. First assume that either S or S q P is consecutive. We consider only the former case as the latter case follows from identical arguments. By using the induction assumption concerning partitioning problems of a set into one can fix S and repartition P u {St : t = 1, ..., m, t t resulting partition of this set, say {S : t = 1, ..., m, t p m - 1 sets, such that the p} will be semi- consecutive and will not reduce the value of the objective function. It easily - 17 - follows that, with partition of Sp = Sp 1,{l' ..., S} is a semi-consecutive optimal N S We next consider the case where neither ISpI 1 P that as and ISp p 1 1 IS+ P , r+l < j and for some 1 n , and p argument shows that for some {St : t = 1, ..., m, t of Theorem 2 for the case into sets Sp and S Now, if either of the sets element, one can apply m = 2 , where {i, r, r+l, j respect to q} = {i |s p S+ p A similar q . fixed and apply the (established) conclusions to repartition {Sp, } S P u S = {i, r, r+l, j} q is a semi-consecutive partition with and where the value of the objective is not reduced. S P ,S q , p , consists of more than a single q a combination of the arguments used earlier to establish > 1 , one can repartition number of nonempty sets among partition is consecutive. at least two elements. r+l We can next hold . the existence of a semi-consecutive optimal partition. example p In particular, i < r , S 1 p, t P = {j} . S Also observe is not consecutive, we have that S As is consecutive. S+ = {r+l} . q and S = {r} P we have that JS+J q S q nor p As into = k k-l sets where k is the where each of the sets in the new ..., S } m {S1 , IN- N- Specifically, if for we have that one of these sets contains By combining the elements in that set and applying the induction assumption concerning partitioning problems of sets consisting of less than n elements into optimal partition. m sets, we conclude the existence of a semi-consecutive We continue by assuming that the sets each S, contains at most a single element, in which case each contains precisely one element. It follows from the semi-consecutiveness of the partition {S , S } p q with respect to {i, r, r+l, j} that either In either case, the set containing r S P = {r, r+l}. or (which also contains S q = {r, rl} . r+l ) is consecutive and our earlier arguments assure the existence of a semi-consecutive optimal O partition. 7"·*"81s-----·-------san-araar-·ll ____1 - 18 - Proof of Theorem 3: Chakravarty, Orlin and Rothblum established a special case of Theorem 3 g' ( where ) Their arguments apply to the general is symmetric and separable. D case considered in Theorem 3 and are omitted. Proof of Theorem 4: Let bl, ..., b3k be the input for a given 3-Partition problem. We will show below how to transform the 3-Partition problem into the Barnes-Hoffman partition problem. The transformation is based on the following elementary lemma. Let Lemma. al, ..., a n and i=l ai . be the data for the m Then the maximum cost partition has Barnes-Hoffman partition problem. value at most dl, ..., d Moreover, it is possible to achieve this bound if and only if it is possible to determine a feasible partition S1, ... S such that a. = a 1 m same subset ai) 2 "( IS and equality holds if and only if a iS. j for all t in the By the Cauchy-Schwarz inequality iS. j . i,t S' Proof of Lemma. a. = a for every two indices t i,t S j . 12 m I I-1 . j=l Thus: n ( I isS. i)C jsj2I 2 i=l and equality holds if and only if a.2 ai = a t for every two indices i,t in the same subset. To complete the proof of the theorem, let b = k -1 subset sum for each subset of the 3-partition). and let di = bii for i = 1, ... , i~~.. ai+jS i+j b = j for 1 < i 3k . b and Let Finally, let O0 j m-l . 3k b ) i=l bi i=l (b bi , let is the m 3k - Thus there are exactly b indices 19 i - for which ai = j By the previous lemma, we can achieve a value of n.l a determine a feasible partition in the same subset. {PO, ..., Pk 1 } 0 k-i bj a 3-partition of "a" bl, ..., b3 k value is partition ..., b1 , ..., b 3 k . t , such that m then let where {S1, ..., S } {b, ... 1' at = i such that , S S for bk 3kt S *D _ = at t for all Then let such that Pi Then ... {P, i,t consists of 1 Pk is {P0 , ... , Pk-1 } consist of elements whose corresponding b. E P . 0 t b. is a 3-partition We have thus seen that there is a feasible such that the "a" value for each Si is constant if Thus the Barnes-Hoffman partition Since the recognition variant is clearly in the class the latter problem is NP-complete. -- a Conversely, if and only if there is a 3-partition for the b's . problem is NP-hard. j = 0, ..., m-1 if only if we can Suppose we can achieve this latter result. be a partition of those elements of S1, for any NP , I - 20 - References Barnes, E.R., and A.J. Hoffman (1982), "Partitioning, spectra, and linear programming". To appear in the Proceedings of the Silver Jubilee Conference on Combinatorics, University of Waterloo, Ontario, Canada. Chakravarty, A.K. (1982a), "Multi-item inventory aggregation into groups", Journal of Operational Research Scciety (U.K.) 32, 19-26. Chakravarty, A.K. (1982b), "Consecutiveness rule for inventory grouping with integer multiple constrained group-review periods", unpublished manuscript. Chakravarty, A.K., J.B. Orlin and U.G. Rothblum (1982), "A partitioning problem with additive objective with an application to optimal inventory groupings for joint replenishment", Operations. Research 30, pp. 1018-1022. DeGroot, M.H. (1975), Probability and Statistics, Addison-Wesley. Denardo, E.V., G. Huberman and U.G. Rothblum (1982), "Optimal locations on a line are interleaved", Operations Research 30, pp. 745-759. Garey, M.R. and D.S. Johnson (1978), "Strong NP-completeness results: examples, and implication," S, Assoc. Comput. Mach. 25, 499-508. motivation, Goyal, S.K. (1974), "Determination of Optimum packing frequence of items jointly replenished", Management Science 21, pp. 436-443. Hwang, F.K. (1981), "Optimal partitions," Journal of ODtimization Theory and Applications 34, pp. 1-10. Hwang, F.K., J. Sun and Y. Yao (1982), "Optimal set partitioning," an unpublished manuscript. Nocturne, D. (1973), "Economic order frequency for several items jointly replenished", Management Science 19, 1093. Silver, E.A. (1976), "A simple method of determining order quantities in joint replenishments under deterministic demands", Management Science 22, 1351. Wagner, H.M. (1969), Principles of Operations Research, Prentice-Hall.