March 2002, December Introduction. We investigate the facial structure of the convex hull of the mixed integer knapsack set

Transcription

1 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON ALPER ATAMTÜRK Abstract. We study the mixed integer knapsack polyhedron, that is, the convex hull of the mixed integer set defined by an arbitrary linear inequality and the bounds on the variables. We describe facet defining inequalities of this polyhedron that can be obtained through sequential lifting of inequalities containing a single integer variable. These inequalities strengthen and/or generalize known inequalities for several special cases. We report computational results on using the inequalities as cutting planes for mixed integer programming. March 2002, December Introduction We investigate the facial structure of the convex hull of the mixed integer knapsack set K = { (x, w) Z I + IR C + : ax + gw b, x u, w v }, where I is the index set of integer variables, C is the index set of continuous variables. The mixed integer knapsack set K is the set of points in Z I + IR C + that satisfy an arbitrary linear inequality and the upper bounds on the variables. We assume that the data is rational, with the exception that u and v may have entries equal to infinite, so that the variables are not necessarily bounded. We impose no sign restriction on a, g, or b. Since each constraint of a mixed integer programming (MIP) formulation defines a mixed integer knapsack set, strong valid inequalities for K can be used as cutting planes for MIP. There are many important polyhedral studies on special cases of the mixed integer knapsack set K. The most studied is probably the 0 1 knapsack set (u = 1 and C = ) for which seminal works [5, 7, 19, 33] date back to 70 s; see also [16, 28, 32, 37]. Crowder, Johnson, and Padberg [13] demonstrate the effectiveness of cutting planes from individual 0 1 knapsack constraints in solving 0 1 programming problems. Date: March 2002, December Alper Atamtürk: Department of Industrial Engineering and Operations Research, University of California at Berkeley, Berkeley, CA atamturk@ieor.berkeley.edu. Supported, in part, by NSF grants DMII and DMII

2 2 ALPER ATAMTÜRK Carrying this line of research to mixed 0 1 programming, Marchand and Wolsey [22] give strong inequalities for the 0 1 knapsack set with a single continuous variable. Recently, Richard et al. [31] study the mixed 0 1 knapsack set with bounded continuous variables. Most of the research on polyhedral analysis of structured sets is done on (mixed) 0 1 problems. Polyhedral studies on problems with integer variables, even for the pure integer case, are rare; see [2, 8, 10, 20, 21, 25] for certain network design problems. Pochet and Weismantel [29] and Pochet and Wolsey [30] study the convex hull of the pure integer knapsack set with divisible coefficients. Ceria et al. [11] give an extension of the 0 1 knapsack cover inequalities for integer knapsacks. Gomory mixed integer cuts [14] or, the equivalent, mixed integer rounding (MIR) cuts [27] are well known valid inequalities for K, and, consequently, for mixed integer programming. They are incorporated in leading optimization software systems after their computational effectiveness has been evidenced in a branch and cut framework [6, 23]. These general algebraic inequalities depend on the representation of the constraints rather than the geometry of the feasible set since multiplying the coefficients of a constraint by a constant may lead to a different MIR inequality [12]. Furthermore, Gomory mixed integer cuts or mixed integer rounding cuts are not only valid for an MIP problem, but also for its group relaxation [15], obtained by dropping the bounds of the basic variables. This suggests that stronger inequalities for K may be identified by studying directly K, rather than its group relaxation, as illustrated in Example 1 in Section 2. Our goal here is to derive strong inequalities based on the geometric structure of the convex hull of K (conv(k)). One difficulty with studying (mixed) integer polyhedra is that simple extensions of combinatorial, disjunctive and/or rounding arguments, that give strong inequalities for (mixed) 0 1 programming, generally do not lead to inequalities that define high dimensional faces for integer programming. For instance, even though for 0 1 knapsacks a minimal cover inequality is facet defining on the space of the variables defining the cover, its extension to integer knapsacks may not define a high dimensional face. An intuitive reason for this is that integer points lie deep in the linear programming (LP) relaxation as opposed to on the surface as in the case for (mixed) 0 1 problems. Recall that all integer points of a 0 1 programming problem are among the extreme points of its LP relaxation. In this study we make use of superadditive functions for defining strong inequalities for the mixed integer knapsack set. It is well known that the convex hull of the feasible region of any MIP problem can be described with inequalities defined by superadditive functions and convex functions [4]. However, from a practical perspective, the challenge is to identify the shape of specific functions that can be used effectively as cutting planes in branch and cut computations. See Gu et. al [17] and Marchand and Wolsey [22] for two successful works in this direction for mixed 0 1 programming. In Section 2 we review recent developments that motivated this study and compute the lifting function of a simple MIR inequality with two variables. Section 3 contains the main

3 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 3 results of the paper. Here we describe facet defining inequalities for conv(k) by building on the results in Section 2. In Section 4 we highlight the connections between the new inequalities and others defined earlier in the literature for certain special cases. In particular, if the coefficients of all integer variables (a i i I) have the same sign, either positive or negative, then the new inequalities dominate mixed integer rounding inequalities [14, 27]. For the special case of 0 1 knapsack with a single continuous variable, the inequalities reduce to the ones given in [22]. For the knapsack set with bounded integer variables, they generalize and strengthen integer cover inequalities [11] and weight inequalities [24]. In Section 5 we present a summary of computational experiments for testing the effectiveness of the new inequalities as cutting planes. The results indicate that the inequalities may be useful in branch and cut algorithms for MIP. In order to simplify the notation, we assume (wlog) that g i { 1, 1} for all i C since continuous variables can be rescaled. We define C + = {i C : g i = 1}, C = C \ C +, I B = {i I : u i < }, and C B = {i C : v i < }. We assume (wlog) that g i = 1 for all i C B, after complementing variables if necessary; thus C B C +. Throughout we assume that conv(k) is full dimensional. We let a + denote max{a, 0} for a IR. 2. Preliminaries We start with an example to illustrate that inequalities stronger than Gomory mixed integer or mixed rounding inequalities can be identified by studying conv(k) directly, rather through its group relaxation. Example 1. Suppose the mixed integer knapsack set is given as K = { (x, w) Z 2 IR : x 1 + ax 2 w 1 + ε, x 1 0, x 2 0, w 0 }, where 0 < ε < 1 and a > 2. Although it is not necessary, in order to keep the example simple, we assume that a is integer. The Gomory mixed integer inequality or mixed integer rounding (MIR) inequality [6, 23] (1) x 1 + ax 2 w 1 ε 1 cuts off the fractional vertex (x 1, x 2, w) = (1 + ε, 0, 0) of the LP relaxation of K. Inequality (1) defines a facet of the convex hull of the group relaxation [15] of K conv(k G) = conv{x Z 2, w IR : x 1 + ax 2 w 1 + ε, x 2 0, w 0}, obtained by dropping the nonnegativity constraint x 1 0; however, it is not facet defining for conv(k ). On the other hand, inequality (2) x 1 + a 2ε 1 ε x 2 w 1 ε 1

4 4 ALPER ATAMTÜRK is stronger than (1) since a < (a 2ε)/(1 ε) for a > 2 and 0 < ε < 1. Indeed, inequality (2) defines a facet of conv(k ). Notice that the difference between the coefficients of x 2 in inequalities (1) and (2) becomes arbitrarily large as ε approaches to one. Inequality (2) is the special form of (15) for K. This study is motivated by the following recent developments: the knowledge of a complete linear description of the convex hull of the restriction of K with a single integer variable, the existence of a polynomial time separation algorithm for this restriction, and the possibility of sequence independent lifting for general mixed integer programming. In the rest of Section 2 we review these developments and compute the lifting function of a simple MIR inequality with two variables as building blocks for studying conv(k) The convex hull of the restriction with a single integer variable. We start by describing a property of the facets of the convex hull of the mixed integer knapsack set K = (x, w) ZI + IR C + : a i x i + w i w i b, x u, w v i I i C + i C regarding the unbounded continuous variables. We call the nonnegativity constraints on the variables, the knapsack constraint, and the upper bound constraints as the trivial inequalities of conv(k). The following property is useful. Proposition 1. Any non trivial facet defining inequality πx+µw π o of conv(k) satisfies (1) µ i = 0 for all i C + \ C B, (2) µ i = α for all i C, where α is a negative scalar. Proof. Let πx + µw π o be a non trivial inequality defining facet F of conv(k). For i C + \ C B, since πx + µw π o differs from w i 0, there exists a point (x, w ) on F with w i > 0. Since reducing w i by a small ɛ > 0 gives a feasible point, validity of the inequality implies µ i 0. Also since the inequality differs from ax + gw b, there exists ( x, w) on F with a x + g w < b. Since by increasing w i by small ɛ > 0, we maintain feasibility, validity of πx + µw π o implies µ i 0. On the other hand for i C, since w i can be increased without violating feasibility, validity of πx + µw π o implies that µ i 0. However if µ i = 0, then the feasible point (ˆx, ŵ) with πˆx > π 0, ŵ i large enough (as C B C + ), and ŵ j = 0 for j C \ {i} is violated by πx+µw π o. Notice that point (ˆx, ŵ) exists, since otherwise πx+µw π o is dominated by bound constraints, hence cannot define a facet. Thus we have µ i < 0. Since πx+µy π o is different from w i 0, there exists a point ( x, w) on F with w i > 0. The point obtained from ( x, w) by decreasing w i by small ɛ > 0 and increasing w j for j C \ {k} by ɛ is feasible. Since ( x, ỹ) is on F, we have µ i µ j. Finally µ i µ j follows from symmetry.

5 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 5 The first part of Proposition 1 is stated in [22] without proof. Now for some l I, consider the restriction of K obtained by fixing all integer variables but x l to zero: T = (x l, w) Z + IR C + : a l x l + w i w i b, x l u l, w i v i i C B. i C + i C It follows from Proposition 1 and [21] that conv(t ) is given by the inequalities in the description of T and (3) (a l r)x + w i w i b ηr S C B if a l > 0, i S i C (4) rx + w i w i v S + ηr S C B if a l < 0, i S i C where η = (b v S )/a l, r = b v S (b v S )/a l a l, and v S = i S v i. Moreover, an exact linear time algorithm is given for separating inequalities (3) (4) in [3]. This suggests that inequalities for K lifted from (3) (4) may be potentially useful as cutting planes for K Sequence independent lifting. In this section we review a lifting technique for (general) integer variables. Consider a mixed integer set P = {x Z I, y IR C : Ax +Gy b}, where A, G, and b are rational matrices with m rows. Let (L, U, R) be a partition of I and P L,U,R (d) = {x R Z R, y IR C : A R x R + Gy d} be a nonempty restriction of P, obtained by fixing x i = l i for i L and x i = u i for i U, where l i > and u i < + are the minimum and maximum values x i attains in P, respectively. Let (5) π R x R + µy π o be a valid inequality for P L,U,R (d) and the lifting function Φ : IR m IR { } of π R x R +µy π o be defined as Φ(a) = π o max {π R x R + µy : (x R, y) P L,U,R (d a)}. We let Φ(a) = if P L,U,R (d a) =. Since (5) is valid for P L,U,R (d a), the maximization problem above is bounded and consequently Φ(a) >. Definition 1. ϕ : IR m IR is superadditive on D IR m if ϕ(a) + ϕ(b) ϕ(a + b) for all a, b D such that a + b D. A valid inequality for P can be obtained from (5) by sequential lifting, i.e., introducing the fixed variables x i i L U to the inequality one at a time in some sequence [34]. One difficulty with this approach is that it requires the solution of a nonlinear (fractional) mixed integer problem for each fixed variable, as opposed to a linear mixed integer problem as in 0 1 programming.

6 6 ALPER ATAMTÜRK For monotone (A 0) 0 1 programming and monotone mixed 0 1 programming, Wolsey [35] and Gu et al. [18] show that superadditive lifting functions lead to sequence independent lifting of valid inequalities, which reduces the computational burden of lifting significantly. The theorem below states that this property holds for general mixed integer programming as well if lower dimensional restrictions are obtained by fixing integer variables to a bound rather than to some intermediate value. Theorem 2. [1] Let Φ be defined as before and let φ : IR m IR be a superadditive function such that φ Φ. Then inequality (6) π R x R + i L φ(a i )(x i l i ) + i U φ( A i )(u i x i ) + µy π o is valid for P. In addition, if φ(a i ) = Φ(A i ) for all i L, φ( A i ) = Φ( A i ) for all i U, and inequality (5) defines a k dimensional face of conv(p L,U,R (d)), then inequality (6) defines an at least k + L + U dimensional face of conv(p ). We note that if the lifting function is superadditive, nonlinearity of the lifting problems is resolved easily and lifting (5) in any sequence leads to a unique inequality for P. We use Theorem 2 for deriving strong inequalities for K in Section Lifting function of a simple MIR inequality. Here we compute the lifting function of a simple MIR inequality for a two variable mixed integer restriction of K as a building block for studying conv(k). Let S = {x Z, y IR + : cx y d, l x u} with c, d Q and l, u Z {, + }, l < u. Let η = d/c and r = d d/c c. Observe that the LP relaxation of S has a fractional vertex (d/c, 0) if and only if d/c Z (or equivalently r 0) and l < d/c < u. If c > 0, the fractional vertex (d/c, 0) is cut off by the simple mixed integer rounding (SMIR) inequality [27, 36] (7) (c r)x y d ηr. On the other hand, if c < 0, it is cut off by inequality (8) rx y ηr. As illustrated in Figure 1 inequalities (7) and (8) are sufficient to describe conv(s) when added to the original inequalities of S in either case. Lifting these inequalities amounts to maximizing a linear function over S as a function of d. Maximizing an arbitrary linear function over S is easy. Without loss of generality, we assume that the objective coefficient of y is negative and by scaling is 1, since otherwise the problem is unbounded, and write the optimization problem as (9) ζ( d) = max{ex y : cx y d, l x u, x Z, y IR + }.

7 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 7 y y c r r l η 1 η u x l η 1 η u x (c > 0) d/c d/c (c < 0) Figure 1. SMIR inequality and the convex hull of S. If e 0 or e > c, problem (9) has a trivial optimal solution with x = l or with x = u. Otherwise, an optimal solution, which is an extreme point of conv(s), can be found graphically in Figure 1 as stated in the following simple lemma. Lemma 3. If 0 < e c, then problem (9) has an optimal solution (x, y) with objective value ζ( d) that can be expressed as (10) (x, y, ζ( d)) = where η = d/c and r = d d/c c. (u, 0, eu) if d/c u, ( η, c r, (e c) η + d) if c r < e c & l < d/c < u, ( η 1, 0, e( η 1)) if 0 < e c r & l < d/c < u, (l, lc d, (e c)l + d) if d/c l, Now we can compute the lifting function Φ of the SMIR inequality (7) over S. Lifting function of (8) can be computed similarly. Let Φ(a) = d ηr max{(c r)x y : cx y d a, l x u, x Z, y IR + }. Theorem 4. The lifting function Φ of inequality (7) can be expressed as (η u 1)(c r) if a < d uc, k(c r) if kc a < kc + r, (11) Φ(a) = a (k + 1)r if kc + r a < (k + 1)c, a (η l)r if a d lc. k Z

8 8 ALPER ATAMTÜRK Proof. The result follows from setting d in (10) equal to d a and evaluating the objective function. (1) If a d uc, or equivalently (d a)/c u, then Φ(a) = d ηr (c r)u = (η u 1)(c r). (2) Let r = d a (d a)/c c and η = (d a)/c. If kc a < kc + r, or equivalently r r, then Φ(a) = d ηr (c r)( η 1). Using η = η k in this case, we get Φ(a) = k(c r). (3) If kc + r a < (k + 1)c, or equivalently r < r, then Φ(a) = d ηr ( r η + d a). Using η = η (k + 1) in this case, we obtain Φ(a) = a (k + 1)r. (4) If a d lc, or equivalently (d a)/c l, then Φ(a) = d ηr ( lr + d a) = a (η l)r. Φ φ 3(c r) 3c (η u 1)c d uc (η l)(c r) c (c r) r c d lc 3c a (c r) (η l)c (u η + 1)(c r) 3(c r) Figure 2. Lifting function Φ and its superadditive approximation φ (η = u 1 = l + 2). A particular realization of Φ is depicted in Figure 2. Φ is superadditive on IR + and on IR separately. However, it is superadditive on IR if and only if l = and u = +. The function Φ depicted in Figure 2 is not superadditive on IR, as Φ( c) + Φ(3c) > Φ(2c). Observe from (11) that the function values on intervals a < (η u 1)c and a > (η l)c are due to the finite upper bound and lower bound on x. If we let l = and u = +, Φ equals its superadditive lower bound (12) φ(a) = { k(c r) if kc a < kc + r, a (k + 1)r if kc + r a < (k + 1)c. k Z

9 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 9 It is shown in [1] that lifting (7) with the superadditive approximation φ gives exactly the MIR inequality [26] i I ( a i /c + (f i f) + 1 f where f i = a i /c a i /c and f = d/c d/c. )x i 3. Facets of conv(k) y c(1 f) d/c, In this section we describe valid inequalities for K defined by the exact lifting function Φ, rather than its superadditive approximation φ. Let l I and U I B \ {l}. Defining z i = u i x i for i U and z i = x i for i I \ U, we rewrite K as K = (z, w) ZI + IR C + : ā i z i + w i w i b, z u, w v, i I i C + i C where ā i = a i for i U, ā i = a i for i I \ U, and b = b i U a iu i. Now by fixing all integer variables except z l to zero, we obtain the following restriction of K K U,l = (z l, w) Z + IR C + : a l z l + w i w i b, z l u l, w v. i C + i C If C =, we update u l as min{u l, b/a l } and require that u l 1, so that conv(k U,l ) is full dimensional. The facets of this restriction have been presented in Section 2.1. Here we extend them to facets of conv(k) using Theorem 2. Since the families of inequalities for K U,l depend on the sign of a l, we consider these cases separately in Sections 3.1 and Case 1: a l > 0. In order to obtain facets of conv(k) we lift the facet defining inequalities of conv(k U,l ) (13) (a l r)z l + w i w i b ηr, i S i C where η = ( b v S )/a l and r = b v S ( b v S )/a l a l for S C B. Letting y = i C w i + i S (v i w i ), we write (13) as (14) (a l r)z l y d ηr, where d = b v S. Since this aggregation of variables has no impact on the objective value of the lifting problems, the lifting function of (13) over K U,l is equivalent to the lifting function of (14) over K U,l = {z l Z +, y IR + : a l z l y d, z l u l }. Hence, lifting inequality (13) reduces to lifting SMIR inequality (14).

10 10 ALPER ATAMTÜRK Inequality class I. Recall that the lifting function Φ of the SMIR inequality is not superadditive on IR. However, it is superadditive on IR + and on IR separately, which allows us to use Theorem 2 in two phases. Let I + = {i I \ {l} : ā i > 0} and I = {i I \ {l} : ā i < 0}. In order to simplify the notation, we define li + = {l} I + and li = {l} I. Since Φ is superadditive on IR +, by Theorem 2, inequality (14) can be lifted to (15) Φ(ā i )z i y d ηr i li + since Φ(ā l ) = a l r. In order to lift (15) with x i i I, we compute (16) Ω(a) = d ηr max Φ(ā i )z i y : (z, y) K I,lI +(a), i li + where K I,lI +(a) = (z, y) ZlI+ + IR + : ā i z i y d a, z i u i i li + i li + for a IR. Lemmas 5 and 6 below are the central results on the structure of optimal solutions to problem (16) that lead to an explicit description of Ω in a special case, a superadditive lower bound on Ω, and the valid inequalities described in Theorem 9. Let I ++ = {i I + : ā i d} = {1, 2,..., n } be indexed in nonincreasing order of ā i, ties broken arbitrarily, and let n = min{i I ++ : u i = } (if I ++ =, then let n = 0; if u i < for all i I ++, then let n = n ). Also if I ++, let J ++ = {i I ++ : ā i ηa l } = {1, 2,..., s }, s = min{i J ++ : u i = } (if u i < for all i J ++, then let s = s ). Since J ++ I ++, we have s n and if u i = for some i J ++, then n = s. Lemma 5. For a 0 the maximization problem (16) has an optimal solution (z, y) such that (i) z i = u i for all i {1, 2,..., j 1} and z i = 0 for all i {j + 1, j + 2,..., n } for some j {1, 2,..., n }, (ii) z i = 0 for all i I + with 0 ā i r, and (iii) z i = 0 for all i I + with r < ā i < ηa l if the constraint x l u l is removed from problem (16). Proof. (i) For i I ++ we have Φ(ā i ) = ā i ηr. Let ā h > ā i for h, i I ++ and suppose z h < u h and z i > 0. Decreasing z i by one, increasing z h by one and y by ā h ā i, we obtain a feasible solution with the same objective value. (ii) Follows from Φ(ā i ) = 0 for i I + if 0 ā i r. (iii) Suppose z i = q > 0 for some i I + such that r < ā i < ηa l. If ka l ā i < ka l + r, then Φ(ā i ) = (a l r)k. Since x l u l is removed, decreasing z i to 0 and increasing z l by kq, we obtain a feasible solution with the same objective value. Otherwise

11 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 11 if ka l + r ā i < (k + 1)a l, then Φ(ā i ) = ā i (k + 1)r. Similarly, decreasing z i to 0 and increasing z l by (k + 1)q and y by ((k + 1)a l ā i )q we obtain a feasible solution with the same objective value. Lemma 6. If I + = I ++, then for a 2 < a 1 0 the maximization problem (16) has optimal solutions (z(a 1 ), y(a 1 )) and (z(a 2 ), y(a 2 )) that satisfy z i (a 1 ) z i (a 2 ) for all i I ++. Proof. Let (z(a 1 ), y(a 1 )) and (z(a 2 ), y(a 2 )) be two solutions satisfying Lemma 5 (i) for a 1 > a 2. Suppose z k (a 1 ) > z k (a 2 ) for some k {1, 2,..., n }. Then, since z k (a 1 ) > 0 and z k (a 2 ) < u k, Lemma 5 (i) implies that z i (a 1 ) = u i for all i {1, 2,..., k 1} and z i (a 2 ) = 0 for all i {k + 1, k + 2,..., n }. Thus z i (a 1 ) z i (a 2 ) holds for all i I ++. Next we show that the knapsack constraint i I ++ ā iz i (a 2 ) + a l z l (a 2 ) y(a 2 ) d a 2 has a slack of at most r. Suppose the slack is r + ɛ with ɛ > 0. If z l (a 2 ) < u l, then by increasing z l (a 2 ) by one and y(a 2 ) by (a l r ɛ) +, we obtain a solution with objective value min{a l r, ɛ} larger than for (z(a 2 ), y(a 2 )). On the other hand, if z l (a 2 ) = u l, then by decreasing z l (a 2 ) by η 1, and increasing z k (a 2 ) by one and y(a 2 ) by (a k d ɛ) +, the objective value is increased by min{a k d, ɛ}. Both cases either contradict the optimality of (z(a 2 ), y(a 2 )) or give an alternative optimal solution in which z k (a 2 ) is increased. Therefore, we may assume that the slack of the knapsack constraint for (z(a 2 ), y(a 2 )) is at most r. Then, since a 1 > a 2 as well, feasibility of (z(a 1 ), y(a 1 )) requires that (17) y(a 1 ) > y(a 2 ) + ā i (z i (a 1 ) z i (a 2 )) + a l z l (a 1 ) a l z l (a 2 ) r i I ++ (18) ā k + a l z l (a 1 ) a l z l (a 2 ) r. The second inequality follows from y(a 2 ) 0, z i (a 1 ) z i (a 2 ) for i I ++, and z k (a 1 ) > z k (a 2 ). Now let δ k = ā k ηa l. We consider two cases depending on the sign of δ k. In each case we either obtain a contradiction or change the value of z k (a 1 ) or z k (a 2 ) toward satisfying z k (a 1 ) z k (a 2 ). First suppose that δ k 0. If z l (a 2 ) η, then the solution obtained from (z(a 2 ), y(a 2 )) by increasing z k (a 2 ) by one and y(a 2 ) by δ k, and decreasing z l (a 2 ) by η is another optimal solution in which z k (a 2 ) is increased. Therefore, we may assume that z l (a 2 ) η 1. Then from (18) we have y(a 1 ) > ā k + a l z l (a 1 ) a l (η 1) r = ā k d + a l z l (a 1 ). Thus y(a 1 ) = ā k d + a l z l (a 1 ) + ɛ, where ɛ > 0. If z l (a 1 ) 1, the solution obtained from (z(a 1 ), y(a 1 )) by decreasing z l (a 1 ) by one and y(a 1 ) by a l has an objective value r larger than for (z(a 1 ), y(a 1 )). Otherwise z l (a 1 ) = 0, and the solution obtained from (z(a 1 ), y(a 1 )) by increasing z l (a 1 ) by η 1, decreasing z k (a 1 ) by one and y(a 1 ) by ā k d + min{r, ɛ} has an objective value min{r, ɛ} larger than for (z(a 1 ), y(a 1 )). Both cases contradict the optimality of (z(a 1 ), y(a 1 )).

12 12 ALPER ATAMTÜRK Now suppose that δ k < 0. If z l (a 1 ) u l η, then the solution obtained from (z(a 1 ), y(a 1 )) by decreasing z k (a 1 ) by one and increasing z l (a 1 ) by η and y(a 1 ) by δ k is an alternative optimal solution in which z k (a 1 ) is decreased. Therefore, we may assume that z l (a 1 ) u l η + 1. Let κ = z l (a 1 ) (u l η + 1). From (18) we have y(a 1 ) > ā k + a l (u l η + 1) + a l κ a l z l (a 2 ) r = ā k a l (η 1) r + a l (u l z l (a 2 )) + a l κ ā k d + a l κ. So y(a 1 ) = ā k d + a l κ + ɛ, where ɛ > 0. If κ 1, then the solution obtained from (z(a 1 ), y(a 1 )) by decreasing z l (a 1 ) by one and y(a 1 ) by a l has an objective value r larger than for (z(a 1 ), y(a 1 )). Otherwise, κ = 0 or z l (a 1 ) = u l (η 1), and the solution obtained from (z(a 1 ), y(a 1 )) by increasing z l (a 1 ) by η 1, decreasing z k (a 1 ) by one and y(a 1 ) by ā k d + min{r, ɛ} has an objective value min{r, ɛ} larger than for (z(a 1 ), y(a 1 )). Both cases contradict the optimality of (z(a 1 ), y(a 1 )). Theorem 7. If ā i r or ā i d for all i I +, then u ih ηr + a if m ih δ i a m ih, (u ih η + k)r + a if m ih δ i (k + 1)a l + r a m ih δ i ka l, u ih ηr + m ih δ i (k + 1)(a l r) if m ih δ i (k + 1)a l a m ih δ i (k + 1)a l +r, u sus ηr + pr + a if m J ++ (p + 1)a l + r a m J ++ pa l, Ω(a)= m J ++ + u sus ηr (p + 1)(a l r) if m J ++ (p + 1)a l a m J ++ (p + 1)a l + r, u iui ηr + (η u l 1)r + a if m ih δ i (a l r) a m ih, u iui ηr+(η u l )r + m ih δ i (k+1)a l if m ih δ i (k+1)a l a m ih δ i (k+1)a l +r, u iui ηr + (η u l 1)r + (a l r) + a if m ih δ i (k + 1)a l + r a m ih δ i ka l, m I ++ + u nun ηr + (u l η + 1)r if a m I ++, where δ i = ā i ηa l for i I ++, u ih = i 1 k=1 u k + h, m ih = m (i 1)ui 1 hā i for h {0, 1,..., u i }, i {1, 2,..., n} with m 0u0 = 0, and m ih = m ih (u l η + 1)a l for h {0, 1,..., u i }, i {s, s + 1,..., n}, and m J ++ = m sus, m I ++ = m nun, k {0, 1,..., η 1}, and p {0, 1,..., u l η}. Proof. From Lemma 5 (ii), we may assume that z i = 0 for all i I + with ā i r. Consequently, the condition of Lemma 6 is satisfied. If i I ++ =, then (15) equals (14); hence Ω(a) = Φ(a) for a 0 as m J ++ = 0 and m I ++ = (η u l 1)a l. Otherwise, from Lemma 5 (i) and Lemma 6, there exist optimal solutions to (16), in which z i i I ++ increase monotonically in nonincreasing order of ā i as a decreases. That is, as a decreases from 0 there exists optimal solutions, where first z 1 is incremented from 0 to u 1 and then z 2 and so on. Thus by fixing z i i I ++ in the order described in Lemma 5 (i) and Lemma 6, the lifting problem reduces to optimizing over the remaining variables z l and y. Suppose z i = u i for i {1, 2,..., j 1}, z j = ρ, and z i = 0 for i {j + 1, j + 2,..., n }. So the right hand side of the knapsack constraint for the reduced problem in two variables is d m j(ρ 1) a.

13 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 13 Then, if δ j 0, similar to the discussion in Section 2.3, for m jρ a + m j(ρ 1) m j(ρ 1) an optimal solution for the maximization problem (16) is given by (z j, z l, y) = (ρ, k, 0) (ρ, 0, a + δ j + a l r) if δ j a 0, (ρ, k, a + δ j + (k + 1)a l r) if δ j ka l (a l r) a δ j ka l, if δ j (k + 1)a l a δ j ka l (a l r), where k {0, 1,..., η 1}. Since (ρ, η 1, 0) and (ρ + 1, 0, δ j + a l r) both have the same objective value, as a decreases further, the structure of optimal values of z l and y repeats for z j {ρ+1, ρ+2,..., u j }. On the other hand if δ j < 0, then for m jρ a+ m j(ρ 1) m j(ρ 1) (z j, z l, y) = (ρ, u l η + 1, a + δ j + a l r) if (a l r) δ j a 0, (ρ, u l η k, 0) if δ j (k + 1)a l a δ j (k + 1)a l + r, (ρ, u l η k, a + (k + 1)a l r + δ j ) if δ j (k + 1)a l + r a ka l δ j, where k {0, 1,..., η 1} is optimal for the lifting problem (16). Since (ρ, u l, 0) and (ρ + 1, u l η + 1, δ j + a l r) have the same objective value, as a decreases further, z j increases to ρ + 1 and the structure of optimal values of z l and y repeats. Evaluating Ω for these optimal solutions by incrementing z 1, z 2,..., z n one at a time in the order given in Lemma 5 (i) and Lemma 6 we obtain the expression in the statement of the theorem for Ω. An example lifting function Ω is depicted in Figure 4. Observe that the last case in the definition of Ω applies if u n < and the three cases before that apply if u l <. Also note that if I ++ =, Ω(a) = Φ(a) for a 0 as m J ++ = 0 and m I ++ = (η u l 1)a l. Giving an explicit description of Ω is difficult in general, because the properties described in Lemma 5 (i) and Lemma 6 do not hold for x i i I + with r < ā i < d. Therefore, instead, we give a lower bound on Ω, which equals Ω over a significant part of its domain. The lower bound is obtained by dropping the upper bound constraint z l u l from the lifting problem (16) so that there is an easy description of the optimal solutions to this relaxed problem as described in part (iii) of Lemma 5. Dropping z l u l from (16), we obtain the following

14 14 ALPER ATAMTÜRK lower bound on Ω: ω(a) = d ηr max Φ(ā i )z i y : i li + ā i z i y d a, z i u i i I +, z Z li+ +, y IR + i li + = d ηr max (a l r)z l + (ā i ηa l )z i y : i J ++ a l z l + ā i z i y d a, z i u i i J ++, z Z lj ++ +, y IR +. i J ++ The last equality follows from part (iii) of Lemma 5. Since ā i d for all i J ++ and the upper bound on z l is dropped, it follows from Theorem 7 that ω can be expressed as ω(a)= u ih ηr + a if m ih δ i a m ih, (u ih η + k)r + a if m ih δ i (k+1)a l + r a m ih δ i ka l, u ih ηr + m ih δ i (k+1)(a l r) if m ih δ i (k+1)a l a m ih δ i (k+1)a l +r, u sus ηr + pr + a if m J ++ (p+1)a l + r a m J ++ pa l, m J ++ + u sus ηr (p + 1)(a l r) if m J ++ (p+1)a l a m J ++ (p+1)a l + r, where k {0, 1,..., η 1} and p Z +. Proposition ω is a superadditive lower bound on Ω on IR. 2. Ω(a) = ω(a) for m J ++ a 0; hence Ω is superadditive on [ m J ++, 0], where m J ++ = m J ++ (u l η + 1)a l. 3. Ω is superadditive on IR under any of the following conditions: (i) u l =, (ii) η = 1, (iii) i I + s.t. r < ā i < ηa l, (iv) i I ++ s.t. u i =. Proof. 1. Since δ i 0 for all i J ++, ω is a special case of the superadditive function ψ introduced in Section 3.3 with parameters b i = δ i, e = a l, ρ = r, and τ = η. As ω is obtained by dropping the constraint z l u l from the lifting problem (16), ω is a lower bound on Ω. 2. Follows from the descriptions of the functions Ω and ω. 3. In case (i) Ω = ω on IR. In case (ii) Ω is a special case of the superadditive function ψ in Section 3.3 with parameters b i = δ i + a l r, e = r, ρ = r, and τ = 1. In cases (iii) and (iv) Ω is a special case ψ with the same parameters as in part 1. Remark 1. Observe that whenever x l {0, 1} and (13) is facet defining for conv(k U,l ), we have r = d and η = 1. Therefore the condition of Theorem 7 is satisfied and, from Proposition 8, Ω is superadditive on IR. It is possible to construct Ω that is not superadditive on IR if none of the conditions of part 3 of Proposition 8 is satisfied.

15 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 15 From Theorem 2 and Proposition 8 we obtain the valid inequalities described in Theorem 9. Theorem 9. For l I such that a l > 0, U I B \ {l}, and S C B let Φ, Ω, and ω be defined as before. Then inequality (19) Φ(a i )x i + Φ( a i )(u i x i ) + ω(a i )x i + i li + \U i I + U i I \U i I Uω( a i )(u i x i ) + i S w i i C w i b ηr is valid for K. It is facet defining for conv(k) if inequality (13) is facet defining for conv(k U,l ) and ā i m J ++ for all i I. Moreover, when any one of the conditions of part 3 of Proposition 8 is satisfied, ω may be replaced with the exact lifting function Ω so that (19) is facet defining for conv(k). Example 2. Let K = {x Z 3 +, w IR 2 + : 3x x 2 4x 3 + w 1 w 2 8, x 1 3, w 1 1}. For l = 1 consider the restriction K,1 = {x 1 Z +, w IR 2 + : 3x 1 + w 1 w 2 8, x 1 3, w 1 1}. From Section 2.1 the two additional inequalities needed to describe conv(k,1 ) are (20) 2x 1 + w 1 w 2 5, (21) x 1 w 2 2 with S = {1} and S =, respectively. Lifting (20) first with x 2 using Φ, and then with x 3 using Ω gives us the facet defining inequality (19) (22) 2x 1 + 7x 2 3x 3 + w 1 w 2 5. Note that here η = (h v 1 )/a l = 3, r = 1, and consequently Φ(10) = 7, δ 1 = a 2 ηa l = 1, and Ω( 4) = 3. The lifting functions Φ and Ω for (20) are drawn in Figures 3 and 4. Observe that the MIR inequality (23) 2x 1 + 6x 2 3x 3 + w 1 w 2 5, obtained by lifting (20) using the lower bound φ is weaker than (22). Similarly, lifting (21) first with x 2 using Φ, and then with x 3 using Ω gives us the facet defining inequality (19) (24) x 1 + 4x 2 2x 3 w 2 2 as η = h/a l = 3, r = 2, and consequently Φ(10) = 4 and since δ 1 = 1 we have Ω( 4) = 2. Lifting functions for this inequality are not drawn. Again the corresponding MIR inequality (25) x 1 + 3x 2 2x 3 w 2 2 obtained by lifting (21) using φ is weaker than (24).

16 16 ALPER ATAMTÜRK Φ φ (η u l 1)a l r a l r 5 8 ηa l r a l r r a l r r a l r r 2 4 Figure 3. Lifting function Φ in Example 2 (η = u l = 3). δ 1 r a l r r a l r r a l r δ 1 ā Ω Figure 4. Lifting functions Ω in Example 2 (η = 3).

17 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON Inequality class II. Since Φ is superadditive also on IR, this time we lift the SMIR inequality (14) first with x i i I to obtain the intermediate inequality (26) i li Φ(ā i )z i y d ηr since Φ(ā l ) = a l r. Next we lift (26) with x i i I +. Let us define the lifting function of (26) as (27) Γ(a) = d ηr max Φ(ā i )z i y : (z, y) K I +,li (a), i li where K I +,li (a) = (z, y) ZlI + IR + : ā i z i y d a, z i u i i li i li for a IR +. Lemmas 10 and 11 below are the central results on the structure of optimal solutions to problem (27). They lead to an explicit description of Γ in a special case, a superadditive lower bound on Γ, and the valid inequalities described in Theorem 14. Let I = {i I : ā i d u l a l } = {1, 2,..., n } be indexed in nondecreasing order of ā i, ties broken arbitrarily. Let n = min{i I : u i = } (if I =, then let n = 0; if u i < for all i I, then let n = n ). If I, let J = {i I : ā i (η 1 u l )a l } = {1, 2,..., s }, s = min{i J : u i = } (if u i < for all i J, then let s = s ). Since J I, we have s n and if u i = for some i J, then n = s. Lemma 10. For a 0 the maximization problem (27) has an optimal solution (z, y) such that (i) z i = u i for all i {1, 2,..., j 1} and z i = 0 for all i {j + 1, j + 2,..., n } some j {1, 2,..., n }, (ii) z i = 0 for all i I with r a l ā i 0, and (iii) z i = 0 for all i I with (η 1 u l )a l < ā i < r a l if z l 0 is removed from problem (27).

18 18 ALPER ATAMTÜRK Proof. (i) For i I, we have Φ(ā i ) = (η 1 u l )(a l r). Let ā h < ā i for h, i I and suppose z h < u h and z i > 0. Decreasing z i by one, increasing z h by one, we obtain a feasible solution with the same objective value. (ii) Since Φ(ā i ) = ā i for i I with r a l ā i 0, if z i = p > 0, we obtain a feasible solution, with the same objective value by increasing y by ā i p and decreasing z i to zero. (iii) Suppose z i = p > 0. If ka l ā i < ka l + r, then Φ(ā i ) = (a l r)k. Notice that since ā i < 0 and ā l > 0, k is a negative integer. Since z l 0 is removed, by decreasing z i to 0 and decreasing z l by kp, we obtain a feasible solution with the same objective value. Else if ka l + r ā i < (k + 1)a l, then Φ(ā i ) = ā i (k + 1)r. Similarly, decreasing z i to 0 and decreasing z l by (k + 1)p and y by ((k + 1)a l ā i )p, we obtain a feasible solution with the same objective value. Lemma 11. If I = I, then for a 1 > a 2 0 the maximization problem (27) has optimal solutions (z(a 1 ), y(a 1 )) and (z(a 2 ), y(a 2 )) that satisfy z i (a 1 ) z i (a 2 ) for all i I. Proof. Let a 1 > a 2 and (z(a 1 ), y(a 1 )) and (z(a 2 ), y(a 2 )) be two optimal solutions satisfying Lemma 10 (i). Suppose z k (a 1 ) < z k (a 2 ) for some k {1, 2,..., n }. Then z k (a 1 ) < u k, z k (a 2 ) > 0, and from Lemma 10 (i), we have z i (a 1 ) = 0 for all i {k + 1, k + 2,..., n } and z i (a 2 ) = u i for all i {1, 2,..., k 1}, implying z i (a 1 ) z i (a 2 ) for all i I. Let ϕ be the slack of the knapsack constraint i I ā iz i (a 2 )+a l z l (a 2 ) y(a 2 ) d a 2. Since a 1 > a 2, feasibility of (z(a 1 ), y(a 1 )) requires that (28) (29) y(a 1 ) > y(a 2 ) + i I ā i (z i (a 1 ) z i (a 2 )) + a l z l (a 1 ) a l z l (a 2 ) ϕ ā k + a l z l (a 1 ) a l z l (a 2 ) ϕ. The last inequality follows from y(a 2 ) 0, z k (a 1 ) < z k (a 2 ), ā i < 0, and z i (a 1 ) z i (a 2 ) for all i I. Let δ k = (η u l 1)a l ā k. The upper bound we give on the slack ϕ is a function of the sign of δ k and the value of z l (a 2 ). If z l (a 2 ) < u l, then ϕ r. Since otherwise ϕ = r + ɛ with ɛ > 0 and increasing z l (a 2 ) by one and increasing y(a 2 ) by (a l r ɛ) + gives a solution with objective value min{a l r, ɛ} larger than for (z(a 2 ), y(a 2 )). On the other hand, if z l (a 2 ) u l η, then ϕ r + δ k. Since otherwise ϕ = δ k + r + ɛ, with ɛ > 0 and the solution obtained from (z(a 2 ), y(a 2 )) by decreasing z l (a 2 ) by u l η and y(a 2 ) by (a l r ɛ) + and increasing z k (a 2 ) by one has an objective value min{a l r, ɛ} larger than for (z(a 2 ), y(a 2 )). Hence, we conclude that when δ k 0, we have ϕ r + δ k if z l (a 2 ) = u l and ϕ r otherwise; and when δ k < 0, we have ϕ r + δ k if z l (a 2 ) u l η and ϕ r otherwise. Next we use the bounds on ϕ either to obtain a contradiction or to change the value of z k (a 1 ) or z k (a 2 ) toward satisfying z k (a 1 ) z k (a 2 ). First consider the case δ k 0. If z l (a 1 ) η 1, then the solution obtained from (z(a 1 ), y(a 1 )) by increasing z k (a 1 ) by one and increasing z l (a 1 ) by u l η + 1 is feasible since δ k 0 and has the same objective value.

19 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 19 So we may assume that z l (a 1 ) η. Then, using ϕ r + δ k, from (29) we get y(a 1 ) > ā k a l (u l η) δ k r a l r. That is, y(a 1 ) = a l r+ɛ, where ɛ > 0. But then the solution obtained from (z(a 1 ), y(a 1 )) by decreasing z l (a 1 ) by one and decreasing y(a 1 ) by a l r+min{r, ɛ} is feasible since z l (a 1 ) 1 and has an objective value min{r, ɛ} larger than for (z(a 1 ), y(a 1 )), which contradicts the optimality of (z(a 1 ), y(a 1 )). Now consider the case δ k < 0. If z l (a 2 ) u l η + 1, then the solution obtained from (z(a 2 ), y(a 2 )) by decreasing z k (a 2 ) by one and decreasing z l (a 2 ) by u l η + 1 is feasible has the same objective value. Therefore, we may assume that z l (a 2 ) u l η. We need to consider three subcases depending on the values of z l (a 1 ) and z l (a 2 ). (i) z l (a 1 ) > 0: Using z l (a 2 ) u l η and ϕ r, from (29) we get y(a 1 ) > ā k + a l z l (a 1 ) a l (u l η) r a l z l (a 1 ) r a l r. So y(a 1 ) = a l r + ɛ with ɛ > 0. Since z l (a 1 ) 1, the solution obtained from (z(a 1 ), y(a 1 )) by decreasing z l (a 1 ) by one and decreasing y l by a l r + min{r, ɛ} gives a feasible solution with objective value min{r, ɛ} larger than for (z(a 1 ), y(a 1 )), contradicting its optimality. (ii) z l (a 1 ) = 0 and z l (a 2 ) < u l η: Let κ = u l η z l (a 2 ). Using ϕ r, (29) gives y(a 1 ) > ā k a l (u l η κ) r (a l (η u l ) ā k ) + a l κ r a l κ r a l r. Thus again y(a 1 ) = a l r+ɛ with ɛ > 0. Since z l (a 1 ) = 0, in this case the solution obtained from (z(a 1 ), y(a 1 )) by increasing z k (a 1 ) by one and increasing z l (a 1 ) by u l η and y(a 1 ) by a l r+min{ɛ, δ k +r} is feasible and improves the objective value by min{ɛ, δ k + r}. If δ k + r > 0, this contradicts the optimality of (z(a 1 ), y(a 1 )). If δ k + r = 0, we have an alternative solution in which z k (a 1 ) is one larger. (iii) z l (a 1 ) = 0 and z l (a 2 ) = u l η: Using ϕ δ k + r from (29) we have y(a 1 ) > ā k a l u l + a l η δ k r = ( ā k + (η u l 1)a l δ k ) + a l r = a l r. Since and z l (a 1 ) = 0 and y(a 1 ) = a l r + ɛ with ɛ > 0 the case reduces to case 2 above.

20 20 ALPER ATAMTÜRK Theorem 12. If ā i d u l a l or ā i r a l for all i I, then u ih (u l η+1)(a l r) if m ih a m ih +δ i, (u ih (u l η+1) + k)(a l r) if m ih +δ i +ka l a m ih +δ i + ka l +r, u ih (u l η+1)(a l r)+a m ih δ i (k+1)r if m ih +δ i +ka l +r a m ih +δ i + (k+1)a l, (u sus (u l η+1) + p)(a l r) if m J +pa l a m J +pa l + r, Γ(a)= u sus (u l η+1)(a l r)+a m J (p+1)r if m J +pa l + r a m J +(p+1)a l, (u iui (u l η+1)+η)(a l r) if m ih a m ih +δ i + r (u iui (u l η+1)+η)(a l r)+a m ih δ i (k+1)r if m ih +δ i +ka l +r a m ih +δ i +(k+1)a l (u iui (u l η+1)+η+k)(a l r) if m ih +δ i + ka l a m ih + δ i + ka l +r u nun (u l η+1)(a l r)+a m I ηr if a m I, where δ i = (η u l 1)a l ā i for i I, u ih = i 1 k=1 u k + h, m ih = m (i 1)ui 1 hā i for h {0, 1,..., u i } and i {1, 2,..., n} with m 0u0 = 0, m ih = m ih +ηa l for h {0, 1,..., u i } and i {s, s + 1,..., n}, m J = m sus, m I = m nun, k {0, 1,..., u l η}, and p {0, 1,..., η 1}. Proof. From Lemma 10 (ii), we may assume that z i = 0 for all i I such that ā i r a l. Hence the condition of Lemma 11 is satisfied. Observe that if I =, inequality (26) equals (14). Consequently Γ(a) = Ω(a) for a IR + as m J = 0 and m I = ηa l. Otherwise, from Lemma 10 (i) and Lemma 11, as a increases, there exist optimal solutions in which z i i I is incremented monotonically in nondecreasing order of ā i. Thus after fixing z 1, z 2,..., z n in this order, the problem reduces to one with two variables z l and y as in Section 2.3. Suppose z i = u i for i {1, 2,..., j 1}, z j = ρ, and z i = 0 for i {j + 1, j + 2,..., n }. For the restricted problem the right hand side of the knapsack constraint becomes d m j(ρ 1) a. Then, if δ j 0, for m j(ρ 1) m j(ρ 1) + a m jρ an optimal solution for the restricted problem is given by (ρ, u l, 0) if 0 a δ j (z j, z l, y) = (ρ, u l k, 0) if δ j + ka l a δ j + ka l + r (ρ, u l k, a δ j ka l r) if δ j + ka l + r a δ j + (k + 1)a l, where k {0, 1,..., u l η}. Since (ρ, η, a l r) and (ρ + 1, u l, 0) are alternative optimal solutions when a increases further, the values of z l and y repeat the same pattern for z j {ρ + 1, ρ + 2,..., u j }. On the other hand, if δ j < 0, then similarly for m j(ρ 1) m j(ρ 1) + a m jρ (ρ, u l η, 0) if 0 a δ j + r (z j, z l, y) = (ρ, u l η k, a δ j ka l r) if δ j + ka l + r a δ j + (k + 1)a l (ρ, u l η k, 0) if δ j + ka l a δ j + ka l + r, where k {0, 1,..., u l η} is an optimal solution to the restricted lifting problem. Since (ρ, 0, a l r) and (ρ + 1, u l η, 0) have the same objective value, when a increases further, the structure of optimal solutions is repeated for z j {ρ + 1, ρ + 2,..., u j }. Using the

21 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 21 equality d ηr = (η 1)(a l r) in (27) and evaluating Γ for these optimal solutions by in incrementing z 1, z 2,..., z n one at a time in the order described in Lemma 10 (i) and Lemma 11 gives the expression of the theorem for Γ. An example lifting function Γ is depicted in Figure 5. Observe that the last case in the definition of Γ applies if u n <. Also note that if I =, Γ(a) = Φ(a) for a 0 as m J = 0 and m I = ηa l. The properties described in Lemma 10 (i) and Lemma 11 do not hold for x i i I with d u l a l < ā i < r a l, which makes it hard to characterize Γ in general. Therefore, we give a lower bound γ on Γ by dropping the nonnegativity constraint z l 0 from (27) so that an optimal solution can be easily described based on part (iii) of Lemma 10. So consider the relaxation of problem (27) γ(a) = d ηr max Φ(ā i )z i + (a l r)z l y : i I ā i z i + a l z l y d a, z i u i i li, z l Z, z Z I +, y IR + i I = (η 1)(a l r) max (η 1)(a l r)z i + (a l r)z l y : i J z i u i i lj, z l Z, z Z J +, y IR + ā i z i + a l z l y (η 1)a l + r a. i J The second equality follows from Lemma 10 (iii). Since ā i d u l a l for all i J and the lower bound on z l is dropped, from Theorem 12 we get u ih (u l η + 1)(a l r) if m ih a m ih + δ i, (u ih (u l η + 1) + k)(a l r) if m ih +δ i +ka l a m ih + δ i +ka l +r, u γ(a) = ih (u l η+1)(a l r)+a m ih δ i (k + 1)r if m ih +δ i +ka l +r a m ih +δ i +(k + 1)a l, (u sus (u l η + 1) + p)(a l r) if m J + pa l a m J + pa l + r, u sus (u l η + 1)(a l r)+a m J (p + 1)r if m J + pa l + r a m J + (p + 1)a l, where k {0, 1,..., u l η} and p Z +.

22 22 ALPER ATAMTÜRK Proposition γ is a superadditive lower bound on Γ on IR Γ(a) = γ(a) for 0 a m J ; hence Γ is superadditive on [0, m J ], where m J = m J + ηa l. 3. Γ is superadditive on IR + under any of the following conditions: (i) η = u l, (ii) i I s.t. (η u l 1)a l < ā i < r a l, (iii) i I s.t. u i =. Proof. 1. Since δ i 0 for all i J, γ is a special case of the superadditive function χ in Section 3.3 with parameters b i = δ i, e = a l, ρ = r, and τ = u l η + 1. Since γ is obtained by solving a relaxation of the lifting problem obtained by dropping the constraint z l 0, it is a lower bound on Γ. 2. Immediate from the descriptions of Γ and γ. 3. In case (i) Γ is a special case of the superadditive function χ in Section 3.3 with b i = δ i + r, e = a l r, ρ = 0, and τ = 1. In cases (ii) and (iii) Γ is a special case of χ with the same parameters as in part 1. Remark 2. Observe that if x l {0, 1} and (13) is facet defining for conv(k U,l ), then r = d and η = u l. Therefore the condition of Theorem 12 is satisfied and by Proposition 13 Γ is superadditive on IR +. It is possible to construct Γ that is not superadditive on IR + if none of the conditions of part 3 of Proposition 13 is satisfied. Finally Theorem 2 and Proposition 13 lead to the valid inequalities described in Theorem 14. Theorem 14. For l I s.t. a l > 0, U I B \ {l}, and S C B let Φ, Γ, and γ be defined as before. Then inequality (30) γ(a i )x i + γ( a i )(u i x i ) + Φ(a i )x i + i I + \U i I + U i li \U i I UΦ( a i )(u i x i ) + i S w i i C w i h ηr is valid for K. It is facet defining for conv(k) if inequality (13) is facet defining for conv(k U,l ) and ā i m J for all i I +. Moreover, when any one of the conditions of part 3 of Proposition 13 is satisfied, γ may be replaced with Γ so that (30) is facet defining for conv(k). Example 2 (cont.) When we lift (20) first with x 3 using Φ, and then with x 2 using Γ we obtain the facet defining inequality (30) (31) 2x 1 + 4x 2 2x 3 + w 1 w 2 5 since Φ( 4) = 2 and Γ(10) = 4 for η = 3, r = 1, and δ 1 = (η u l 1)a l a 3 = 1. The lifting functions Φ and Γ for (20) are depicted in Figures 3 and 5.

23 ON THE FACETS OF THE MIXED INTEGER KNAPSACK POLYHEDRON 23 On the other hand lifting (21) first with x 3 using Φ, then with x 2 using Γ gives us the facet defining inequality (30) (32) x 1 + 2x 2 x 3 w 2 2 as η = 3, r = 2 and consequently Φ( 4) = 1 and since δ 1 = 1, we have Γ(10) = 2. Γ ā δ 1 r a l r δ 1 r a l r δ 1 r a l r δ 1 Figure 5. Lifting function Γ in Example 2 (cont.) (η = 3) Case 2: a l < 0. In this case the inequalities (33) rz l + w i w i v S ηr i S i C S C B where η = ( b v S )/a l and r = v S b+ ( b v S )/a l a l are sufficient to describe conv(k U,l ) when added to formulation with a l z l + i C + w i i C w i b and the bounds. Lifting them in a similar way as in Section 3.1, we obtain the inequalities described in Theorems 15 and 16. Theorem 15. For l I such that a l < 0, U B I \ {l}, and S C B let Φ and ω be defined as before. Then inequality i li \U (a i + Φ( a i ))x i + i I + U i li U ( a i + ω(a i ))(u i x i ) + ( a i + Φ(a i ))(u i x i )+ i I + \U(a i + ω( a i ))x i + i S w i i C w i v S ηr is valid for K. It is facet defining for conv(k) if inequality (33) is facet defining for conv(k U,l ) and ā i m J for all i I +.

24 24 ALPER ATAMTÜRK Theorem 16. For l I such that a l < 0, U B I \{l}, and S C B let Φ and γ be defined as before. Then inequality i li + \U (a i + Φ( a i ))x i + i I \U i li + U (a i + γ( a i ))x i + ( a i + Φ(a i ))(u i x i )+ i I U( a i + γ(a i ))(u i x i ) + i S w i i C w i v S ηr is valid for K. It is facet defining for conv(k) if inequality (33) is facet defining for conv(k U,l ) and ā i m J ++ for all i I Four superadditive functions. Here we prove the superadditivity of four general piecewise linear continuous functions of which the lifting functions Ω, Γ, ω, and γ introduced in Sections and are particular cases. Let b i IR + i = 1, 2,..., m be such that b i b i+1. Let e ρ 0 and a i = τe + b i for some nonnegative integer τ. Define the partial sums A 0 = 0, A i = i k=1 a k, and B i = A i 1 + b i for 1 i m. Let χ : [0, A m ] IR + be defined as iτ(e ρ) if A i a B i+1, χ(a)= (iτ + k)(e ρ) if B i+1 + ke a B i+1 + ke + ρ, iτ(e ρ) + a B i+1 (k+1)ρ if B i+1 + ke + ρ a B i+1 + (k+1)e, where k {0, 1,..., τ 1} and ψ : [ A m, 0] IR be ψ(a) = a + χ( a) for a [ A m, 0]. Also let χ : IR + IR + be defined as { χ(a) if 0 a A m, χ(a) = χ(a m ) + a A m if A m a. and ψ : IR IR as ψ(a) = a + χ( a) for a IR. As stated in Proposition 8, ω is a special case of ψ and and Ω is a special case of ψ under the conditions of part 3 of the same proposition. Also γ is a special case of χ as described in Proposition 13 and Γ is a spacial case of χ under the conditions of part 3 of the same proposition. By the following lemmas, χ, χ, ψ, and ψ are superadditive on their domain. Lemma 17. χ is superadditive on [0, A m ]. Proof. Throughout we will make use of the following observations: 1. χ is nondecreasing. 2. A i+j A i + A j for i, j {0, 1,..., m} such that i + j m (since 0 b i+1 b i ). The proof consists of verifying that χ(a) + χ(b) χ(a + b) for all a, b 0 such that a + b A m, which reduces to the verification of the following six cases by symmetry.