This action might not be possible to undo. Are you sure you want to continue?
BY R. E. GOMORY
THOMAS J. WATSON RESEARCH CENTER, YORKTOWN HEIGHTS, NEW YORK
Communicated by R. Courant, November 1, 1966
In reference 5 a connection was given between the integer and noninteger solutions to the linear programing problem (1) maximize z = cx Ax = b, x> 0. In (1) x is an m + n vector, b is an integer m-vector, c an m + n vector, and A an m X (m + n) integer matrix containing an m X m identity matrix. A is assumed to be rearranged and partitioned into an m X m optimal basis matrix B for the noninteger problem and a collection of nonbasic columns forming the matrix N with A = (B,N). An alternative form of (1) that is useful here for geometric interpretation is to revert to inequalities, writing A as (A', I). Then (1) becomes (la) maximize z = c'x' A'x' . b
where x' and c' are n-vectors. Under suitable conditions, given in reference 5, the integer solution to (1) could be obtained from the noninteger one by solving the optimization problem
subject to the conditions
£ gat9 = go t=1
where the t, are required to be nonnegative integers. The c* (which are not important here) are the relative cost coefficients associated with the columns of N, and g9 is the element of the factor module g = M(I)/M(B) corresponding to the ith column of N. Here M(I) is the module of all integer m-vectors, and M(B) the module generated over the integers by the columns of B. The connection between integer and noninteger solutions established by (2) held only under certain conditions. One way to develop this approach into a general integer programing algorithm would be to develop from (2) new inequalities or "cutting planes" for a method similar to that of reference 6. The geometrical interpretation of the solutions to (2) suggests that this is possible and this approach is outlined here. To see this, consider the cone P' in the space of the variables x' of (la) formed by using only the inequalities corresponding to the nonbasic variables or equivalently, P' is the cone obtained if in (la) the nonnegativity condition for the basic variables is dropped. Within P' is the polyhedron P' which is the convex hull of the integer points of P'. P' is an interesting object of study in itself. In addition, its faces
whose proof is not given here. 0 g. 3. tn). equations (3) lead to the proof of the following theorems which involve considering the tree of shortest paths over the group G..) This calculation will not be described here.. Let gi. and 7. tn) for which Ztigi = 0.. and hence a face of P'.. .. The remaining gi and the element go corresponding to b i=p can then be represented as p-vectors with respect to this basis. Since the variables t of (2) determine a corresponding x satisfying the equations of (1) by t -* x = (B'-(b .Nt). . . gp. In addition to providing a method of computing faces of P" by linear programing. . E. .VOL. GOMORY 17 clearly provide the strongest inequalities or cutting planes for the general integer programing problem that can be deduced locally.. Fifth: Duplication. THEOREM 1. without using the nonlocal information available from the nonnegativity condition on the basic variables. . 57. can easily be taken care of and simply reduces the size of the problem to be dealt with. i. those that are simply faces of the original problem (1). i. 5.. Third: The multiplicity of rows in (3) can be dealt with by a row-generating method similar to the methods of references 1. i.e. gp) so that G = gi91) 02 . Or alternatively one can work only with those T that satisfy (2) for some nonnegative c*..t) and hence also determine the x' of (la). . by solving a single problem like (2) with an additional side calculation that less than doubles the work. satisfying the equations of (2). 1967 MATHEMATICS: R. THEOREM 2.4. Fourth: There is a simple way of getting a first feasible solution to (3).. The inequality 2irrti > 7ro is a face of P' if and only if the Ore are a basic feasible solution of the system of inequalities. see ref. A number of remarks can be made about the feasibility of this computation. so k = E Ykigi 1=1 and the following theorem. . . essentially a shortest path problem over the group G. the direct sum. Iffor some basis g9. .e.s then the 7ri given by . Second: The trivial faces of P". can be discarded by choosing lro $ 0.. Because of this. The needed row at each simplex step can be generated (at worst) by solving a problem approximately equivalent to (2). gn be the group elements corresponding to the columns of N. inequalities on the tj yield inequalities on the x'. ... go = 2yojgj has a component s for which 70. First: Although there are an infinity of T satisfying (2) and hence an infinity of inequalities. an n X n basis matrix is the most that is required at any time. many columns of N mapping into the same group element..5s . holds. Choose g ) from among the gj a basis (we can assume it is g1.e. 7rT > lro (3) made up from all vectors T = (ti. and in this sense one can talk about an inequality on the tj being a face of P". those T not containing as a vector T' = (ti. (For an estimate of the work involved in solving (2). . it is easy to reduce this to a finite number by considering only the irreducible T. the representation of go. Faces to Pa can be characterized by the following easily proved theorem which allows their computation by linear programing. maxk > p Yk.
"Decomposition principle for linear programs. 6 Gomory. 2 Edmonds. Hu. independent of the particular columns present in the matrix M.1. different tie-breaking methods produce different trees. and T... Fulkerson. The 7rO' corresponding to b' can be obtained by adding the tree distance from go to g' to the original Tro. E. . and P. i = 1." SIAM J. "A suggested computation for maximal multi-commodity network flows. R. G. R." Operations Res. yield groups G that are direct sums of cyclic groups of order 2. J. N.. . for some component s. "A linear programing approach to the cutting-stock problem. 5. . In what follows. then any basis of g satisfies the conditions of Theorem 2. ed. . .. "On the relation between integer and noninteger solutions to linear programs. "Synthesis of a communication network. 348369 (1964). 4 Gilmore. NR 047040. Jr. T2. Wolfe. * This work was supported in part by the Office of Naval Research under contract Nonr 3775(00). 9.2 In the next theorem we refer directly to the shortest path tree.. . 269-302. pp.is TO = MATHEMATICS: R.. "Maximum matching and a polyhedron with 0. COROLLARY. If g is the direct sum of cyclic groups of order 2. i = 1. (4) TD-. 1 Dantzig. R. E.. Let the faces Fj be all faces of the higher-dimensional integer polyhedron P* obtained from (2) by letting the index i in (2) range over all group elements." in Recent Advances in Mathematical Programming." Management Sci. Res. we have a graph and hence for any choice of 7ri. then the 7ri are also a face for the polyhedron P" resulting from the right-hand side b'. 12. R. b' denotes a right-hand side in (1) and g' is the group element fb' under the natural mappingf which sends M(I) onto G. the shortest paths from one point to all the others will form a spanning tree.. 101-111 (1960).. Bur.. We can now state Theorem 3. obtained by deleting from (4) all 7rij whose corresponding group element is not a column of N. C. and R. GOMORY YOs. 53. "An algorithm for integer solutions to linear programs. a shortest path tree. If R is a shortest path tree for the 7rT forming a face of P". Graves and Philip Wolfe (New York: McGraw-Hill. The faces Fj'. or . C. B. P. 7 Gomory..ij. YOys is exactly one less than the order of g. 7k = w=1. Gomory. E.. 5 Gomory. In a graph. 260-265 (1965). n. Tk = 'Yksk > p. PROC. 125-130 (1965). Tii. L. 849-859 (1961). and if g' = fb' is separated from 0 in R by go. Natl. . we have the corollary.. Std." Operations Res. 1-vertices. R. B. and 0 otherwise give a face of P". It is easily shown that all A consisting of columns with at most two nonzero entries that are restricted to be 1. S. and D. include all faces of the original polyhedron P". and if the gi. 69B. The condition of the theorem is always met whenever. 8. However.i))Fj = (Troi... 1963). if any unambiguous method of breaking ties among paths is used so that there is a unique shortest path between two points.. Finally we state a theorem that allows the computation of faces on a once-andfor-all basis." J. This connects with the work of Edmonds. n." these PROCEEDINGS. A. E. 97-101 (1958).. E. Robert L. THEOREM 3. THEOREM 4. In particular. If the elements of g are taken as nodes of a graph. are taken as directed arcs connecting each g' to the point g' + gi. 3 Ford.