Martin Grötschel Institute of Mathematics, Technische Universität Berlin (TUB) DFG-Research Center “Mathematics for key technologies” (M ATHEON ) Konrad-Zuse-Zentrum für Informationstechnik Berlin (ZIB) Independence Systems, Matroids, the Greedy Algorithm, and related Polyhedra Martin Grötschel Summary of Chapter 4 of the class Polyhedral Combinatorics (ADM III) May 25, 2010 & June 1, 2010
Matroids and Independence Systems Let E be a finite set, I a subset of the power set of E. The pair (E,I ) is called independence system on E if the following axioms are satisfied: (I.1) The empty set is in I. (I.2) If J is in I and I is a subset of J then I belongs to I. Let (E,I ) satisfy in addition: (I.3) If I and J are in I and if J is larger than I then there is an element j in J, j not in I, such that the union of I and j is in I. Then M=(E,I ) is called a matroid.
Notation Let (E,I ) be an independence system. Every set in I is called independent. Every subset of E not in I is called dependent. For every subset F of E, a basis of F is a subset of F that is independent and maximal with respect to this property. The rank r(F) of a subset F of E is the cardinality of a largest basis of F. Important property, submodularity: The lower rank of F is the cardinality of a smallest basis of F.
The Largest Independent Set Problem Problem: Let (E,I ) be an independence system with weights on the elements of E. Find an independent set of largest weight. We may assume w.l.o.g. that all weights are nonnegative (or even positive), since deleting an element with nonpositive weight from an optimum solution, will not decrease the value of the solution.
The Greedy Algorithm Let (E,I ) be an independence system with weights c(e) on the elements of E. Find an independent set of largest weight. The Greedy Algorithm: 1. Sort the elements of E such that 2. Let 3. FOR i=1 TO n DO: 4. OUTPUT A key idea is to interprete the greedy solution as the solution of a linear program.
The greedy algorithm works for matroids Proof using axiom (I.3) on the blackboard. Martin Grötschel 6
Polytopes and LPs Let M=(E,I ) be an independence system with weights c(e) on the elements of E.
The Dual Greedy Algorithm Let (E,I ) be an independence system with weights c(e) for all e. After sorting the elements of E so that set Then is a feasible solution of the dual LP by construction. (integral if the weights are integral))
Observation Let (E,I ) be an independence system with weights c(e) for all e. After sorting the elements of E so that We can express every greedy and optimum solution as follows:
Rank Quotient Let (E,I ) be an independence system with weights c(e) for all e. The number q is between 0 and 1 and is called rank quotient of (E,I ). Observation: q = 1 iff (E,I ) is a matroid.
The General Greedy Quality Guarantee a quality guarantee
Consequences Let M=(E,I ) be an independence system with weights c(e) on the elements of E.
Consequences Theorem. For every independence system (E,I with weights c(e) for all elements e of E, max c T x, xP(I) ≥ c(I opt ) ≥ c(I greedy ) ≥ q max c T x, xP(I) ≥ q c(I opt ), in other words, the greedy solution value is bounded from below by q times the maximum value of the LP relaxation and not only by q times the optimum value of the weighted independent set problem. Martin Grötschel 13
More Proofs Another proof of the completeness of the system of nonnegativity constraints and rank inequalities will be given in the class on the blackboard, see further slides. Martin Grötschel 14
Completeness Proof of the Matroid Polytope Martin Grötschel 15
Completeness Proof of the Matroid Polytope (continued) Martin Grötschel 16 The proof above is from (GLS, pages ), see
Facets of the matroid polytope It will also be shown on the blackboard that a rank inequality x(F) ≤ r(F) defines a facet of the matroid polytope if and only if the set F is closed and inseparable, see Martin Grötschel, Facetten von Matroid-Polytopen, Operations Research Verfahren XXV, 1977, , downloadable from A subset F of E is closed if r(F U {e})>r(F) for all e in E\F. A subset F of E is separable if there exist two nonempty disjoint subsets F 1 and F 2 of F whose union is F and such that r(F)= r(F 1 )+ r(F 2 ). Martin Grötschel 17
The Forest Polytope Martin Grötschel 18
A Partition Matroid Polytope Martin Grötschel 19
The 1-Tree Polytope 1-trees come up as relaxations of the symmetric travelling salesman problem. Given a complete graph K n =(V,E), V={1,2,…,n}. A 1-tree is the union of the edge set of a spanning tree of the complete graph on the node set {2,3,…,n} and two edges with endnode 1. Every 1-tree has n edges and contains exactly one cycle. This means that every travelling salesman tour is a 1-tree. The set of 1-trees is the set of bases of a matroid on E. A complete description of the convex hull of the incidence vectors of all 1-trees in K n is given by: 0≤ x e ≤ 1 for all e in E x(E(W)) ≤ |W| - 1 for all node sets W in V not containing 1 x(δ(1)) = 2 x(E) = n Martin Grötschel 20
The Branching and the Arborescence Polytope Martin Grötschel 21
The Branching and the Arborescence Polytope Martin Grötschel 22
The Matroid Intersection Polytope Martin Grötschel 23
The Matroid Intersection Polytope Martin Grötschel 24
The Matroid Intersection Polytope Martin Grötschel 25
The Matroid Intersection Polytope Martin Grötschel 26 The proof above is from (GLS, pages ), see
Submodular Functions and Polymatroids The whole polyhedral and algorithmic theory developed so far can be generalized to submodular functions and polymatroids. This is worked out in detail in Chapter 10 of (GLS), see Martin Grötschel 27
Claude Berge (perfect graphs) and Jack Edmonds (matching and matroids) Martin Grötschel 28
Martin Grötschel Institute of Mathematics, Technische Universität Berlin (TUB) DFG-Research Center “Mathematics for key technologies” (M ATHEON ) Konrad-Zuse-Zentrum für Informationstechnik Berlin (ZIB) Independence Systems, Matroids, the Greedy Algorithm and related Polyhedra Martin Grötschel Summary of Chapter 4 of the class Polyhedral Combinatorics (ADM III) May 18, 2010 The End