A3Q5 In this question you will use linear programming to prove the *mex-flow min-cut theorem* which states that the maximum flow from s tot in a grap is equal to the minimum size of a cut thet disconnects s from t. This question is reasonably challenging and there's no shame in skipoing it However, its an opportunity to use nearly everything we have seen so far in the course to show how linear programming duality generalizes an important resuit in graph theory. For full credit you must provide mathematically rigorous proofs of all the parts. Recall that given a directed graph G = (V U (s, f), E) with a designated source vertex s and designated target vertex f, the maximum flow is defined to be the maximum flow that can be routed through the graph from s to t (subject to edge capacity constraints and the restriction that at every vertex w € V, the out-flow must be equal to the in-flow). The vertex-arc incidence matrix of G is a (1, 0}-matric M with one row for each vertex u € V and one column for each arc (v, w) € E, The entries of the matrix are such that M[u, vw] := —Lifu =v, M{u, vw] == 1iftu= wand M[u, vw] := 0 otherwise Recall that the max-flow from s to tin G (assuming unit capacities on the edges) is given by the linear program max{ DY xa |Mx=0,x<1,x20) (P) wR tee where I denotes the vector with all ones as entries. Ans ~ t cut of Gis a partition of V U (5,1) into sets S, T such that s € S, 1 € T and so that removing all edges in E that go from a vertex in 5 to a vertex in T will disconnects from tin the graph G (je. there is no path from s tot after removing these edges). Recall the minimum s ~ ¢ cut of Gis defined tobe the smallest size of an 5 ~ fcut in G (equivalently itis the minimum number of edges the need to be removed to disconnects from 0. {@) Prove that the optimal value of (P) is equal to. Xow st @) WWW WENEB IHS Ft Yj B1— uy, MS, NEE Jn 2, MEN EE veo min wR” eRe ‘Assume above that there's no edge from stot in E (if there is such an edge itis easy to show that this changes the value by exactly 1). To prove this, you must use as a blackbox the following version of strong LP duality: for all mneN,AER™, DER" cER", max{(c, x) | Ax max 6, x20) = min((b,y)| ATy Ze) (ovhen both optima exis. (b) Extend the inductive proof of proposition 4.3 in chapter- (which was also shown in lecture) to conclude that the following matrix is totally unimodular (TU) . wef oo. {c) From part (b) we obtain that the feasibility region of (0) (i.e. a set of the form (M’z > e') for some integer vector c’) is a polyhedron al of whose extreme points are integral and thus there is an optimal solution to (D) that is integral, Use this integrality fact to conclude that (D) equals the minimum size of a cut. Hint: use the S, T partition definition of a cut given above and think about how to define such a partition using an integer vector u satisfying the constraints in (D) (and vice-versa for the easier direction).‘Theorem (Veinott and Dantzig): An integral matrix A is unimodular if and only i the polyhedron {x | Ax = b, x > 0} isintegral for every integral vector b. ‘Theorem (Hoffman and Kruskal): An integral matrix Ais totally unimodular if and oniy if the polyhedron {x | Ax 0} is integral (.e. all extreme points are integer vectors). (6) Prove Hoffman-Kruskal theorem using (a), (b) and the Veinott-Dantzig theorem. One of the key advantages of the formalism of linear programming (and convex programming more generally) is that it provides 2 unified framework with which to study a diverse lst of problems. inthe previous assignment we proved (in Q5) that the max-flow min-cut theorem can be derived as a consequence of linear programming duality. Several important theorems can be proved using linear programming duality. In this manner, it provides a Unified treatment ofa ist of important theorems that were originally proved using ad-hoc case-specitic methods. see another such example by proving that the maximum size matching in a bipartite Inthis question you ‘graph equals the minimum size vertex cover. {@) Given an undirected graph G, we define the (undirected) incidence matrix M to be the matrix with a row for each vertex and a column for each edge such that forall u € V ande € E, the entry M[v, e] is equal to 1if vertex v and edge e are incident, and 0 otherwise. Use the Ghoulla-Houri Theorem to prove that the incidence ‘matrix of a graph is TU if and only ifthe graph is bipartite (Le. its vertex set can be partitioned into two sets such that every edge in the graph must go from one set to the other) (b) Let 1 denote the vector with all ones as entries. By strong LP duality and the Hoffman-Kruskal Theorem we have the following result for any TU matrix M: max {(1,x)| Mx <1, x20} = (1,y) | MTy 21, y2 0} min { ster integra Let G=(V/£) be an undirected bipartite graph. A matching is a subset E’ CE of the edges such that no two edges share a vertex. A vertex covers a subset V’ C V of the vertices such that every edge in E has an endpoint in V’, Use the above duality result to prove that the maximum size of a matching in a bipartite graph equals the minimum size of a vertex cover. Exercise 5.2. In section 5.3, we discussed how to analyze the impact of introducing a new variable to a linear program. Suppose instead, we are no longer able to produce a product and we remove a variable from the LP. How can we determine if an optimal basis for the original problem remains ‘optimal for the new problem? (Consider the two cases of when the removed variable is basic or non-basic). max(—xne1 | 2G + x2 +o tan) +m =m m1 € (0,1) VEE [n+ LI} (a) What is the optimal solution when n is even and what is it when n is odd? (b) Show that when n is odd, any branch-and-bound algorithm (i.e. no matter what order the branch and bound solves LPs or chooses fractional variables) will need to solve exponentially (i.e. at least 2" for some universal constant C > 1) many LP subproblems when n is odd, in order to find the optimal solution.