Massachusetts Institute of Technology Michel X. Goemans 18.433: Combinatorial Optimization Spring 2015 Solutions to problem set 3 3-8 We know that every extreme point of a polytope is the unique solution for a system of equations obtained by selecting some of the inequalities defining the polytope and setting them as equalities. In particular, if x∗ is a vertex of Q, then there exists a subset I of {1, . . . , m}, where m is the number of rows of A, such that x∗ is the unique solution of the system (Ax)i = bi , for all i ∈ I Cx = d. But this implies that x∗ is the unique solution for a system of equations obtained by selecting some of the inequalities that define P and setting them as equalities. Since x∗ ∈ P , it follows that x∗ is an extreme point of P . Therefore, all the vertices of Q are vertices of P . 3-9 First consider the situation in which M1 and M2 are such that M1 ∆M2 have more than one connected component. Consider one of these connected components, say S ⊆ V , and partition M1 and M2 into M1 = M1s ∪ M1t and M2 = M2s ∪ M2t where M1s and M2s correspond to the edges within S. By definition M1s ∪ M2s 6= ∅. Now define two other matchings by M3 = M1s ∪ M2t and M4 = M2s ∪ M1t . Observe that χ(M1 ) + χ(M2 ) = χ(M3 ) + χ(M4 ) which implies that any face that contains M1 and M2 will also contain M3 and M4 , and thus cannot be an edge. Conversely, suppose that M1 ∆M2 has only one connected component, and say that this component has k1 edges from M1 and k2 edges from M2 . We must have that |k1 − k2 | ≤ 1. Now consider the following cost function: 1 e ∈ M1 ∩ M2 −1 e ∈ / (M1 ∪ M2 ) ce = k2 e ∈ M1 \ M2 k1 e ∈ M2 \ M1 . Notice that c(M1 ) = c(M2 ) = b where b := |M1 ∩ M2 | + 2k1 k2 and for any other matching M we have that c(M ) < b. Thus the valid inequality cT x ≤ b induces a face with only the incidence vectors of M1 and M2 has vertices. Thus the line segment between M1 and M2 defines an edge. 3-11 Consider the family of permutations {σ1 , . . . , σn } given by: σi (1) = i, σi : σi (i) = 1, σi (j) = j, j ∈ / {1, i}. Solutions to problem set 3 Spring 2015 2 In other words, σ1 is the identity permutation and σi is the permutation that transpose 1 and i. Let us prove that these n permutations are affinely independent. For that suppose that λ1 , . . . , λn are real numbers such that: n X n X λi σi = 0, i=1 λi = 0. i=1 Then for every j 6= 1: 0= n X ! λi σi (j) = X i=1 λi · j + λj · 1 i6=j = n X ! λi · j − λj · j + λj · 1 i=1 = X n λi · j + λj · (1 − j). i=1 | {z } =0 Hence, for every j 6= 1, λj (1 − j) = 0, which implies that λj = 0. Using that Pn i=1 λi = 0 we deduce that λ1 is also 0. Therefore, the family {σ1 , . . . , σn } is a set of n affinely independent permutations. P P 3-13 (a) Fix a ∈ {−1, 0, 1}n satisfying ni=1 ai = 1 and 0 ≤ ji=1 ai ≤ 1 for each j = 1, . . . , n − 1. It is enough to show that −1 ≤ n X ai (ek )i ≤ 1 i=1 for each k = 0, . . . , n − 1 (it is symmetric for −ek ’s). Note that (ek )i = 1 if i ≤ k and (ek )i = −1 if i > k. We have n X i=1 Since Pk i=1 ai (ek )i = k X i=1 ai − n X ai = 2 i=k+1 k X ai − 1. i=1 ai is 0 or 1, it is between −1 and 1. P (b) Fix a ∈ {−1, 0, 1}n as in the previous part. Let bj = ji=1 ai for j = 1, . . . , n. Then, bj ∈ {0, 1} for any j = 1, . . . , n − 1 and bn = 1 by definition of a. On the other hand, if we are given b ∈ {0, 1}n with bn = 1, we can find the corresponding a ∈ {−1, 0, 1}n by letting a1 = b1 and ai = bi − bi−1 for i = 2, . . . , n. This is a bijection between a’s and b’s. Hence, there are 2n−1 such a’s and 2n inequalities. Solutions to problem set 3 Spring 2015 3 P (c) First note that aT ek is either −1 or 1, since aT ek = 2 ki=1 ai − 1. Let b as defined in (b). Then, aT ek = 1 if and only if bk = 1 (we say b0 = 0). Thus, {x ∈ S | aT x = 1} = {ek | bk = 1} ∪ {−ek | bk = 0} {x ∈ S | aT x = −1} = {ek | bk = 0} ∪ {−ek | bk = 1}. So each inequality defines distinct hyperplanes, because they contain different set of extreme points. Moreover, if we choose exactly one vector from each {ek , −ek }, then they are affinely independent. For, note that it is enough to show that {e0 , . . . , en−1 } are linearly independent, and they are indeed linearly independent since { 12 (e1 − e0 ), 12 (e2 − e1 ), . . . , 21 (en−1 − en−2 ), − 21 (en−1 + e0 )} is the standard basis of Rn . (d) Note that 0 is in the interior of conv(S). Hence, no facet can contain {ek , −ek } for any k = 0, . . . , n − 1 (otherwise it will contain 0). Since conv(S) is fulldimensional, any facet should contain at least n extreme points, i.e., it contains exactly one from each {ek , −ek }. So there are at most 2n facets of conv(S). On the other hand, we showed in (c) that each of 2n inequalities defines distinct facet. Hence they define conv(S).