Kernel type results can be proved on other structures than posets. LetM1 = (E,I1) and M2 = (E,I2) be matroids and let ≤1 and ≤2 be a linear order on their common ground set E. Common independent set K of these matroids are called an M1M2-kernel if for any elemente∈E\K there exist a cycleC of matroidMi for somei∈ {1,2}such that C ⊆K∪ {e}and c≤i e holds for each c∈C−e.
Theorem 2.9 (Fleiner [20]). For any matroids M1 = (E,I1) and M2 = (E,I2) and for any linear orders ≤1 and ≤2 on E, there exists an M1M2-kernel. If K1, K2 ⊆ E are M1M2-kernels then the greedy algorithm on Mi for processing order ≤i selects an M1M2-kernel from K1 ∪K2 for any i = 1,2. These two operations determine a lattice on M1M2-kernels with property 1.4. Moreover, spanM1K1 = spanM1K2 and spanM2K1 =spanM2K2 holds for any M1M2-kernels K1 and K2.
The keys to Theorem 2.9 are Corollary 1.17 and the fact that the choice function defined by the greedy algorithm is substitutable and increasing.
Proof. Define choice functions F and G on E as follows. For subset X of E let x1 <1 x2 <1 x3 <1 . . . be the ≤1 order of the elements of X and let F0(X) = ∅ and for
CHAPTER 2. KERNEL TYPE RESULTS 23 i= 1,2, . . . define
Fi(X) =
Fi−1(X)∪ {xi} if Fi−1(X)∪ {xi} ∈ I1
Fi−1(X) if Fi−1(X)∪ {xi} 6∈ I1 ,
and let F(X) := F|X|(X). That is, Fi(X) is the basis of X that the greedy algorithm selects from X in matroid M1 for order ≤1. Define choice function G similarly for matroid M2 and order ≤2. Clearly,F and G are choice functions and |F(X)|=rk1(X) and |G(X)| = rk2(X), so by the monotone property of the rank function, both F and G are increasing. A standard property of matroids is that choice function F defined by F(X) = X \ F(X) as the nonselected elements of F is monotone. Hence DF(X) :=
E \ F(X) = F(X)∪(E\X) is an antitone determinant of F, showing that F (and G for a similar reason) is an increasing substitutable choice function.
To show thatM1M2-kernels are exactly F G-kernels, assume first that K is an F G-kernel, i.e. K = X ∩Y where X = DG(Y) and Y = DF(X). Now K is a common independent set of M1 and M2 asF(X) =X∩ DF(X) = X∩Y =DG(Y)∩Y =G(Y).
Let e∈E\K and assume no cycle C ⊆K∪ {e} of M1 exist such that c≤1 e holds for each element c of C−e. This means e 6∈ X as otherwise e ∈ F(X) = K would hold.
However, then e∈ DF(X) =Y ande6∈K =G(Y), hence the greedy algorithm does not pick e from Y, i.e. there must exist some cycle C0 if M2 in Y such that C0 ⊂ K ∪ {e}
and c≤2 e holds for each element cof C0−e. Consequently,K is an M1M2-kernel.
Assume now thatK is an M1M2-kernel and define
X :=K∪ {e∈E :e6∈ F(K∪ {e}} and Y :=DF(X) .
Set K is independent in M1 hence K = F(X) = X ∩Y holds. So Y = DF(X) = K ∪(E \X) = {e ∈ E : e ∈ F(K ∪ {e}}. As K is an M1M2-kernel, for any element e of Y \K there exists a cycle C ⊂ K ∪ {e} of M2 such that c ≤2 e holds for any c∈C. Consequently,G(Y) =K and hence DG(Y) = E\(X)∪K =X, that is,K is an F G-kernel.
Hence the existence ofM1M2-kernels and the structural properties ofM1M2-kernels directly follow from Corollary 1.17 and the definition of choice functions F and G.
It is interesting to see that Theorem 2.9 generalizes the following well-known result.
Theorem 2.10. If graph G is bipartite then dM ≡ dM0 holds for any two stable b-matching M and M0. Moreover,M(v) = M0(v) whenever |M(v)|< b(v) holds.
Note that the famous Rural Hospitals Theorem of Roth [57] is the special case of the above Theorem 2.10 where G has bipartiton (A, B) and b(a) = 1 holds for each vertex a ∈A.
Proof. Stableb-matchings ofGare exactlyM1M2-kernels whereM1 andM2 are parti-tion matroids on E (the partitions are given by the stars of the two color classes), linear orders ≤1 and ≤2 are compatible with the preferences of the vertices, and b determines the rank of the parts in the partiton. If M and M0 are stable b-matchings of G then spanM1(M) = spanM1(M0) and spanM2(M) = spanM2(M0) holds by Theorem 2.9 and this directly implies Theorem 2.10.
Chapter 3
The structure of kernels
In this chapter, we study the structure of various F G-kernels. We shall see that the lattice structure of kernels (due to Corollary 1.17) allows us to prove various interesting structural results.
3.1 Uncrossing of kernels and median kernels
First, we deduce as a corollary of Corollary 1.17 that F G-kernels can be efficiently un-crossed.
Theorem 3.1 (Fleiner [20]). Let F,G : 2E →2E be w-increasing substitutable choice functions for some modular and strictly monotone functionw:E →R+and letK1, K2, . . . , Km by the second part of Corollary 1.17. This means that if we replace Ki and Kj by F G-kernelsKi∧FKj andKi∨FKj, then this does not change Pm
i=1χ(Ki), while it increases Pm
i=1|Ki|2. So at some point no such replacement is possible any more and hence we have a chain of F G-kernels as required by the theorem. It follows from Pm
i=1χ(Ki) =
In case of stable b-matchings, Theorem3.1 has an especially simple form that relies on the following generalization by Ba¨ıou and Balinski of the Comparability Theorem [59]
by Roth and Sotomayor.
CHAPTER 3. THE STRUCTURE OF KERNELS 25 Theorem 3.2 (Ba¨ıou and Balinski [7]). Let G= (V, E) be a bipartite graph and let v be a linear order on the set E(v) of edges incident to vertex v and let b : V → N. If S1 and S2 are stable matchings and v ∈V then one of the two properties below must hold.
• S1(v) = S2(v) or
• |S1(v)|=|S2(v)|=b(v) and the b(v) v-best element of set S1(v)∪S2(v) is either S1(v) or S2(v).
It is not difficult to prove Theorem 3.2 with elementary tools, i.e. by studying the domination in the symmetric difference of two stable b-matchings. A corollary of Theo-rem 3.2 is that for any vertex v of G there is a linear order on those subsets of edges of E(v) that a stableb-matching can contain. As it (hopefully) does not cause ambiguity’s, this linear order is denoted also by v. The theorem below in particular implies that if each agent in one part of the bipartite graph select the ith best assignment out of k given stable b-matchings then these choices result in another stable b-matching.
Theorem 3.3 (Fleiner [19]). Let G= (V, E)be a bipartite graph with partsA andB and let v be a linear order on the setE(v)of edges incident withv for each vertexv and let b :V →N. Then for any stable b-matchings S1, S2, . . . , Sk, and for any 1≤i, j ≤k, sets SAi := S
{Si(v) : v ∈ A} and SBj := S
{Si(v) : v ∈ B} are stable b-matchings with the property that SAi =SBk+1−i holds for 1≤i≤k where S1(v), S2(v), . . . , Sk(v) denotes edge sets S1(v), S2(v), . . . , Sk(v) in v-order.
Proof. For vertexv ofGand stable matchingsSand S0 let S v S0 denote thatv prefers S to S0. According to Theorem 3.2, we can listS1, S2, . . . , Sn as Sv1 v Sv2 v . . .v Svn for each vertex v. Observe that SAi = W
a∈A
Vi
`=1Sa` and SBj = V
b∈B
Wj
`=1Sb` where we use the choice function of A for operation ∧ and operation ∨ is calculated with the choice function of B. Chains SA1, SA2, . . . , SAk and SB1, SB2, . . . , SBk are opposite, hence SAi =SBk+1−i .
Theorem 3.3 above has been proved for stable matchings by Teo and Sethuraman with the help of linear programming tools [66]. Later, not being aware of Theorem 3.3, Klaus and Klijn gave a very similar short proof for a special case [50].