ae . Jee AA 527 / _ Appl. Math, Lett. Vol. 1, No. t, pp. 45-48, 1988 0893-9659/88 $3.00 + 0.00 Printed in the U.S.A. All rights reserved, Copyright (c) 1988 Pergamon Journals Ltd 1 2 3 ad Ss | 2 4 292 series ASoa The Number of Perfect Matchings in a Hypercube Niall Graham and Frank Harary Computing Research Laboratory New Mexico State University Las Cruces, NM 88003 ABSTRACT A perfect matching or a 1-factor of a graph G is a spanning subgraph that is regular of degree one. Hence a perfect matching is a set of independent edges which matches all the nodes of G in pairs. Thus in a hypercube parallel processor, the number of perfect matchings evaluates the number of different ways that all the processors can pairwise exchange information in parallel. Making use of matrices and their permanents one can write a straightforward formula which we evaluate forn <5. A perfect matching or a 1-factor of a graph G is a regular spanning subgraph f degree one. In other words a perfect matching is a set of independent edges in E(G) that spans V(G). Define f,{G) as the number of 1-factors of G. The bipartite adjacency matrix (ba-matraz) B = B(G) of a bipartite graph G=(V,E) where V = UUW, |U| =m, |W|=n, is the mXn matrix that indicates the presence or absence of an edge between each (u,w) pair of nodes by a one or zero, respectively. The following theorem was found independently by both Fisher [1] and Kasteleyn [3]. Theorem A. The number f;(G) of perfect matchings in an n Xn bipartite graph G is perB, the permanent of the ba-adjacency matrix B of graph G. {(G) = perB(G) (1) Proof. The permanent of a square binary matrix is simply the number of ways of choosing exactly one 1 from each row and each column. Hence, there exists a one-to-one correspondence between perfect matchings and the unit contribu- tions to this permanent. [] The hypercube Q, may be recursively defined [2,p.23] in terms of carte- sian product: © Ke n=l Qn = > 2 (2) Qn-1XK_g 2 Using this definition, the ba-matrix B, of a hypercube Q, may be conveniently written recursively, with I denoting the identity matrix of order 2”~?; 4546 N, GRAHAM and F. HARARY B=(1), Baa = E A (3) By Theorem A the value of perB, is the number of perfect matchings in Q,. A submatrix X of a matrix A is the matrix formed by choosing a subset of the rows of A and a subset of the columns of A. It is convenient to give an expression for counting the perfect matchings of Q, 41. Theorem 1. The number of perfect matchings of Q,,, is given by Si(@Qnai1) = perByy = ¥ (perX)?. (4) XCB, Proof. Consider any k Xk matrix X in the upper left B, in B,,, as in (1) such that perX 0. The permanent of X is the number of perfect matchings in the subgraph of Q, induced by the nodes corresponding to the rows and columns of X. Call this induced subgraph Gy. Now match all nodes of V(Q,) — V(Gy) with their neighbors in the other copy of Q,. Clearly, the unmatched nodes in the second copy of Q, induce a graph isomorphic to Gy and its permanent is perX. Thus, for each square submatrix X in B, there are (perX)? perfect matchings in B, 4). [] 1110 Forexample B, = [23] soby(1), Bg = torn and O111 =. 2 11 |? fi(Q3) = perBy = 1? + 4per[1]? + per i1i1| =9 Note that the empty matrix with unit permanent is an admissible submatrix of B, and contributes 1 to the above sum in the 1? term. It is convenient to introduce some additional notation. For any square matrix A, let [t] = 3 (verx) (5) XCA where the summation is over all k xk submatrices X of A. Thus (1) may be rewritten f(Qnn) = erByy = *| (6) k=0 B B ¢ Obviously, [>| = 1 and || = (perB,)? = f,°(Q,). The number IDThe number of perfect matchings in a hypercube 47 © B. ones in B, is | Jae is the number of edges in Q,, so | = n2"!, B To derive a closed form for . it is convenient to identify all dissimilar pairs of columns of B,. These correspond to all dissimilar pairs of nodes in Q,. Any pair of nodes may be completely specified by their distance because of symmetry, and the number of pairs at distance 2k is 2P Yeg) = 2-(4) | 1 Jlee 2k When k = 1 any two nodes at distance 2 are mutually adjacent to exactly two nodes and are each individually adjacent to n-1 other nodes. Therefore, any pair of columns in B, corresponding to nodes at distance 2 consist of two rows of the form 1 1 and n—1 rows of the form 10 and n-1 rows of the form 0 1 and the other 2""!~ 2n rows 00. Thus, the sum of the squares of all permanents formed by selecting two nodes at distance 2 is } if? 11/7 11/7 10/7 ©,.,[} 1] + 2(n-2)per| } ‘| + 2(n-2)per[ § 1] + (n-2)?per 01| = n? Any two nodes at distance greater than two obviously have disjoint neighbor- hoods and their corresponding pair of columns in B, contain n copies of 10 and of 01 in their rows, the other rows consisting of 0 0 entries. Thus the per- manents of the 2X2 matrices formed in these columns contribute a factor of n*, giving Jeon eo Sy) = 2" 1, hence *| = n22"-2(9%-1_ 4) k=0 1110 #1000 1101 0100 1011 0010 0111 0001 To compute perB, = per 1000 1110 0100 1101 © 0010 1011 " 0001 O11148 N. GRAHAM and F. HARARY - Bs| . or we first evaluate 3 Since all sets of three nodes of even weight in Q3 are similar, any three columns of Bz may be chosen. Within any three columns there are two dissimilar (with respect to the automorphism group of Q3) 3x3 submatrices giving Bs 4 1107 1107 3 = (3) per}101 | +3perl101 | $=124, so O11 111 i=4| B f(Q4) = : = 143.2? + 3°.2(2?-1) + 12449? = 272 1=0 ‘We conclude by only mentioning the result that fi(Qs) = 589,185 which was similarly calculated by a computer program. An even more difficult unsolved problem in graphical enumeration is the exact determination of the number fi*( Q,) of equivalence classes of perfect matchings in hypercube Q, with respect to its automorphism group ['(Q,). It as been shown that [(Q,) = [Ss] Sn the exponentiation group [2,p.177] oft the two symmetric groups Sy raised to the power S,. It is also known that, in principle, the number fi'(Q,) of these similarity classes can be calculated from he group of the graph Q, with respect to the group of the subgraph 2"-1K, which is a perfect matching of Q,). But this approach has not yet proved help- ul. Obviously f,"(Q@:) = :*(Q@2) = 1, f,°(@s3) = 2 and we have also ound that /,*(Q4) = 8. References 1] M.E. Fisher, Statistical mechanics of dimers on a plane lattice Phys. Rev. 124 (1961) 278-286. 2) F.Harary, Graph Theory. Addison-Wesley, Reading (1969). 3] P.W. Kasteleyn, The statistics of dimers on a lattice I: The number of dimer arrangements on a quadratic lattice. Physica 27 (1961) 1209-1225.