Chapter 4 Families of Partitions

4.1 A partial order on partitions

We have seen that the group-divisible association scheme is dened by a partition of into blocks of equal size. A particularly tidy basis for the Bose-Mesner algebra A in this case is {I, B, J}, where B is the adjacency matrix for the relation is in the same block as. The main object of this chapter is to give a single general construction for association schemes dened by families of partitions of . Let F be a partition of into nF subsets, which I shall call F-classes. (I use the letter F because statisticians call partitions factors.) Dene the relation matrix RF in R by RF (, ) = 1 0 if and are in the same F-class otherwise, (4.1)

just as B was dened in Section 3.1. Then, as in Section 3.4, dene the subspace VF of R by VF = v R : v() = v() whenever and are in the same F-class . Then dimVF = nF . Let PF be the orthogonal projector onto VF . Then (PF v)() is equal to the average of the values v() for in the same F-class as . Denition The partition F is uniform if all classes of F have the same size. (Unfortunately, there is no consensus about what to call such a partition. The words uniform, regular, balanced and proper are all in use.)
1 Proposition 4.1 If F is uniform with classes of size kF then PF = kF RF .

79

80

CHAPTER 4. FAMILIES OF PARTITIONS

So much for a single partition. Now we consider what happens when we have two partitions. If F and G are both partitions of , write F G if every F-class is contained in a G-class; write F G if F G and F = G. We may pronounce F G as F is ner than G or G is coarser than F. (Statisticians often say F is nested in G.) Example 4.1 In the association scheme 3/(2 4) the set consists of 24 elements, divided into three rectangles as follows.

The set is partitioned in three waysinto rows, columns and rectangles. Identifying the names of the partitions with the names of their classes, we have rows rectangles and columns rectangles. Lemma 4.2 If F G then VG VF .

Proof Any function which is constant on G-classes must be constant on Fclasses too if each F-class is contained in a G-class. There are two special, but trivial, partitions on every set with more than one element. The universal partition U consists of a single class containing the whole of (the universe). At the other extreme, the equality partition E has as its classes all the singleton subsets of ; in other words, and are in the same E-class if and only if = . Note that E F U for all partitions F. The relation satises: (i) (reexivity) for every partition F, F (ii) (anti-symmetry) if F (iii) (transitivity) if F G and G F; F then F = G; H.

G and G

H then F

This means that is a partial order. Partial orders are often shown on Hasse diagrams: there is a dot for each element (partition in this case); if F G then F is drawn below G and is joined to G by a line or sequence of lines, all going generally upwards. Example 4.1 revisited The Hasse diagram for this example is in Figure 4.1.

4.1. A PARTIAL ORDER ON PARTITIONS U

v

81 U

v rectangles d d d dv columns rows v d d d dv

v d d d v vrowsdv columns letters d d d dv

E Figure 4.1: Hasse diagram for Example 4.1 Figure 4.2: Hasse diagram for Example 4.2

Example 4.2 The set of n2 cells in a n n Latin square has three non-trivial partitionsinto rows, columns and letters. The Hasse diagram in shown in Figure 4.2. A set with a partial order on it is often called a poset. Every poset has a zeta function and a M bius function, which I shall now dene for our particular partial o order. Let F be any set of partitions of . Dene in RF F by (F, G) = 1 if F G 0 otherwise.

In fact, the partial order is often considered to be the subset {(F, G) : F G} of F F . From this viewpoint, is just the adjacency matrix of . The elements of F can be written in an order such that F comes before G if F G. Then is an upper triangular matrix with 1s on the diagonal, so it has an inverse matrix which is also upper triangular with integer entries and 1s on the diagonal. The matrix is called the M bius function of the poset (F , ). o Example 4.1 revisited Here E rows columns rectangles U E 1 1 1 1 1 rows 0 1 0 1 1 columns 0 0 1 1 1 rectangles 0 0 0 1 1 U 0 0 0 0 1

82 so

CHAPTER 4. FAMILIES OF PARTITIONS

E rows columns rectangles U E 1 1 1 1 0 1 0 1 0 rows 0 columns 0 0 1 1 0 0 0 1 1 rectangles 0 U 0 0 0 0 1

Example 4.2 revisited For the Latin square, E rows columns letters U E rows columns letters U 1 1 1 1 1 0 1 0 0 1 0 0 1 0 1 0 0 0 1 1 0 0 0 0 1

and so E rows columns letters U E 1 1 1 1 2 0 rows 1 0 0 1 0 columns 0 1 0 1 letters 0 0 0 1 1 U 0 0 0 0 1

If F and G are partitions, let F G be the partition whose classes are the non-empty intersections of F-classes with G-classes. Then (i) F G (ii) if H F and F G F and H G, and F G,

G then H

so F G is the inmum (or greatest lower bound) of F and G. Dually, the supremum (or least upper bound) F G of F and G is dened by (i) F (ii) if F F G and G H and G F G, and H then F G H.

To construct F G, draw a coloured graph whose vertex set is . There is a red edge between and if and are in the same F-class, and a blue edge if and are in the same G-class. Then the classes of F G are the connected components of the red-and-blue graph.

4.2. ORTHOGONAL PARTITIONS

83

Example 4.1 revisited In the example with rectangular arrays, we have rows columns = E and rows columns = rectangles. Example 4.2 revisited In the Latin square, the supremum of any two of the nontrivial partitions is U and the inmum of any two is E. Of course, if F G then F G = F and F G = G. Moreover, just as the empty sum is 0 and the empty product is 1, we make the convention that the empty inmum is U and the empty supremum is E. Lemma 4.3 If F and G are partitions of , then VF VG = VFG . Proof Let v be a vector in R . Then v VF VG v is constant on each component of the blue graph and v is constant on each component of the red graph v is constant on each component of the redand-blue graph v VFG .

4.2

Orthogonal partitions

Denition Let F and G be partitions of . Then F is orthogonal to G if VF is geometrically orthogonal to VG , that is, if PF PG = PG PF . Note that this implies that F is orthogonal to G if F PG PF = PG in that case. Lemma 4.4 If F is orthogonal to G, then (i) PF PG = PFG ; (ii) VF (VFG ) is orthogonal to VG . Proof These both follow immediately from Lemma 2.2, using the fact that VF VG = VFG . Corollary 4.5 Partitions F and G are orthogonal to each other if and only if, (i) within each class of F G, each F-class meets every G-class, and G, because PF PG =

84 (ii) for each element of ,

CHAPTER 4. FAMILIES OF PARTITIONS

|F G-class containing | |F-class containing | = . |F G-class containing | |G-class containing | Corollary 4.6 If F is orthogonal to G and if F, G and F G are all uniform then F G is also uniform and kFG kFG = kF kG . Example 4.1 revisited Here rows are orthogonal to columns even though no row meets every column. Example 4.2 revisited As in Example 1.6, write R, C and L for Rrows , Rcolumns and Rletters . Proposition 4.1 shows that Prows = n1 R, Pcolumns = n1C, Pletters = n1 L and PU = n2 J. We saw in Example 1.6 that RC = CR = RL = LR = CL = LR = J. Thus the partitions into rows, columns and letters are pairwise orthogonal. Theorem 4.7 Let F be a set of pairwise orthogonal partitions of which is closed under . For F in F , put WF = VF Then (i) the spaces WF and WG are orthogonal to each other whenever F and G are different partitions in F ; (ii) for each F in F , VF =
G F

G F

VG

WG .

Proof

(i) If F = G then F G must be different from at least one of F and G. Suppose that F G = F. Then F G F and F G F so WF
VF VFG ,

while WG VG . Lemma 4.4 (ii) shows that VF VFG is orthogonal to VG , because F is orthogonal to G. Hence WF is orthogonal to WG .

4.2. ORTHOGONAL PARTITIONS

85

(ii) Since the spaces WG , for G F, are pairwise orthogonal, their vector space sum is direct, so it sufces to prove that VF =

G F

WG.

(4.2)

We do this by induction. If there is no G such that G F then WF = VF and Equation (4.2) holds. The denition of WF shows that VF = WF +
H F

VH .
F then

If the inductive hypothesis is true for every H with H

H F

VH = WG.
H FG H

If G

F then G

F, and if G

F then G

F so

H FG H

WG = WG.
G F

Thus VF = WF +

G F

WG = WG.
G F