Using graphs to find the best block designs

R. A. Bailey and Peter J. Cameron

arXiv:1111.3768v1 [math.ST] 16 Nov 2011

R. A. Bailey obtained a DPhil in group theory from the University of

Oxford. She worked at the Open University, and then held a post-doctoral
research fellowship in Statistics at the University of Edinburgh. This was fol-
lowed by ten years in the Statistics Department at Rothamsted Experimental
Station, which at that time came under the auspices of the Agriculture and
Food Research Council. She returned to university life as Professor of Mathe-
matical Sciences at Goldsmiths’ College, University of London, and has been
Professor of Statistics at Queen Mary, University of London, since 1994.
Peter J. Cameron is Professor of Mathematics at Queen Mary, University
of London, where he has been since 1986, following a position as tutorial
fellow at Merton College, Oxford. Since his DPhil in Oxford, he has been
interested in a variety of topics in algebra and combinatorics, especially their
interactions. He has held visiting positions at the University of Michigan,
California Institute of Technology, and the University of Sydney. He is cur-
rently chair of the British Combinatorial Committee.

Contents
1 What makes an incomplete-block design good for experi-
ments? 3

2 Graphs from block designs 5

2.1 The Levi graph . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 The concurrence graph . . . . . . . . . . . . . . . . . . . . . . 6
2.3 The Laplacian matrix of a graph . . . . . . . . . . . . . . . . 8
2.4 Laplacians of the concurrence and Levi graphs . . . . . . . . . 9

3 Statistical issues 10
3.1 Estimation and variance . . . . . . . . . . . . . . . . . . . . . 10
3.2 Optimality criteria . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3 Questions and an example . . . . . . . . . . . . . . . . . . . . 12

1
4 Highly patterned block designs 13
4.1 Balanced incomplete-block designs . . . . . . . . . . . . . . . 13
4.2 Other special designs . . . . . . . . . . . . . . . . . . . . . . . 14

5 Graph concepts linked to D-optimality 15

5.1 Spanning trees of the concurrence graph . . . . . . . . . . . . 15
5.2 Spanning trees of the Levi graph . . . . . . . . . . . . . . . . 16

6 Graph concepts linked to A-optimality 16

6.1 The concurrence graph as an electrical network . . . . . . . . 16
6.2 The Levi graph as an electrical network . . . . . . . . . . . . . 18
6.3 Spanning thickets . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.4 Random walks and electrical networks . . . . . . . . . . . . . 22
6.5 Foster’s formula and generalizations . . . . . . . . . . . . . . . 23
6.6 Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

7 Graph concepts linked to E-optimality 24

7.1 Measures of bottlenecks . . . . . . . . . . . . . . . . . . . . . 24
7.2 Variance balance . . . . . . . . . . . . . . . . . . . . . . . . . 26

8 Some history 26

9 Block size two 28

9.1 Least replication . . . . . . . . . . . . . . . . . . . . . . . . . 28
9.2 One fewer treatment . . . . . . . . . . . . . . . . . . . . . . . 30
9.3 More blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
9.4 A little more history . . . . . . . . . . . . . . . . . . . . . . . 34

10 Very low average replication 34

10.1 Least replication . . . . . . . . . . . . . . . . . . . . . . . . . 35
10.2 One fewer treatment . . . . . . . . . . . . . . . . . . . . . . . 35

11 Further reading 38

Abstract
A statistician designing an experiment wants to get as much infor-
mation as possible from the data gathered. Often this means the most
precise estimate possible (that is, an estimate with minimum possible
variance) of the unknown parameters. If there are several parame-
ters, this can be interpreted in many ways: do we want to minimize
the average variance, or the maximum variance, or the volume of a
confidence region for the parameters?

2
 In the case of block designs, these optimality criteria can be cal-
culated from the concurrence graph of the design, and in many cases
from its Laplacian eigenvalues. The Levi graph can also be used. The
various criteria turn out to be closely connected with other proper-
ties of the graph as a network, such as number of spanning trees,
isoperimetric number, and the sum of the resistances between pairs of
vertices when the graph is regarded as an electrical network.
In this chapter, we discuss the notions of optimality for incomplete-
block designs, explain the graph-theoretic connections, and prove some
old and new results about optimality.

1 What makes an incomplete-block design good

for experiments?
Experiments are designed in many ways: for example, Latin squares, block
designs, split-plot designs. Combinatorialists, on the other hand, have a
much more specialized usage of the term “design”, as we remark later. We
are concerned here with incomplete-block designs, more special than the
statistician’s designs and more general than the mathematician’s.
To a statistician, a block design has two components. There is an under-
lying set of experimental units, partitioned into b blocks of size k. There is a
further set of v treatments, and also a function f from units to treatments,
specifying which treatment is allocated to which experimental unit; that is,
f (ω) is the treatment allocated to experimental unit ω. Thus each block
defines a subset, or maybe a multi-subset, of the treatments.
In a complete-block design, we have k = v and each treatment occurs once
in every block. Here we assume that blocks are incomplete in the sense that
k < v.
We assume that the purpose of the experiment is to find out about the
treatments, and differences between them. The blocks are an unavoidable
nuisance, an inherent feature of the experimental units. In an agricultural
experiment the experimental units may be field plots and the blocks may
be fields or plough-lines; in a clinical trial the experimental units may be
patients and the blocks hospitals; in process engineering the experimental
units may be runs of a machine that is recalibrated each day and the blocks
days. See [5] for further examples.
In all of these situations, the values of b, k and v are given. Given
these values, not all incomplete-block designs are equally good. This chapter
describes some criteria that can be used to choose between them.
For example, Fig. 1 shows two block designs with v = 15, b = 7 and

3
 1 1 2 3 4 5 6 1 1 1 1 1 1 1
2 4 5 6 10 11 12 2 4 6 8 10 12 14
3 7 8 9 13 14 15 3 5 7 9 11 13 15

(a) (b)

Figure 1: Two block designs with v = 15, b = 7 and k = 3

k = 3. We use the convention that the treatments are labelled 1, . . . , v, that

columns represent blocks, and that the order of the entries in each column
is not significant. Where necessary, we use the notation Γj to refer to the
block which is shown as the jth column, for j = 1, . . . , b.
The replication ri of treatment i is defined to be |f −1 (i)|, which is the
number of experimental units to which it is allocated. For the design in
Fig. 1(a), ri ∈ {1, 2} for all i. As we see later, statisticians tend to prefer
designs in which all the replications are as equal as possible. If ri = rj for
1 ≤ i < j ≤ v then the design is equireplicate: then the common value of ri
is usually written as r, and vr = bk.
The design in Fig. 1(b) is a queen-bee design because there is (at least)
one treatment that occurs in every block. Scientists tend to prefer such
designs because they have been taught to compare every treatment to one
distinguished treatment, which may be called a control treatment.

1 1 1 1 2 2 2 1 1 1 1 2 2 2
2 3 3 4 3 3 4 1 3 3 4 3 3 4
3 4 5 5 4 5 5 2 4 5 5 4 5 5

(a) (b)

Figure 2: Two block designs with v = 5, b = 7 and k = 3

Fig. 2 shows two block designs with v = 5, b = 7 and k = 3. The design

in Fig. 2(b) shows a new feature: treatment 1 occurs on two experimental
units in block Γ1 . A block design is binary if f (α) 6= f (ω) whenever α and ω
are experimental units in the same block. The design in Fig. 2(a) is binary.
It seems to be obvious that binary designs must be better than non-binary
ones, but we shall see later that this is not necessarily so. However, if there
is any block on which f is constant, then that block provides no information
about treatments, so we assume from now on that there are no such blocks.

4
 1 2 3 4 5 6 7 1 2 3 4 5 6 7
2 3 4 5 6 7 1 2 3 4 5 6 7 1
4 5 6 7 1 2 3 3 4 5 6 7 1 2

(a) (b)

Figure 3: Two block designs with v = 7, b = 7 and k = 3

Fig. 3 shows two equireplicate binary block designs with v = 7, b = 7

and k = 3. A binary design is balanced if every pair of distinct treatments
occurs together in the same number of blocks. If that number is λ, then
r(k − 1) = (v − 1)λ. Such designs are also called 2-designs or BIBDs. The
design in Fig. 3(a) is balanced with λ = 1; the design in Fig. 3(b) is not
balanced.
Pure mathematicians usually assume that, if they exist, balanced designs
are better than non-balanced ones. (Indeed, many do not call a structure a
‘design’ unless it is balanced.) As we shall show in Section 4.1, this assump-
tion is correct for all the criteria considered here. However, for given values
of v and k, a non-balanced design with a larger value of b may produce more
information than a balanced design with a smaller value of b.

2 Graphs from block designs

2.1 The Levi graph
A simple way of representing a block design is its Levi graph, or incidence
graph, introduced in [40]. This graph has v + b vertices, one for each block
and one for each treatment. There are bk edges, one for each experimental
unit. If experimental unit ω is in block j and f (ω) = i, then the correspond-
ing edge ẽω joins vertices i and j. Thus the graph is bipartite, with one
part consisting of block vertices and the other part consisting of treatment
vertices. Moreover, the graph has multiple edges if the design is not binary.
Fig. 4 gives the Levi graph of the design in Fig. 2(b).
We regard two block designs as the same if one can be obtained from
the other by permuting the experimental units within each block. Since the
vertices of the Levi graph are labelled, there is a bijection between block
designs and their Levi graphs.
Let nij be the number of edges from treatment-vertex i to block-vertex j;
that is, treatment i occurs on nij experimental units in block j. The v × b

5
 Γ1 Γ2 Γ3 Γ4 Γ5 Γ6 Γ7
❍
❍
① P
❍
①
❍PP P
①
PP ✟ ❅❍❍
①
✟
① ✏① ✘✘✘
✟
✏✟ ✟✘
✟
①
❅
❅❅❍❍ ❍❍PP❅❅✟ PP ❅❍
✟ ✟ ✏✏✟ ✘ ✟
❅ ✟❍ ✏✏ ✟ ✘✘ ✟
❅ ❍❍ ❍✟ P
❍✟P❅PP✟✏
P
✟P✏ ✏❍
❅P ✘✘✟
❍
✟ ✘✘ ✟✟
❅ ❍❍ ✘PP❍ ✟
❅
❅ ✟✟ ❍❍ ✟❅
✏
✟✏
✏ P✘✘✟❅
✘✘ P PP
✟ PP❍✟
❅ ✟ ❍ ✟
✏
✘ ✘❍ ✟ ①✟ P ❍P
❅✟
① ✟
❍✘
✏
① ❅✟
❍ ① ❅✟
P ❍①
1 2 3 4 5

Figure 4: The Levi graph of the design in Fig. 2(b)

matrix N whose entries are the nij is the incidence matrix of the block design.
If the rows and columns of N are labelled, we can recover the block design
from its incidence matrix.

2.2 The concurrence graph

In a binary design, the concurrence λij of treatments i and j is ri if i = j
and otherwise is the number of blocks in which i and j both occur. For non-
binary designs we have to count the number of occurrences of the pair {i, j}
in blocks according to multiplicity, so that λij is the (i, j)-entry of Λ, where
Λ = NN⊤ . The matrix Λ is called the concurrence matrix of the design.
The concurrence graph of the design has the treatments as vertices. There
are no loops. If i 6= j, then there are λij edges between vertices i and j. Each
such edge corresponds to a pair {α, ω} of experimental units in the same
block, with f (α) = i and f (ω) = j: we denote this edge by eαω . (This edge
does not join the experimental units α and ω; it joins the treatments applied
to these units.) It follows that the degree di of vertex i is given by
X
di = λij . (1)
j6=i

Figs. 5 and 6 show the concurrence graphs of the designs in Figs. 1 and 2,
respectively.
If k = 2, then the concurrence graph is effectively the same as the block
design. Although the block design cannot be recovered from the concurrence
graph for larger values of k, we shall see in Section 3.2 that the concurrence
graphs contain enough information to decide between two block designs on
any of the usual statistical criteria. They were introduced as variety concur-
rence graphs in [44], but are so useful that they may have been considered
earlier.

6
 8 9
13 7 8 11 ① ① 10
7
① ① ① ① ① ❉ ☎ ①
❜❜ 4 ✧✧ ❜❜ 5 ✧✧ ❏ ❉ ☎ ✡❅
❜✧ ✧
❜ ❅①
◗ ❏ ❉ ☎ ✡ ✑ 11
① ①
✧ ❜ ✧ ❜ 6 ①
✧
① ❜① ✧
① ❜① ◗ ❏ ❉ ☎ ✡ ✑
10 1 ❚❚ ✔✔ 2 14 5 ❵❵❵◗◗❏ ❉ ☎ ✡✑✑✥✥✥①12
①
❵❵◗❵ ✥
✑
☎✡
❏❉❵ ✥✥
✥✥✥
①❵
❚✔①3 ✥ ✥ ✑ ✡ ◗
❏ ❵❵❵
✔ ✥ ✑ 1 ◗ ❵①13
✔✔ ❚ 4 ①
✑ ✡ ❏ ◗
12 ① ✔
✔
① ❚①9 ✑ ✡ ❏ ◗◗①
❚❚ ✔ 6 3 ✑① 14
❅ ✡ ❏
❚①
✔ ❅✡
① ❏①
15 2 15
(a) (b)

Figure 5: The concurrence graphs of the designs in Fig. 1

3 3
✟✟ ✟ ❍
①❍❍ ✟✟❍ ① ❍❍
✟
✟ ✟ ✟✟
✟ ✂❍❇
✂ ❇❍❍❍
❍❍ ✟ ✟✟ ✂❇✂❇❍ ❍❍
✟
✟
✟✟ ✂✂❇❇ ❍ ❍❍
❍① ✟
✟✟
✟ ✂ ✂ ❇ ❇ ❍❍ ❍①
1 ✟✟
◗ ✂ ❇ ❍
✑❍ ✟
◗ ✂ ❇ ❍
✑
✁✁ 2 1 ✁✁ 2
① ①
◗❆ ◗ ✂
❆ ❆◗ ✂ ❇ ✑ ✑ ❆❆❆◗
◗ ◗ ✂ ✂ ❇ ❇ ✑ ✑
◗
❆ ◗ ✂ ❇ ✑ ✁ ❇✑ ✑ ✁ ◗
❆ ◗ ✂ ❇ ✑ ✁ ✑ ✑ ✁
❆❆ ◗ ✂◗
✂ ✑ ◗
✑ ✑❇ ❇ ✁✁ ❆❆ ◗ ✂◗
✂ ✑ ◗
✑✑❇ ❇ ✁✁
❆ ✂ ✑ ◗ ❇ ❇ ✁✁
❆ ✂ ✑ ◗ ✑ ◗ ❆ ✂ ✑ ◗ ❇ ❇ ✁✁
❆ ✂ ✑ ◗ ✑ ◗
❆✑
✂ ✂✑ ❇✁❇①
◗ ❆✂✑ ◗❇✁❇①
4❆ ① ◗ ✁5 4❆ ① ✂✑ ◗ ✁5

(a) (b)

Figure 6: The concurrence graphs of the designs in Fig. 2

   
8 −1 −3 −2 −2 8 −2 −2 −2 −2

 −1 8 −3 −2 −2 


 −2 8 −2 −2 −2 


 −3 −3 10 −2 −2 


 −2 −2 8 −2 −2 

 −2 −2 −2 8 −2   −2 −2 −2 8 −2 
−2 −2 −2 −2 8 −2 −2 −2 −2 8

(a) (b)

Table 1: The Laplacian matrices of the concurrence graphs in Fig. 6

7
2.3 The Laplacian matrix of a graph
Let H be an arbitrary graph with n vertices: it may have multiple edges, but
no loops. The Laplacian matrix L of H is defined to be the square matrix
with rows and columns indexed by the vertices of H whose (i, i)-entry Lii
is the valency of vertex i and whose (i, j)-entry Lij is the negativeP of the
number of edges between vertices i and j if i 6= j. Then Lii = j6=i Lij for
1 ≤ i ≤ n, and so the row sums of L are all zero. It follows that L has
eigenvalue 0 on the all-1 vector; this is called the trivial eigenvalue of L. We
show below that the multiplicity of the zero eigenvalue is equal to the number
of connected components of H. Thus the multiplicity is 1 if and only if H is
connected.
Call the remaining eigenvalues of L non-trivial. They are all non-negative,
as we show in the following theorem (see [7]).

Theorem 1 (a) If L is a Laplacian matrix, then L is positive semi-definite.

(b) If L is a Laplacian matrix of order n and x is any vector in Rn , then

X
x⊤ Lx = (xi − xj )2 .
edges ij

(c) If L1 and L2 are the Laplacian matrices of graphs H1 and H2 with the
same vertices, and if H2 is obtained from H1 by inserting one extra
edge, then L2 − L1 is positive semi-definite.

(d) If L is the Laplacian matrix of the graph H, then the multiplicity of the
zero eigenvalue of L is equal to the number of connected components
of H.

Proof Each edge between vertices i and j defines a v × v matrix whose

entries are all 0 apart from the following submatrix:

 i j 
i 1 −1
j −1 1 .

The Laplacian is the sum of these matrices, which are all positive semi-
definite. This proves (a), (b) and (c).
From (b), the vector x is in the null space of the Laplacian if and only
if x takes the same value on both vertices of each edge, which happens if
and only if it takes a constant value on each connected component. This
proves (d).

8
 Theorem 1 shows that the smallest non-trivial eigenvalue of a connected
graph is positive. This eigenvalue is sometimes called the algebraic connec-
tivity of the graph. The statistical importance of this is shown in Section 3.2.
In Section 3.1 we shall need the Moore–Penrose generalized inverse of L−
of L (see [45]). Put P0 = n−1 Jn , where Jn is the n × n matrix whose entries
are all 1, so that P0 is the matrix of orthogonal projection onto the space
spanned by the all-1 vector. If H is connected then L + P0 is invertible, and

L− = (L + P0 )−1 − P0 ,

so that LL− = L− L = In − P0 , where In is the identity matrix of order n.

2.4 Laplacians of the concurrence and Levi graphs

There is a relationship between the Laplacian matrices of the concurrence
and Levi graphs of a block design ∆. Let N be the incidence matrix of
the design, and R the diagonal matrix (with rows and columns indexed by
treatments) whose (i, i) entry is the replication ri of treatment i. If the design
is equireplicate, then R = rIv , where r is the replication number.
For the remainder of the paper, we will use L for the Laplacian matrix of
the concurrence graph G of ∆, and L̃ for the Laplacian matrix of the Levi
graph G̃ of ∆. Then it is straightforward to show that
 
⊤ R −N
L = kR − NN , L̃ = .
−N⊤ kI

The Levi graph is connected if and only if the concurrence graph is con-
nected; thus 0 is a simple eigenvalue of L̃ if and only if it is a simple eigenvalue
of L, which in turn occurs if and only if all contrasts between treatment pa-
rameters are estimable (see Section 3.1). A block design with this property
is itself called connected : we consider only connected block designs.
In the equireplicate case, the above expressions for L and L̃ give a relation-
ship between their Laplacian eigenvalues, as follows. Let x be an eigenvector
of L with eigenvalue φ 6= rk. Then, for each of the two solutions θ of the
quadratic equation
rk − φ = (r − θ)(k − θ),
there is a unique vector z in Rb such that [ x⊤ z⊤ ]⊤ is an eigenvector of L̃
with eigenvalue θ. Conversely, any eigenvalue θ 6= k of L̃ arises in this way.
The Laplacian matrices of the concurrence graphs in Fig. 6 are shown in
Table 1.

9
3 Statistical issues
3.1 Estimation and variance
As part of the experiment, we measure the response Yω on each experimental
unit ω. If ω is in block Γ, then we assume that

Yω = τf (ω) + βΓ + εω ; (2)

here, τi is a constant depending on treatment i, βΓ is a constant depending

on block Γ, and εω is a random variable with expectation 0 and variance σ 2 .
Furthermore, if α 6= ω, then εα and εω are uncorrelated.
It is clear that we can add a constant to every block parameter, and sub-
tract that constant from every treatment parameter, without changing (2).
It is therefore impossible to estimate the individual treatment parameters.
However, if the design is connected, then we can estimate all contrastsPin the
treatment P parameters: that is, all linear combinations of the form i xi τi
for which i xi = 0. In particular, we can estimate all the simple treatment
differences τi − τj .
An estimator is a function of the responses Yω , so it is itself a random
variable. An estimator of a value is unbiased if its expectation is equal to
the true value; it is linear if it is a linear function of the responses. Amongst
linear unbiased estimators, the best one (the so-called BLUE), is the one with
the least variance. Let Vij be the variance of the BLUE for τi − τj .
If all the experimental units form a single block, then the BLUE of τ1 − τ2
is just the difference between the average responses for treatments 1 and 2.
It follows that  
1 1
V12 = + σ2.
r1 r2
When v = 2, this variance is minimized (for a given number of experimental
units) when r1 = r2 . Moreover, if the responses are normally distributed
then the length of the √ 95% confidence interval for τ1 − τ2 is proportional
to t(r1 + r2 − 2, 0.975) V12 , where t(d, p) is the 100p-th percentile of the
t distribution on d degrees of freedom. The smaller the confidence interval,
the more likely is our estimate to be close to the true value. This length can
be made smaller by increasing r1 + r2 , decreasing |r1 − r2 |, or decreasing σ 2 .
However, matters are not so simple when k < v and v > 2. The following
result can be found in any statistical textbook about block designs (see the
section on further reading for recommendations).

Theorem 2 Let L be the Laplacian

P matrix of the concurrence graph of a
connected block design. If x
i i = 0, then the variance of the BLUE of

10
 xi τi is equal to (x⊤ L− x)kσ 2 . In particular, the variance Vij of
the BLUE
P
i
of the simple difference τi − τj is given by Vij = L− − −
ii + Ljj − 2Lij kσ .
2

3.2 Optimality criteria

We want all of the Vij to be as small as possible, but this is a multi-
dimensional problem if v > 2. Let V̄ be the average of the variances Vij
over all treatments i, j with i 6= j. Theorem 2 shows that, for each fixed i,
X X
2
Vij = (L− − −
ii + Ljj − 2Lij )kσ
j6=i j6=i
2
= [(v − 1)L− − − −
ii + (Tr(L ) − Lii ) + 2Lii ]kσ
2
= [vL− −
ii + Tr(L )]kσ ,

because the row sums and column sums of L are all 0. It follows that V̄ =
2kσ 2 Tr(L− )/(v − 1).
Let θ1 , . . . , θv−1 be the non-trivial eigenvalues of L, now listed according
to multiplicity and in non-decreasing order. Then
1 1
Tr(L− ) = +···+ ,
θ1 θv−1
and so
1
V̄ = 2kσ 2 × .
harmonic mean of θ1 , . . . , θv−1
A block design is defined to be A-optimal (in some given class of designs
with the same values of b, k and v) if it minimizes the value of V̄ ; here ‘A’
stands for ‘average’. Thus a design is A-optimal if and only if it maximizes
the harmonic mean of θ1 , . . . , θv−1 .
For v > 2, the generalization of a confidence interval is a confidence
ellipsoid centered at the point (τ̂1 , . . . , τ̂v ) which gives the estimated
P value of
(τ1 , . . . , τv ) in the (v − 1)-dimensional subspace of Rv for which τi = 0. A
block design is called D-optimal if it minimizes the p volume of this confidence
ellipsoid. Since this volume is proportional to det(L− + P0 ), a design is
D-optimal if and only if it maximizes the geometric mean of θ1 , . . . , θv−1 .
Here ‘D’ stands for ‘determinant’.
Rather than looking at averages, we might consider the worst case. If all
the entries in the vector P x are multiplied by2 a constant c, then the variance
of the estimator of xi τi is multiplied by c . Thus, those contrast vectors x
which give the largest variance relative to their own length are those which
maximize x⊤ L− x/x⊤ x; these are precisely the eigenvectors of L with eigen-
value θ1 . A design is defined to be E-optimal if it maximizes the value of θ1 ;
here ‘E’ stands for ‘extreme’.

11
 More generally, for p in (0, ∞), a design is called Φp -optimal if it minimizes
Pv−1 !1/p
−p
i=1 θi
.
v−1

Thus A-optimality corresponds to p = 1, D-optimality corresponds to the

limit as p → 0, and E-optimality corresponds to the limit as p → ∞.
Let L1 and L2 be the Laplacian matrices of the concurrence graphs of
block designs ∆1 and ∆2 for v treatments in blocks of size k. If L2 − L1 is
positive semi-definite, then ∆2 is at least as good as ∆1 on all the Φp -criteria.
Theorem 1(c) shows that adding an extra block to a design cannot decrease
its performance on any Φp -criterion.
There are even more general classes of optimality criteria (see [28] and
[48] for details). Here we concentrate on A-, D- and E-optimality.

3.3 Questions and an example

A first obvious question to ask is: do these criteria agree with each other?
Our optimality properties are all functions of the concurrence graph.
What features of this graph should we look for if we are searching for opti-
mal, or near-optimal, designs? Symmetry? (Nearly) equal degrees? (Nearly)
equal numbers of edges between pairs of vertices? Distance-regularity? Large
girth (ignoring cycles within a block)? Small numbers of short cycles (ditto)?
High connectivity? Non-trivial automorphism group?
Is it more useful to look at the Levi graph rather than the concurrence
graph?

Example 1 Fig. 7 shows the values of the A- and D-criteria for all equi-
replicate block designs with v = 8, b = 12 and k = 2: of course, these are
just regular graphs with 8 vertices and degree 3. The harmonic mean is
shown on the A-axis, and the geometric mean on the D-axis. (Note that this
figure includes some designs that were omitted from Figure 3 of [4].) The
rankings on these two criteria are not exactly the same, but they do agree
at the top end, where it matters. The second-best graph on both criteria
is the cube; the best is the Möbius ladder, whose vertices are the elements
of Z8 and whose edges are {i, i + 1} and {i, i + 4} for i in Z8 . These two
graphs are so close on both criteria that, for practical purposes, they can be
regarded as equally good.
The plotting symbols show the edge-connectivity of the graphs: edge-
connectivity 3, 2, 1 is shown as ×, +, ◦ respectively. This does suggest that
the higher the edge-connectivity the better is the design on the A- and D-

12
 D
Möbius ladder
× ×
×
× cube
3.0 ++
++
++ s
K2,6
++
◦ +
◦ +
◦
2.4 ◦

1.8
0.6 1.2 1.8 2.4 3.0 A

Figure 7: Values of two optimality criteria for all equireplicate block designs
with v = 8, b = 12, and k = 2, and for K2,6

criteria. This is intuitively reasonable: if k = 2, then the edge-connectivity

is the minimum number of blocks whose removal disconnects the design. In
this context, it has been called breakdown number : see [39].
The four graphs with edge-connectivity 3 have no double edges, so con-
currences differ by at most 1. The only other regular graph with no double
edges is ranked eighth (amongst regular graphs) by the A-criterion. This
suggests that (near-)equality of concurrences is not sufficient to give a good
design.
The symbol s shows the non-regular graph K2,6 , which also has eight
vertices and twelve edges. It is not as good as the regular graphs with edge-
connectivity 3, but it beats many of the other regular graphs.
This pattern is typical of the block designs investigated by statisticians
for most of the 20th century. The A- and D-criteria agree closely at the top
end. High edge-connectivity appears to show good designs. Many of the best
designs have a high degree of symmetry.

4 Highly patterned block designs

4.1 Balanced incomplete-block designs
BIBDs are intuitively appealing, as they seem to give equal weight to all
treatment comparisons. They were introduced for agricultural experiments

13
by Yates in [54].
In [38], Kshirsagar proved that, if there exists a BIBD for given values
of v, b and k, then it is A-optimal. Kiefer generalized this in [35] to cover
Φp -optimality for all p in (0, ∞), including the limiting cases of D- and E-
optimality. The core of Kiefer’s
Pv−1 proof is as follows: binary designs maximize
Tr(L), which
P is−pequal to i=1 θi ; for any fixed −p
value T of this sum of positive
numbers, θi is minimized at [T /(v − 1)] when θ1 = · · · = θv−1 =
T /(v − 1); and T −p is minimized when T is maximized.

4.2 Other special designs

Of course, it frequently occurs that the values of b, v, k available for an ex-
periment are such that no BIBD exists. (Necessary conditions for the exist-
ence of a BIBD include the well-known divisibility conditions v | bk and
v(v − 1) | bk(k − 1), which follow from the elementary results in Section 1,
and Fisher’s inequality asserting that b ≥ v.)
In the absence of a BIBD, various other special types of design have been
considered, and some of these have been proved optimal. Here is a short
sample.
A design is group-divisible if the treatments can be partitioned into “groups”
all of the same size, so that the number of blocks containing two treatments is
λ1 if they belong to the same group and λ2 otherwise. Chêng [16, 17] showed
that if there is a group-divisible design with two groups and λ2 = λ1 + 1 in
the class of designs with given v, b, k, then it is Φp -optimal for all p, and in
particular is A-, D- and E-optimal.
A regular-graph design is a binary equireplicate design with two possible
concurrences λ and λ+1. It is easily proved that, in such a design, the number
of treatments lying in λ + 1 blocks with a given treatment is constant; so the
graph H whose vertices are the treatments, two vertices joined if they lie in
λ + 1 blocks, is regular.
Now Chêng [17] showed that a group-divisible design with λ2 = λ1 +1 = 1,
if one exists, is Φp -optimal in the class of regular-graph designs for all p.
Cheng and Bailey [20] showed that a regular-graph design for which the
graph is strongly regular (see [15]) and which has singular concurrence matrix
is Φp -optimal, for all p, among binary equireplicate designs with given v, b,
k.
Designs with the property described here are particular examples of par-
tially balanced designs with respect to an association scheme: see Bailey [3].
Another class which has turns out to be optimal in many cases, but
whose definition is less combinatorial, consists of the variance-balanced de-
signs, which we consider later in the chapter.

14
5 Graph concepts linked to D-optimality
5.1 Spanning trees of the concurrence graph
Let G be the concurrence graph of a connected block design, and let L be
its Laplacian matrix. A spanning tree for G is a spanning subgraph which is
a tree. Kirchhoff’s famous Matrix-tree theorem in [36] states the following:

Theorem 3 If G is a connected graph with v vertices and Laplacian ma-

trix L, then the product of the non-trivial eigenvalues of L is equal to v
multiplied by the number of spanning trees for G.

Thus we have a test for D-optimality:

A design is D-optimal if and only if its concurrence graph has the

maximal number of spanning trees.

Note that Theorem 3 gives an easy proof of Cayley’s theorem on the num-
ber of spanning trees for the complete graph Kv . The non-trivial eigenvalues
of its Laplacian matrix are all equal to v, so Theorem 3 shows that it has
v v−2 spanning trees.
If G is sparse, it may be much easier to count the number of spanning
trees than to compute the eigenvalues of L. For example, if G has a single
cycle, which has length s, then the number of spanning trees is s, irrespective
of the remaining edges in G.
In the context of optimal block designs, Gaffke discovered the importance
of Kirchhoff’s theorem in [26]. Cheng followed this up in papers such as
[16, 17, 19]. Particularly intriguing is the following theorem from [21].

Theorem 4 Consider block designs with k = 2 (connected graphs). For each

given v there is a threshold b0 such that if b ≥ b0 then any D-optimal design
for v treatments in b blocks of size 2 is nearly balanced in the sense that

• no pair of replications differ by more than 1;

• for each fixed i, no pair of concurrences λij differ by more than 1.

In fact, there is no known example with b0 > v − 1, which is the minimal

number of blocks required for connectivity.

15
5.2 Spanning trees of the Levi graph
In [27] Gaffke stated the following relationship between the numbers of span-
ning trees in the concurrence graph and the Levi graph.

Theorem 5 Let G and G̃ be the concurrence graph and Levi graph for a
connected incomplete-block design for v treatments in b blocks of size k. Then
the number of spanning trees for G̃ is equal to k b−v+1 times the number of
spanning trees for G.

Thus, an alternative test for D-optimality is to count the number of span-

ning trees in the Levi graph. For binary designs, the Levi graph has fewer
edges than the concurrence graph if and only if k ≥ 4.

6 Graph concepts linked to A-optimality

6.1 The concurrence graph as an electrical network
We can consider the concurrence graph G as an electrical network with a
1-ohm resistance in each edge. Connect a 1-volt battery between vertices i
and j. Then current flows in the network, according to these rules.

Ohm’s Law: In every edge, the voltage drop is the product of the current
and the resistance.

Kirchhoff’s Voltage Law: The total voltage drop from one vertex to any
other vertex is the same no matter which path we take from one to the
other.

Kirchhoff’s Current Law: At each vertex which is not connected to the

battery, the total current coming in is equal to the total current going
out.

We find the total current from i to j, and then use Ohm’s Law to define
the effective resistance Rij between i and j as the reciprocal of this current.
It is a standard result of electrical network theory that the linear equations
implicitly defined above for the currents and voltage differences have a unique
solution.
Let T be the set of treatments and Ω the set of experimental units.
Current flows in each edge eαω , where α and ω are experimental units in the
same block which receive different treatments; let I(α, ω) be the current from
f (α) to f (ω) in this edge. Thus I is a function I: Ω × Ω 7→ R such that

16
 (a) I(α, ω) = 0 if α = ω or if f (α) = f (ω) or if α and ω are in different
blocks.

(b) I(α, ω) = −I(ω, α) for (α, ω) in Ω × Ω.

This defines a further function Iout : T 7→ R by

X X
Iout (l) = I(α, ω) for l in T .
α:f (α)=l ω∈Ω

Voltage is another function V : T 7→ R. The following two conditions ensure

that Ohm’s and Kirchhoff’s Laws are satisfied.

(c) If there is any edge in G between f (α) and f (ω), then

I(α, ω) = V (f (α)) − V (f (ω)).

(d) If l ∈
/ {i, j}, then Iout (l) = 0.

If G is connected and different voltages V (i) and V (j) are given for a
pair of distinct treatments i and j, then there are unique functions I and V
satisfying conditions (a)–(d). Moreover, Iout (j) = −Iout (i) 6= 0. Then Rij is
defined by
V (i) − V (j)
Rij = .
Iout (i)
It can be shown that the value of Rij does not depend on the choice of values
for V (i) and V (j), so long as these are different. In practical examples, it is
usually convenient to take V (i) = 0 and let I take integer values.
What has all of this got to do with block designs? The following theorem,
which is a standard result from electrical engineering, gives the answer.

Theorem 6 If L is the Laplacian matrix of a connected graph G, then the

effective resistance Rij between vertices i and j is given by

Rij = L− − −


ii + Ljj − 2Lij .

Comparing this with Theorem 2, we see that Vij = Rij × kσ 2 . Hence we

have a test for A-optimality:

A design is A-optimal if and only if its concurrence graph, re-

garded as an electrical network, minimizes the sum of the pairwise
effective resistances between all pairs of vertices.

17
 [−25]
✈
✔
✔ ❚❚
5 ✔ ⑦ ❚ 5
❃✔ ❚
[−15] [−20] ✔ ❚ [−30] [−47]
✈ ✲ ✔
✈ ✲ ❚✈ ✲ ✈
❚ 5 ✔ 10 ❚ 17 j ✔
❚ ✔ ❚ ✔
5❚⑥ ❃✔10 2⑥❚
❃✔19
❚ ✔ ❚ ✔
❚ ✈✔ ❚ ✔
[−10] ✔❚ ✔ [−28]
❚✈
✔ ❚ ✔ ❚
5 ✔ ❚ 10 14 ✔ ❚ 7
❃✔ ❚⑥ ❃✔ ❚⑥
✔ ❚ ✔ ❚
✔
✈ 5✛ ❚✈i ✲14 ✔
✈ ✲7 ❚✈
[−5] [0] ❚ ✔
[−14] [−21]
❚ ✔
⑦❚ ✔7
7 ❚ ❃ ✔
❚ ✔
❚✔✈
[−7]

Figure 8: The current between vertices i and j in a concurrence graph

Effective resistances are easy to calculate without matrix inversion if the

graph is sparse.
Figure 8 shows the concurrence graph of a block design with v = 12,
b = 6 and k = 3. Only vertices i and j are labelled. Otherwise, numbers
beside arrows denote current and numbers in square brackets denote voltage.
It is straightforward to check that conditions (a)–(d) are satisfied. Now
V (i) − V (j) = 47 and Iout (i) = 36, and so Rij = 47/36. Therefore Vij =
(47/12)σ 2 . Moreover, for graphs consisting of b triangles arranged in a cycle
like this, it is clear that average effective resistance, and hence the average
pairwise variance, can be calculated as a function of b.

6.2 The Levi graph as an electrical network

The Levi graph G̃ of a block design can also be considered as an electrical
network. Denote by B the set of blocks. Now current is defined on the
ordered edges of the Levi graph. Recall that, if ω is an experimental unit
in block Γ, then the edge ẽω joins Γ to f (ω). Thus current is defined on
(Ω × B) ∪ (B × Ω) and voltage is defined on T ∪ B. Conditions (a)–(d) in
Section 6.1 need to be modified appropriately.
The next theorem shows that a current–voltage pair (I, V ) on the con-
˜ Ṽ ) on
currence graph G can be transformed into a current–voltage pair (I,

18
 ˜ Γ) flows in edge ẽα from vertex
the Levi graph G̃. In G̃, the current I(α,
f (α) to vertex Γ, where α ∈ Γ. Hence the pairwise variance Vij can also be
calculated from the effective resistance R̃ij in the Levi graph.

Theorem 7 Let G be the concurrence graph and G̃ be the Levi graph of a

connected block design with block size k. If i and j are two distinct treat-
ments, let Rij and R̃ij be the effective resistance between vertices i and j in
the electrical networks defined by G and G̃, respectively. Then R̃ij = kRij ,
and so Vij = R̃ij σ 2 .

Proof Let (I, V ) be a current–voltage pair on G. For (α, Γ) ∈ Ω × B, put

˜ Γ) = −I(Γ,
˜ α) =
X
I(α, I(α, ω)
ω∈Γ

˜ Γ) = I(Γ,
if α ∈ Γ; otherwise, put I(α, ˜ α) = 0. Put Ṽ (i) = kV (i) for all i
in T , and X
Ṽ (Γ) = V (f (ω))
ω∈Γ

for all Γ in B. It is clear that I˜ satisfies the analogues of conditions (a)

and (b).
If α ∈ Γ, then
˜ Γ) =
X
I(α, I(α, ω)
ω∈Γ
X
= [V (f (α)) − V (f (ω))]
ω∈Γ

= kV (f (α)) − Ṽ (Γ) = Ṽ (f (α)) − Ṽ (Γ),

so the analogue of condition (c) is satisfied.

If Γ ∈ B, then

I˜out (Γ) = ˜ α) = −
X XX
I(Γ, I(α, ω) = 0,
α∈Γ α∈Γ ω∈Γ

because I(α, α) = 0 and I(α, ω) = −I(ω, α). If l ∈ T then

I˜out (l) = ˜ Γ) =
X X X X
I(α, I(α, ω) = Iout (l).
α:f (α)=l Γ∈B α:f (α)=l ω∈Ω

In particular, I˜out (l) = 0 if l ∈ / {i, j}, which shows that the analogue of
˜ Ṽ ) is the current–voltage pair on
condition (d) is satisfied. It follows that (I,
G̃ defined by Ṽ (i) and Ṽ (j).

19
 Now
Ṽ (i) − Ṽ (j) k(V (i) − V (j))
R̃ij = = = kRij .
˜
Iout (i) Iout (i)
Then Theorems 2 and 6 show that Vij = Rij σ 2 .

When k = 2 it seems to be easier to use the concurrence graph than the

Levi graph, because it has fewer vertices, but for larger values of k the Levi
graph may be better, as it does not have all the within-block cycles that the
concurrence graph has. Fig. 9 gives the Levi graph of the block design whose
concurrence graph is in Fig. 8, with the same two vertices i and j attached to
the battery. This gives R̃ij = 47/12, which is in accordance with Theorem 7.
j
[−5] [−7] [−21] [−47] [−25] [−15]
treatments ① ① ① ① ① ①
12 ✻
[−5] ①[−7] [−21] ① [−35] ① [−25] ✥①[−15]
blocks ① ①
✥✥ ✥
❅ ❅ ❅ ❅ ✥✥✥ ❅
❅ ❅ ❅
❘7 ✥✥❅✥ ❅
❄ 5❅ 7 ❘
❅ ✥
✿ ❅✥✥ ❅ ❄ ❅ ❄
5 ■ 7
✻ ✥✥✥ ✻7 ✥ 7
✻ ■ 5 5 ■5 5
❅
✥ ✥ ✥ ❅ 5 ❅ ❅ ❅
treatments ✥✥✥ ❅①
① ❅① ❅① ❅① ❅①
[−10] [0] [−14] [−28] [−30] [−20]
i

Figure 9: Current between i and j for the Levi graph corresponding to the
concurrence graph in Fig. 8

Here is another way of visualizing Theorem 7. From the block design we

construct a graph G0 with vertex-set T ∪ Ω ∪ B: the edges are {α, Γ} for
α ∈ Γ ∈ B and {α, f (α)} for α ∈ Ω. Let (I0 , V0 ) be a current–voltage pair
on G0 for which both battery vertices are in T . We obtain the Levi graph
G̃ from G0 by becoming blind to the vertices in Ω. Thus the resistance in
each edge of G̃ is twice that in each edge in G0 , so this step multiplies each
effective resistance by 2.
Because none of the battery vertices is in B, we can now obtain G from
G̃ by replacing each path of the form (i, Γ, j) by an edge (i, j). There is
no harm in scaling all the voltages by the same amount, so we can obtain
(I, V ) on G from (I, ˜ Ṽ ) on G̃ by putting V (i) = Ṽ (i)/k for i in T , and
I(α, ω) = V (f (α)) − V (f (ω)) for α, ω in the same block. If Γ is a block, then

˜ Γ) =
X X X
0= I(α, [Ṽ (f (α)) − Ṽ (Γ)] = k V (f (α)) − k Ṽ (Γ),
α∈Γ α∈Γ α∈Γ

20
 P
and so Ṽ (Γ) = V (f (α)). Also, if α ∈ Γ, then
α∈Γ
X X
I(α, ω) = [V (f (α)) − V (f (ω))]
ω∈Γ ω∈Γ
X
= kV (f (α)) − V (f (ω))
ω∈Γ

= Ṽ (f (α)) − Ṽ (Γ)
˜ Γ).
= I(α,

Therefore, this transformation reverses the one used in the proof of Theo-
rem 7.
There is yet another way of obtaining Theorem 7. If we use the re-
sponses Yω to estimate the block parameters βΓ in (2) as well as the treat-
ment parameters τi , then standard theory of linear models shows that, if the
design
Pv is connected,
Pb then we canPestimate
P linear combinations of the form
i=1 xi τi + j=1 zj βj so long as xi = zj . Moreover, the variance of the
BLUE of this linear combination is
   
− x 2 R N
[ x⊤ z⊤ ]C σ , where C=
z N⊤ kIb

and R is the diagonal matrix of replications.

If we reparametrize equation (2) by replacing βj by −γj for j = 1, . . . ,
b, then the estimable quantities are the contrasts in τ1 , . . . , τv , γ1 , . . . , γb .
The so-called information matrix C must be modified by multiplying the
last b rows and the last b columns by −1: this gives precisely the Laplacian
L̃ of the Levi graph G̃. Just as for L, but unlike C, the null space is spanned
by the all-1 vector.

6.3 Spanning thickets

We have seen that the value of the D-criterion is a function of the number
of spanning trees of the concurrence graph G. It turns out that the closely
related notion of a spanning thicket enables us to calculate the A-criterion;
more precisely, the value of each pairwise effective resistance in G.
A spanning thicket for the graph is a spanning subgraph that consists of
two trees (one of them may be an isolated vertex).

Theorem 8 If i and j are distinct vertices of G then

number of spanning thickets with i, j in different parts
Rij = .
number of spanning trees

21
 This is also rather easy to calculate directly when the graph is sparse.
Summing all the Rij and using Theorem 8 gives the following result from
[49].

Theorem 9 If F is a spanning thicket for the concurrence graph G, denote

by F1 and F2 the sets of vertices in its two trees. Then
X
|F1 | |F2 |
X spanning thickets F
Rij =
i<j
number of spanning trees

6.4 Random walks and electrical networks

It was first pointed out by Kakutani in 1945 that there is a very close con-
nection between random walks and electrical networks. In a simple random
walk, a single step works as follows: starting at a vertex, we choose an edge
containing the vertex at random, and move along it to the other end. This
definition accommodates multiple edges, and is easily adapted to graphs with
edge weights (where the probability of moving along an edge is proportional
to the weight of the edge).
If we are thinking of an edge-weighted graph as an electrical network, we
take the weights to be the conductances of the edges (the reciprocals of the
resistances).
The connection is simple to state:

Theorem 10 Let i and j be distinct vertices of the connected edge-weighted

graph G. Apply voltages of 1 at i and 0 at j. Then the voltage at a vertex l
is equal to the probability that the random walk, starting at l, reaches i before
it reaches j.

From this theorem, it is possible to derive a formula for the effective

resistance between two vertices. Here are two such formulas. Given two
vertices i and j, let Pesc (i → j) be the probability that a random walk
starting at i reaches j before returning to i; and let Si (i, j) be the expected
number of times that a random walk starting at i visits i before reaching j.
Then the effective resistance between i and j is given by either of the two
expressions
1 Si (i, j)
and ,
di Pesc (i → j) di
where di is the degree of i. (If the edge resistances are not all 1, then the
term di should be replaced by the sum of the reciprocals of the resistances
of all edges incident with vertex i.)

22
 The random walk approach gives alternative proofs of some of the main
results about electrical networks. We discuss this further in the guide to the
literature.

6.5 Foster’s formula and generalizations

In 1948, Foster [25] discovered that the sum of the effective resistances be-
tween all adjacent pairs of vertices of a connected graph on v vertices is equal
to v − 1. Thirteen years later, he found a similar formula for pairs of vertices
at distance 2: X Rij
= v − 2.
dh
i∼h∼j

Further extensions have been found, but require a stronger condition on the
graph. The sum of resistances between all pairs of vertices at distance at
most m can be written down explicitly if the graph is walk-regular up to
distance m; this means that the number of closed walks of length k starting
and finishing at a vertex i is independent of i, for k ≤ m. The formula was
discovered by Emil Vaughan, to whom this part of the chapter owes a debt.
In particular, if the graph is distance-regular (see [14]), then the value
of the A-criterion can be written down in terms of the so-called intersection
array of the graph.

6.6 Distance
At first sight it seems obvious that pairwise variance should decrease as con-
currence increases, but there are many counter-examples to this. However,
the following theorem is proved in [3].

Theorem 11 If the Laplacian matrix L has precisely two distinct non-trivial

eigenvalues, then pairwise variance is a decreasing linear function of concur-
rence.

It does appear that effective resistance, and hence pairwise variance, gen-
erally increases with distance in the concurrence graph. In [7, Question 5.1]
we pointed out that this is not always exactly so, and asked if it is nevertheless
true that the maximal value of Rij is achieved for some pair of vertices {i, j}
whose distance apart in the graph is maximal. Here is a counter-example.

Example 2 Let k = 2, so that the block design is the same as its concurrence
graph. Take v = 10 and b = 14. The graph consists of a cube, with two
extra vertices 1 and 2 attached as leaves to vertex 3. The vertex antipodal

23
to 3 in the cube is labelled 4. It is straightforward to check (either using an
electrical network, or by using the fact that the cube is distance-regular) that
the effective resistance between a pair of cube vertices is 7/12, 3/4 and 5/6
for vertices at distances 1, 2 and 3. Hence R1j ≤ 11/6 for all cube vertices j,
while R12 = 2. On the other hand, the distance between vertices 1 and 2 is
only 2, while that between either of them and vertex 4 is 4.

There are some ‘nice’ graphs where pairwise variance does indeed increase
with distance. The following result is proved in [6]. Biggs gave the equivalent
result for effective resistances in [12].

Theorem 12 Suppose that a block design has just two distinct concurrences,
and that the pairs of vertices corresponding to the larger concurrence form
the edges of a distance-regular graph H. Then pairwise variance increases
with distance in H.

7 Graph concepts linked to E-optimality

7.1 Measures of bottlenecks
A ‘good’ graph (for use as a network) is one without bottlenecks: any set of
vertices should have many edges joining it to its complement. So, for any
subset S of vertices, we let ∂(S) (the boundary of S) be the set of edges
which have one vertex in S and the other in its complement, and then define
the isoperimetric number ι(G) by
 
|∂S| v
ι(G) = min : S ⊆ V (G), 0 < |S| ≤ .
|S| 2

The next result shows that the isoperimetric number is related to the E-
criterion. It is useful not so much for identifying the E-optimal designs as for
easily showing that large classes of designs cannot be E-optimal: any design
whose concurrence graph has low isoperimetric number performs poorly on
the E-criterion.

Cutset Lemma 1 Let G have an edge-cutset of size c whose removal sep-

arates the graph into parts S and G \ S with m and n vertices respectively,
where 0 < m ≤ n. Then
 
1 1 2 |∂S|
θ1 ≤ c + ≤ .
m n |S|

24
ProofP We know that θ1 is the minimum of x⊤ Lx/x⊤ x over real vectors x
with i xi = 0. Put 
n if i ∈ S
xi =
−m otherwise.
Then x⊤ x = nm(m + n) and
X
x⊤ Lx = (xi − xj )2 = c(m + n)2 .
edges ij

Hence
x⊤ Lx c(m + n)2
 
1 1 2c 2 |∂S|
θ1 ≤ ⊤ = =c + ≤ = .
x x nm(m + n) m n m |S|
Corollary 1 Let θ1 be the smallest non-trivial eigenvalue of the Laplacian
matrix L of the connected graph G. Then θ1 ≤ 2ι(G).
There is also an upper bound for the isoperimetric number in terms of
θ1 , which is loosely referred to as a ‘Cheeger-type inequality’; for details, see
the further reading.
We also require a second cutset lemma, phrased in terms of vertex cutsets.
Cutset Lemma 2 Let G have a vertex-cutset C of size c whose removal
separates the graph into parts S and T with m, n vertices respectively (so
nm > 0). Let m′ and n′ be the number of edges from vertices in C to vertices
in S, T respectively. Then
m′ n2 + n′ m2
θ1 ≤ .
nm(m + n)
In particular, if there are no multiple edges at any vertex of C then θ1 ≤ c,
with equality if and only if every vertex in C is joined to every vertex in
S ∪ T.
Proof Put 
 n if i ∈ S
xi = −m if i ∈ T
0 otherwise.


Then x⊤ x = nm(m + n) and x⊤ Lx = m′ n2 + n′ m2 , and so

m′ n2 + n′ m2
θ1 ≤ .
nm(m + n)
If there are no multiple edges at any vertex in C then m′ ≤ cm and n′ ≤ cn
and the result follows.

25
7.2 Variance balance
A block design is variance-balanced if all the concurrences λij are equal for
i 6= j. In such a design, all of the pairwise variances Vij are equal. Morgan
and Srivastav proved the following result in [43].

Theorem 13 If the constant concurrence λ of a variance-balanced design

satisfies (v − 1)λ = ⌊(bk/v)⌋(k − 1) then the design is E-optimal.

A block with k different treatments contributes k(k − 1)/2 edges to the

concurrence graph. Let us define the defect of a block to be
k(k − 1)
− the number of edges it contributes to the graph.
2
The following result is proved in [7].

Theorem 14 If k < v, then a variance-balanced design with v treatments is

E-optimal if the sum of the block defects is less than v/2.

Table 1(b) shows that the design in Fig. 2(b) is variance-balanced. Block
Γ1 has defect 1, and each other block has defect 0, so the sum of the block
defects is certainly less than 5/2, and Theorem 14 shows that the design is E-
optimal. It is rather counter-intuitive that the non-binary design in Fig. 2(b)
can be better than the design in Fig. 2(a); in fact, in his contribution to the
discussion of Tocher’s paper [51], which introduced this design, David Cox
said
I suspect that . . . balanced ternary designs are of no practical
value.
Computation shows that the design in Fig. 2(a) is Φp -better than the one
in Fig. 2(b) if p < 5.327. In particular, it is A- and D-better.

8 Some history
As we have seen, if the experimental units form a single block and there are
only two treatments then it is best for their replications to be as equal as
possible. Statisticians know this so well that it is hard for us to imagine
that more information may be obtained, about all treatment comparisons, if
replications differ by more than 1.
In agriculture, or in any area with qualitative treatments, A-optimality
is the natural criterion. If treatments are quantities of different substances,

26
then D-optimality is preferable, as the ranking on this criterion is invariant
to change of measurement units. Thus industrial statisticians have tended
to prefer D-optimality, although E-optimality has become popular among
chemical process engineers. Perhaps the different camps have not talked to
each other as much as they should have.
For most of the 20th century, it was normal practice in field experiments
to have all treatments replicated three or four times. Where incomplete
blocks were used, they typically had size from 3 to 20. Yates introduced his
square lattice designs with v = k 2 in [55]. He used uniformity data and two
worked examples to show that these designs can give lower average pairwise
variance than a design using a highly replicated control treatment, but both
of his examples were equireplicate with r ∈ {3, 4}.
In the 1930s, 1940s and 1950s, analysis of the data from an experiment
involved inverting the Laplacian matrix without a computer: this is easy
for BIBDs, and only slightly harder if the Laplacian matrix has only two
distinct non-trivial eigenvalues. The results in [35] and [38] encouraged the
beliefs that the optimal designs, on all Φp -criteria, are as equireplicate as
possible, with concurrences as equal as possible, and that the same designs
are optimal, or nearly so, on all of these criteria.
Three short papers in the same journal in 1977–1982 demonstrate the
beliefs at that time. In [29], John and Mitchell did not even consider designs
with unequal replication. They conjectured that, if there exist any regular-
graph designs for given values of v, b and k, then the A- and D-optimal designs
are regular-graph designs. For the parameter sets which they had examined
by computer search, the same designs were optimal on the A- and D-criteria.
In [33], Jones and Eccleston reported the results of various computer searches
for A-optimal designs without the constraint of equal replication. For k = 2
and b = v ∈ {10, 11, 12} (but not v = 9) their A-optimal design is almost
a queen-bee design, and their designs are D-worse than those in [29]. The
belief in equal replication was so ingrained that some readers assumed that
there was an error in their program.
John and Williams followed this with the paper [30] on conjectures for
optimal block designs for given values of v, b and k. Their conjectures in-
cluded:

• the set of regular-graph designs always contains one that is optimal

without this restriction;

• among regular-graph designs, the same designs are optimal on the A-

and D-criteria.

They endorsed Cox’s dismissal of non-binary designs, strengthening it to the

27
statement that they “are inefficient”, and declared that the three unequally
replicated A-optimal designs in [33] were “of academic rather than of prac-
tical interest”. These conjectures and opinions seemed quite reasonable to
people who had been finding good designs for the sizes needed in agricultural
experiments.
At the end of the 20th century, there was an explosion in the number of
experiments in genomics, using microarrays. Simplifying the story greatly,
these are effectively block designs with k = 2, and biologists wanted A-
optimal designs, but they did not know the vocabulary ‘block’ or ‘A-optimal’,
‘graph’ or ‘cycle’. Computers were now much more powerful than in 1980,
and researchers in genomics could simply undertake computer searches with-
out the benefit of any statistical theory. In 2001, Kerr and Churchill [34]
published the results of a computer search for A-optimal designs with k = 2
and v = b ≤ 11. For v ∈ {10, 11, 12}, their results were completely consistent
with those in [33], which they did not cite. They called cycles loop designs.
Mainstream statisticians began to get involved. In 2005, Wit, Nobile and
Khanin published the paper [53] giving the results of a computer search for
A- and D-optimal designs with k = 2 and v = b. The results are shown
in Fig. 10. The A-optimal designs differ from the D-optimal designs when
v ≥ 9, but are consistent with those found in [34].
What is going on here? Why are the designs so different when v ≥ 9?
Why is there such a sudden, large change in the A-optimal designs? We
explain this in the next section.

9 Block size two

9.1 Least replication
If k = 2, then the design is the same as its concurrence graph, and connec-
tivity requires that b ≥ v − 1. If b = v − 1, then all connected designs are
trees, such as those in Fig. 11. Theorem 3 shows that the D-criterion does
not differentiate between them.
In a tree, the effective resistance Rij is just the length of the unique path
between vertices i and j. Theorems 2 and 6 show that the only A-optimal
designs are the stars, such as the graph on the right of Fig. 11.
In a star with v vertices, the contrast between any two leaves is an eigen-
vector of the Laplacian matrix L with eigenvalue 1, while the contrast be-
tween the central vertex and all the other vertices is an eigenvector with
eigenvalue v. If v ≥ 5 and G is not a star then there is an edge whose
removal splits the graph into two components of sizes at least 2 and 3. Cut-

28
 r r r r
r r r
r
r r r r r
r
r r r
r r r r

r r r r r
r r r
r r
r r r r
r r
r r r r
r r
r r r r
r r r r

(a) D-optimal designs

r r r r
r r r
r
r r r r r
r
r r r
r r r r
r r r r r r r r
r r r r
r r r r r r r r r
r r r
r r r r r r

(b) A-optimal designs

Figure 10: D-and A-optimal designs with k = 2 and 6 ≤ v = b ≤ 11

✟✉ ✉
✟
✉ ✟ ✉ ✉ ✉
❅ ❆ ❅
❅ ❆ ❅
✉ ❅✉ ❆✉ ✉ ❅✉ ✉
❆ ❅ ✁ ❅
❆ ❅ ✁ ❅
❆✉
❍❍ ✁
❅✉ ✉ ❅✉
❍✉ ✉

Figure 11: Two trees with v = 9, b = 8 and k = 2

29
set Lemma 1 then shows that θ1 ≤ 5/6 < 1. The only other tree which is
not a star√ is the path of length 3, for which direct calculation shows that
θ1 = 2 − 2 < 1. Hence the E-optimal designs are also the stars.

9.2 One fewer treatment

If b = v and k = 2, then the concurrence graph G contains a single cycle: such
graphs are called unicyclic. Let s be the length of the cycle. All the remaining
vertices are in trees attached to various vertices of the cycle. Fig. 12 shows
two unicyclic graphs with v = 12 and s = 6.

10 ① 10 ①

①12 8 ① ① ① ①12 ①8 ① ①
❚ ✔ ❚ ❚ ✔
❚ ✔ 9 11 ❚ ❚ ✔9 11
❚ ✔ ❚ ❚ ✔
❚① ✔
① ①7 ❚① ❚①
✔ ①7
✔1 6 ❚ ✔1 6❚
✔ ❚ ✔ ❚
✔ ❚ ✔ ❚
✔
①2 5 ❚① ✔2
① 5 ❚①
❚ ✔ ❚ ✔
❚ ✔ ❚ ✔
❚ ✔ ❚ ✔
3
❚① 4 ✔
① 3
❚① 4①
✔

(a) (b)

Figure 12: Two unicyclic graphs with b = v = 12 and s = 6

As we remarked in Section 5, the number of spanning trees in a unicyclic

graph is equal to the length of the cycle. Hence, Theorem 3 gives the following
result.

Theorem 15 If k = 2 and b = v ≥ 3, then the D-optimal designs are

precisely the cycles.

For A-optimality, we first show that no graph like the one in Fig. 12(a)
can be optimal. If vertex 12 is moved so that it is joined to vertex 6, instead
of vertex 1, then the sum of the variances Vi,12 for i in the cycle is unchanged
and the variances Vi,12 for the remaining vertices i are all decreased. This
argument shows that all the trees must be attached to the same vertex of
the cycle.

30
 V̄ /σ 2
5
+ v = 13
+ +
+ × v = 12
+ ×
+ × × ◦ v = 11
4 + ×
× ◦ ◦
◦ • v = 10
+
× ◦
+
× ◦ • • • ⋄ v=9
+
×
◦ + +
× ◦ •
•
⋄ +
×
◦ ×
◦ ◦
• • ⋄ ⋄ ⋄ ∗ v=8
∗ •
⋄ • ⋄ ⋄
⊙ ∗ ⋄ ∗ ⊙ v=7
3 ⋆ ⊙ ∗ ∗ ∗ ∗
⊙ ⊙ ⋆ v=6
⋆ ⊙
⋆ ⊙
⋆
⋆
2 s
3 6 9 12 15

Figure 13: Average pairwise variance, in a unicylic graph with v vertices, as

a function of the length s of the cycle

Now consider the tree on vertices 6, 8, 9, 10 and 11 in Fig. 12(a). If the

two edges incident with vertex 8 are modified to those in Fig. 12(b), then the
set of variances between these five vertices are unchanged, as are all others
involving vertex 8, but those between vertices 9, 10, 11 and vertices outside
this tree are all decreased. This argument shows that, for any given length s
of the cycle, the only candidate for an A-optimal design has v − s leaves
attached to a single vertex of the cycle.
The effective resistance between a pair of vertices at distance d in a cycle
of length s is d(s − d)/s, while that between a leaf and the cycle vertex to
which it is attached is 1. Hence a short calculation shows that the sum of
the pairwise effective resistances is equal to g(s)/12, where

g(s) = −s3 + 2vs2 + 13s − 12sv + 12v 2 − 14v.

Now V̄ /σ 2 = g(s)/[3v(v − 1)] and we seek the minimum of g(s) for integers
s in the interval [2, v].
Fig. 13 plots g(s)/[3v(v − 1)] for s in [2, v] and 6 ≤ v ≤ 13. When v ≤ 7,
the function g is monotonic decreasing, so it attains its minimum on [2, v]
at s = v. For all larger values of v, the function g has a local minimum in
the interval [3, 5]: when v ≥ 9, the value at this local minimum is less than
g(v). This change from the upper end of the interval to the local minimum
explains the sudden change in the A-optimal designs. Detailed examination
of the local minimum gives the following result.

31
Theorem 16 If k = 2 and b = v ≥ 3 then the A-optimal designs are:

• a cycle, if v ≤ 8;

• a square with v − 4 leaves attached to one vertex, if 9 ≤ v ≤ 11;

• a triangle with v − 4 leaves attached to one vertex, if v ≥ 13;

• either of the last two, if v = 12.

What about E-optimality? The smallest eigenvalue of the Laplacian ma-

trix of the triangle with one or more leaves attached to one vertex is 1, as
is that of the digon with two or more leaves attached to one vertex. We
now show that almost all other unicyclic graphs have at least one non-trivial
eigenvalue smaller than this.
Suppose that vertex i in the cycle has a non-empty tree attached to it, so
that {i} is a vertex cutset. If s ≥ 3 then there are no double edges, so Cutset
Lemma 2 shows that θ1 < 1 unless all vertices are joined to i, in which case
s = 3. If s = 2 and there are trees attached to both vertices of the digon,
then applying Cutset Lemma 2 at each of these vertices shows that θ1 < 1
unless v =√4 and there is one leaf at each vertex of the digon: for this graph,
θ1 = 2 − 5 < 1. A digon with leaves attached to one vertex is just a star
with one edge doubled.
The cycle of size v is a cyclic design. The smallest eigenvalue of its
Laplacian matrix is 2(1 − cos(2π/v)), which is greater than 1 when v ≤ 5,
is equal to 1 when v = 6, and is less than √ 1 when v ≤ 7. When v = 3 it
is equal to 3, which is greater than 3 − 3, which is the smallest Laplacian
eigenvalue of the digon with one leaf.
Putting all of this together proves the following result.

Theorem 17 If k = 2 and b = v ≥ 3, then the E-optimal designs are:

• a cycle, if v ≤ 5;

• a triangle with v − 3 leaves attached to one vertex, or a star with one

edge doubled, if v ≥ 7;

• either of the last two, if v = 6.

Thus, for v ≥ 9, the ranking on the D-criterion is essentially the opposite

of the ranking on the A- and E-criteria. The A- and E-optimal designs are
far from equireplicate. The change is sudden, not gradual. These findings
were initially quite shocking to statisticians.

32
9.3 More blocks
What happens when b is larger than v but still has the same order of mag-
nitude? The following theorems show that the A- and E-optimal designs are
very different from the D-optimal designs when v is large. The proofs of
Theorems 19 and 20 are in [4] and [7] respectively.

Theorem 18 Let G be the concurrence graph of a connected block design ∆

with k = 2 and b ≥ v. If ∆ is D-optimal then G does not contain any bridge
(an edge cutset of size one): in particular, G contains no leaves.

Proof Suppose that {i, j} is an edge-cutset for G. Let H and K be the

parts of G containing i and j, respectively.
Since G is not a tree, we may assume that H is not a tree, and so there is
some edge e in H that is not in every spanning tree for H. Let n1 and n2 be
the numbers of spanning trees for H that include and exclude e, respectively,
and let m be the number of spanning trees for K. Every spanning tree for G
consists of spanning trees for H and K together with the edge {i, j}. Hence
G has (n1 + n2 )m spanning trees.
Let ℓ be a vertex on e with ℓ 6= i. Form G′ from G by removing edge e
and inserting the edge e′ , where e′ = {ℓ, j}.
Let T and T ′ be spanning trees for H and K respectively. If T does not
contain e then T ∪ {{i, j}} ∪ T ′ and T ∪ {e′ } ∪ T ′ are both spanning trees
for G′ . If T contains e then (T \ {e}) ∪ {{i, j}} ∪ {e′ } ∪ T ′ is a spanning tree
for G′ . Hence the number of spanning trees for G′ is at least (2n2 + n1 )m,
which is greater than (n1 + n2 )m because n2 ≥ 1. Hence G does not have
the maximal number of spanning trees and so ∆ is not D-optimal.

Theorem 19 Let c be a positive integer. Then there is a positive integer

vc such that if b − v = c and v ≥ vc then all A-optimal designs with k = 2
contain leaves.

Theorem 20 If 20 ≤ v ≤ b ≤ 5v/4 then the concurrence graph for any

E-optimal design with k = 2 contains leaves.

Of course, to obtain a BIBD when k = 2, b needs to be a quadratic

function of v. What happens if b is merely a linear function of v? In [7] we
conjectured that if b = cv for some constant c then there is a threshold result
like the one in Theorem 19. However, current work by Robert Johnson and
Mark Walters [32] suggests something much more interesting—that there is
a constant C with 3 < C < 4 such that if b ≥ Cv and k = 2 then all
A-optimal designs are (nearly) equireplicate, and that random such graphs

33
(in a suitable model) are close to A-optimal with high probability. On the
other hand if b ≤ Cv then a graph consisting of a large almost equireplicate
part (all degrees 3 and 4 with average degree close to Cv) together with a
suitable number of leaves joined to a single vertex is strictly better than any
queen-bee design.

9.4 A little more history

The results on D- and A-optimality in Sections 9.1 and 9.2 were proved in
[4], partly to put to rest mutterings that the results of [33, 34, 53] found by
computer search were incorrect. The results on E-optimality are in [7].
In spite of the horror with which these results were greeted, it transpired
that they were not new. The D- and E-optimal designs for b = (v −1)/(k −1)
were identified in [11] in 1991. The A-optimal designs for k = 2 and b = v −1
had been given in [41] in 1991. Also in 1991, Tjur gave the A-optimal designs
for k = 2 and b = v in [52]: his proof used the Levi graph as an electrical
network.
A fairly common response to these unexpected results was ‘It seems to
be just block size 2 that is a problem.’ Perhaps those of us who usually
deal with larger blocks had simply not thought that it was worth while to
investigate block size 2 before the introduction of microarrays.
However, as we sketch in the next section, the problem is not block size 2
but very low average replication. The proofs there are similar to those in
this section; they are given in more detail in [8, 46]. Once again, it turns out
that these results are not all new. The D-optimal designs for v/(k −1) blocks
of size k were given by Balasubramanian and Dey in [10] in 1996—but their
proof uses a version of Theorem 5 with the wrong value of the constant. The
A-optimal designs for v/(k − 1) blocks of size k were published by Krafft and
Schaefer in [37] in 1997—but those authors are not blameless either, because
they apparently had not read [52]!
Our best explanation is that agricultural statisticians are so familiar with
average replication being at least 3 that when we saw these papers we decided
that they had no applicability and so forgot them.

10 Very low average replication

In this section we once again consider general block size k. A block design
is connected if and only if its Levi graph is connected. The Levi graph has
v + b vertices and bk edges, so connectivity implies that bk ≥ b + v − 1; that
is, b(k − 1) ≥ v − 1.

34
10.1 Least replication
If b(k −1) = v −1 and the design is connected, then the Levi graph G̃ is a tree
and the concurrence graph G looks like those in Fig. 5. Hypergraph-theorists
do not seem to have an agreed name for such designs.
For both D- and A-optimality, it turns out to be convenient to use the
Levi graph. Since all the Levi graphs are trees, Theorem 5 shows that the
D-criterion does not distinguish among connected designs.
By Theorem 7, Vij = R̃ij σ 2 . When G̃ is a tree, R̃ij = 2 when i and j
are in the same block; otherwise, R̃ij = 4 if any block containing i has a
treatment in common with any block containing j; and otherwise, R̃ij ≥ 6.
The queen-bee designs are the only ones for which R̃ij ≤ 4 for all i and j,
and so they are the A-optimal designs.
The non-trivial eigenvalues of a queen-bee design are 1, k and v, with
multiplicities b − 1, b(k − 2) and 1, respectively. If the design is not a queen-
bee design, then there is a treatment i that is in more than one block but
not in all blocks. Thus vertex i forms a cutset for the concurrence graph G
which is not joined to every other vertex of G. Cutset Lemma 2 shows that
θ1 < 1. Hence the E-optimal designs are also the queen-bee designs.

10.2 One fewer treatment

If b(k − 1) = v, then the Levi graph G̃ has bk edges and bk vertices, and so it
contains a single cycle, which must be of some even length 2s. If 2 ≤ s ≤ b,
then the design is binary; if s = 1, then there is a single non-binary block,
whose defect is 1. In this case, k ≥ 3, because each block must have more
than one treatment.
For 2 ≤ s ≤ b, let C(b, k, s) be the class of designs constructed as follows.
Start with a loop design for s treatments. Insert k − 2 extra treatments into
each block. The remaining b − s blocks all contain the same treatment from
the loop design, together with k − 1 extra treatments. Figs. 8 and 9 show
the concurrence graph and Levi graph, respectively, of a design in C(6, 3, 6).
For k ≥ 4, the designs in C(b, k, 1) have one treatment which occurs twice
in one block and once in all other blocks, with the remaining treatments all
replicated once. The class C(b, 3, 1) contains all such designs, and also those
in which the treatment in every block is the one which occurs only once in
the non-binary block.

Theorem 21 If b(k − 1) = v, then the D-optimal designs are those in

C(b, k, b).

35
Proof The Levi graph G̃ is unicyclic, so its number of spanning trees is
maximized when the cycle has maximal length. Theorem 5 shows that the
D-optimal designs are precisely those with s = b.

Theorem 22 If b(k−1) = v then the A-optimal designs are those in C(b, k, s),
where the value of s is given in Table 2.

k b 2 3 4 5 6 7 8 9 10 11 12 13
2 2 3 4 5 6 7 8 4 4 4 3 or 4 3
3 2 3 4 5 6 3 3 3 3 3 2 2
4 2 3 4 5 3 2 2 2 2 2 2 2
5 2 3 4 5 2 2 2 2 2 2 2 2
6 2 3 4 2 2 2 2 2 2 2 2 2

Table 2: Value of s for A-optimal designs for b(k − 1) treatments in b blocks

of size k: see Theorem 22

Proof The Levi graph G̃ has one cycle, whose length is 2s, where 1 ≤ s ≤ b.
A similar argument to the one used at the start of the proof of Theorem 16
shows that this cannot be A-optimal unless the design is in C(b, k, s). If
s ≥ 2 or k ≥ 4, then each block-vertex in the cycle has k − 2 treatment-
vertices attached as leaves; all other block-vertices are joined to the same
single treatment-vertex in the cycle, and each has k − 1 treatment vertices
attached as leaves. In C(b, 3, 1) the first type of design has a Levi graph
like this, and the other type has the same multiset of effective resistances
between treatment-vertices, because their concurrence graphs are identical.
The following calculations use the first type.
Let V1 be the set of treatment-vertices in the cycle, V2 the set of other
treatment-vertices joined to blocks in the cycle, and V3 the set of remaining
treatment-vertices. For 1 ≤ i ≤ j ≤ 3, denote by Rij the sum of the pairwise
resistances between vertices in Vi and Vj .
Put
s−1
X 2d(2s − 2d) s2 − 1
R1 = =
d=1
2s 3
and
s−1
X (2d + 1)(2s − 2d − 1) 2s2 + 1
R2 = = .
2s 6
d=0

36
Then R11 = sR1 /2, R12 = s(k − 2)(R2 + s), R13 = (b − s)(k − 1)(R1 + 2s),
R22 = s(k − 2)(k − 3) + s(k − 2)2 [R1 + 2(s − 1)]/2, R23 = (b − s)(k − 1)(k −
2)(R2 +3s), and R33 = (b−s)(k −1)(k −2)+2(b−s)(b−s−1)(k −1)2. Hence
the sum of the pairwise effective resistances between treatment-vertices in the
Levi graph is g(s)/6, where

g(s) = −(k − 1)2 s3 + 2b(k − 1)2 s2 − [6bk(k − 1) − 4k 2 + 2k − 1]s + c

and c = b(k − 1)[12b(k − 1) − 5k − 4].

If s = 1 then the design is non-binary. However,

g(1) − g(2) = (3k − 9 + 6b)(k − 1) − 3,

which is positive, because k ≥ 2 and b ≥ 2. Therefore the non-binary designs

are never A-optimal.
Direct calculation shows that g(2) > g(3) when b = 3, and that g(2) >
g(3) > g(4) when b = 4. These inequalities hold for all values of k, even
though g is not decreasing on the interval [2, 4] for large k when b = 4.
If b = 5 and k ≥ 6, then g(3) > g(2) and g(5) > g(2). Thus the local
minimum of g occurs in the interval (1, 3) and is the overall minimum of g
on the interval [1, 5].
Differentiation gives

g ′ (b) = b(k − 1)[(b − 6)(k − 1) − 6] + 4k 2 − 2k + 1.

If g ′ (b) > 0 then g has a local minimum in the interval (1, b). If, in addition,
g(3) > g(2), then the minimal value for integer s occurs at s = 2. These
conditions are both satisfied if k = 3 and b ≥ 12, k = 4 and b ≥ 8, k ≥ 5 and
b ≥ 7, or k ≥ 9 and b ≥ 6.
Given Theorem 16, there remain only a finite number of pairs (b, k) to
be checked individually to find the smallest value of g(s). The results are in
Table 2.

Theorem 23 If b(k − 1) = v, b ≥ 3 and k ≥ 3, then the E-optimal designs

are those in C(b, k, b) if b ≤ 4, and those in C(b, k, 2) and C(b, k, 1) if b ≥ 5.

Proof If 2 < s < b then the concurrence graph G has a vertex which forms
a vertex-cutset and which is not joined to all other vertices; moreover, G has
no multiple edges. Thus Cutset Lemma 2 shows that θ1 < 1.
Direct calculation shows that θ1 = 1 if s = 1 or s = 2. For k ≥ 4, all
contrasts between singly replicated treatments in the same block are eigen-
vectors of the Laplacian matrix L with eigenvalue k. When k ≥ 3 and s = b

37
the contrast between singly and doubly replicated treatments has eigenvalue
2(k − 1). For s = b, a straightforward calculation shows that the remaining
eigenvalues of L are
  s  
2πn 2 2πn
k − cos ± (k − 1)2 − sin
b b

for 1 ≤ n ≤ b − 1. The smallest of these is

q
k − cos(2π/b) − (k − 12 ) − sin2 (2π/b) :

this is greater than 1 if b = 3 or b = 4, but less than 1 if k ≥ 3 and b ≥ 5.

11 Further reading
The Laplacian matrix of a graph, and its eigenvalues, are widely used, espe-
cially in connection with network properties such as connectivity, expansion,
and random walks. A good introduction to this material can be found in the
textbook by Bollobás [13], especially Chapters II (electrical networks) and
IX (random walks). Connection between the smallest non-zero eigenvalue
and connectivity is described in surveys by de Abreu [1] and by Mohar [42].
In this terminology, a version of Theorem 17 is in [24].
The basic properties of electrical networks can be found in textbooks of
electrical engineering, for example Balabanian and Bickart [9]. A treatment
connected to the multivariate Tutte polynomial appears in Sokal’s survey [50].
Bollobás describes several approaches to the theory, including the fact (which
we have not used) that the current flow minimises the power consumed in
the network, and explains the interactions between electrical networks and
random walks in the network. See also Deo [22].
The connection with optimal design theory was discussed in detail by the
authors in their survey [7]. Further reading on optimal design can be found
in John and Williams [31], Schwabe [47], or Shah and Sinha [48]. For general
principles of experimental design, see Bailey [5].

Acknowledgement This chapter was written at the Isaac Newton Insti-

tute for Mathematical Sciences, Cambridge, UK, during the 2011 programme
on Design and Analysis of Experiments.

38
References
[1] Nair Maria Maia de Abreu, Old and new results on algebraic connectiv-
ity of graphs. Linear Algebra and its Applications, 423 (2007), 53–73.

[2] N. Alon and V. Milman, λ1 , isoperimetric inequalities for graphs, and

superconcentrators. J. Combinatorial Theory Series B 38 (1985), 73–88.

[3] R. A. Bailey, Association Schemes: Designed Experiments, Algebra and

Combinatorics, Cambridge Studies in Advanced Mathematics 84, Cam-
bridge University Press, Cambridge, 2004.

[4] R. A. Bailey, Designs for two-colour microarray experiments, Applied

Statistics 56 (2007), 365–394.

[5] R. A. Bailey, Design of Comparative Experiments. Cambridge Series

in Statistical and Probabilistic Mathematics 25, Cambridge University
Press, Cambridge, 2008.

[6] R. A. Bailey, Variance and concurrence in block designs, and distance in

the corresponding graphs, Michigan Mathematical Journal, 58, (2009),
105–124.

[7] R. A. Bailey and Peter J. Cameron, Combinatorics of optimal designs. In

Surveys in Combinatorics 2009 (ed. S. Huczynska, J. D. Mitchell and C.
M. Roney-Dougal), London Math. Soc. Lecture Notes 365, Cambridge
University Press 2009, pp. 19–73.

[8] R. A. Bailey and Alia Sajjad, Optimal inccomplete-block designs with

nearly minimal replication, in preparation.

[9] N. Balabanian and T. A. Bickart, Electrical Network Theory, Wiley,

1969.

[10] K. Balasubramanian and Aloke Dey, D-optimal designs with minimal

and nearly minimal number of units, Journal of Statistical Planning
and Inference 52 (1996), 255–262.

[11] R. B. Bapat and A. Dey, Optimal block designs with minimal number
of observations, Statist. Probab. Lett. 11 (1991), 399–402.

[12] N. L. Biggs, Potential theory on distance-regular graphs, Combinatorics,

Probability and Computing, 2 (1993), 107–119.

39
[13] B. Bollobás, Modern Graph Theory. Graduate Texts in Mathematics
184, Springer, New York, 1998.

[14] A. E. Brouwer, A. M. Cohen and A. Neumaier, Distance-Regular Graphs,

Ergebnisse der Mathematik und ihrer Grenzgebiete (3), 18, Springer-
Verlag, Berlin, 1989.

[15] Peter J. Cameron, Strongly regular graphs. In Topics in Algebraic Graph

Theory (ed. L. W. Beineke and R. J. Wilson), Cambridge University
Press, Cambridge, 2004, pp. 203–221.

[16] C.-S. Chêng, Optimality of certain asymmetrical experimental designs,

Annals of Statistics 6 (1978), 1239–1261.

[17] C.-S. Chêng, Maximizing the total number of spanning trees in a graph:
two related problems in graph theory and optimum design theory, J.
Combinatorial Theory Series B 31 (1981), 240–248.

[18] C.-S. Cheng, Graph and optimum design theories—some connections

and examples, Bull. Internat. Statistical Institute 49 (1) (Proceedings
of the 43rd Session, Buenos Aires) (1981), 580–590.

[19] C.-S. Cheng, On the optimality of (M.S)-optimal designs in large sys-

tems, Sankhyā 54 (1992), 117–125.

[20] C.-S. Cheng and R. A. Bailey, Optimality of some two-associate-class

partially balanced incomplete-block designs, Annals of Statistics 19
(1991), 1667–1671.

[21] C.-S. Cheng, J. C. Masaro and C. S. Wong, Do nearly balanced multi-

graphs have more spanning trees?, J. Graph Theory 8 (1985), 342–345.

[22] N. Deo, Graph Theory with Applications to Engineering and Computer

Science, Prentice Hall, New Delhi, 1980.

[23] J. Dodziuk, Difference equations, isoperimetric inequality and transience

of certain random walks, Trans. American Math. Soc. 284 (1984), 787–
794.

[24] Shaun M. Fallat, Steve Kirkland and Sukanta Pati, Maximizing alge-
braic connectivity over unicyclic graphs. Linear and Multilinear Algebra
3 (2003), 221–241.

40
[25] R. M. Foster, The average impedance of an electrical network, in Reiss-
ner Anniversary Volume, Contributions to Applied Mathematics (ed.
J. W. Edwards), Ann Arbor, Michigan, 1948, pp. 333–340.

[26] N. Gaffke, Optimale Versuchsplanung für linear Zwei-Faktor Modelle,

PhD thesis, Rheinisch-Westfälische Technische Hochschule, Aachen,
1978.

[27] N. Gaffke, D-optimal block designs with at most six varieties, Journal
of Statistical Planning and Inference 6 (1982), 183–200.

[28] Radoslav Harman, Minimal efficiency of designs under the class of or-
thogonally invariant information criteria, Metrika 60 (2004), 137–153.

[29] J. A. John and T. J. Mitchell, Optimal incomplete block designs, J.

Royal Statistical Soc. Series B 39 (1977), 39–43.

[30] J. A. John and E. R. Williams, Conjectures for optimal block designs,

J. Royal Statistical Soc. Series B 44 (1982), 221–225.

[31] J. A. John and E. R. Williams, Cyclic and Computer Generated Designs

(2nd edition), Monographs on Statistics and Applied Probability 38,
Chapman and Hall, London, 1995.

[32] J. R. Johnson and M. Walters, Optimal resistor networks, in prepara-

tion.

[33] B. Jones and J. A. Eccleston, Exchange and interchange procedures to

search for optimal designs, J. Royal Statistical Soc. Series B 42 (1980),
238–243.

[34] M. K. Kerr and G. A. Churchill, Experimental design for gene expression

microarrays, Biostatistics 2 (2001), 183–201.

[35] J. Kiefer, Construction and optimality of generalized Youden designs. In

A Survey of Statistical Design and Linear Models (ed. J. N. Srivastava),
North-Holland, Amsterdam (1975), pp. 333–353.

[36] G. Kirchhoff, Über die Auflösung der Gleichenung, auf welche man bei
der Untersuchung der linearen Verteilung galvanischer Ströme gefürht
wird, Annals of Physical Chemistry 72 (1847), 497–508.

[37] Olaf Krafft and Martin Schaefer, A-optimal connected block designs
with nearly minimal number of observations, Journal of Statistical Plan-
ning and Inference 65 (1997), 357–386.

41
[38] A. M. Kshirsagar, A note on incomplete block designs, Annals of Math-
ematical Statistics 29 (1958), 907–910.

[39] A. H. M. M. Latif, F. Bretz and E. Brunner, Robustness considerations

in selecting efficient two-color microarray designs, Bioinformatics 25
(2009), 2355–2361.

[40] F. W. Levi, Finite Geometrical Systems, University of Calcutta, Cal-

cutta, 1942. iii + 51 pp.

[41] N. K. Mandal, K. R. Shah and B. K. Sinha, Uncertain resources and

optimal designs: problems and perspectives, Calcutta Statistical Asso-
ciation Bulletin 40 (1991), 267–282.

[42] Bojan Mohar, The Laplacian spectrum of graphs. Graph Theory, Com-
binatorics and Applications, Vol. 2 (Kalamazoo, MI, 1988), 871–898,
Wiley-Interscience, New York, 1991.

[43] J. P. Morgan and S. K. Srivastav, The completely symmetric designs

with blocksize three, J. Statistical Planning and Inference 106 (2002),
21–30.

[44] H. D. Patterson and E. R. Williams, Some theoretical results on general

block designs, Congressus Numerantium 15 (1976), 489–496.

[45] R. M. Pringle and A. A. Rayner, Generalized Inverse Matrices with

Applications to Statistics, Griffin’s Statistical Monographs and Courses,
28, Griffin, London, 1971.

[46] Alia Sajjad, Optimality in sparse block designs, PhD thesis, Quaid-i-
Azam University, Islamabad, 2011.

[47] R. Schwabe, Optimum Designs for Multi-Factor Models, Lecture Notes

in Statistics 113, Springer, New York, 1996.

[48] K. R. Shah and B. K. Sinha, Theory of Optimal Designs, Lecture Notes

in Statistics 54, Springer-Verlag, New York, 1989.

[49] L. W. Shapiro, An electrical lemma, Math. Magazine 60 (1987) 36–38.

[50] Alan D. Sokal, The multivariate Tutte polynomial (alias Potts model) for
graphs and matroids. Surveys in Combinatorics 2005 (ed. B. S. Webb),
London Mathematical Society Lecture Notes 327, Cambridge University
Press, 2005, pp. 173–226.

42
[51] K. D. Tocher, The design and analysis of block experiments, J. Royal
Statistical Soc. Series B 14 (1952), 45–100.

[52] T. Tjur, Block designs and electrical networks, Annals of Statistics 19

(1991), 1010–1027.

[53] E. Wit, A. Nobile and R. Khanin, Near-optimal designs for dual channel
microarray studies, Applied Statistics 54 (2005), 817–830.

[54] F. Yates, Incomplete randomized designs, Annals of Eugenics 7 (1936),

121–140.

[55] Frank Yates, A new method of arranging variety trials involving a large
number of varieties, Journal of Agricultural Science 26 (1936), 424–455.

43