The Annals of Statistics

2007, Vol. 35, No. 2, 747–771

DOI: 10.1214/009053606000001253
c Institute of Mathematical Statistics, 2007

RESOLVABLE DESIGNS WITH LARGE BLOCKS

arXiv:0708.1896v1 [math.ST] 14 Aug 2007

By J. P. Morgan1 and Brian H. Reck2

Virginia Tech and University of Pittsburgh

Dedicated to the memory of I. M. Chakravarti

Resolvable designs with two blocks per replicate are studied from
an optimality perspective. Because in practice the number of repli-
cates is typically less than the number of treatments, arguments can
be based on the dual of the information matrix and consequently
given in terms of block concurrences. Equalizing block concurrences
for given block sizes is often, but not always, the best strategy. Suffi-
cient conditions are established for various strong optimalities and a
detailed study of E-optimality is offered, including a characterization
of the E-optimal class. Optimal designs are found to correspond to
balanced arrays and an affine-like generalization.

1. Introduction. Block designs arise in comparative experimentation as

fundamental devices for improving efficiency when working with heteroge-
neous experimental units. The blocks are simply a partition of the units into
(say) b sets exhibiting homogeneity within sets. Given the blocks, of sizes
k1 , k2 , . . . , kb , a block design is an assignment of v treatments to the bj=1 kj
P

units. Optimality theory for block designs attempts to determine which of

the many possible assignments is in some sense best.
In some applications there are restrictions on the collection of possible as-
signments. A block design is resolvable if the blocks can be partitioned into
replicates, defined as sets of blocks with the property that each treatment
is assigned to one unit in each set. The practical impact of, and motivation
for, resolvability is to gain orthogonality between treatments and nuisance
factors of concern. For instance, resolvability in sequential experimentation,

Received February 2005; revised February 2006.

1
Supported by NSF Grant DMS-01-04195.
2
Supported by NIMH Training Grant 5T32MH020053-05.
AMS 2000 subject classifications. Primary 62K05; secondary 62K10, 05B05.
Key words and phrases. Resolvable block design, optimal design, dual design, balanced
array.

This is an electronic reprint of the original article published by the

Institute of Mathematical Statistics in The Annals of Statistics,
2007, Vol. 35, No. 2, 747–771. This reprint differs from the original in pagination
and typographic detail.
1
2 J. P. MORGAN AND B. H. RECK

with replicates corresponding to time periods, is used to mitigate time ef-

fects. Resolvability can likewise be useful in multi-site experiments and in
experiments with multiple individuals handling experimental runs. Notably,
the United Kingdom has for some time required the use of resolvable designs
in agricultural field trials (see [20]).
The combinatorial study of resolvability in block designs goes back at
least as far as the well-known Kirkman’s [14] schoolgirl problem. The notion
entered the statistical lexicon with Yates’ work on square lattice designs [29,
30], although the term “resolvable design” was introduced by Bose [4]. Yates’
lattice designs were extended to rectangular lattices by Harshbarger [11, 12];
see also [2]. Williams [27] and Patterson and Williams [21] introduced a
large family of resolvable designs which they termed α-designs. Williams,
Patterson and John [28] derived resolvable designs with two replicates from
BIBDs (balanced incomplete block designs). Bailey, Monod and Morgan
[1] proved strong optimality for the affine resolvable designs introduced by
Bose [4]. Resolvable BIBDs have received significant attention in both the
combinatorial and the statistical literature; for a summary, see Morgan [17],
who surveys the major classes of resolvable designs with many references.
With α-designs, Patterson and Williams [21] provided a flexible method
for obtaining reasonably efficient resolvable designs for a wide range of values
v, constant block size k and replication r. They also adapted their method
to obtain resolvable designs with two different block sizes, k and k − 1, a
first attempt at addressing the obvious restriction that resolvability with
equal block sizes can be achieved only for v a multiple of k. John, Russell,
Williams and Whitaker [13], in revisiting that idea, concluded that the α-
technique for two block sizes could produce relatively inefficient designs for
small v and recommended an interchange algorithm for construction of de-
signs with better efficiency. John et al. [13] also discussed the practical need
for resolvable designs with unequal block sizes; for example, about half of
245 experiments examined by Patterson and Hunter [19] had unequal block
sizes. See also [20].
To the authors’ knowledge, the literature contains no systematic work on
determining optimal resolvable block designs when block sizes need not be
equal. This paper will undertake such work, for the special case of two blocks
per replicate. Only with two blocks in each replicate must a block be large in
the sense of containing at least half of the treatments. And with two blocks
per replicate, the block sizes must be unequal for any odd v (although this
work is not restricted to odd v).
Let D(v, r; k1 , k2 ) denote the class of all resolvable block designs with r
replicates of v treatments, each replicate consisting of two blocks, of sizes
k1 and k2 , where k1 + k2 = v. Of special interest is k1 = k, k2 = k − 1 for
odd v = 2k − 1, but here no restrictions are placed on the two block sizes for
the general theory. Let k1 denote the larger block size, k1 ≥ k2 , so that with
 RESOLVABLE DESIGNS 3

Table 1
A resolvable design in D(9, 5; 5, 4)

1 6 1 2 1 4 1 3 1 2
2 7 4 3 2 5 2 4 3 5
3 8 5 6 3 6 5 6 4 6
4 9 7 8 7 8 8 7 7 9
5 9 9 9 8

no loss of generality v2 ≤ k1 ≤ v − 2. In the most common applications of

resolvable designs, the number of treatments is large relative to the number
of replicates; here, r ≤ v − 1 is required, allowing optimality problems to
be more easily attacked using the dual of the information matrix. These
considerations frame the goal of this paper: to determine the best design
d ∈ D(v, r; k1 , k2 ). An example of a resolvable design in D(9, 5; 5, 4) is shown
in Table 1, with the blocks written as columns. Later, this design will be
proven optimal with respect to many useful criteria.
For any resolvable design d ∈ D, ignoring the replicate grouping leaves an
underlying simple block design for v treatments in 2r blocks. If the roles of
blocks and treatments are reversed in this underlying design, another simple
block design with 2r treatments in v blocks of size r is produced, called the
dual design.
Section 2 provides optimality background and establishes equivalences
with the dual problem. Section 3 finds conditions for global optimality of
designs having equality of block concurrences; for those who wish to jump
ahead, the main results there are Theorems 4–7. Section 4 characterizes
the E-optimal designs (Theorem 14) and finds the Schur-best of the E-
optimal designs (Theorem 15). The main results are applied to special cases
with k1 − k2 ≤ 2 in Section 5 and designs for these cases are constructed in
Section 6. Concluding comments appear in Section 7.

2. Model, information and optimality criteria. Let yhjl denote the yield
from the lth experimental unit in block j of replicate h. Thus, the triples
(h, j, l) identify the experimental units and the design d corresponds to a map
d[h, j, l] from the units to the set of treatments. The standard linear model for
the yields incorporates a mean effect µ, replicate effects ρh , block effects βhj ,
treatment effects τd[h,j,l] and mean zero, uncorrelated, equivariable random
error terms ehjl :
yhjl = µ + ρh + βhj + τd[h,j,l] + ehjl ,
h = 1, . . . , r; j = 1, 2; l = 1, . . . , kj . This model may be written in matrix
terms as
y = µ1 + (Ir ⊗ 1v )ρ + Lβ + Xd τ + e,
4 J. P. MORGAN AND B. H. RECK

where, with the yhjl lexicographically ordered in the vr × 1 yield vector y,

the block incidence matrix is Lvr×2r = Ir ⊗ 10k1 01k1 and the design matrix


k2 k2
is the vr × v incidence matrix Xd , for which row (h, j, l) has a 1 in column
i if and only if d[h, j, l] = i [i.e., unit (h, j, l) receives treatment i] and all
other entries are zero. Replicate effects, block effects, treatment effects and
error vectors are ρr×1 , β2r×1 , τv×1 and evr×1 , respectively. Choice of design
is equivalently choice of Xd .
Linear model theory says that the information matrix Cd for estimation
of the treatment effects τ is
(1) Cd = Xd′ [I − L(L′ L)−1 L′ ]Xd = rI − Nd Ds−1 Nd′ ,
where Ds = L′ L = Diag(k1 , k2 , k1 , k2 , . . . , k1 , k2 ) is the diagonal matrix of
block sizes. The v × 2r matrix Nd is the treament/block incidence matrix.
The general off-diagonal element (Nd Nd′ )i,i′ of Nd Nd′ is the number of blocks
to which both treatments i and i′ are assigned, called a treatment concur-
rence count. Notice that the replicate incidence matrix (Ir ⊗ 1v ) plays no
role in (1); the same form of information matrix is obtained for any sim-
ple block design. It follows immediately that the information matrix for the
corresponding dual design is
1
(2) Cdual = Ds − Nd′ Nd .
r
The off-diagonal elements of Nd′ Nd are block concurrence counts.
All treatment contrasts are estimable with design d if and only if Cd has
rank v − 1. Any such d is said to be connected; only connected designs are
considered here. Most (but not all) commonly employed optimality criteria,
including those to be used here, are functions of the v − 1 nonzero eigenvalues
of Cd . These will be ordered and labeled zd1 ≤ zd2 ≤ · · · ≤ zd,v−1 .
Much of the optimality work below focuses on minimizing functions of
the form
v−1
X
(3) ψf (zd ) = f (zdi ),
i=1

where f is convex and zd is the vector of nonzero eigenvalues. If f in (3) is

Schur-convex (Schur-convex functions include the convex functions; see [3],
Section II.3), then (3) is said to be a Schur-criterion. Design d1 is Schur-
better than design d2 (d2 is Schur-inferior to d1 ) if ψf (zd1 ) ≤ ψf (zd2 ) for
all Schur-convex f with strict inequality for at least one such f . A design
optimal with respect to all (i.e., minimizing all) Schur-convex criteria is said
to be Schur-optimal. If f in (3) satisfies (i) f is continuously differentiable
on (0, maxd∈D tr(Cd )) with f ′ < 0, f ′′ > 0 and f ′′′ < 0 and (ii) limx→0 f (x) =
∞, then (3) is said to be a type-1 criterion (see [7]). A design optimal
 RESOLVABLE DESIGNS 5

with respect to all type-1 criteria is said to be type-1-optimal. One popular

criterion belonging to both families just defined is the A-criterion specified
by f (x) = x1 . A criterion not of the form (3) (though it can be written as a
limit of such criteria) is

1
(4) ψE (zd ) = ,
zd1

called the E-criterion. An E-optimal design minimizes (4) or, equivalently,

maximizes zd1 . For a broader discussion of optimality criteria and their sta-
tistical meanings, see [24].
For the current endeavor, it is advantageous to approach the zdi , and con-
sequently (3) and (4), through Cdual . How this is done is shown in Lemma 1
below. Let φdhh′ be the block concurrence for (i.e., the number of treat-
ments common to) the blocks of size k1 in replicates h and h′ . For h 6= h′ ,
k12
let φ∗dhh′ = φdhh′ − k1 +k 2
and define the (symmetric) optimality matrix Md
by
k1 k2
 
 v φ∗d12 φ∗d13 ··· φ∗d1r 
k1 k2
 
φ∗d23 ∗
 
 ··· φd2r 
 v 
k1 k2
 
(5) Md = 
 ··· ∗
φd3r 
,
 v 

.. .. 
. . 
 

 k1 k2 
v
with eigenvalues ed1 ≥ ed2 ≥ · · · ≥ edr .

Lemma 1. For any d ∈ D(v, r; k1 , k2 ), the eigenvalues of Cd are 0, max{0,

v − r − 1} copies of r and (r − kve1 kd12 , r − kve1 kd22 , . . . , r − kve1 dr
k2 ). Consequently,
a ψf -optimal design minimizing (3) will equivalently minimize rh=1 f (r −
P
vedh
k1 k2 ) and an E-optimal design minimizing (4) will equivalently minimize
ed1 .

Lemma 1 is proved in Appendix A.1. Provided r ≤ v − 1 (as earlier re-

quired in Section 1), Cd has v − r − 1 eigenvalues fixed at r. Working with the
dual through Md not only makes this evident, but allows these structurally
fixed eigenvalues to be easily set aside.
With the optimality problem recast in terms of Md and its eigenvalues, a
crucial concept for proofs of Schur-optimality is now defined.
6 J. P. MORGAN AND B. H. RECK

Definition. Let {x }n and {yi }ni=1 be nonincreasing sequences of real

Pn i i=1 Pn
numbers such that i=1 xi = i=1 yi . If
l
X l
X
xi ≤ yi for all 1 ≤ l ≤ n
i=1 i=1

or, equivalently,
n−l+1
X n−l+1
X
xi ≥ yi for all 1 ≤ l ≤ n,
i=n i=n

then {yi }ni=1 is said to majorize {xi }ni=1 .

The importance of majorization is evident in the following result (see,

e.g., [3], page 40).

Theorem 2. Let {x }n and {yi }ni=1 be P

Pn i i=1 Pn
nonincreasingPsequences of real
numbers such that i=1 xi = i=1 yi . Then ni=1 f (xi ) ≤ ni=1 f (yi ) for all
real-valued convex functions f if and only if {yi }ni=1 majorizes {xi }ni=1 .

If the sequences {xi }ni=1 and {yi }ni=1 are written as the elements of vectors
x and y, then the fact that y majorizes x is written as y ≻ x or x ≺ y.
Let ed and ed∗ be the vectors of eigenvalues for the optimality matrices
for designs d and d∗ , respectively. Lemma 3 lists simple majorization facts
([3], page 30) used in subsequent sections. For majorization comparisons,
the third of these allows work directly with the edh rather than the r − kve1dh k2
(see Lemma 1). Thus, if ed ≻ ed∗ and the two vectors are not identical, then
d∗ is Schur-better than d. Design d∗ is Schur-optimal if ed ≻ ed∗ for every
d ∈ D.
Pn Pn
Lemma 3. For real numbers {xi }ni=1 and {yi }ni=1 with i=1 xi = i=1 yi ,

(i) if x1 ≥ x2 = x3 = · · · = xn and y1 ≥ x1 , then {yi }ni=1 majorizes {xi }ni=1 ;

(ii) if x1 = x2 = · · · = xn−1 ≥ xn and xn ≥ yn , then {yi }ni=1 majorizes
{xi }ni=1 ;
(iii) if {yi }ni=1 majorizes {xi }ni=1 , then {a − ybi }ni=1 majorizes {a − xbi }ni=1
for any real a, b.

3. Equal concurrence designs and global optimality. Among simple block

designs, the BIBDs are Schur-optimal, a result which follows from equality of
treatment concurrences inducing complete symmetry of the information ma-
trix. The analogous notion for duals is equality of block concurrences, which
this section explores for utility with resolvable designs. A resolvable design
 RESOLVABLE DESIGNS 7

d ∈ D(v, r; k1 , k2 ) having block concurrence counts φd12 = φd13 = φd23 = · · · =

φd,r−1,r = θ for some k1 − k2 ≤ θ ≤ k1 is called an equal concurrence design
with common concurrence θ, or ECD(θ). The subsections below will explore,
in turn, (i) ECDs with equality of eigenvalues in the optimality matrix, (ii)
other ECDs which can be proven Schur-optimal and (iii) Schur-domination
of designs with unequal block concurrences by one or more ECDs.

3.1. Schur-optimality via equality of eigenvalues. For an ECD(θ), the

optimality matrix (5) is
p k2 k2
    
(6) Md = − θ− 1 I + θ − 1 J,
v v v
where I is the r × r identity matrix, J is the r × r matrix of ones and
p = k1 k2 is the product of the block sizes. Like a BIBD information matrix,
it is completely symmetric, but unlike that matrix, Md for an ECD(θ) is
nonsingular and thus can have two distinct, relevant eigenvalues, rather than
just one.

Theorem 4. Suppose that D(v, r; k1 , k2 ) is a resolvable design setting

for which (k1 + k2 )|k12 and define
k12 k2
(7) θ∗ = = 1.
k1 + k2 v
Then ECD(θ ∗ )’s in D are Schur-optimal whenever they exist.

k2
Proof. Given the conditions on D, the inequalities k1 − k2 ≤ k1 +k 1
2
≤ k1
∗
imply that θ = θ is an admissible value for the common block concurrence
of an ECD(θ). For θ = θ ∗ , Md in (6) is vp I. Since the eigenvalues of Md are all
identical, they are majorized by the eigenvalues of every competing design.


Corollary 3.4 of [1] established that affine-resolvable designs are Schur-

optimal. Theorem 4 generalizes that result when there are two blocks per
replicate. Here, the optimality condition is that all pairs of blocks of size
k1 have the same concurrence (7). When k1 = k2 and 2|k1 , ECD(θ ∗ )’s are
affine-resolvable designs. It is obvious that (6) has two distinct eigenvalues
for any θ 6= θ ∗ .

Example. Consider the setting D(9, 4; 6, 3). Then (k1 + k2 )|k12 with θ ∗ =
4 and if an ECD(4) exists, it is Schur-optimal. In fact, an ECD(4) does exist
and is shown in Table 2.
The settings for which k12 is a multiple of k1 + k2 are relatively sparse
(a situation much like that of BIBDs relative to all simple block design
8 J. P. MORGAN AND B. H. RECK

Table 2
A Schur-optimal ECD( 4)
in D(9, 4; 6, 3)

1 7 1 5 1 3 3 1
2 8 2 6 2 4 4 2
3 9 3 9 5 9 5 9
4 4 6 6
5 7 7 7
6 8 8 8

settings). For the 1225 pairs 2 ≤ k2 < k1 ≤ 51, only 23 meet the divisibil-
ity requirement implied by (7). Theorem 4 is thus only a start, albeit an
important one.

3.2. Global optimality of other ECDs. Good designs are expected to have
eigenvalue structures “close” to that of ECD(θ ∗ )’s, suggesting this question:
Is some equal concurrence design Schur-optimal when (k1 + k2 )|/k12 ? To in-
vestigate this question, define the block concurrence parameter θ̄ by
k12
 
(8) θ̄ = int
k1 + k2
and write
k12
(9) γ= − θ̄.
k1 + k2
Then 0 ≤ γ < 1 and a necessary condition for existence of ECD(θ ∗ ) is γ = 0.
Consequently, γ is called the block discordancy coefficient; it measures the
departure of the block sizes from that required for equality of all eigenvalues.
The parameter γ will play a pivotal role in the remainder of this paper, as
will the concurrence discrepancies defined next.
Define the block concurrence discrepancy matrix ∆d = (δdhh′ ), where
if h 6= h′ ,

φdhh′ − θ̄,
δdhh′ =
0, if h = h′ .
For h 6= h′ , the off-diagonal elements δdhh′ will be referred to as block con-
currence discrepancies. Rewritten in terms of block discrepancies and the
discordancy coefficient, the general optimality matrix (5) becomes
p
(10) Md = I − γ(J − I) + ∆d .
v
For ECD(θ ∗ ), γ = 0, ∆d = 0 and Md = vp I. If γ 6= 0, (10) shows that the form
of optimal design may depend both on the magnitude of γ and the values
 RESOLVABLE DESIGNS 9

of the discrepancies. Matrix (10) is completely symmetric for, and only for,
ECDs, in which case ∆d is an integer multiple of J − I.
The ECDs which are combinatorially closest to ECD(θ ∗ ) are those with
either θ = θ̄ or θ = θ̄ + 1. Not surprisingly, these are strong competitors in the
optimality race. ECD(θ̄)s have φdhh′ = θ̄ for each 1 ≤ h 6= h′ ≤ r, so ∆d = 0
and the eigenvalues of Md are
p p
(11) ξ1 (γ) = + γ and ξ2 (γ) = − (r − 1)γ,
v v
with frequencies r − 1 and 1, respectively. The two eigenvalues satisfy ξ1 (γ) ≥
ξ2 (γ). If φdhh′ = θ̄ + 1 for every h 6= h′ , then ECD(θ̄ + 1)’s have ∆d = (J − I)
and the eigenvalues of Md are
p
ξ1 (γ − 1) = − (1 − γ) and
v
(12)
p
ξ2 (γ − 1) = + (r − 1)(1 − γ),
v
with frequencies r − 1 and 1, respectively, and with ξ2 (γ − 1) ≥ ξ1 (γ − 1).

Theorem 5. For 0 ≤ γ ≤ 21 , ECD(θ̄)’s are type-1-optimal and E-optimal.

Proof. The eigenvalues of the information matrix for any design d ∈

D(v, r; k1 , k2 ) are 0 < zd1 ≤ zd2 ≤ · · · ≤ zdr and v − r − 1 copies of r, and
Pr ¯
i=1 zdi = r(r − 1) is constant for all designs in D. For an ECD(θ̄), call it d,
¯ , following from (11), have the form zd1
the zdi ¯ = zd2
¯ = · · · = zd,v−2
¯ ≤ zd,v−1
¯ .
Theorem 2.3 of [7] thus gives the result if it can also be shown that d ¯
minimizes v−1 2
P
i=1 zdi over D.
For d ∈ D(v, r; k1 , k2 ) with optimality matrix (Md )hh′ = (δdhh′ − γ) having
trace tr Md = pr v ,
v−1 r  2
vedh
tr Cd2 2 2
X X
= zdi = (v − r − 1)r + r−
i=1 h=1
p

2vr v2
= (v − 1)r 2 − tr Md + 2 tr Md2
p p
2v 2 X X
= (v − 3)r 2 + r + (δdhh′ − γ)2 ,
p2 h<h′

so that tr Cd2 is minimized by designs that minimize 2

PP
h<h′ (δdhh′ − γ) .
1
Since δdhh′ is integral, the unique minimum of tr Cd2 on 0 ≤ γ < 2 is at
δdhh′ ≡ 0, achieved only by ECD(θ̄). For γ = 12 , any values δdhh′ ∈ {0, 1}
minimize tr Cd2 . 
10 J. P. MORGAN AND B. H. RECK

Now define the F-criterion as the value of the largest eigenvalue of Cd

that is not constrained by the setting to equal r, that is,
vedr
ψF (Cd ) = zdr = r − .
p
Minimizing ψF (Cd ) over D is equivalent to maximizing edr . This criterion
can be important in establishing Schur-optimality, as shown next.

Theorem 6. An ECD(θ̄) is Schur-better than a competitor with a dif-

ferent set of eigenvalues if and only if it is F-equivalent or better than that
competitor. Consequently, ECD(θ̄)’s are Schur-optimal if and only if they
are F-optimal.

Proof. Follows from (11) and Lemma 3(i). 

A result of a similar flavor holds for ECD(θ̄ + 1)’s using the E-criterion.
As pointed out by Kunert [15], page 385, designs with eigenvalues zdi in the
form of Lemma 3(ii) are Schur-best whenever they are E-optimal. For the
current problem, this is stated as follows.

Theorem 7. An ECD(θ̄ + 1) is Schur-better than a competitor with a

different set of eigenvalues if and only if it is E-equivalent or better than
that competitor. Consequently, ECD(θ̄ + 1)’s are Schur-optimal if and only
if they are E-optimal.

In establishing necessary and sufficient conditions for Schur-optimality

for their respective ECDs in terms of a single eigenvalue, Theorems 6 and 7
provide simple tests for comparing these designs to any other design. An im-
mediate consequence is that an ECD(θ) with θ ∈ / {θ̄, θ̄ + 1} is Schur-inferior
to at least one of these two competitors. Thus, among ECDs, only these
two competitors remain. They are compared to one another in the following
corollary.

Corollary 8. ECD(θ̄)’s are Schur-better than ECD(θ̄ + 1)’s if and

only if γ ≤ 1r , and ECD(θ̄ + 1)’s are Schur-better than ECD(θ̄)’s if and only
if γ ≥ r−1
r .

Proof. Simply use (11) and (12) in applying Theorems 6 and 7. 

Though design nonexistence can play a role, Corollary 8 says that one
does not expect to find a Schur-optimal design for 1r ≤ γ ≤ r−1
r (Theorem 5
makes for an interesting juxtaposition). Schur-domination can nonetheless
be used to eliminate many competitors, as shown next.
 RESOLVABLE DESIGNS 11

3.3. Schur-inferiority of designs lacking equality of block concurrences.

Given the results of the preceding subsection, the remaining question from a
global optimality perspective is if (and when) designs outside the ECD class
can be preferable. This question can be effectively pursued by application
of Theorems 6 and 7, once workable bounds for the largest and smallest
eigenvalues ed1 and edr of the optimality matrix are in place.

Lemma 9. Let d ∈ D(v, r; k1 , k2 ) have concurrence discrepancy matrix

∆d = (δdhh′ ) and optimality matrix Md . Then:
p p
(i) δd12 ≤ 0 ⇒ ed1 ≥ v + γ − δd12 and edr ≤ v − γ + δd12 ;
p p
(ii) δd12 > 0 ⇒ ed1 ≥ v − γ + δd12 and edr ≤ v + γ − δd12 .

Proof. The leading 2× 2 minor of Md , which is Md11 = ( vp + γ − δd12 )I −

(γ − δd12 )J , has eigenvalues vp + γ − δd12 and vp − γ + δd12 . A Sturmian sep-
aration theorem ([22], page 64) provides the bounds. 

Define design d to be an almost equal concurrence design, or AECD, if

φdhh′ ∈ {θ̄, θ̄ + 1} for all h 6= h′ and if each value is attained for some h, h′ .
If any φdhh′ is not in {θ̄, θ̄ + 1}, then d is an unequal concurrence design, or
UCD. In terms of discrepancies, AECDs have all δdhh′ ∈ {0, 1}, while UCDs
have some δdhh′ ≤ −1 or ≥ 2. Depending on γ, AECDs can be optimal in at
least some senses (see Section 4), necessarily ruling out global optimality of
ECDs for some γ. The next few results will show that ECDs often dominate
UCDs.

Corollary 10. Suppose d ∈ D(v, b; k1 , k2 ) is a UCD with δdhh′ ≤ −α

for some 1 ≤ h 6= h′ ≤ r and some integer α ≥ 1. Then:
α
(i) ECD(θ̄)’s are Schur-better than d if γ ≤ r−2 ;
(ii) ECD(θ̄ + 1)’s are Schur better than d if γ ≥ r−α−1
r .

Proof. For the UCD d as described in the corollary, take δd12 ≤ −α.
Then from Lemma 9, ed1 ≥ pv + γ − α and vp − γ + α ≥ edr . By Theorem 6,
an ECD(θ̄) is Schur-better than d if ξ2 (γ) ≥ pv − γ + α ≥ edr ⇐⇒ γ ≤ r−2
α
.
p
By Theorem 7, an ECD(θ̄ + 1) is Schur-better than d if ed1 ≥ v + γ − α ≥
ξ2 (γ − 1) ⇐⇒ γ ≥ r−α−1
r . 

Corollary 11. When r ≤ 4, all UCDs with some δdhh′ ≤ −1 are Schur-
inferior to an ECD, and when r = 5 or 6, UCDs with some δdhh′ ≤ −2 are
Schur-inferior to an ECD.

The next two corollaries are similarly shown.

12 J. P. MORGAN AND B. H. RECK

Corollary 12. Suppose d ∈ D(v, b; k1 , k2 ) is a UCD with δdhh′ ≥ α for

some 1 ≤ h 6= h′ ≤ r and some integer α ≥ 2. Then:
(i) ECD(θ̄)’s are Schur-better than d if γ ≤ αr ;
r−α−1
(ii) ECD(θ̄ + 1)’s are Schur better than d if γ ≥ r−2 .

Corollary 13. When r ≤ 4, all UCDs with some δdhh′ ≥ 2 are Schur-
inferior to an ECD, and when r = 5 or 6, UCDs with some δdhh′ ≥ 3 are
Schur-inferior to an ECD.

Corollaries 11 and 13 say that optimal (with respect to any convex cri-
terion) designs in settings D(v, r; k1 , k2 ) with r ≤ 4 must be an ECD(θ̄), an
ECD(θ̄ + 1) or an AECD. Optimal designs in settings with r = 5 or 6 must
have block concurrence discrepancies δdhh′ ∈ {−1, 0, 1, 2} for all 1 ≤ h 6= h′ ≤
r. It is unlikely that such global statements can be much improved. The im-
portance of these results lies in the prevalence of small r in the application
of resolvable designs.

4. E-optimality. Sufficient conditions, which are necessary given exis-

tence, will be developed for E-optimality of designs in D(v, r; k1 , k2 ). The
main results, Theorems 14 and 15, will be stated after first defining a sub-
class where E-optimal designs will be found.
Corollary 16 below will remove the UCDs from E-contention, so that only
ECD(θ̄)’s, ECD(θ̄ + 1)’s and AECDs need be considered. These designs have
all δdhh′ ∈ {0, 1} and so have ∆d which is the adjacency matrix of a simple,
undirected graph on r vertices. Any of these designs for which ∆d − J is
(with suitable ordering of replicates) of the form
−Jt1
 
 −Jt2 
(13)
 
 .. 
 . 
−Jtn
Pn
for positive integers n and t1 ≤ t2 ≤ · · · ≤ tn ( i=1 ti = r) is said to be group-
affine. For group-affine designs, concurrences φdhh′ are constant (= θ̄) within
groups of sizes t1 , . . . , tn and are constant (= θ̄ + 1) between groups. A group-
affine design is said to be uniform if, for given n, the range of group sizes
ti is at most one. For any group-affine design with the number of groups
1
n ≤ nγ = int( 1−γ ), let t(d) denote the vector of its group sizes ti arranged
in increasing order and with nγ elements, padding with zeros as necessary.
For example, if r = 7, γ = 79 and d has ti values 1, 3 and 3, then nγ = 4 and
t(d) = (0, 1, 3, 3). Now the main results can be stated. The phrase “E-Schur-
optimal” means “Schur-optimal within the class of all E-optimal designs.”
 RESOLVABLE DESIGNS 13

Theorem 14. Any group-affine design is E-optimal if n ≤ nγ and γ ≤

r−1
r . If any such design exists, then all such designs form the class DE of
all E-optimal designs. ECD(θ̄ + 1)’s are Schur-optimal for γ ≥ r−1
r .

Theorem 15. For d1 , d2 ∈ DE , if t(d1 ) ≺ t(d2 ) , then d1 is Schur-better

than d2 . Design d∗ ∈ DE is E-Schur-optimal if it is uniform with n = nγ .

Theorem 14 characterizes the class of E-optimal block concurrence struc-

tures for D(v, r; k1 , k2 ). Although not much discussed in the literature, the
class of E-optimal designs in a given simple block design setting often con-
tains a variety of designs with different information matrices (see, e.g., [18]
and [26]). Theorem 14 reveals the same situation for resolvable design set-
tings. This allows other design criteria to be brought to bear: one should
choose the best of the E-optimal designs. Theorem 15 does this. In a very
strong sense, E-Schur-optimal designs are the best of the E-optimal designs.
The proofs of Theorems 14 and 15 depend on a series of technical results
that are developed in the remainder of this section. So as not to overly
disrupt the flow of the main line of proof, the longer of the “sub-proofs” are
delayed until Appendix A. The first task is to rule out UCDs.

Corollary 16. For all r ≥ 2 and 0 ≤ γ < 1, ECD(θ̄)’s are E-better

than UCDs.

Proof. The maximum eigenvalue of the optimality matrix for an ECD(θ̄)

is ξ1 (γ) = pv + γ. For the UCD d, suppose that δd12 ≤ −α for some integer
α ≥ 1. Then by Lemma 9, ed1 ≥ vp + γ − δd12 > ξ1 (γ) and ECD(θ̄)’s are E-
better than d. If δd12 ≥ α for some integer α ≥ 2, then another application of
the lemma gives ed1 ≥ pv − γ + δd12 > ξ1 (γ) and, again, ECD(θ̄)’s are E-better
than d. 

Completing the proof of Theorem 14 is a matter of minimizing ed1 over

all d for which every δdhh′ ∈ {0, 1}. Before providing the details, here is a
sketch of what will be done. First, a lower bound for ed1 is found in terms
of the largest eigenvalue of a specific projection P ∆d P of the discrepancy
matrix. This bound implies that ECD(θ̄)’s are E-superior to all designs
for which P ∆d P has a positive eigenvalue. The next step is thus to find
a necessary and sufficient condition on ∆d so that P ∆d P is nonpositive
definite; this condition turns out to be exactly the definition of group-affine
design. Finally, the general form of Md for group-affine designs is examined
in detail to determine which of these competitors are E-optimal, producing
the conditions of Theorem 14.

Lemma 17. For d ∈ D(v, r; k1 , k2 ) and with P = (I − 1r J), let ud1 and udr
be the maximum and minimum eigenvalues of P ∆d P , respectively. Then:
14 J. P. MORGAN AND B. H. RECK

p
(i) if ud1 > 0 then ed1 ≥ v + γ + ud1 ;
(ii) edr ≤ vp + γ + udr .

The proof of Lemma 17 appears in Appendix A.2. Lemma 17 combined

with Theorems 6 and 7 immediately gives the following corollary:

Corollary 18. For d ∈ D(v, r; k1 , k2 ) and with P = (I − 1r J), let ud1

and udr be the maximum and minimum eigenvalues of P ∆d P .
(i) If γ < − urdr , then ECD(θ̄)’s are Schur-better than d.
(ii) If ud1 > 0 and γ > ( r−urd1 −1 ), then ECD(θ̄ + 1)’s are Schur-better
than d.
(iii) If ud1 > 0, then ECD(θ̄)’s are E-better, but not necessarily Schur-
better, than d.

Parts (i) and (ii) of Corollary 18 provide alternative routes (cf. Corollar-
ies 10 and 12) for establishing Schur-domination of ECDs. Part (iii) is the
key part of the proof of Theorem 14. If d is not to be eliminated by ECD(θ̄),
then P ∆d P cannot have a positive eigenvalue, that is, must be nonpositive
definite. Recall that the current E-competitors are all d for which ∆d is the
adjacency matrix of a simple, undirected graph.

Lemma 19. Let A be the adjacency matrix for a simple undirected graph
of r vertices. PAP is nonpositive definite if and only if A − J may be writ-
ten (possibly after vertex permutation) in the form (13) for some positive
integers n and t1 ≤ t2 ≤ · · · ≤ tn with ni=1 ti = r.
P

Combining Lemma 19 with Corollaries 16 and 18(iii), it has now been

shown that existence of an ECD(θ̄) implies that an E-optimal design is a
group-affine design. Group-affine designs include the ECD(θ̄)’s (put n = 1,
∆d = 0) and the ECD(θ̄ + 1) (put n = r, ∆d = J − I) at the extremes for n.
The proof of Lemma 19 appears in Appendix A.3.
It remains to determine which of the group-affine designs are actually E-
best, requiring knowledge of ed1 for this class. For any group-affine design,
write Md = ( vp + γ)I + ∆d − γJ , where, for given t1 ≤ · · · ≤ tn ,
γJt1 ,t1 (γ − 1)Jt1 ,t2 ··· (γ − 1)Jt1 ,tn
 
 (γ − 1)Jt2 ,t1 γJt2 ,t2 ··· (γ − 1)Jt2 ,tn 
∆d − γJ = (−1)  .. .. ..
 
.. 
 . . . . 
(γ − 1)Jtn ,t1 (γ − 1)Jtn ,t2 ··· γJtn ,tn
(14)
≡ (−1)Hd .
 RESOLVABLE DESIGNS 15

An E-optimal design will maximize the minimum eigenvalue of Hd . Clearly,

Hd has ni=1 (ti − 1) = r − n eigenvalues of zero (corresponding to eigen-
P

vectors which are orthogonal contrasts within groups of sizes t1 , . . . , tn ). So

all of these designs have Hd with at least one eigenvalue of zero, except
ECD(θ̄ + 1), for which the eigenvalues of Hd are 1 (frequency r − 1) and
1 + r(γ − 1). ECD(θ̄ + 1) is therefore E-optimal if 1 + r(γ − 1) ≥ 0, that is,
if γ ≥ r−1
r (in which case it is Schur-optimal; see Theorem 7). Otherwise, all
designs for which Hd is nonnegative definite are E-optimal. Needed now are
the eigenvalues of Hd other than zero. Let Dt be the diagonal matrix with
diagonal elements (t1 , . . . , tn ). Lemma 20 is proved in Appendix A.4.

Lemma 20. The eigenvalues of Hd specified by (14) are 0 (with fre-

1/2 1/2
quency r − n) and the eigenvalues of Dt EDt , where E = In − (1 − γ)Jn .

1/2 1/2
Consequently, Hd is nonnegative definite if and only if Dt EDt is non-
negative definite. This is so if and only if E is nonnegative definite, that is,
if and only if 1 − γ ≤ n1 . Thus, for any γ ≤ r−1 r , the E-optimal designs are
all group-affine designs for which the number of groups n is no larger than
1
1−γ , that is, for which n ≤ nγ . This completes the proof of Theorem 14.
To prove Theorem 15, recall the definition of t(d) given just prior to Theo-
rem 14 and note that now only d ∈ DE is being considered. It is shown in Ap-
pendix A.4 that the eigenvalues of Hd , aside from r − nγ zeros, are the eigen-
values of E 1/2 Dt E 1/2 . Thus, the problem is to show that t(d1 ) ≺ t(d2 ) implies
the eigenvalues of E 1/2 Dt(d1 ) E 1/2 are majorized by those of E 1/2 Dt(d2 ) E 1/2 .
Now, t(d1 ) ≺ t(d2 ) is Pequivalent to t(d1 ) = St(d2 ) , where S is a doubly
stochastic matrix; S = m i=1 ai Qi for permutation matrices Qi and positive
numbers ai summing to 1 ([3], pages 33 and 37). Thus,
E 1/2 Dt(d1 ) E 1/2 = E 1/2 DSt(d2 ) E 1/2
" #
= E 1/2 ai Qi Dt(d2 ) Qi E 1/2
X

ai Qi [E 1/2 Dt(d2 ) E 1/2 ]Qi ,

X
=
i
q
the last equality following because E 1/2 = (I − n1 J) + 1
n − (1 − γ)J com-
mutes with any permutation matrix. This shows that E 1/2 Dt(d1 ) E 1/2 is a
symmetrization of E 1/2 Dt(d2 ) E 1/2 and thus the eigenvalues of the first ma-
trix are majorized by those of the second (this follows from [3], page 69).
Since t(d∗ ) ≺ t(d) for every d ∈ DE , d∗ is E-Schur-optimal and Theorem 15 is
proved.
16 J. P. MORGAN AND B. H. RECK

5. Special cases: (k1 − k2 ) ≤ 2. In this section, the three important

special cases of k1 and k2 being equal or nearly so, k2 ∈ {k1 , k1 − 1, k1 − 2},
are investigated in light of the results in Sections 3 and 4. Put k2 = k1 − m
so that (k1 − k2 ) ≤ 2 says m = 0, 1, or 2, and for any m,
k12 k12 k1 m m2
(15) = = + + .
k1 + k2 2k1 − m 2 4 4(2k1 − m)
k12
Recall that θ̄ is the integer part of (15). The values for γ = k1 +k2 − θ̄ in the
corollaries below are easily found using (15).

Corollary 21. For k1 = k2 ,

(i) if 2|k1 , then γ = 0, (k1 + k2 )|k12 and ECD(θ ∗ )’s are Schur-optimal;
(ii) if 2|/k1 , then γ = 12 and ECD(θ̄)’s are E-Schur and type-1-optimal.

When k1 = k2 and 2|k1 , Schur-optimality also follows from [1], Corollary

3.4. For 2|/k1 , the result follows from Theorems 5 and 15.

Corollary 22. For k1 − k2 = 1,

(i) if 2|k1 , then γ = v+1
4v and ECD(θ̄)’s are E-Schur and type-1-optimal.
(ii) if 2|/k1 , then γ = 3v+1
4v and uniform group-affine designs with four
groups are E-Schur-optimal if r ≥ 5. For r ≤ 4, ECD(θ̄ + 1)’s are Schur-
optimal.

As an example, the design in Table 1 is E-Schur-optimal for 9 treatments

in 5 replicates with block sizes 4 and 5. The design consisting of the first
four replicates is Schur-optimal.

Corollary 23. For k1 − k2 = 2,

(i) if 2|k1 , then γ = v+2
2v , and uniform group-affine designs are E-Schur-
optimal if the number of groups is 3 for v = 6 and 2 for v ≥ 10. For r = 2,
ECD(θ̄ + 1)’s are Schur-optimal;
(ii) if 2|/k1 , then γ = v1 , and ECD(θ̄)’s are Schur-optimal.

The Schur-optimality in part (ii) of Corollary 23 follows from Theorem 6

and Lemma 9.

6. ECDs, balanced arrays and design construction. Balanced arrays were

introduced by Chakravarti [5, 6] as a useful device for fractional factorial
designs and they have since been investigated by a plethora of authors,
including, of late, Kuriki [16], Fuji-Hara and Miyamoto [9], Sinha, Dhar
 RESOLVABLE DESIGNS 17

and Kageyama [25] and Ghosh and Teschmacher [10]. Here, only strength-
two arrays on two symbols will be needed. A balanced array of strength 2,
BA(N, m, 2), on the symbols 0 and 1, is an N × m array with the property
that for any selection of two columns, the N pairs formed by the rows are
(0, 0), (0, 1), (1, 0) and (1, 1), with frequencies µ0 , µ1 , µ1 and µ2 , respec-
tively. Two-symbol orthogonal arrays of strength two are the special case
µ0 = µ1 = µ2 . When the number of 0’s in each column is specified (as is
the case below), the µi ’s are all determined by θ = µ0 and the array will be
denoted BA(N, m, 2; θ).
Bailey, Monod and Morgan [1] demonstrate the combinatorial equivalence
between affine resolvable designs and orthogonal arrays. Their method can
be used to express any resolvable design as a combinatorial array, as follows.
Given a resolvable design for v treatments in r replicates, each consisting of
s blocks, label the blocks within a replicate 0, . . . , s − 1. Construct a v × r
array by identifying rows of the array with treatments of the resolvable
design, and columns with replicates: symbol j ∈ {0, 1, . . . , s − 1} is placed
in row i, column h if and only if treatment i is in block j of replicate h.
Theorem 24 is evident.

Theorem 24. Each ECD(θ) in D(v, r; k1 , k2 ) is equivalent to a BA(v, r, 2; θ).

A group-affine design with n groups of sizes t1 , . . . , tn is a juxtaposition

(BA1 , BA2 , . . . , BAn ) of balanced arrays BAi = BA(v, ti , 2; θ) so that any
two columns from different BAi ’s form a BA(v, 2, 2; θ + 1). Call such a juxta-
position a grouped balanced array, denoted by GBA(v, (t1 , . . . , tn ), 2; θ). Con-
structions for the designs in Corollaries 21–23 of Section 5 are now listed.
These are stated in terms of Hadamard matrices, that is, orthogonal ma-
trices for which every element is 1 or −1. The Hadamard conjecture says
that a Hadamard matrix exists for every order a multiple of four. Existence
is known for all such orders up to 664 and for infinitely many other values
(see [8]). A Hadamard matrix is said to be standardized if the first row and
column are all ones; this can always be achieved.

Theorem 25. If k1 = k2 is even, then a Schur-optimal ECD(θ ∗ ) cor-

responds to an OA(v, r, 2; v2 ). Existence of a Hadamard matrix of order v
implies existence of the OA for every r ≤ v − 1.

Theorem 26. If k1 = k2 is odd, then a type-1-optimal ECD(θ̄) corre-

sponds to a BA(v, r, 2; v−2
4 ). Existence of a Hadamard matrix of order v + 2
implies existence of the BA for every r ≤ v2 .
2
Proof. The value of θ̄ is int( (v/2) v−2
v ) = 4 . Given the standardized
Hadamard matrix, permute columns (except the first) so that the second
18 J. P. MORGAN AND B. H. RECK

row has 1 in the first v+22 columns. Delete the first two rows and the
first v+2
2 columns, then replace −1 by 0 throughout. Clearly, the result is
v v−2
BA(v, 2 , 2; 4 ). 

Theorem 27. If k1 = k2 + 1 is even, then a type-1-optimal ECD(θ̄)

corresponds to a BA(v, r, 2; v+1
4 ). Existence of a Hadamard matrix of order
v + 1 implies existence of the BA for every r ≤ v.
2
Proof. The value of θ̄ is int( ((v+1)/2)
v ) = v+1
4 . Given the standardized
Hadamard matrix, delete the first row and column, then replace −1 by 0
throughout. The result is BA(v, v, 2; v+1
4 ). 

Theorem 28. If k1 = k2 + 1 is odd, then a type-1-optimal ECD(θ̄ + 1)

with r ≤ 4 corresponds to a BA(v, r, 2; v+3 4 ), which always exists. For r ≥ 5
and v ≥ 9, an E-Schur-optimal group-affine design corresponds to a
GBA(v, (t1 , t2 , t3 , t4 ), 2; v−1
4 ) with t4 − t1 ≤ 1. Existence of a Hadamard ma-
trix of order v + 3 with a 4 × r submatrix of the form
1t1 1t2 1t3 −1t4
 
 1t 1t2 −1t3 1t4 
 1 
 1t −1t2 1t3 1t4 
1
−1t1 1t2 1t3 1t4
implies existence of the GBA.
2
Proof. The value of θ̄ is int( ((v+1)/2)
v ) = v−1
4 . For r ≤ 4, take OA(v −
1, r, 2) on {0, 1} and add one row of 0’s to get BA(v, r, 2; v+3
4 ). Given the
assumed Hadamard matrix, delete v + 3 − r columns not containing the
submatrix, then delete the submatrix and add a row of 1’s. Replacing −1
by 0 throughout gives the GBA. 

The maximum r admitted for a given v by the Hadamard construction in

Theorem 28 depends on the particular Hadamard matrix chosen. All noniso-
morphic Hadamard matrices are known up through order 28 and, for these,
a complete search has produced (v, r) = (9, 5), (13, 5), (17, 9), (21, 9), (25, 13);
the first four of these appear in Appendix B. A search of a few known
Hadamard matrices (compiled by N. J. A. Sloane at www.research.att.com/˜njas/hadamard/)
of orders up to 48 has further produced (v, r) = (29, 13), (33, 16), (37, 18), (41, 24), (45, 19).
Clearly there is room for more work to be done here.

Theorem 29. If k1 = k2 + 2 is even, for v ≥ 10 an E-Schur-optimal

group-affine design corresponds to a GBA(v, (t1 , t2 ), 2; v+2
4 ) with t2 − t1 ≤ 1.
Existence of a Hadamard matrix of order v + 2 implies existence of the GBA
for every t1 + t2 = r ≤ v2 + 1.
 RESOLVABLE DESIGNS 19
2
Proof. The value of θ̄ is int( ((v+2)/2)
v ) = v+2
4 . Given the standardized
Hadamard matrix of order v + 2 = 4h, permute columns (except the first)
so that the first three rows are
1h 1h 1h 1h
 
 1h −1h 1h −1h  .
1h −1h −1h 1h
Now delete the first 2h columns, delete the first three rows, then add one row
of −1’s and finally replace −1 by 0 throughout. The result is GBA(v, ( v+2 v+2
4 , 4 ), 2;
v+2 v+2 v+2
4 ) with two juxtaposed BA(v, 4 , 2; 4 ) being the first h and last h
r
columns. For smaller r, delete int(h − 2 ) columns from the first component
BA and int(h − r+1 2 ) from the second. 

Theorem 30. If k1 = k2 + 2 is odd, a Schur-optimal ECD(θ̄) corre-

sponds to a BA(v, r, 2; v+4
4 ). Existence of a Hadamard matrix of order v
implies existence of the BA for every r ≤ v − 1.
2
Proof. The value of θ̄ is int( ((v+2)/2)
v ) = v+4
4 . Given the standardized
Hadamard matrix, delete the first row and column, add a row of −1’s and
then replace −1 by 0 throughout. The result is BA(v, v − 1, 2; v+4
4 ). 

7. Comments. By exploiting properties of the dual, optimality theory

for resolvable designs with two blocks per replicate has been developed.
Section 3 establishes conditions for Schur-optimality of equal concurrence
designs depending on the block discordancy coefficient γ. For γ where Schur-
optimality is not established, the class of competitors has been significantly
narrowed via Schur-ordering for small r. Section 4 characterizes the class of
all E-optimal designs whenever any group-affine design exists and further
determines the Schur-best of the E-optimal designs. Sections 5 and 6 apply
these results for the important cases k1 − k2 ≤ 2, including explicit design
constructions employing an equivalence with balanced arrays. It is evident
from the constructions that many other designs, eliminated by the Schur-
domination argument in Theorem 15, do exist.
For large (r ≥ v) replication, the problem can be quite different and op-
timality work would proceed based on treatment, rather than block, con-
currences. Here is a simple construction for that case: given a BIBD for v
treatments in b blocks of size k, there is a Schur-optimal resolvable design in
D(v, b; k, v − k) formed by the blocks of the BIBD and their complements.
While it is possible that these designs are equivalent to balanced arrays
(when the starting BIBD is symmetric), it is also true that the resulting
φdhh′ can be widely dispersed, for they are determined by the block concur-
rence counts for the underlying BIBD.
20 J. P. MORGAN AND B. H. RECK

The optimality route is clear for r ≤ v − 1 and γ ≤ 12 , but for larger γ

there are still unanswered questions. For instance, A- and E-optimality need
not coincide and the problem of determining an A-best design remains open.
For r ≤ 4, the authors have solved the A-optimality problem in its entirety,
including construction, and for larger r, have done this for the special cases
of k1 − k2 ≤ 2. These and related results will be reported elsewhere.

APPENDIX A: PROOFS
1/2
A.1. Proof of Lemma 1. Let Ds be the diagonal matrix of square roots
−1/2
of block sizes and write Bd = Nd Ds . Multiplying Cd by 1r and right and
−1/2
left multiplying Cdual by Ds , equations (1) and (2) become
1 1 1
(16) Cd = I − Nd Ds−1 Nd′ = I − Bd Bd′ = Cd∗
r r r
and
1 1
(17) Ds−1/2 Cdual Ds−1/2 = I − Ds−1/2 Nd′ Nd Ds−1/2 = I − Bd′ Bd = Cdual
∗
.
r r
Since (1) and (16) differ only by a constant and since the nonzero eigenvalues
of Bd Bd′ and Bd′ Bd are identical, the eigenvalues of Cd can be found from
∗
those of Cdual in (17), which depends on block concurrences and block sizes.
−1/2 −1/2
∗
Continuing, write Cdual = I − 1r Ad for Ad = Bd′ Bd = Ds Nd′ Nd Ds .
1/2 1/2
Regardless of the design d, Ad Ds 1 = rDs 1; r is an eigenvalue of Ad
corresponding to the zero eigenvalue common to Cdual and Cd . One term of
the spectral decomposition of Ad is then
1/2 1/2 √
r(Ds 1)(Ds 1)′ 1
  
(18) = J⊗ √ k1 k1 k2
,
1/2 ′ 1/2
(Ds 1) (Ds 1) (k1 + k2 ) k1 k2 k2

where J is an r × r matrix of ones. Subtract (18) from Ad and denote the

result A∗d . Then a bit of manipulation, employing the fact that all four block
concurrence counts for blocks in replicate h with blocks in replicate h′ are
determined by φdhh′ , gives
√
1
 
(19) A∗d = Md ⊗ √k2 − k1 k2
.
k1 k2 − k1 k2 k1

Since the eigenvalues of the 2 × 2 matrix in (19) are 0 and k1 + k2 , the

b = 2r eigenvalues of A∗d are r copies of 0 and k1vk2 times the r eigenvalues
of Md . The eigenvalues of Cd as stated in the lemma are now immediate.
 RESOLVABLE DESIGNS 21

A.2. Proof of Lemma 17. The lower bound on ed1 follows from
ed1 = max
′
x′ Md x
x x=1
p
  
′
= max x + γ I − γJ + ∆d x
′
x x=1 v
p
  
′
≥ max x + γ I − γJ + ∆d x
x′ x=1 v
x′ 1=0
p
= + γ + max x′ ∆ d x
v x′ x=1
x′ 1=0
p
= + γ + max x′ P ∆ d P x
v x′ x=1
x′ 1=0
p
= + γ + max x′ P ∆ d P x
v x′ x=1
p
= + γ + ud1 .
v
The penultimate equality holds since ud1 > 0 and 1 is an eigenvector of
P ∆d P with eigenvalue 0. Likewise,
edr = min
′
x′ Md x
x x=1
p
  
′
= min x + γ I − γJ + ∆d x
′ x x=1 v
p
≤ + γ + min x′ P ′ ∆ d P x
v x′ x=1
x′ 1=0
p
= + γ + udr .
v
The last equality holds provided udr < 0, for reasons similar to those above.
If udr > 0, the bound still holds, since edr ≤ tr(M
r
d)
= vp ≤ vp + γ + udr .

A.3. Proof of Lemma 19. Suppose that A − J is of the suggested form.

Write t(i) for t1 + t2 + · · · + ti and t(0) = 0. Then PAP = P (A − J)P , implying
that
max eig(PAP ) = max
′
x′ P (A − J)P x
x x=1

= max y ′ (A − J)y
y=P x,x′ x=1

t t
(i)
!2
X X
= max − yj ,
y=P x,x′ x=1
i=1 j=t(i−1) +1
22 J. P. MORGAN AND B. H. RECK

which is clearly nonpositive.

Now, subscripting by the dimension, suppose that Pr Ar Pr is nonpositive
definite (npd). Exhaustive enumeration shows that Ar − Jr must have the
form (13) for r = 3, 4, 5. Assuming nonpositivity implies that the form must
hold for a given r, it will be shown that the same implication holds for r + 1.
Denoting the upper left-hand side r × r submatrix of Ar+1 by Ar (which
is itself an adjacency matrix), it is claimed that Pr1 Ar+1 Pr+1 is npd ⇒
Pr Ar Pr is npd. If not, then there exists x such that x′ Pr Ar Pr x > 0. Write
y ′ = (x′ Pr , 0). Then y ′ 1 = 0 ⇒ y ′ Pr+1 Ar+1 Pr+1 y = y ′ Ar+1 y = x′ Ar x > 0, a
contradiction.
Since Pr Ar Pr is npd, the induction hypothesis says that
−Jt1
 
 −Jt2 
a 
 
Ar+1 − Jr+1 = 
 ..
 . 

 −Jtn 
a′ −1
for some vector ar×1 of 0’s and −1’s. Indeed, by the induction hypothesis,
every s × s principal minor of Ar+1 − Jr+1 must have the form (13), so a
may be partitioned as a′ = (a′1 , a′2 , . . . , a′n ), where ai is either −1ti or 0ti . If
n = 1, the proof is done. If n = 2 and a2 = −1t2 , then
−Jt1 0 a1
 

Ar+1 − Jr+1 =  0 −Jt2 −1t2  .

a′1 −1′t2 −1
If now a1 = −1t1 , then Ar+1 − Jr+1 contains a principal minor,
1 0 1
 

M3 = (−1)  0 1 1  ,
1 1 1
which is not of the form (13), contradicting the fact that the result holds
for r = 3. Thus, at most one of a1 , a2 is nonzero. It follows immediately
that for n > 2, the same statement holds for any pair ai , ai′ . Permuting
rows and columns as needed, it can be assumed that ai = 0ti for i < n and
consequently that Ar+1 − Jr+1 has the form (13).

A.4. Eigenvalue equations for Hd and proof of Lemma 20. Any vector of
the form (c′1 , c′2 , . . . , c′n ), where ci ∈ ℜti is either a contrast vector or the zero
vector, is an eigenvector of Hd with eigenvalue 0. Consequently, all other
eigenvectors are of the form e′ = (x1 1′t1 , x2 1′t2 , . . . , xn 1′tn ) for some scalars
x1 , . . . , xn . The first equation in the system Hd e = λe is γt1 x1 − (1 − γ)t2 x2 −
· · · − (1 − γ)tn xn = λx1 ; the other equations may be written similarly in
order to see that Hd e = λe are equivalent to [Dt − (1 − γ)1n t′ ]x = λx, where
 RESOLVABLE DESIGNS 23

t = (t1 , . . . , tn )′ and x = (x1 , . . . , xn )′ . Thus, the remaining eigenvalues of Hd

are the right eigenvalues of Dt − (1 − γ)1n t′ . Now, Dt is positive definite and
E is nonnegative definite for γ ≥ n−1 n , so both have symmetric square root
matrices and Dt is invertible. Thus,

|Dt − (1 − γ)1n t′ − λI| = 0

⇐⇒ |Dt − (1 − γ)1n t′ − λI||Dt−1 | = 0
⇐⇒ |E − λDt−1 | = 0
1/2 1/2
⇐⇒ |Dt EDt − λI| = 0 (proving Lemma 20)
⇐⇒ |E 1/2 Dt E 1/2 − λI| = 0 (needed in Theorem 15),

the last step following because GG′ and G′ G have the same eigenvalues for
any square G. Now, in the proof of Theorem 15 some elements of t are
allowed to be zero (without loss of generality, t1 = · · · = tz = 0 for integer
z ≥ 1), in which case Dt is not invertible. In this case there are z additional
zero eigenvalues plus a reduced system of n − z equations in tz+1 , . . . , tn . It
1/2 1/2 1/2 1/2
is easy to see that |Dt EDt − λI| = |0z,z − λIz ||D̃t Ẽ D̃t |, where Ẽ and
D̃t are the lower-right submatrices of E and Dt of order n − z. Thus, the n
1/2 1/2
eigenvalues sought are still those of Dt EDt and thus of E 1/2 Dt E 1/2 .

APPENDIX B: UNIFORM GBAS WITH FOUR GROUPS

GBA(9, (1, 1, 1, 2), 2; 2) GBA(13, (1, 1, 1, 2), 2; 3)

11111 11111
00000 11111
01001 01100
01010 00010
00110 10010
10010 00001
00101 10001
11100 00110
10001 10100
00101
01010
01001
11000
24 J. P. MORGAN AND B. H. RECK

GBA(17, (2, 2, 2, 3), 2; 4) GBA(21, (2, 2, 2, 3), 2; 5)

111111111 111111111
000000000 111111111
100100101 110000110
001101100 101010100
100011001 011010001
101110010 000101100
010010110 001010010
010111100 100000011
100001110 010101010
111001001 010000101
101010100 100101001
000101011 011001001
011001010 101001001
110100010 010011010
010110001 011100100
001010011 001001110
011000101 101100010
100011100
000110101
110110000
000110011

Acknowledgments. This research is partially based on work in Reck’s

doctoral dissertation [23], for which Morgan was director. Morgan com-
pleted his Ph.D. under the direction of I. M. Chakravarti. In dedicating
this paper to the memory of Professor Chakravarti, we humbly acknowledge
the ongoing impact of his work on our own and on that of the statistical
and mathematical communities at large. We wish to thank the referees for
comments leading to decided improvement of the presentation.

REFERENCES
[1] Bailey, R. A., Monod, H. and Morgan, J. P. (1995). Construction and optimality
of affine-resolvable designs. Biometrika 82 187–200. MR1332849
[2] Bailey, R. A. and Speed, T. P. (1986). Rectangular lattice designs: Efficiency
factors and analysis. Ann. Statist. 14 874–895. MR0856795
[3] Bhatia, R. (1997). Matrix Analysis. Springer, New York. MR1477662
[4] Bose, R. C. (1942). A note on the resolvability of balanced incomplete block designs.
Sankhyā 6 105–110. MR0008064
[5] Chakravarti, I. M. (1956). Fractional replication in asymmetrical factorial designs
and partially balanced arrays. Sankhyā 17 143–164. MR0085677
[6] Chakravarti, I. M. (1961). On some methods of construction of partially balanced
arrays. Ann. Math. Statist. 32 1181–1185. MR0130770
 RESOLVABLE DESIGNS 25

[7] Cheng, C. S. (1978). Optimality of certain asymmetrical experimental designs. Ann.

Statist. 6 1239–1261. MR0523760
[8] Craigen, R. and Kharaghani, H. (2007). Hadamard matrices and Hadamard de-
signs. In Handbook of Combinatorial Designs, 2nd ed. (C. J. Colbourn and
J. H. Dinitz, eds.) 273–280. Chapman and Hall/CRC, Boca Raton, FL.
[9] Fuji-Hara, R. and Miyamoto, N. (2000). Balanced arrays from quadratic functions.
J. Statist. Plann. Inference 84 285–293. MR1747509
[10] Ghosh, S. and Teschmacher, L. (2002). Comparisons of search designs using search
probabilities. J. Statist. Plann. Inference 104 439–458. MR1906265
[11] Harshbarger, B. (1946). Preliminary report on the rectangular lattices. Biometrics
2 115–119. MR0026784
[12] Harshbarger, B. (1949). Triple rectangular lattices. Biometrics 5 1–13. MR0030486
[13] John, J. A., Russell, K. G., Williams, E. R. and Whitaker, D. (1999). Re-
solvable designs with unequal block sizes. Aust. N. Z. J. Stat. 41 111–116.
MR1705928
[14] Kirkman, T. P. (1851). Query VI on “Fifteen young ladies . . . .” Lady’s and Gen-
tleman’s Diary No. 148 48.
[15] Kunert, J. (1985). Optimal repeated measurements designs for correlated observa-
tions and analysis by weighted least squares. Biometrika 72 375–389. MR0801776
[16] Kuriki, S. (1993). On existence and construction of balanced arrays. Discrete Math.
116 137–155. MR1222122
[17] Morgan, J. P. (1996). Nested designs. In Design and Analysis of Experi-
ments (S. Ghosh and C. R. Rao, eds.) 939–976. North-Holland, Amsterdam.
MR1492587
[18] Morgan, J. P. (2007). Optimal incomplete block designs. J. Amer. Statist. Assoc.
102 655–663.
[19] Patterson, H. D. and Hunter, E. A. (1983). The efficiency of incomplete block
designs in National List and Recommended List cereal variety trials. J. Agricul-
tural Sci. 101 427–433.
[20] Patterson, H. D. and Silvey, V. (1980). Statutory and recommended list trials of
crop varieties in the United Kingdom. J. Roy. Statist. Soc. Ser. A 143 219–252.
[21] Patterson, H. D. and Williams, E. R. (1976). A new class of resolvable incomplete
block designs. Biometrika 63 83–92. MR0403109
[22] Rao, C. R. (1973). Linear Statistical Inference and Its Applications, 2nd ed. Wiley,
New York. MR0346957
[23] Reck, B. H. (2002). Nearly balanced and resolvable block designs. Ph.D. dissertation,
Old Dominion Univ.
[24] Shah, K. R. and Sinha, B. K. (1989). Theory of Optimal Designs. Lecture Notes in
Statist. 54. Springer, New York. MR1016151
[25] Sinha, K., Dhar, V. and Kageyama, S. (2002). Balanced arrays of strength two
from block designs. J. Combin. Des. 10 303–312. MR1918802
[26] Uddin, N. and Morgan, J. P. (1992). Optimal block designs with maximum block-
size and minimum replication constraints. Comm. Statist. Theory Methods 21
179–195. MR1159160
[27] Williams, E. R. (1975). A new class of resolvable designs. Ph.D. dissertation, Univ.
Edinburgh.
[28] Williams, E. R., Patterson, H. D. and John, J. A. (1976). Resolvable designs
with two replications. J. Roy. Statist. Soc. Ser. B 38 296–301. MR0654162
[29] Yates, F. (1936). A new method or arranging variety trials involving a large number
of varieties. J. Agricultural Sci. 26 424–455.
26 J. P. MORGAN AND B. H. RECK

[30] Yates, F. (1940). Lattice squares. J. Agricultural Sci. 30 672–687.

Department of Statistics Genetic Analysis

Virginia Tech GlaxoSmithKline
Blacksburg, Virginia 24061-0439 Research Triangle Park, North Carolina 27709
USA USA
E-mail: [email protected] E-mail: [email protected]