Sudoku, gerechte designs, resolutions, ane

space, spreads, reguli, and Hamming codes

R. A. Bailey, Peter J. Cameron
School of Mathematical Sciences, Queen Mary, University of London,
London E1 4NS, UK
Robert Connelly

Department of Mathematics, Malott Hall, Cornell University,

Ithaca, NY 14853, USA
Abstract
Solving a Sudoku puzzle involves putting the symbols 1, . . . , 9 into
the cells of a 9 9 grid partitioned into 3 3 subsquares, in such
a way that each symbol occurs just once in each row, column, or
subsquare. Such a solution is a special case of a gerechte design, in
which an nn grid is partitioned into n regions with n squares in each,
and each of the symbols 1, . . . , n occurs once in each row, column, or
region. Gerechte designs originated in statistical design of agricultural
experiments, where they ensure that treatments are fairly exposed to
localised variations in the eld containing the experimental plots.
In this paper we consider several related topics. In the rst sec-
tion, we dene gerechte designs and some generalizations, and explain
a computational technique for nding and classifying them. The sec-
ond section looks at the statistical background, explaining how such
designs are used for designing agricultural experiments, and what ad-
ditional properties statisticians would like them to have.
In the third section, we focus on a special class of Sudoku solutions
which we call symmetric. They turn out to be related to some
important topics in nite geometry over the 3-element eld, and to

This research partially supported by NSF Grant Number DMS-0510625.

1
error-correcting codes. We explain all of these connections, and use
them to classify the symmetric Sudoku solutions (there are just two,
up to the appropriate notion of equivalence). In the nal section,
we construct some further Sudoku solutions with desirable statistical
properties, and briey consider some generalizations.
1 Gerechte designs
1.1 Introduction
A Latin square of order n is an n n array containing the symbols 1, . . . , n
in such a way that each symbol occurs once in each row and once in each
column of the array. We say that two Latin squares L
1
and L
2
of order n are
orthogonal to each other if, given any two symbols i and j, there is a unique
pair (k, l) such that the (k, l) entries of L
1
and L
2
are i and j respectively.
In 1956, W. U. Behrens [4] introduced a specialisation of Latin squares
which he called gerechte. The n n grid is partitioned into n regions
S
1
, . . . , S
n
, each containing n cells of the grid; we are required to place the
symbols 1, . . . , n into the cells of the grid in such a way that each symbol
occurs once in each row, once in each column, and once in each region.
The row and column constraints say that the solution is a Latin square,
and the last constraint restricts the possible Latin squares.
By this point, many readers will recognize that solutions to Sudoku puz-
zles are examples of gerechte designs, where n = 9 and the regions are the
33 subsquares. (The Sudoku puzzle was invented, with the name number
place, by Harold Garns in 1979.)
Here is another familiar example of a gerechte design. Let L be any Latin
square of order n, and let the region S
i
be the set of cells containing the
symbol i in the square L. A gerechte design for this partition is precisely
a Latin square orthogonal to L. (This shows that there is not always a
gerechte design for a given partition. A simpler negative example is obtained
by taking one region to consist of the rst n1 cells of the rst row and the
nth cell of the second row.)
We might ask: given a grid, and a partition into regions, what is the
complexity of deciding whether a gerechte design exists?
For another example, consider the partitioned grid shown in Figure 1:
this example was considered by Behrens in 1956. (Ignore the triples to the
2
right of the grid for a moment.) Six solutions are shown. Up to rotations of
the grid and permutations of the symbols 1, . . . , 5, these are all the solutions,
as we will explain shortly. (The complete set of fteen solutions is given
in [3].)
r
1
, c
1
, s
1
r
1
, c
2
, s
1
r
1
, c
3
, s
1
r
1
, c
4
, s
2
r
1
, c
5
, s
2

r
2
, c
1
, s
1
r
2
, c
2
, s
1
r
2
, c
3
, s
5
r
2
, c
4
, s
2
r
2
, c
5
, s
2

r
3
, c
1
, s
4
r
3
, c
2
, s
5
r
3
, c
3
, s
5
r
3
, c
4
, s
5
r
3
, c
5
, s
2

r
4
, c
1
, s
4
r
4
, c
2
, s
4
r
4
, c
3
, s
5
r
4
, c
4
, s
3
r
4
, c
5
, s
3

r
5
, c
1
, s
4
r
5
, c
2
, s
4
r
5
, c
3
, s
3
r
5
, c
4
, s
3
r
5
, c
5
, s
3

3 4 5 1 2
5 1 2 3 4
2 3 4 5 1
4 5 1 2 3
1 2 3 4 5
3 5 2 1 4
2 1 4 5 3
4 3 5 2 1
5 4 1 3 2
1 2 3 4 5
3 1 5 2 4
2 5 4 1 3
4 3 2 5 1
5 4 1 3 2
1 2 3 4 5
2 1 5 3 4
3 5 4 2 1
4 3 1 5 2
5 4 2 1 3
1 2 3 4 5
3 1 5 2 4
2 5 4 1 3
4 3 1 5 2
5 4 2 1 3
1 2 3 4 5
2 4 1 5 3
3 1 5 2 4
5 3 4 1 2
4 5 2 3 1
1 2 3 4 5
Figure 1: A partitioned 5 5 grid (top left), its representation as a block
design (top right), and all inequivalent gerechte designs (bottom)
1.2 Resolvable block designs
A block design is a structure consisting of a set of points and a set of blocks,
with an incidence relation between points and blocks. Often we identify a
block with the set of points incident to it, so that a block design is represented
by a family of sets; however, the same set may occur more than once.
A block design is said to be resolvable if the set of blocks can be par-
titioned into subsets C
1
, . . . , C
r
(called replicates) such that each point is
3
incident with just one block in any replicate C
i
. The partition of the block
set is called a resolution of the design.
The search for gerechte designs for a given partitioned grid can be trans-
formed into a search for resolutions of a block design, as we now show.
The basic data for a gerechte design is an n n grid partitioned into n
regions S
1
, . . . , S
n
, each containing n cells. We can represent this structure
by a block design as follows:
the points are 3n objects r
1
, . . . , r
n
, c
1
, . . . , c
n
, s
1
, . . . , s
n
;
for each of the n
2
cells of the grid, there is a block r
i
, c
j
, s
k
, if the
cell lies in the ith row, the jth column, and the kth region.
Proposition 1.1 Gerechte designs on a given partitioned grid correspond,
up to permuting the symbols 1, . . . , n, in one-to-one fashion with resolutions
of the above block design.
Proof Given a gerechte design, let C
i
be the set of cells containing the
symbol i. By denition, the blocks corresponding to these cells contain each
row, column, or region object exactly once, and so form a partition of the
point set. Any cell contains a unique symbol i, so every block occurs in just
one class C
i
. Thus we have a resolution. The converse is proved in the same
way.
The GAP [10] share package DESIGN [20] can nd all resolutions of a
block design, up to isomorphisms of the block design. In our case, isomor-
phisms of the block design come from symmetries of the partitioned grid, so
we can use this package to compute all gerechte designs up to permutation
of symbols and symmetries of the partitioned grid.
For example, the partition of the 5 5 grid discussed in the preceding
section is represented as a block design with 15 points and 25 blocks of size 3,
also shown in Figure 1. The automorphism group of the design is the cyclic
group of order 4 consisting of the rotations of the grid through multiples
of /2. The DESIGN program quickly nds that, up to automorphisms,
there are just six resolutions of this design, corresponding to six inequivalent
gerechte designs; these are shown in the gure.
The same method shows that, for a 6 6 square divided into 3 2
rectangles, there are 49 solutions up to symmetries of the corresponding
block design and permutations of the symbols. (The number of symmetries
of the block design in this case is 3456; the group consists of all row and
column permutations preserving the appropriate partitions.)
4
1.3 Orthogonal and multiple gerechte designs
We saw earlier the denition of orthogonality of Latin squares. A set of
mutually orthogonal Latin squares is a set of Latin squares in which every
pair is orthogonal. It is known that the size of a set of mutually orthogonal
Latin squares of order n is at most n 1.
Similar denitions and results apply to gerechte designs. We say that two
gerechte designs with the same partitioned grid are orthogonal to each other
if they are orthogonal as Latin squares, and a set of mutually orthogonal
gerechte designs is a set of such designs in which each pair is orthogonal.
Proposition 1.2 Given a partition of the nn grid into regions S
1
, . . . , S
n
each of size n, the size of a set of mutually orthogonal gerechte designs for
this partition is at most nd, where d is the maximum size of the intersection
of a region S
i
and a line (row or column) L
j
,= S
i
.
Proof Take a cell c L
j
S
i
. By permuting the symbols in each square,
we may assume that all the squares have entry 1 in the cell c. Now, in each
square, the symbol 1 occurs exactly once in the region S
i
and not in the
line L
j
; and all these occurrences must be in dierent cells, since for each
pair of squares, the pair (1, 1) of entries already occurs in cell c. So there are
at most [S
i
L
j
[ squares in the set.
This bound is not always attained. Consider the 5 5 gerechte designs
given earlier. The maximum intersection size of a line and a region is clearly 3,
so the bound for the number of mutually orthogonal designs is 2. But by
inspection, each design has the property that the entries in cells (2, 3) and
(3, 5) are equal. (The reader is invited to discover the simple argument to
show that this must be so, independent of the classication of the designs.)
Hence no pair of orthogonal designs is possible. Similarly, for the 66 square
divided into 3 2 rectangles, there cannot exist two orthogonal gerechte
designs, since it is well known that there cannot exist two orthogonal Latin
squares of order 6.
Proposition 1.2 gives an upper bound of 6 for the number of mutually
orthogonal Sudoku solutions. In Section 3.4, we will see that this bound is
attained.
The concept of a gerechte design can be generalized. Suppose that we are
given a set of r partitions of the cells of an n n grid into n regions each of
5
size n. A multiple gerechte design for this partition is a Latin square which
is simultaneously a gerechte design for all of the partitions.
For example, given a set of (mutually orthogonal) Latin squares, the
symbols in each square dene a partition of the n n array into regions. A
Latin square is a multiple gerechte design for all of these partitions if and
only if it is orthogonal to all the given Latin squares.
The problem of nding a multiple gerechte design can be cast into the
form of nding a resolution of a block design, in the same way as for a single
gerechte design. The block design has (r + 2)n points, and each cell of the
grid is represented by a block containing the objects indexing its row, its
column, and the region of each partition which contain it. Again, we can use
the DESIGN program to classify such designs up to symmetries of the grid.
For example, Federer [9], in a section which he attributed to G. M. Cox,
called a m
1
m
2
m
1
m
2
Latin square magic if it is a gerechte design for the
regions forming the obvious partition into m
1
m
2
rectangles, and super
magic if it is simultaneously a gerechte design for the partition into m
2
m
1
rectangles, where m
1
,= m
2
. He considered the problem of nding multiple
gerechte designs (which he called super magic Latin squares) for the 6 6
square partitioned into 3 2 rectangles and 2 3 rectangles. The DESIGN
package nds that there are 26 such designs up to symmetries.
We can also dene a set of mutually orthogonal multiple gerechte designs
in the obvious way, and prove a similar bound for the size of such a set.
We will see examples of these things in Section 3.4.
2 Statistical considerations
In this section, we consider the use of gerechte designs in statistical design
theory, and some additional properties which are important there.
2.1 Agricultural experiments in Latin squares
The statistician R. A. Fisher suggested the use of Latin squares in agricultural
experiments. If n treatments (crop varieties, quantities of fertilizer, etc.)
are to be compared on plots forming an n n grid in a eld, then arranging
the treatments as the symbols of a Latin square ensures that any systematic
change in fertility, drainage, etc. across the eld aects all treatments equally.
Figure 2 shows two experiments laid out as Latin squares.
6
Figure 2: Two experiments using Latin squares. Left: a 5 5 forestry
experiment in Beddgelert in Wales, to compare varieties of tree; designed by
Fisher, laid out in 1929, and photographed in about 1945. Right: a current
6 6 experiment to compare methods of controlling aphids; conducted by
Lesley Smart at Rothamsted Research, photographed in 2004.
If a Latin square experiment is to be conducted on land that has recently
been used for another Latin square experiment, it is sensible to regard the
previous treatments as relevant and so to use a Latin square orthogonal to
the previous one. As explained above, this is technically a sort of gerechte
design, but no agricultural statistician would call it that.
The purpose of a gerechte design in agricultural experimentation is to
ensure that all treatments are fairly exposed to any dierent conditions in
the eld. In fact, gerecht(e) is the German for fair in the sense of just.
Rows and columns are good for capturing dierences such as distance from
a wood but not for marking out stony patches or other features that tend to
clump in compact areas. Thus, in the statistical and agronomic literature,
the regions of a gerechte design are always taken to be spatially compact
areas.
7
2.2 Randomization
Before a design is used for an experiment, it is randomized. This means that
a permutation of the cells is chosen at random from among all those that
preserve the three partitions: into rows, into columns, and into regions. It is
by no means common for the cells to be actually square plots on the ground;
when they are, it is also possible to transpose rows and columns, if the regions
are unchanged by this action. This random permutation is applied to the
chosen gerechte design before it is laid out in the eld.
One important statistical principle is lack of bias. This means that every
plot in the eld should be equally likely to be matched, by the randomization,
to each abstract cell in the gerechte design, so that any individual plot with
strange characteristics is equally likely to aect any of the treatments. To
achieve this lack of bias, the set of permutations used for randomizing must
form a transitive group, in the sense that there is such a permutation carrying
any nominated cell to any other. The allowable permutations of the 5 5
grid in Figure 1 do not have this property, but those for magic Latin squares
do. There are others, but no complete classication as far as we know.
For the remainder of this section we assume that n = m
1
m
2
and the
regions are m
1
m
2
rectangles. Then the rows, columns and regions dene
some other areas: a large row is the smallest area that is simultaneously a
union of regions and a union of rows; a minirow is the non-empty intersection
of a row and region; large columns and minicolumns are dened similarly.
A pair of distinct cells in such a grid is in one of eight relationships,
illustrated in Figure 3 for the 6 6 grid with 3 2 regions. For i = 1,
. . . , 8, the cell labelled is in relationship i with the cell labelled i. Thus
a pair of distinct cells is in relationship 1 if they are in the same minirow;
relationship 2 if they are in the same minicolumn; relationship 3 if they are
in the same region but in dierent rows and columns; relationship 4 if they
are in the same row but in dierent regions; relationship 5 if they are in the
same column but in dierent regions; relationship 6 if they are in the same
large row but in dierent rows and regions; relationship 7 if they are in the
same large column but in dierent columns and regions; relationship 8 if they
are in dierent large rows and large columns.
The group of permutations used for randomization has the property that
a pair of distinct cells can be mapped to another pair by one of the permuta-
tions if and only if they are in the same relationship. If, in addition, we can
transpose the rows and columns (not possible in Figure 3) then relationships
8
1 4
2 6
3
5 8
7
Figure 3: Eight relationships between pairs of distinct cells in the 6 6 grid
1 and 2 are merged, as are 4 and 5, and 6 and 7.
The simple-minded analysis of data from an experiment in a gerechte
design assumes that the response (such as yield of grain, or the logarithm of
the number of aphids) on each cell is the sum of four unknown parameters,
one each for the row, column and region containing the cell, and one for
the treatment (symbol) applied to it. In addition, there is random variation
from cell to cell. This is explained in [2]. The statistician is interested in
the treatment parameters, not only in their values but also in whether their
dierences are greater than can be explained by cell-to-cell variation.
However, one school of statistical thought holds that if the innate dif-
ferences between rows, between columns and between regions are relevant,
then so potentially are those between minirows, minicolumns, large rows and
large columns. Yates took this view in his 1939 paper [24], whose discussion
of a 4 4 Latin square with balanced corners may be the rst published
reference to gerechte designs. Thus the eight relationships all have to be
considered when the gerechte design is chosen.
2.3 Orthogonality and the design key
Two further important statistical properties often conict with each other.
One is ease of analysis, which means not ease of performing arithmetic but
ease of explaining the results to a non-statistician. So-called orthogonal de-
signs, like the one in Figure 4, have this property.
A gerechte design with rectangular regions is orthogonal if the arrange-
ment of symbols in each region can be obtained from the arrangement in
any other region just by permuting minirows and minicolumns. In Figure 4,
9
5 2 6 3 4 1
6 3 4 1 5 2
4 1 5 2 6 3
2 5 3 6 1 4
3 6 1 4 2 5
1 4 2 5 3 6
Figure 4: An orthogonal design for the 6 6 grid with 3 2 regions
each minicolumn contains either treatments 1, 2 and 3 or treatments 4, 5
and 6. When the statisticain investigates whether there is any real dierence
between the average eects of these two sets of treatments, (s)he compares
their dierence (estimated from the data) with the underlying variability
between minicolumns within regions and columns (also estimated from the
data). Similarly, dierences between the average eects of the three sets of
two treatments 1, 4, 2, 5 and 3, 6 are compared with the variability of
minirows within regions and rows. Treatment dierences orthogonal to all of
those, such as the dierence between the average of 1, 5 and the average
of 2, 4, are compared with the residual variability between the cells after
allowing for the variability of all the partitions.
An orthogonal design for an m
1
m
2
m
1
m
2
square with m
1
m
2
regions
may be constructed using the design key method [21, 22], as recommended
in [3]. The large rows are labelled by A
1
, which takes values 1, . . . , m
2
.
Within each large row, the rows are labelled by A
2
, which takes values 1,
. . . , m
1
. Similarly, the large columns are labelled by B
1
, taking values 1,
. . . , m
1
, and the columns within each large column by B
2
, taking values 1,
. . . , m
2
. Then put N
1
= A
1
+B
2
modulo m
2
and N
2
= A
2
+B
1
modulo m
1
.
The ordered pairs of values of N
1
and N
2
give the m
1
m
2
symbols. In Figure 4,
the rows are numbered from top to bottom, the columns from left to right,
and the correspondence between the ordered pairs and the symbols is as
follows.
N
2
N
1
1 2 3
1 1 2 3
2 4 5 6
10
(When explaining this construction to non-mathematicians we usually take
the integers modulo m to be 1, . . . , m rather than 0, . . . , m1.)
Variations on this construction are possible, especially when m
1
and m
2
are both powers of the same prime p. For example, if m
1
= 4 and m
2
= 2
then we can work modulo 2, using A
1
to label the large rows, A
2
and A
3
to
label the rows within large rows, B
1
and B
2
to label the large columns, and
B
3
to label the columns within large columns. Numbers can be allocated by
putting N
1
= A
1
+B
3
, N
2
= A
2
+B
1
and N
3
= A
3
+B
2
. All that is required
is that no non-zero linear combination (modulo 2) of N
1
, N
2
and N
3
contains
only A
1
, B
1
and B
2
, or a subset thereof.
2.4 Eciency and concurrence
The other important statistical property is eciency, which means that the
estimators of the dierences between treatments should have small variance.
At one extreme, we might decide that the innate dierences between mini-
columns are so great that the design in Figure 4 provides no information at
all about the dierence between the average of treatments 1, 2, 3 and the
average of treatments 4, 5, 6; and similarly for minirows. In this case, it
can be shown (see [1, Chapter 7]) that the relevant variances can be deduced
from the matrix
M = m
1
m
2
I
1
m
2

R

1
m
1

C
+J.
Here I is the n n identity matrix and J is the n n all-1 matrix. The
concurrence of symbols i and j in minirows is the number of minirows con-
taining both i and j (which is n when i = j): the matrix
R
contains these
concurrences. The matrix
C
is dened similarly, using concurrences in mini-
columns. It is known that if the o-diagonal entries in the matrix M are all
equal then the average variance is as small as possible for the given values
of m
1
and m
2
, so the usual heuristic is to choose a design in which the o-
diagonal entries dier as little as possible. If m
1
= m
2
, this means that the
sums of the concurrences are as equal as possible. We explore this property
for Sudoku solutions in Section 4.1.
A compromise between these two statistical properties is general balance
[13, 16, 17], which requires that the concurrence matrices
R
and
C
com-
mute with each other. A special case of general balance is adjusted orthogo-
nality [8, 14], for which
R

C
= n
2
J. It can be shown that a gerechte design
11
with rectangular regions is orthogonal in the sense of Section 2.3 if it has
adjusted orthogonality and
2
R
= nm
2

R
and
2
C
= nm
1

C
. This property
is also explored further in Section 4.1.
3 Some special Sudoku solutions
Our main aim in this section is to consider some very special Sudoku solutions
which we call symmetric. We state our main results rst. The proofs will
take us on a tour through parts of nite geometry and coding theory; we
have included brief introductions to these topics, for readers unfamiliar with
them who want to follow us through the proofs of the theorems. Later in
the section, we show how to construct other Sudoku solutions having some
of the statistical properties introduced in the preceding section.
We have seen that a Sudoku solution is a gerechte design for the 9 9
array partitioned into nine 3 3 subsquares. To dene symmetric Sudoku
solutions, we need a few more types of region.
As dened in the last section, a minirow consists of three cells forming
a row of a subsquare, and a minicolumn consists of three cells forming a
column of a subsquare. We dene a broken row to be the union of three
minirows occurring in the same position in three subsquares in a column,
and a broken column to be the union of three minicolumns occurring in the
same position in three subsquares in a row. A location is a set of nine cells
occurring in a xed position in all of the subsquares (for example, the centre
cells of each subsquare).
Now a symmetric Sudoku solution is an arrangement of the symbols
1, . . . , 9 in a 9 9 grid in such a way that each symbol occurs once in each
row, column, subsquare, broken row, broken column, and location. In other
words, it is a multiple gerechte design for the partitions into subsquares,
broken rows, broken columns, and locations. Figure 5 shows a symmetric
Sudoku solution. The square shown has the further property that each of
the 3 3 subsquares is semi-magic, that is, its row and column sums (but
not necessarily its diagonal sums) are 15 (John Bray [6]).
As in the rst section, two Sudoku solutions are equivalent if one can be
obtained from the other by a combination of row and column permutations
(and possibly transposition) which preserve all the relevant partitions, and
re-numbering of the symbols.
The main result of this section asserts that, up to equivalence, there are
12
8 1 6 2 4 9 5 7 3
3 5 7 6 8 1 9 2 4
4 9 2 7 3 5 1 6 8
7 3 5 1 6 8 4 9 2
2 4 9 5 7 3 8 1 6
6 8 1 9 2 4 3 5 7
9 2 4 3 5 7 6 8 1
1 6 8 4 9 2 7 3 5
5 7 3 8 1 6 2 4 9
Figure 5: A semi-magic symmetric Sudoku solution
precisely two symmetric Sudoku solutions. This theorem can be proved by
a computation of the type described in the rst section. However, we give a
more conceptual proof, exploiting the links with the other topics of the title.
We also consider mutually orthogonal sets; we show that the maximum
number of mutually orthogonal Sudoku solutions is 6, and the maximum
number of mutually orthogonal symmetric Sudoku solutions is 4. Moreover,
there is a set of six mutually orthogonal Sudoku solutions of which four are
symmetric. These are exhibited in Figure 10.
Throughout this section will will use GF(3) to denote the nite eld with
three elements (the integers modulo 3).
3.1 Preliminaries
In this subsection we describe briey the notions of ane and projective
geometry and coding theory. Readers familiar with this material may skip
this subsection.
Ane geometry An ane space is just a vector space with the distin-
guished role of the origin removed. Its subspaces are the cosets of the vector
subspaces, that is, sets of the form U +v, where U is a vector subspace and
v a xed vector, the coset representative. This coset is also called the trans-
late of U by v. Two ane subspaces which are cosets of the same vector
subspace are said to be parallel, and the set of all cosets of a given vector
13
subspace forms a parallel class. A transversal for a parallel class of ane
subspaces is a set of coset representatives for the vector subspace.
We use the terms point, line and plane for ane subspaces of
dimension 0, 1, 2 respectively. We denote the n-dimensional ane space over
a eld F by AG(n, F); if [F[ = q, we write AG(n, q).
We will use the fact that a subset of AG(n, F) is an ane subspace if
(and only if) it contains the unique ane line through each pair of its points.
In ane space over the eld GF(3), a line has just three points, and the third
point on the line through p
1
and p
2
is the midpoint (p
1
+p
2
)/2 = (p
1
+p
2
).
Projective geometry Much of the argument in the proof of the main
theorem of this section will be an examination of collections of subspaces
of a vector space. This can also be cast into geometric language, that of
projective geometry.
The n-dimensional projective space over a eld F is the geometry whose
points, lines, planes, etc. are the 1-, 2-, 3-dimensional (and so on) subspaces
of an (n + 1)-dimensional space V (which we can take to be F
n+1
). A point
P lies on a line L if P L (as subspaces of F
n+1
).
For example, a point of the projective space PG(n, F) is a 1-dimensional
subspace of the vector space F
n+1
, and so it corresponds to a parallel class
of lines in the ane space AG(n + 1, F). The points of the projective space
can therefore be thought of as points at innity of the ane space.
We will mostly be concerned with 3-dimensional projective geometry; we
refer to [7, 12]. We will use the following notions:
Two lines are said to be skew if they are not coplanar. Skew lines
are necessarily disjoint. Conversely, since any two lines in a projective
plane intersect, disjoint lines are skew. So the terms disjoint and
skew for lines in projective space are synonyms. We will normally
refer to disjoint lines. Note that disjoint lines in PG(n, F) arise from
2-dimensional subspaces in F
n+1
meeting only in the origin.
A hyperbolic quadric is a set of points satisfying an equation like x
1
x
2
+
x
3
x
4
= 0. Any such quadric contains two rulings, each of which is
a set of pairwise disjoint lines covering all the points of the quadric
(Figure 6). Such a set of lines is called a regulus, and the other set is
the opposite regulus. Any three pairwise disjoint lines of the projective
space lie in a unique regulus. The lines of the opposite regulus are all
the lines meeting the given three lines (their common transversals).
14
A spread is a family of pairwise disjoint lines covering all the points
of the projective space. A spread is regular if it contains the regulus
through any three of its lines. (Any three lines of a spread are pairwise
disjoint, and so lie in a unique regulus.) It can be shown that, if the
eld F is nite, then there exists a regular spread. In particular, this
holds when F = GF(3).

i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i

f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f

e
e
e
e
e
e
e
e
e
e
e
e
e
e e
e
e
e
e
e
e
e
e
e
e
e
e
e
e e

t
t
t
t
t
t
t
t
t
t
t
t
t
t

t
t
t
t
t
t
t
t
t
t
t
t
t
t
e
e
e
e
e
e
e
e
e
e
e
e
e
e
7
7
7
7
7
7
7
7
7
7
7
7
7
7
e
e
e
e
e
e
e
e
e
e
e
e
e
e
7
7
7
7
7
7
7
7
7
7
7
7
7
7
d
d
d
d
d
d
d
d
d
d
d
d
d
d

d
d
d
d
d
d
d
d
d
d
d
d
d
d

Figure 6: A hyperbolic quadric and its two rulings

The fact that any pair of lines in a projective plane intersect is a con-
sequence of the dimension formula of linear algebra. The points and lines
of the plane are 1- and 2-dimensional subspaces of a 3-dimensional vector
space; and if two 2-dimensional subspaces U
1
and U
2
are unequal, then
dim(U
1
U
2
) = dim(U
1
) + dim(U
2
) dim(U
1
+U
2
) = 2 + 2 3 = 1.
The second and third bullet points are most easily proved using coordinates.
We will see an example of a regulus and its opposite in coordinates later.
In cases where regular spreads exist, any three pairwise disjoint lines are
contained in a regular spread.
In the nal section of the paper we briey consider higher dimensions,
and use the fact that PG(2m 1, F) has a spread of (m 1)-dimensional
subspaces.
Coding theory A code of length n over a xed alphabet A is just a set
of n-tuples of elements of A; its members are called codewords. The Ham-
ming distance between two n-tuples is the number of positions in which they
15
dier. The minimum distance of a code is the smallest Hamming distance
between distinct codewords. For example, if the minimum distance of a code
is 3, and the code is used in a communication channel where one symbol in
each codeword might be transmitted incorrectly, then the received word is
closer to the transmitted word than to any other codeword (by the triangle
inequality), and so the error can be corrected; we say that such a code is
1-error-correcting.
A 1-error-correcting code of length 4 over an alphabet of size 3 contains
at most 9 codewords. For, given any codeword, there are 1 +4 2 = 9 words
which can be obtained from it by making at most one error; these sets of nine
words must be pairwise disjoint, and there are 3
4
= 81 words altogether, so
there are at most 9 such sets. If the bound is attained, the code is called
perfect, and has the property that any word is distant at most 1 from a unique
codeword.
It is known that there is, up to a suitable notion of equivalence, a unique
perfect code of length 4 over an alphabet of size 3, the so-called Hamming
code. We do not assume this uniqueness; we will determine all perfect codes
in the course of our proof (see Proposition 3.2).
If the alphabet is a nite eld F, the code C is linear if it is a subspace
of the vector space F
n
. The Hamming code is a linear code. Note that
translation by a xed vector preserves Hamming distance; so, for example, if
a linear code is perfect 1-error-correcting, then so is each of its cosets.
A linear code C of dimension k can be specied by a generator matrix, a
k n matrix whose row space is C. The code with generator matrix
_
0 1 1 1
1 0 1 2
_
(1)
is a Hamming code. Of course, permutations of the rows and columns of this
matrix, and multipication of any column by 1, give generator matrices for
other Hamming codes.
See Hill [11] for further details.
3.2 Sudoku and geometry over GF(3)
Following the idea of the design key described in Section 2.3, we coordinatize
the cells of a Sudoku grid using GF(3) = 0, 1, 2. Each cell c has four
coordinates (x
1
, x
2
, x
3
, x
4
), where
16
x
1
is the number of the large row containing c;
x
2
is the number of the minirow of this subsquare which contains c;
x
3
is the number of the large column containing c;
x
4
is the number of the minicolumn of this subsquare which contains c.
(In each case we start the numbering at zero. We number rows from top to
bottom and columns from left to right.)
Now the cells are identied with the points of the four-dimensional vector
space V = GF(3)
4
. The origin of the vector space is the top left cell. However,
there is nothing special about this cell, so we should think of the coordinates
as forming an ane space AG(4, 3).
Somem regions of the Sudoku grid which we have already discussed are
cosets of 2-dimensional subspaces, as shown in the following table. Each 2-
dimensional subspace corresponds to a line in PG(3, 3); we name these lines
for later reference.
Equation Description of cosets Line in PG(3, 3)
x
1
= x
2
= 0 Rows L
1
x
3
= x
4
= 0 Columns L
2
x
1
= x
3
= 0 Subsquares L
3
x
1
= x
4
= 0 Broken columns L
5
x
2
= x
3
= 0 Broken rows L
6
x
2
= x
4
= 0 Locations L
4
Table 1: Some subspaces of GF(3)
4
In addition, the main diagonal is the subspace dened by the equations
x
1
= x
3
and x
2
= x
4
, and the antidiagonal is x
1
+x
3
= x
2
+x
4
= 2, a coset of
the subspace x
1
= x
3
, x
2
= x
4
. (The other cosets of these two subspaces
are not so obvious in the grid.)
Now, in a Sudoku solution, each symbol occurs in nine positions forming
a transversal for the cosets of the subspaces dening rows, columns, and
subsquares as above (this condition translates into one position in each
row, column, or subsquare). A Sudoku solution is symmetric if it also has
the analogous property for broken rows, broken columns, and locations.
17
We call a Sudoku solution linear if, for each symbol, its nine positions
form an ane subspace in the ane space. All the Sudoku solutions in this
subsection and the next are linear. We will say that a linear Sudoku solution
is of type A if all nine ane subspaces are cosets of the same vector subspace,
and of type B otherwise.
3.3 Symmetric Sudoku solutions
In this section we classify, up to equivalence, the symmetric Sudoku solutions.
We show that there are just two of them; both are linear, and one is of type
A (dened by the nine cosets of a xed subspace), while the other is of type
B (involving cosets of dierent subspaces).
Consider the set of positions where a given symbol occurs in a symmetric
Sudoku solution, regarded as a subset of V = GF(3)
4
. These positions form
a code of length 4 containing nine codewords. Given any two coordinates
i and j, and any two eld elements a and b, there is a unique codeword p
satisfying p
i
= a and p
j
= b (see Table 1). The minimum distance of this
code is thus at least 3, since distinct codewords cannot agree in two positions.
Conversely, if S is a set of points with minimum distance at least 3, then for
any given a and b, there is at most one p S with p
i
= a and p
j
= b; so, if
[S[ = 9, there must be exactly one such point p. So we have shown:
Proposition 3.1 A symmetric Sudoku solution is equivalent to a partition
of V into nine perfect codes.
It is clear from this Proposition that the partition into cosets of a Ham-
ming code gives a symmetric Sudoku solution. We prove that there is just
one further partition, up to equivalence.
Proposition 3.2 Any perfect 1-error correcting code in V = GF(3)
4
is an
ane subspace.
Proof Let H be such a perfect code. Then H consists of 9 vectors, any two
agreeing in at most one coordinate. As above, given distinct coordinates i, j
and eld elements a, b, there is a unique p H with p
i
= a and p
j
= b.
Any two vectors of H have distance at least 3; so

p,qH
d(p, q) 9 8 3 = 216,
18
where d denotes Hamming distance. On the other hand, if we choose any
coordinate position (say the rst), and suppose that the number of vectors of
H having entries 0, 1, 2 there are respectively n
0
, n
1
, n
2
, then the contribution
of this coordinate to the above sum is
n
0
(9 n
0
) +n
1
(9 n
1
) +n
2
(9 n
2
) = 81 (n
2
0
+n
2
1
+n
2
3
) 81 27 = 54,
and so the entire sum is at most 4 54 = 216. So equality must hold, from
which we conclude that any pair of vectors have distance 3 (that is, agree in
one position).
Now take p, q H. Suppose, without loss, that they agree in the rst
coordinate; say p = (a, b
1
, c
1
, d
1
) and q = (a, b
2
, c
2
, d
2
). Since b
1
,= b
2
, the
remaining element of GF(3) is (b
1
+ b
2
). There is a unique element r of
H having rst coordinate a and second coordinate (b
1
+ b
2
); since it must
disagree with each of p and q in the third and fourth coordinates, it must
be r = (a, (b
1
+ b
2
), (c
1
+ c
2
), (d
1
+ d
2
)) = (p + q). This is the third
point on the ane line through p and q. So H is indeed an ane subspace,
as required (cf. p. 14).
Any translate of a perfect code is a perfect code; so any perfect code is
a coset of a vector subspace which is itself a perfect code. We call such a
subspace allowable. Our next task is to nd the allowable subspaces.
Lemma 3.3 The vectors p = (a
1
, a
2
, a
3
, a
4
), and q = (b
1
, b
2
, b
3
, b
4
) are two
linearly independent vectors in an allowable subspace X of V if and only if
the four ratios a
i
/b
i
, for i = 1, 2, 3, 4 are distinct, where 1/0 = is one
ratio that must appear, and the indeterminate form 0/0 does not appear.
Proof The vectors p, q and any two of the standard basis vectors, with just
one non-zero coordinate (equal to 1), must be linearly independent. So the
determinant of the corresponding matrix, which is a
i
b
j
a
j
b
i
, is not zero.
Then the result follows.
Given Lemma 3.3, we see that when a basis for an allowable subspace is
put into row-reduced echelon form, it takes one the following eight possibili-
ties.
_
_
_
_
_
_
1011
or
1022
_
_
_
and
_
_
_
0112
or
0121
_
_
_
_
_
_
or
_
_
_
_
_
_
1012
or
1021
_
_
_
and
_
_
_
0111
or
0122
_
_
_
_
_
_
(2)
19
These are the only allowable subspaces. So any perfect code in V is a coset
of one of those eight vector subspaces.
Our conclusion for symmetric Sudoku solutions so far can be summarised
as follows:
Any symmetric Sudoku solution is linear;
In a symmetric Sudoku solution, the positions of each symbol form a
coset of one of the eight subspaces given above.
Next we come to the question of how such subsets can partition V . One
simple way is just to take all cosets of one of the above 2-dimensional vector
subspaces; this gives the solutions we described above as Type A. Another
choice is the following. Extend an allowable subspace X to an appropriate
3-dimensional vector subspace Y of V . The three cosets of Y partition V ,
and we can look for another allowable subspace X

of Y which can be used to

partition one or two of these cosets. For this to work, it is necessary that the
linear span of X and X

be 3-dimensional. For each choice of an allowable

X, it is easy to check that there are four other allowable X

such that the

span of X and X

is 3-dimensional, but there is no set of three allowable

subspaces such that the span of each pair is 3-dimensional.
Conversely, take any symmetric Sudoku solution, and consider the corre-
sponding partition of V into cosets of allowable 2-dimensional subspaces. If
any pair of such subspaces are distinct and span the whole of V , then any
of their cosets will intersect, contradicting the Sudoku property. Thus their
span must be a 3-dimensional vector subspace Y and hence they are two
subspaces X and X

as in the previous paragraph. Thus X and X

are the
only two allowable subspaces parallel to any set in the partition for a Sudoku
solution. Furthermore, in each of the three cosets of Y , cosets of only one of
X or X

can appear. Thus the Sudoku solutions described in the previous

paragraph are the only ones possible.
Using this analysis we can see that for each choice of one of the 8 allowable
planes, since there are exactly 4 choices for another such that their span is
3-dimensional, there are 8 4/2 = 16 possible choices of such pairs. For
each pair, we want to use each plane to partition at least one of the three 3-
dimensional ane spaces determined by the pair of planes: there are 6 ways
of doing this. Thus there are 6 16 = 96 possible Sudoku solutions of this
sort. In addition, there are 8 possible Sudoku solutions comprising the cosets
of a single plane. This gives 96+8 = 104 total number of symmetric Sudoku
20
solutions, falling into just two classes up to equivalence under symmetries of
the grid.
In the spirit of the Sudoku puzzle, we give in Figure 7 a partial symmetric
Sudoku which can be uniquely completed (in such a way that each row,
column, subsquare, broken row, broken column or location contains each
symbol exactly once). The solution is of type B; that is, it is not equivalent
to the one shown in Figure 5.
7
7
6
4 3
1 5 8
2 7
1 4
4
1
Figure 7: A Sudoku-type puzzle
The fact that there are just two inequivalent symmetric Sudoku solutions,
proved in the above analysis, can be conrmed with the DESIGN program,
which also shows that, if we omit the condition on locations, there are 12
dierent solutions; and, if we omit both locations and broken columns, there
are 31021 dierent solutions. The total number of Sudoku solutions up to
equivalence (that is, solutions with only the conditions on rows, columns,
and subsquares) is 5472730538; this number was computed by Ed Russell
and Frazer Jarvis [18].
3.4 Mutually orthogonal Sudoku solutions
In this section we construct sets of mutually orthogonal Sudoku solutions of
maximum size. The results of the construction are shown in Figure 10.
Theorem 3.4 (a) There is a set of six mutually orthogonal Sudoku solu-
tions. These squares are also gerechte designs for the partition into
21
locations, and have the property that each symbol occurs once on the
main diagonal and once on the antidiagonal. Each of the Sudoku solu-
tions is linear of type A.
(b) There is a set of four mutually orthogonal multiple gerechte designs
for the partitions into subsquares, locations, broken rows and broken
columns; they also have the property that each symbol occurs once on
the main diagonal and once on the antidiagonal. Each of the Sudoku
solutions is linear of type A.
Remark We saw already that the number 6 in part (a) is optimal. The
number 4 in (b) is also optimal. For, given such a set, we can as before
suppose that they all have the symbol 1 in the cell in the top left corner.
Now the 1s in the subsquare in the middle of the top row cannot be in its
top minirow or its left-hand minicolumn, so just four positions are available;
and the squares must have their ones in dierent positions.
Proof (a) Our six Sudoku solutions will all be linear of type A; that is,
they will be given by six parallel classes of planes in the ane space. The
orthogonality of two solutions means that each plane of the rst meets each
plane of the second in a single point. This holds precisely when the two
vector subspaces meet just in the origin (so that their direct sum is the
whole space). In other words, the vector subspaces correspond to disjoint
lines in the projective space PG(3, 3).
In our situation, the ane planes x
1
= x
2
= 0 and x
3
= x
4
= 0 whose
cosets dene rows and columns correspond to two disjoint lines L
1
and L
2
of
PG(3, 3); and the ane plane x
1
= x
3
= 0 whose cosets dene the subsquares
to a line L
3
which intersects both L
1
and L
2
(in the points (0, 0, 0, 1)) and
(0, 1, 0, 0)) respectively). So we have to nd six pairwise disjoint lines which
are disjoint from the given three lines.
Now there is a regulus 1 containing L
1
and L
2
, whose opposite regulus
contains L
3
. Moreover, 1 is contained in a regular spread. Then the six lines
of the spread not in 1 are disjoint from L
3
, and have the required property.
(See Figure 8.)
Calculation shows that the remaining lines of 1 are x
1
x
3
= x
2
x
4
= 0
and x
1
+ x
3
= x
2
+ x
4
= 0, and the other three lines of the opposite regulus
are x
1
x
2
= x
3
x
4
= 0, x
1
+x
2
= x
3
+x
4
= 0, and x
2
= x
4
= 0, which is the
Locations line L
4
(the line such that the cosets of the corresponding vector
22
L
1
L
2
L
3
Figure 8: A regulus, the opposite regulus, and a spread
subspace dene the partition into locations). The main diagonal and the
antidiagonal are cosets of the subspaces corresponding to the other two lines
of 1. Since the remaining six lines of the spread are disjoint from these, our
claim about locations and diagonals follows. It is clear from the construction
that all the corresponding Sudoku solutions are linear of type A.
A dierent set of six mutually orthogonal Sudoku solutions can be ob-
tained by choosing a regulus 1

disjoint from 1 and contained in the spread,

and replacing it by the opposite regulus. This also gives linear solutions of
type A.
(b) For the second part, it is more convenient to work in the ane space
AG(4, 3). As we have seen, a type A symmetric Sudoku solution is given by
the cosets of one of the eight admissible subspaces of V . It is easily checked
that the following four matrices span subspaces with the property that any
two of them meet only in the zero vector, from which it follows that the
corresponding symmetric Sudoku solutions are orthogonal.
_
0 1 1 1
1 0 1 2
_
,
_
0 1 2 2
1 0 2 1
_
,
_
0 1 2 1
1 0 1 1
_
,
_
0 1 1 2
1 0 2 2
_
Another set of four mutually orthogonal symmetric Sudoku solutions is ob-
tained by using the other four admissible subspaces (obtained by changing
the sign of the coordinates in the nal column).
We can use the solution to (b) to nd an explicit construction for (a).
Recall that we seek six lines of the projective space disjoint from the lines
L
1
, L
2
and L
3
. All of these must be disjoint from L
4
also.
Four of these are also disjoint from the lines L
5
and L
6
dened by x
1
=
x
4
= 0 and x
2
= x
3
= 0; these are the four Hamming codes H
1
, . . . , H
4
that we constructed. Now, there is a unique regulus 1

containing L
1
and
23
L
2
and having L
5
and L
6
in the opposite regulus; the other two lines of 1

can be added to the four lines arising from the Hamming codes to produce
the required set of six lines. They have equations x
1
+x
4
= x
2
+x
3
= 0 and
x
1
x
4
= x
2
x
3
= 0. See Figure 9. The resulting six mutually orthogonal
Sudoku solutions are shown in Figure 10; the last four are symmetric.
It can be shown that the four lines H
1
, . . . , H
4
disjoint from the two reguli
themselves form a regulus; they and the lines of the opposite regulus are the
eight Hamming codes.
r
r
r
r
r

L
1
L
2
L
3
L
4
L
5
L
6
H
1
H
2
H
3
H
4
R
..
. .
R

Figure 9: Two reguli in the construction of mutually orthogonal gerechte

designs
111
111
222
222
333
333
749
658
857
469
968
547
475
896
586
974
694
785
444
444
555
555
666
666
173
982
281
793
392
871
718
239
829
317
937
128
777
777
888
888
999
999
416
325
524
136
635
214
142
563
253
641
361
452
326
589
134
697
215
478
952
734
763
815
841
926
687
342
498
153
579
261
659
823
467
931
548
712
385
167
196
248
274
359
921
675
732
486
813
594
983
256
791
364
872
145
628
491
439
572
517
683
354
918
165
729
246
837
238
965
319
746
127
854
864
273
945
381
756
192
593
427
671
538
482
619
562
398
643
179
451
287
297
516
378
624
189
435
836
751
914
862
725
943
895
632
976
413
784
521
531
849
612
957
423
768
269
184
347
295
158
376
Figure 10: Six mutually orthogonal Sudoku solutions
24
This analysis can also be used to dene and count orthogonal symmetric
Sudoku solutions. First we note that, if two symmetric Sudoku solutions are
orthogonal, then both must be of type A. For, as we saw earlier, orthogonality
means that each coset in the rst solution meets each coset in the second in
a single ane point (so the corresponding lines in the projective space are
disjoint). A type B Sudoku solution involves cosets of two 2-dimensional
spaces with non-zero intersection, corresponding to two intersecting lines in
PG(3, 3). But the only lines available are the eight lines of a regulus and its
opposite, and no such line is disjoint from two intersecting lines in the set.
Now two type A symmetric Sudoku solutions are orthogonal if and only
if the corresponding lines belong to the same regulus. So there are 8 3 = 24
such ordered pairs.
4 Further special Sudoku solutions and gen-
eralizations
In the rst subsection of this section, we construct some Sudoku solutions
having some of the desirable statistical properties dened in Section 2. In
the second, we give some generalizations to gerechte designs of other sizes,
using other nite elds.
4.1 The block design in minirows and minicolumns
The cells in the minirows and minicolumns form lines of the ane space
AG(4, 3). In any type A symmetric Sudoku solution comprising all cosets of
a xed vector subspace S, such a line together with S spans a 3-dimensional
subspace which contains three cosets of S. So all the nine lines in this
subspace contain the same three symbols. This means that the 27 minirows
dene just three triples from1, . . . , 9, each triple occuring in nine minirows.
The same condition holds for the minicolumns. Thus the design is orthogonal,
in the sense of Section 2.3. Moreover, the block design on 1, . . . , 9 formed
by the minirows and minicolumns is a 33 grid with each grid line occurring
nine times as a block. Each pair of symbols lies in either 0 or 9 blocks of the
design. (These properties are easily veried by inspection of Figure 5.)
In general, a block design is said to be balanced if every pair of symbols lies
in the same number of blocks. Since the average number of blocks containing
a pair of symbols from1, . . . , 9 in this design is 2273/
_
9
2
_
= 9/2, the design
25
cannot be balanced. But we could ask whether there is a Sudoku solution
which is better balanced than a type A symmetric solution; for example, one
in which each pair occurs in either 4 or 5 blocks. Such solutions exist; the
rst example was constructed by Emil Vaughan [23].
Given such a design with pairwise concurrences 4 and 5, we obtain a
regular graph of valency 4 on the vertex set 1, . . . , 9 by joining two vertices
if they occur in ve blocks of the design. The nicest such graph is the 33
grid, the line graph of K
3,3
. (This graph is strongly regular, and the resulting
design would be partially balanced with respect to the Hamming association
scheme consisting of the graph and its complement: see [1].) Vaughans
solution does not realize this graph, but we subsequently found one which
does. An example is given in Figure 11. (Two vertices in the same row or
column of the 3 3 grid are adjacent.)
1 5 2 6 8 9 7 4 3
3 8 7 1 2 4 9 6 5
9 4 6 3 5 7 1 8 2
2 1 4 8 7 6 3 5 9
6 9 5 4 1 3 2 7 8
8 7 3 5 9 2 6 1 4
5 6 1 9 3 8 4 2 7
7 3 8 2 4 1 5 9 6
4 2 9 7 6 5 8 3 1
t
t
t
t
t
t
t
t
t
1 2 4
9 6 5
7 8
3
Figure 11: A Sudoku solution in which the block design in minirows and
minicolumns has concurrences 4 and 5, and its corresponding graph
We could ask whether even more is true: is there a Sudoku solution in
which each pair of symbols occur together 2 or 3 times in a minirow, 2 or 3
times in a minicolumn, and 4 or 5 times altogether? (We saw in Section 2.4
that balancing concurrences in minirows and minicolumns separately is a
desirable statistical property.) A computation using GAP showed that such
a solution cannot exist; one cannot place more than ve symbols satisfying
these constraints without getting stuck. It is not clear what the best
compromise is.
26
We further found that there exist Sudoku solutions in which the design
in minirows and minicolumns is partially balanced with respect to the 3 3
grid with concurrences (4, 5), (3, 6), (2, 7) or (0, 9), but not (1, 8) (for which
at most four symbols can be placed). The type A linear Sudoku solution in
Figure 5 realizes the case (0, 9).
We also considered another special type of Sudoku solution based on the
properties of the minirows and minicolumns: those for which the designs
formed by minirows and minicolumns have adjusted orthogonality, in the
sense that their concurrence matrices
R
and
C
satisfy
R

C
= 81J, where
J is the all-one matrix. (Here the (i, j) entry of
R
counts the number of
minirows in which i and j both occur, and similarly for
C
.) The special
Sudoku solution of Figure 5 has this property, but it is not unique. (In this
solution, all entries of each concurrence matrix are 0 or 9.) We found that
there are, up to symmetry, 194 Sudoku solutions for which the minirows and
minicolumns have adjusted orthogonality in this sense, of which 104 have the
property that both
R
and
C
have entries dierent from 0 and 9. One of
these solutions is shown in Figure 12.
1 2 3 4 5 6 7 8 9
7 8 9 1 3 2 6 5 4
4 5 6 7 8 9 1 3 2
3 1 2 6 4 5 9 7 8
9 7 8 2 1 3 4 6 5
6 4 5 9 7 8 2 1 3
8 9 1 5 6 4 3 2 7
2 3 7 8 9 1 5 4 6
5 6 4 3 2 7 8 9 1
Figure 12: Minirows and minicolumns form designs with adjusted orthogo-
nality, but the overall design is not orthogonal
A word about the computations reported in this section. The strategy
is to place the symbols 1, . . . , 9 in the grid successively to satisfy the con-
straints. The positions of a single symbol in the grid subject to the Sudoku
constraints that it occurs once in each row, column and subsquare can be
described by a permutation of the set 1, . . . , 9, where the set of positions
27
is (i, (i)) : 1 i 9. There are 6
6
of these Sudoku permutations.
We say that two Sudoku permutations are compatible if they place their
symbols in disjoint cells satisfying the appropriate conditions (for example,
for concurrences 4 and 5, that there are either 4 or 5 occurrences of the two
symbols in the same minirow or minicolumn). Then we form a graph as
follows: the vertex set is the set of all Sudoku permutations, and we join two
vertices if they are compatible. We now search randomly for a clique of size 9
in the compatibility graph: this is a set of nine mutually compatible Sudoku
permutations, dening a Sudoku solution with the required properties.
Adjusted orthogonality of the two designs is not captured by any obvious
compatibility condition on the Sudoku permutations, and we proceeded dif-
ferently. Since each of the two concurrence matrices has diagonal entries 9, we
see that adjusted orthogonality implies that two symbols cannot occur both
in the same minirow and in the same minicolumn. Using this as the com-
patibility condition, we built the compatibility graph, and found all cliques
of size 9, using the GAP package GRAPE [19]. Remarkably, it turned out
that all of them actually give designs with adjusted orthogonality; we know
no simple reason for this fact, since our compatibility condition appears not
strong enough to guarantee this.
4.2 Other nite eld constructions
The construction in Section 3.4 can be generalized.
Proposition 4.1 Let q be a prime power, and a and b positive integers. Let
n = q
a+b
. Partition the n n square into q
a
q
b
rectangles. Then we can
nd
q
a+b
1
(q
a
1)(q
b
1)
q 1
.
mutually orthogonal gerechte designs for this partitioned grid.
Remark If a < b, our upper bound for the number of mutually orthogonal
gerechte designs for this grid is q
b
(q
a
1). If a = 1, this bound is equal to
the number in the theorem, so our bound is attained. If a > 1, however,
the bound is not met by the construction. For example, if p = 2, a = 2 and
b = 3, the bound is 24 but the construction achieves 10. If a and b are not
coprime, we can improve the construction by replacing q, a, b by q
d
, a/d, b/d,
where d = gcd(a, b).
28
Proof Represent the cells by points of the ane space AG(2(a+b), q) with
coordinates x
1
, . . . , x
a+b
, y
1
, . . . , y
a+b
. The rows are cosets of the subspace
x
1
= = x
a+b
= 0, the columns are cosets of the subspace y
1
= =
y
a+b
= 0, and the rectangles are cosets of x
1
= = x
a
= y
1
= = y
b
= 0.
As before, we work in the projective space PG(2(a +b) 1, q). The rst
two subspaces are disjoint, and are part of a spread of q
a+b
1 subspaces of
the same dimension. The third subspace meets the rst in (q
b
1)/(q 1)
points and the second in (q
a
1)/(q1) points, and has (q
a
1)(q
b
1)/(q1)
further points. In the worst case, this subspace meets (q
a
1)(q
b
1)/(q 1)
further spaces of the spread, each in one point. This leaves q
a+b
1 (q
a

1)(q
b
1)/(q 1) spread spaces disjoint from it, as required.
Our construction of mutually orthogonal symmetric Sudoku solutions also
generalizes:
Proposition 4.2 Let q be a prime power, and consider the q
2
q
2
grid,
partitioned into q q subsquares, broken rows, broken columns, and locations
as in the preceding section. Then there exist (q 1)
2
mutually orthogonal
multiple gerechte design for these partitions; this is best possible.
Proof We follow the same method as before, working over GF(q). The lines
of PG(3, q) dening rows, columns, subsquares, broken rows, broken columns,
and locations lie in the union of two reguli with two common lines, which
form part of a regular spread. The remaining (q 1)
2
lines of the spread give
the required designs. The upper bound is proved as before.
Acknowledgements
The left-hand photograph in Figure 2 appears in [5], reproduced by permis-
sion of the Forestry Commission. It can also be found on the web at [15]. We
thank Lesley Smart for permission to use the right-hand photograph, which
was taken by Neil Mason of the Plant and Invertebrate Ecology Division of
Rothamsted Research.
References
[1] R. A. Bailey, Association Schemes: Designed Experiments, Algebra and
Combinatorics, Cambridge Studies in Advanced Mathematics 84, Cam-
bridge University Press, Cambridge, 2004.
29
[2] R. A. Bailey, J. Kunert and R. J. Martin, Some comments on gerechte
designs. I. Analysis for uncorrelated errors. J. Agronomy & Crop Science
165 (1990), 121130.
[3] R. A. Bailey, J. Kunert and R. J. Martin, Some comments on gerechte
designs. II. Randomization analysis, and other methods that allow for
inter-plot dependence, J. Agronomy & Crop Science 166 (1991), 101
111.
[4] W. U. Behrens, Feldversuchsanordnungen mit verbessertem Ausgleich
der Bodenunterschiede, Zeitschrift f ur Landwirtschaftliches Versuchs-
und Untersuchungswesen 2 (1956), 176193.
[5] J. F. Box, R. A. Fisher: The Life of a Scientist. John Wiley & Sons,
New York, 1978.
[6] J. N. Bray, personal communication, February 2006.
[7] P. J. Cameron, Projective and Polar Spaces, QMW Maths Notes 13,
Queen Mary and Westeld College, London, 1991; available from
http://www.maths.qmul.ac.uk/
~
pjc/pps/
[8] J. A. Eccleston and K. G. Russell, Connectedness and orthogonality in
multi-factor designs, Biometrika 62 (1975), 341345.
[9] W. T. Federer, Experimental DesignTheory and Applications, Macmil-
lan, New York, 1955.
[10] The GAP Group, GAP Groups, Algorithms, and Programming, Ver-
sion 4.6; Aachen, St Andrews, 2005, http://www.gap-system.org/
[11] R. Hill, A First Course in Coding Theory, Clarendon Press, Oxford,
1986.
[12] J. W. P. Hirschfeld, Finite Projective Spaces of Three Dimensions, Ox-
ford University Press, Oxford, 1985.
[13] A. M. Houtman and T. P. Speed, Balance in designed experiments with
orthogonal block structure, Ann. Statist. 11 (1983) 10691085.
[14] S. M. Lewis and A. M. Dean, On general balance in rowcolumn designs,
Biometrika 78 (1991), 595600.
30
[15] Materials for the History of Statistics.
URL: http://www.york.ac.uk/depts/maths/histstat/
[16] J. A. Nelder, The analysis of randomized experiments with orthogo-
nal block structure. II. Treatment structure and the general analysis of
variance, Proc. Roy. Soc. London A 283 (1965) 163178.
[17] J. A. Nelder, The combination of information in generally balanced de-
signs, J. Roy. Statistic. Soc. B 30 (1968), 303311.
[18] Ed Russell and Frazer Jarvis, There are 5472730538 essentially dierent
Sudoku grids,
http://www.afjarvis.staff.shef.ac.uk/sudoku/sudgroup.html
[19] L. H. Soicher, GRAPE: a system for computing with graphs and groups.
In L. Finkelstein and W. M. Kantor, editors, Groups and Computation,
volume 11 of DIMACS Series in Discrete Mathematics and Theoreti-
cal Computer Science, pages 287291. American Mathematical Society,
1993. GRAPE homepage:
http://www.maths.qmul.ac.uk/
~
leonard/grape/
[20] Leonard H. Soicher, The DESIGN package for GAP,
http://designtheory.org/software/gap design/
[21] H. D. Patterson, Generation of factorial designs, J. Roy. Statist. Soc. B
38 (1976), 175179.
[22] H. D. Patterson and R. A. Bailey, Design keys for factorial experiments,
Applied Statistics 27 (1978), 335343.
[23] Emil Vaughan, personal communication, November 2005.
[24] F. Yates, The comparative advantages of systematic and randomized
arrangements in the design of agricultural and biological experiments,
Biometrika 30 (1939), 440466.
31