Magic Squares and Sudoku

Author(s): John Lorch

Source: The American Mathematical Monthly, Vol. 119, No. 9 (November 2012), pp. 759-770
Published by: Mathematical Association of America
Stable URL: [Link] .
Accessed: 31/03/2013 03:56

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
[Link]

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@[Link].

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access to
The American Mathematical Monthly.

[Link]

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
 Magic Squares and Sudoku
John Lorch

Abstract. We introduce a family of magic squares, called linear magic squares, and show
that any parallel linear sudoku solution of sufficiently large order can be relabeled so that all
of its subsquares are linear magic. As a consequence, we show that if n has prime factoriza-
k k
tion p11 · · · ptkt and q = min{ p j j | 1 ≤ j ≤ t}, then there is a family of q(q − 1) mutually
orthogonal magic sudoku solutions of order n 2 whenever q > 3; such an orthogonal family is
complete if n is a prime power.

1. INTRODUCTION. Our purpose is to investigate the existence and construction of

orthogonal magic sudoku solutions. A sudoku solution of order n 2 is an n 2 × n 2 array
in which the symbols {0, 1, . . . , n 2 − 1} appear exactly once in each row, column, and
n × n subsquare.1 (Anyone finishing a newspaper sudoku puzzle will have produced
a sudoku solution of order nine, with slightly different symbols.) A sudoku solution of
order n 2 becomes a magic sudoku solution if each n × n subsquare is a magic square
of order n: each row, column, and diagonal of the square adds to the same number.2
A magic sudoku solution is shown in Figure 1; we will have more to say about it later.

15 2 1 12 9 4 7 10 6 11 8 5 0 13 14 3
4 9 10 7 2 15 12 1 13 0 3 14 11 6 5 8
8 5 6 11 14 3 0 13 1 12 15 2 7 10 9 4
3 14 13 0 5 8 11 6 10 7 4 9 12 1 2 15
9 4 7 10 15 2 1 12 0 13 14 3 6 11 8 5
2 15 12 1 4 9 10 7 11 6 5 8 13 0 3 14
14 3 0 13 8 5 6 11 7 10 9 4 1 12 15 2
5 8 11 6 3 14 13 0 12 1 2 15 10 7 4 9
6 11 8 5 0 13 14 3 15 2 1 12 9 4 7 10
13 0 3 14 11 6 5 8 4 9 10 7 2 15 12 1
1 12 15 2 7 10 9 4 8 5 6 11 14 3 0 13
10 7 4 9 12 1 2 15 3 14 13 0 5 8 11 6
0 13 14 3 6 11 8 5 9 4 7 10 15 2 1 12
11 6 5 8 13 0 3 14 2 15 12 1 4 9 10 7
7 10 9 4 1 12 15 2 14 3 0 13 8 5 6 11
12 1 2 15 10 7 4 9 5 8 11 6 3 14 13 0

Figure 1. A magic sudoku solution of order 16.

The combination of sudoku and magic squares seems a match made in heaven, but
the relationship gets off to a rocky start due to the scarcity of magic squares of low
[Link]
MSC: Primary 05B15
1 Any n 2 × n 2 array can be canonically partitioned into n × n subarrays as in Figure 1. These subarrays are

referred to as subsquares throughout.

2 Order-n magic squares commonly have symbol set {1, . . . , n 2 }; for our purposes we prefer {0, . . . ,

n 2 − 1}.

November 2012] MAGIC SQUARES AND SUDOKU 759

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
order. There are no magic squares of order two, thus ruling out magic sudoku in order
four. Likewise, if you are working in your favorite book of order nine sudoku puzzles,
such as [13], you will never produce a magic sudoku solution because the restriction on
the centers of magic squares of order three prevents these squares from being cobbled
into a sudoku solution. Several variants of magic sudoku have been proposed in order
nine, including magidoku and quasi-magic sudoku (both described in [5] and the latter
painstakingly counted in [6]), as well as modular magic sudoku [10].
Because magic squares begin to proliferate in order four and grow like weeds there-
after, magic sudoku solutions seem likely in orders larger than nine, and indeed this is
the case. In these higher orders we choose to use orthogonality as a measure of magic
sudoku diversity. Two sudoku solutions of like order are orthogonal if, upon superim-
position, each ordered pair of symbols appears exactly once. For example, in Figure 2
we see two orthogonal sudoku solutions of order four on the left and the corresponding
array of distinct ordered pairs on the right. Orthogonality, more than just a curiosity,
is an important notion with connections to statistical design, coding, finite projective
geometry, and graph theory. (More information on these connections can be found in
[3], [4] or [14].) A set of pairwise mutually orthogonal sudoku solutions of order n 2
has size at most n(n − 1). This bound is achieved when n is a prime power, but for
general composite n the maximum size of an orthogonal family of sudoku solutions is
unknown.3 However, one can construct orthogonal families of sudoku solutions of size
k k
q(q − 1), where q = min{ p j j | 1 ≤ j ≤ t} when p11 · · · ptkt is the prime factorization
of n. (See [1] or [12].) We show that if q > 3 then we lose nothing by replacing the
word “sudoku” with the words “magic sudoku” in the previous sentence.4

0 1 3 2 0 3 2 1 00 13 32 21
2 3 1 0 2 1 0 3 22 31 10 03
3 2 0 1 3 0 1 2 33 20 01 12
1 0 2 3 1 2 3 0 11 02 23 30

Figure 2. Two orthogonal order-four sudoku solutions together with the corresponding array of distinct or-
dered pairs.

The strategy and structure of the paper are as follows. In Section 2 we introduce a
class of magic squares called linear magic squares (Definition 2.1), indicate a method
for constructing these squares that extends De la Loubère’s famous construction
(Proposition 2.3), and show that they exist in all prime power orders larger than three
(Proposition 2.4). In Section 3 we review parallel linear sudoku solutions, a class
of sudoku solutions used in [1] and [12] to construct orthogonal families of sudoku
solutions, and show that if any parallel linear sudoku solution is relabeled so that its
upper left subsquare is linear magic, then all of its subsquares must be linear magic
(Proposition 3.3).5 This, along with a modification of MacNeish’s product construc-
tion [11], gives the result mentioned at the end of the previous paragraph (Theorem
3.6). Results in Section 3 can be specialized to pandiagonal magic sudoku; we show in
Section 4 how these specialized results can be strengthened by considering Keedwell
sudoku (Corollary 4.3). Section 5 contains appendix material.
3 This is a restriction to sudoku of a classical latin squares open problem. A latin square is an array of order

n with entries chosen from n symbols such that each symbol appears once in each row and column.
4 This, of course, establishes the existence of magic sudoku solutions when q > 3.
5 Linear sudoku is not mentioned explicitly in [12], see [9] for more information.

760 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 119

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
2. LINEAR MAGIC SQUARES AND THEIR CONSTRUCTION. We introduce
a family of magic squares that will enable us to transform certain types of sudoku
solutions into a magic sudoku solution via relabeling. Throughout set q = p k where
p is prime, and F = GF(q) (i.e., the field with q elements). Rows and columns of a
square of order q will be labeled by elements of F from top to bottom and from left
to right, respectively, according to the following lexicographic order on F. An element
ν ∈ F can be identified with a polynomial in Z p [x] of degree less than or equal to k − 1
(modulo some irreducible polynomial f of degree k in Z p [x]), which can in turn be
identified with a k-tuple in Zkp by

ν ↔ vk−1 x k−1 + vk−2 x k−2 + · · · + v0 ↔ (vk−1 , vk−2 , . . . , v0 ). (1)

Under this identification, addition in F corresponds to componentwise addition mod-

ulo p, while multiplication is carried out via polynomial multiplication modulo f .
Viewing 0, 1, . . . , p − 1 as the standard representatives of elements of Z p , we put a
lexicographic order on F (regarding F merely as a set) with

(0, . . . , 0, 0) < (0, . . . , 0, 1) < · · · < (0, . . . , 0, p − 1) < (0, . . . , 1, 0) < · · ·

< ( p − 1, p − 1, . . . , p − 1).

Since rows and columns of a square of order q can be labeled by elements of F, lo-
cations within magic squares of order q are identified with elements of F2 . To avoid
instances of notational ambiguity in our subsequent use of (1), we occasionally write
elements of Z p in teletype to distinguish them from elements of F.

Definition 2.1. A square of order q with entries exhausting the set {0, . . . , q 2 − 1} is
linear magic if
• entries within each row and within each column add to q(q 2 − 1)/2,
• entries within each coset of the one-dimensional subspace h(1, 1)i of F2 add to
q(q 2 − 1)/2, and
• entries within each coset of the one-dimensional subspace h(1, −1)i of F2 add to
q(q 2 − 1)/2.

As a first example we consider the square appearing above the winged figure in
Albrecht Dürer’s engraving Melencolia I (Figure 3). Subtracting one from each entry
of the Dürer square to conform to our choice of symbols, we obtain

15 2 1 12
4 9 10 7
, (2)
8 5 6 11
3 14 13 0

which appears as the upper left subsquare in Figure 1’s magic sudoku solution. To see
why the Dürer square (2) is linear magic, we quickly check that each row and column
sum is 4(42 − 1)/2 = 30, and then move on to the conditions on cosets of h(1, 1)i and
h(1, −1)i in F2 , where F = GF(4). Since −1 = 1 in F (char F = 2), we only check
conditions on the cosets of h(1, 1)i. Under our identification of F with Z22 we have

h(1, 1)i = h(01, 01)i = {(00, 00), (01, 01), (10, 10), (11, 11)},

November 2012] MAGIC SQUARES AND SUDOKU 761

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
Figure 3. Albrecht Dürer’s Melencolia I. The date of the engraving, 1514, is cleverly inserted in the square,
which is believed to be the first magic square appearing in European art.

so the sums of entries in the four distinct cosets of h(1, 1)i are:

(00, 00) + h(01, 01)i : 15 + 9 + 6 + 0 = 30,

(00, 01) + h(01, 01)i : 2 + 4 + 11 + 13 = 30,
(3)
(00, 10) + h(01, 01)i : 1 + 7 + 8 + 14 = 30,
(00, 11) + h(01, 01)i : 3 + 5 + 10 + 12 = 30.

Other examples include

0 14 25 69 74 58 48 35 37
46 30 44 7 9 23 67 78 56
0 23 16 14 7 65 76 60 53 28 39 5 16 18
19 12 5 3 21 61 63 77 40 51 29 19 3 17
8 1 24 17 10 and 26 1 12 59 70 72 38 49 33 . (4)
22 15 13 6 4 42 47 31 21 8 10 54 68 79
11 9 2 20 18 32 43 45 11 22 6 80 55 66
75 62 64 27 41 52 15 20 4
13 24 2 73 57 71 34 36 50

The cosets of h(1, 1)i and h(1, −1)i in the left square of (4) are exactly the two di-
agonals together with the broken diagonals (i.e., upward shifts of the main or anti-
diagonal by one or more cells). We conclude that this square is a pandiagonal magic
square, meaning that each row, column, diagonal, and broken diagonal has the same
entry sum. Meanwhile, in the right square of (4), the only cosets that coincide with
the diagonals or broken diagonals are the main and anti-diagonals themselves. For ex-
ample, the entries {46, 76, 25, 59, 8, 29, 15, 36, 66} in the right square lie in the coset
(1, 0) + h(1, 1)i; these locations do not form a broken diagonal.
These examples suggest that linear magic squares are magic. Indeed, under the
lexicographic ordering of F, the cosets of h(1, 1)i include the main diagonal, namely
h(1, 1)i itself, while the cosets of h(1, −1)i include the anti-diagonal. (This latter fact
can be verified in a manner similar to that of [9], Corollary 3.2.) Further, as we pointed

762 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 119

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
out above, the examples exhibit varied behavior with regard to broken diagonals. In the
special case that q = p (i.e., k = 1), the cosets of h(1, 1)i and h(1, −1)i are exactly
the diagonals and broken diagonals of the square; this is because in these cases F is
cyclic with respect to addition. We summarize these facts in the following proposition.

Proposition 2.2. Using the standard lexicographic ordering on F via identification

with Zkp , any linear magic square of order q = p k is magic, and when k = 1 the
square is pandiagonal magic.

We will show by construction that linear magic squares of order q exist for q > 3.
However, we first identify the integer entries {0, 1, . . . , q 2 − 1} of our squares with
elements of F2 . Each λ ∈ {0, 1, . . . , q 2 − 1} can be written in base q as λ = (λq , λ1 )
where λq is the number of q’s and λ1 is the number of ones. In turn, each of λq and λ1
can be written as a k-tuple of integers in base p, say,
λq = (λqk−1 , λqk−2 , . . . , λq0 ) and λ1 = (λ1k−1 , λ1k−2 , . . . , λ10 ), (5)

where λq j and λ1 j denote the number of p j ’s in λq and λ1 , respectively. Each of the

k-tuples in (5) can be regarded as a member of F via the identification of F with Zkp ,
and therefore λ = (λq , λ1 ) may be regarded as an element of F2 .
And now for the construction. Let A, B, C, D be k × k matrices with entries in Z p
and define T : {0, 1, . . . , q 2 − 1} → F2 by
λq
  
A B
T (λ) = . (6)
C D λ1
The purpose of T is to tie a number λ to a location T (λ) in the square. When k = 1 the
mapping T specializes to the well-known magic square construction of De la Loubère
(see [2]).

Proposition 2.3. The mapping T defined in (6) determines a linear
magic square of
order q = p k if the matrices A, B, C, D, A ± C, B ± D, and CA DB are all nonsingu-
lar over Z p .


Proof. The fact that CA DB is nonsingular ensures a one-to-one correspondence be-
tween symbols and locations, so we proceed with checking magic sums in the cor-
responding square S. Let µ ∈ F ∼ = Zkp . The set of numbers in the µ-th row of S are
the solutions λ = (λq , λ1 ) of the equation Aλq + Bλ1 = µ, or rather Aλq = µ − Bλ1 .
Since both A and B are nonsingular there is a unique solution λq to the rightmost equa-
tion for each choice of λ1 and this collection of λq ’s exhausts Zkp as λ1 ranges over Zkp .
Regarding λq and λ1 as numbers via our identification in (5) above, this says that each
of λq and λ1 achieves each of the values {0, 1, . . . , q − 1} exactly once. Therefore,
summing up the row entries gives
(0 + 1 + · · · + (q − 1)) · q + (0 + 1 + · · · + (q − 1)) · 1
q 2 (q − 1) q(q − 1)
= +
2 2
2
q(q − 1)
= ,
2
as required by Definition 2.1. A similar argument with C, D in place of A, B shows
that S is magic in columns.

November 2012] MAGIC SQUARES AND SUDOKU 763

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
 It remains to show that S is magic in cosets of h(1, 1)i and h(1, −1)i. A number λ
lying in a location within a coset (µ1 , µ2 ) + h(1, 1)i must satisfy

µ1 α
   
T (λ) = +
µ2 α

for some α ∈ F ∼
= Zkp , or equivalently

(A − C)λq + (B − D)λ1 = µ1 − µ2 . (7)

Using the fact that both A − C and B − D are nonsingular, we may apply an argument
identical to that given immediately above for the rows of S to conclude that S is magic
on cosets of h(1, 1)i. An analogous argument together with the fact that A + C and
B + D are nonsingular implies that S is magic on cosets of h(1, −1)i.

We can render a 180◦ rotation of the Dürer square (2) using our construction. Here
p = 2, k = 2, and

0 1 1 1
 
 
A B  1 0 1 1 
= .
C D 1 1 1 0 
1 1 0 1

For instance, the fact that

0 1 1 1 1 0
    
 1 0 1 1  0   1 
 1 = (8)
1 1 0  0   1 
1 1 0 1 0 1

confirms that the symbol “8” (represented by the column vector on the left side of (8))
lies in the second row and last column of the rotated Dürer square (represented by the
column vector on the right side of (8)).6
This construction is useful only if there exist matrices A, B, C, D satisfying the
conditions of Proposition 2.3. Except for the case when k = 1 and p = 2, 3, such
matrices exist for all values of p and k.

Proposition 2.4. Linear magic squares exist in all prime power orders q with q > 3.

Proof. For various values of p and k we provide examples of matrices A, B, C, D

satisfying the conditions of Proposition 2.3. Here Ik denotes the k × k identity matrix,
Ak is the k × k matrix with 1’s on the anti-diagonal and 0’s elsewhere, and
 
1 2 2
J =  1 0 2 .
1 2 1

6 Remember that we begin counting rows and columns with 0.

764 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 119

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
 Throughout we set B = C = Ik . When p > 3 put A = 2Ik , and D = −2−1 Ik . In
all remaining cases we set A = D, so it suffices to specify A. When p = 3 we keep in
mind the k × k matrices
0 2A k−3
 
" # 0
2
0 2Ak/2
,
 
and  0 J 0 (9)
Ak/2 0  
A k−3 0 0
2

where “0” represents the zero matrix of the appropriate size. For even values of k we
let A be the left-hand matrix in (9), while if k > 1 is odd we let A be the right-hand
matrix in (9).
If p = 2 we keep in mind the k × k matrices

0 0 1
 
" #
0 1
and  Ik−2 0 1  . (10)
 
Ik−1 1
0 1 0

When k is even we let A be the left-hand matrix in (10), while if k > 1 is odd we
let A be the right-hand matrix in (10).

3. ORTHOGONAL MAGIC SUDOKU SOLUTIONS. In this section, we con-

struct orthogonal families of magic sudoku solutions. These families are complete
(i.e., as large as possible) when the sudoku solutions are of order q 2 , where q > 3 is
a prime power. Parallel linear sudoku solutions, first introduced by Bailey, Cameron,
and Connelly [1], provide a key tool in the construction.

3.1. A primer on parallel linear sudoku solutions. We recall the notion of a par-
allel linear sudoku solution (originally described in [1]). As before, let F be the finite
field of order q. The set of cell locations within a sudoku solution of order q 2 can be
identified with the vector space F4 over F: Each location has an address (x1 , x2 , x3 , x4 )
(denoted x1 x2 x3 x4 hereafter), where x1 and x2 denote the large row and mini row of
the location, respectively, while x3 and x4 denote the large column and mini column
of the location, respectively. (A large row is a row of q subsquares, while a mini row
is a row of q 2 cell locations within a given large row, similarly for columns.) Using
our identification of F with Zkp , large rows can be labeled in increasing lexicographic
order from top to bottom starting from zero, while mini rows can be similarly labeled
from top to bottom within a given large row. The same is true for columns, with labels
increasing as we move from left to right. It is helpful to note that within a given address
x1 x2 x3 x4 the pair (x1 , x3 ) determines a subsquare, while the pair (x2 , x4 ) determines a
location within a given subsquare.
An example is given in Figure 4 with F = Z3 . The asterisked symbol lies in position
0122 because it resides in the 0-th large row and 1-st mini row within that large row,
and it resides within the 2-nd large column and 2-nd mini column with that large col-
umn. Another interpretation is that the asterisked symbol lies in the (0, 2)-subsquare,
and the (1, 2)-location within that subsquare. In this example the lexicographic order-
ing is trivial because F is a prime field.

Definition 3.1. A sudoku solution is parallel linear if for each symbol, the collec-
tion of locations containing that symbol is a coset of a single two-dimensional vector
subspace g of F4 .

November 2012] MAGIC SQUARES AND SUDOKU 765

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
 0 1 2 4 5 3 8 6 7
3 4 5 7 8 6 2 0 1∗
6 7 8 1 2 0 5 3 4
1 2 0 5 3 4 6 7 8
4 5 3 8 6 7 0 1 2
7 8 6 2 0 1 3 4 5
2 0 1 3 4 5 7 8 6
5 3 4 6 7 8 1 2 0
8 6 7 0 1 2 4 5 3

Figure 4. A parallel linear sudoku solution generated by g = h1002, 0212i with asterisked symbol in location
0122.

The word “parallel” in Definition 3.1 refers to the q 2 parallel cosets of g.7 For paral-
lel linear sudoku solutions the requirements that each symbol meets each column, row,
and subsquare exactly once translate to the requirement that the subspace g has trivial
intersection with gc = h1000, 0100i, gr = h0010, 0001i, and gss = h0100, 0001i.
For example, the sudoku solution in Figure 4 is parallel linear, generated by the
two-dimensional subspace g = h1002, 0212i of Z43 . For practice, one might check that
g meets gr , gc , and gss trivially, that locations determined by the elements of g con-
tain the symbol “0”, and that locations determined by 0101 + g contain the symbol
“4”. The sudoku solution in Figure 1 is also parallel linear, generated by the subspace
g = h1010, 0111i ⊂ F4 , where F = GF(4). We identify F with Z22 as in (1), with
lexicographic order 00 < 01 < 10 < 11. Field operations on F can be given by asso-
ciating a pair (α, β) ∈ Z22 with the polynomial αx + β ∈ Z2 [x]/hx 2 + x + 1i. Given
this structure, the reader might verify that the symbol “5” in the top row of Figure 1
lies in the location (00, 00, 10, 11), that the locations in g contain the symbol “15”,
and that locations in (00, 01, 00, 10) + g contain the symbol “10”.

3.2. Magic parallel linear sudoku solutions. For F = GF(4), we assigned labels to
the cosets of g = h1010, 0111i ⊂ F2 in such a way that the Dürer square (2) would
be the upper left subsquare of the magic sudoku solution in Figure 1. All of the other
subsquares of that sudoku solution also turn out to be linear magic; we now see why
this is not a coincidence.

Lemma 3.2. Suppose M is a parallel linear sudoku solution whose symbols lie
in cosets of a two-dimensional subspace g of F4 . The entire set of cosets of g is
{(0, α, 0, β) + g | α, β ∈ F}.

Proof. Since there are q 2 distinct cosets of g and q 2 proposed representatives of these
cosets, we only need show that equal cosets correspond to equal representatives. If
(0, α1 , 0, β1 ) + g = (0, α2 , 0, β2 ) + g then (0, α1 − α2 , 0, β1 − β2 ) ∈ g, but it is also
true that (0, α1 − α2 , 0, β1 − β2 ) ∈ gss . Since g ∩ gss = 0000, it follows that α1 = α2
and β1 = β2 and hence that the representatives are identical.

Proposition 3.3. Let q be a prime power with q > 3. Any parallel linear sudoku so-
lution of order q 2 can be relabeled so that each subsquare is linear magic.
7 One also finds in [1] a description of “nonparallel” linear sudoku solutions in which cosets of two distinct

subspaces of F2 are used to house symbols.

766 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 119

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
Proof. Suppose q > 3 and M is a parallel linear sudoku solution generated (up to
relabeling) by a two-dimensional subspace g of F4 . By Proposition 2.4 we may assign
labels in M so that the upper left subsquare of M is linear magic. We claim that this
labeling scheme forces all of the subsquares of M to be linear magic. The locations of
two symbols b1 and b2 in M are the members of two cosets of g; by Lemma 3.2 these
have the form

(0, α1 , 0, β1 ) + g and (0, α2 , 0, β2 ) + g,

respectively. Because g ∩ gss = 0000, for each pair (µ, δ) ∈ F2 there is a unique pair
(λ, γ ) ∈ F2 such that (µ, λ, δ, γ ) ∈ g. Therefore the locations of b1 and b2 in the
(µ, δ)-subsquare of M are (µ, α1 + λ, δ, β1 + γ ) and (µ, α2 + λ, δ, β2 + γ ), respec-
tively. If b1 and b2 happen to lie in the same row of any one of these subsquares, then
α1 = α2 , which in turn forces b1 and b2 to lie in the same row in every subsquare. Sim-
ilarly, if b1 and b2 lie in the same column of any given subsquare, then they must lie in
the same column in every subsquare. This says that if, say, the upper left subsquare of
M is row magic and column magic, then so is every subsquare of M.
The positions of the symbols b1 , b2 in the upper left subsquare are (0, α1 , 0, β1 ) and
(0, α2 , 0, β2 ), respectively. Another way of saying this is that the positions of b1 , b2
within the (0, 0) subsquare are (α1 , β1 ) and (α2 , β2 ). Note b1 , b2 lie in the same coset
of h(1, 1)i within the (0, 0)-subsquare of M if and only if (α1 , β1 ) = (α2 , β2 ) + (c, c)
for some c ∈ F, or equivalently α1 − α2 = β1 − β2 . Meanwhile, the positions of b1 , b2
within the (µ, δ)-subsquare of M are (α1 + λ, β1 + γ ) and (α2 + λ, β2 + γ ), and since
α1 − α2 = β1 − β2 we know (α1 + λ) − (α2 + λ) = (β1 + γ ) − (β2 + γ ). It follows
that b1 , b2 lie in the same coset of h(1, 1)i within the (µ, δ)-subsquare of M. A similar
argument applies in the case that b1 , b2 lie in a coset of h(1, −1)i within the (0, 0)-
subsquare. From this we conclude that if the upper left subsquare of M is magic on
cosets of h(1, 1)i and on cosets of h(1, −1)i, then so is every subsquare of M.

Two parallel linear sudoku solutions are orthogonal if and only if their generating
two-dimensional subspaces of F4 intersect trivially. Using this idea it is possible [9] to
show that if f (x) = x 2 + a1 x + a2 ∈ F[x] is irreducible then the collection

{h(1, 0, αa1 − β, αa2 ), (0, 1, −α, −β)i | α, β ∈ F, α 6 = 0}

of two dimensional subspaces of F4 will produce, up to relabeling, a complete pairwise

mutually orthogonal family of sudoku solutions of order q 2 . Putting this together with
Proposition 3.3 we have the following.

Corollary 3.4. Let q be a prime power with q > 3. There exists a complete family of
q(q − 1) pairwise mutually orthogonal magic sudoku solutions of order q 2 .

3.3. Magic sudoku solutions in general orders. Thus far our discussion has been
limited to magic sudoku solutions of order q 2 , with a q a prime power. In this section
we produce orthogonal magic sudoku solutions in general orders. The key ingredient
is a famous construction of orthogonal latin squares due to MacNeish [11].
If B = (bi j ) is a latin square of order n and c is a symbol, we let (c, B) denote the
latin square of order n whose (i, j)-th entry is the ordered pair (c, bi j ). MacNeish [11]
asserts that if {A(1) , . . . , A(r ) } and {B (1) , . . . , B (r ) } are families of mutually orthogonal
latin squares of order m and n, respectively, and C (k) = [(auv (k)
, B (k) )] for 1 ≤ u, v ≤ m,
(1) (r )
then {C , . . . , C } is a set of mutually orthogonal latin squares of order mn. The latin
square C (k) is called the MacNeish product of A(k) and B (k) . Unfortunately, problems

November 2012] MAGIC SQUARES AND SUDOKU 767

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
arise when we try to specialize this result to sudoku. The array C (k) need not be a
sudoku solution even if both A(k) and B (k) are sudoku solutions. Pedersen and Vis [12]
smartly overcome this problem by showing that one can reorder the rows/columns of
the C (k) independently of A(k) and B (k) to render a new array C̃ (k) , which we call a
modified MacNeish product, that is guaranteed to be a sudoku solution if both A(k)
and B (k) are sudoku solutions. Further, this reordering does not affect orthogonality.
(See [12] for details.)
Now, let A and B be magic sudoku solutions of orders r 2 and t 2 with symbols
{0, 1, . . . , r 2 − 1} and {0, 1, . . . , t 2 − 1}, respectively. It can be arranged [12] so that
any subsquare S of a modified MacNeish product of A and B is a MacNeish product
of two subsquares X = (xi j ) and Y = (yi j ) of A and B, respectively. That is, we can
write

(x11 , Y ) (x12 , Y ) · · · (x1r , Y )

 
 (x21 , Y ) (x22 , Y ) · · · (x2r , Y ) 
S= .. .. .. .
 . . . 
(xr 1 , Y ) (xr 2 , Y ) · · · (xrr , Y )

We claim that S is a magic square on the symbols {0, 1, . . . , (r t)2 − 1}, where we are
identifying an entry (xi j , ykl ) of S with the number xi j t 2 + ykl ∈ {0, 1, . . . , (r t)2 − 1}.
For example, if we sum the entries of the first row of S we can use the fact that both X
and Y are magic to obtain

t (t 2 − 1) t (t 2 − 1)
   
t x11 t 2 + + · · · + t x1r t 2 +
2 2
r (r 2 − 1) 3 t (t 2 − 1)
= t + r
2 2
r t ((r t)2 − 1)
= ,
2

which is the desired magic sum. A similar computation can be made for any row,
column, or diagonal of S, so S is magic. Therefore, we obtain the following lemma.

Lemma 3.5. A modified MacNeish product of two magic sudoku solutions is again a
magic sudoku solution.

Applying MacNeish’s result [11] together with Lemma 3.5 and Corollary 3.4 yields
the following theorem.

k k
Theorem 3.6. If n has prime factorization p11 · · · ptkt and q = min{ p j j | 1 ≤ j ≤ t},
then there is a family of q(q − 1) pairwise mutually orthogonal magic sudoku solutions
of order n 2 whenever q > 3.

4. MAGIC KEEDWELL SUDOKU. Because linear magic squares of prime order

are pandiagonal (Proposition 2.2), we have the following corollary to Proposition 3.3.

Corollary 4.1. Let p > 3 be prime. Any parallel linear sudoku solution of order p 2
can be relabeled so that each subsquare is pandiagonal.

768 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 119

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
 However, if we are interested in obtaining magic sudoku solutions whose sub-
squares are pandiagonal, we can do much better than Corollary 4.1. It is known [2]
that pandiagonal magic squares exist in order n if and only if n > 3 and n is not
singly even (i.e., not congruent to 2 modulo 4). Methods of constructing pandiago-
nal magic squares abound; several, including De la Loubère’s method, are discussed
in [2]. Meanwhile, Keedwell [7] constructs sudoku solutions of order n 2 by applying
combinations of elementary row and column permutations, denoted α and β, respec-
tively, to a fixed n × n subsquare K containing n 2 distinct symbols. Specifically, α K
is the subsquare one obtains from K by shifting the rows of K up by one row, while
β K is obtained by shifting the columns of K left by one column. Now let M be the
array of order n 2 whose (i, j)-th subsquare is α ci β d j K for some ci , di ∈ Z≥0 . In [7] it
is shown that one can always pick ci , d j , with i, j ∈ {1, . . . , n} so that M is a sudoku
solution, independent of K . This M is known as a Keedwell sudoku solution. Parallel
linear sudoku solutions of order p 2 ( p prime) are special kinds of Keedwell sudoku
solutions.
As an example, the results of Keedwell [5] indicate that for any 4 × 4 array K
consisting of 16 distinct symbols, the order-16 square

K αK α2 K α3 K
αβ K α βK
2
α βK
3
βK
(11)
α2β 2 K α3β 2 K β2 K αβ 2 K
α3β 3 K β3 K αβ 3 K α2β 3 K

is a diagonal sudoku solution.8 If K is a pandiagonal magic square, say

14 9 2 5
3 4 15 8
K = ,
13 10 1 6
0 7 12 11

then the sudoku solution (11) becomes

14 9 2 5 3 4 15 8 13 10 1 6 0 7 12 11
3 4 15 8 13 10 1 6 0 7 12 11 14 9 2 5
13 10 1 6 0 7 12 11 14 9 2 5 3 4 15 8
0 7 12 11 14 9 2 5 3 4 15 8 13 10 1 6
4 15 8 3 10 1 6 13 7 12 11 0 9 2 5 14
10 1 6 13 7 12 11 0 9 2 5 14 4 15 8 3
7 12 11 0 9 2 5 14 4 15 8 3 10 1 6 13
9 2 5 14 4 15 8 3 10 1 6 13 7 12 11 0
.
1 6 13 10 12 11 0 7 2 5 14 9 15 8 3 4
12 11 0 7 2 5 14 9 15 8 3 4 1 6 13 10
2 5 14 9 15 8 3 4 1 6 13 10 12 11 0 7
15 8 3 4 1 6 13 10 12 11 0 7 2 5 14 9
11 0 7 12 5 14 9 2 8 3 4 15 6 13 10 1
5 14 9 2 8 3 4 15 6 13 10 1 11 0 7 12
8 3 4 15 6 13 10 1 11 0 7 12 5 14 9 2
6 13 10 1 11 0 7 12 5 14 9 2 8 3 4 15
8 Here the modifier diagonal means that every symbol appears on each of the two array diagonals.

November 2012] MAGIC SQUARES AND SUDOKU 769

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions
We observe that all of the subsquares of this sudoku solution are pandiagonal magic, as
well they should be, because K is pandiagonal magic and the operators α and β send
diagonals/broken diagonals to diagonals/broken diagonals. Putting all of this together
we have the following proposition.

Proposition 4.2. Suppose n > 3 is not singly even. Any Keedwell sudoku solution of
order n 2 can be relabeled so that each subsquare is pandiagonal magic.

Further, Theorems 4.2 through 4.4 of [8] guarantee the existence of a mutually
orthogonal family of p( p − 1) Keedwell solutions of order n 2 , where p is the smallest
prime factor of n. By Proposition 4.2 each of these solutions can be relabeled to obtain
a pandiagonal magic sudoku solution.

Corollary 4.3. Suppose n > 3 is not singly even. Let p be the smallest prime divisor of
n. There exist a family of p( p − 1) pairwise mutually orthogonal pandiagonal magic
sudoku solutions of order n 2 .

ACKNOWLEDGMENTS. I thank the referees and the Monthly editor for leaving this article far better than
they found it.

REFERENCES

1. R. A. Bailey, P. J. Cameron, R. Connelly, Sudoku, gerechte designs, resolutions, affine space, spreads,
reguli, and Hamming codes, Amer. Math. Monthly 115 (2008) 383–403.
2. W. W. R. Ball, H. S. M. Coxeter, Mathematical Recreations and Essays, thirteenth edition. Dover Publi-
cations, New York, 1987.
3. C. J. Colbourn, J. H. Dinitz, Mutually orthogonal latin squares: a brief survey of constructions,
J. Statist. Plann. Inference 95 (2001) 9–48; available at [Link]
3758(00)00276-7.
4. The CRC Handbook of Combinatorial Designs. Edited by C. J. Colbourn and J. H. Dinitz. CRC Press,
Boca Raton, 1996.
5. T. Forbes, Quasi-magic sudoku puzzles, M500 215 (2007) 1–11.
6. S. K. Jones, S. Perkins, P. A. Roach, Properties, isomorphisms, and enumeration of 2-quasi-magic sudoku
grids, Discrete Math. 313 no. 13 (2011) 1098–1110; available at [Link]
disc.2010.09.026.
7. A. D. Keedwell, On sudoku squares, Bull. Inst. Combin. Appl. 50 (2007) 52–60.
8. J. D. Lorch, Mutually orthogonal families of linear sudoku solutions, J. Aust. Math. Soc. 87 (2009) 409–
420; available at [Link]
9. , Orthogonal diagonal sudoku solutions: An approach via linearity, Australas. J. Combin. 51
(2011) 139–145.
10. J. D. Lorch, E. L. Weld, Modular magic sudoku (to appear).
11. H. F. MacNeish, Euler squares, Ann. Math. 23 (1922) 221–227; available at [Link]
2307/1967920.
12. R. M. Pedersen, T. J. Vis, Sets of mutually orthogonal Sudoku latin squares, College Math. J. 40 (2009)
174–180; available at [Link]
13. P. Riley, L. Taalman (Brainfreeze Puzzles), No-Frills Sudoku. Puzzle Wright Press, Sterling Publishing
Company, New York, 2011.
14. F. S. Roberts, Applied Combinatorics, Prentice-Hall, Englewood Cliffs, New Jersey, 1984.

JOHN LORCH is a graduate of Palmer High School, the University of Colorado at Colorado Springs,
and Oklahoma State University. His mathematical interests mostly lie in topics that can be readily shared
with undergraduate students. Aside from mathematics he enjoys blues guitar, European history, classic hor-
ror/supernatural fiction, and stupid, juvenile, wife-annoying movies.
Department of Mathematical Sciences, Ball State University, Muncie, IN 47306 USA
jlorch@[Link]

770 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 119

This content downloaded from [Link] on Sun, 31 Mar 2013 [Link] AM

All use subject to JSTOR Terms and Conditions