Finite geometric configurations
Citation for published version (APA):
Seidel, J. J. (1976). Finite geometric configurations. In P. Scherk (Ed.), Foundations of Geometry (Selected
papers of a conference, Toronto, Canada, 1974) (pp. 215-250). University of Toronto Press.
Document status and date:
Published: 01/01/1976
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
Proc. Conference
11Foundations of Geometry
11ed. P. Sherk,
Univ
. of Toronto Press (1976),
p.
2
1
5 - 250.
FINITE GEOMETRIC CONFIGURATIONS1
J.J. Seidel
l. INTRODUCTION
We present five elementary examples in 3- and 4-dimensional
spaces over Galois fields GF (q) , over the reals lR, and
over the complexes
a:.
They serve as an introduction for the subsequent sections. These sections treat discretegeometry, that is a mixture (with a geometric flavour) of
discrete mathematics, combinatorics, finite groups,
geo-metry which has certain applications in coding theory,
statistical designs, network theory, and graph theory.
1.1. Hadamard matrices
Given a cube in
m
3 (with side 2, centre at the origin, sides parallel to the coordinate axes), we wish to construct aregular tetrahe?ron by selecting 4 among the 8 vertices of the cube. The answer is provided by the points
(l, l, l)
[11
11]
(l,-1,-l) The matrix 1 l -1 -1
H4
(-1,1,-l) l -1 1 -1 =:
(-1 ,-1 t 1) l -1 -1 1
1. Notes taken by J.G. Sunday .
c-is a Hadamar4 matrix of order 4.
DEFINITION. A Hadamard matrix Hr of the order r is a square matrix of order r, with elements ±1, satisfying
H u1' = ri .
r r r
Necessary conditions for the existence of Hr are
r
=
2, r=
0 (mod 4).It has been conjectured that these conditions are suffi-cient. This conjecture has been verified for r < 188 and for several infinite series. For the state of affairs we
refer toM. Hall {6), J.H. van Lint [18), J.S. Wallis [20].
1.2. A perfect ternary code
In the vector space V(4, 3) of dimension 4 over GF(3) con-sider the plane spanned by
f :; (1, 0, 1, 2) 1 g = (0, 1, 1, 1) o
There are 9 vectors in this plane, all but the origin having 3 non-zero coordinates. Hence each pair of the 9 vectors has 3 coordinates in which they differ (has Ham-ming distance 3). Calling words the 81 vectors in V(4, 3), code words the 9 vectors in the plane, we have found a 1-error-correcting code. This code (the plane) is linear and perfect. This means the following. Around each code word we draw a sphere of (Hamming) radius 1, containing 8 words. The 9 spheres are mutually disjoint, and exhaust V(4, 3) by the count
9 (1 + 8) = 81.
216
1.3. The binary projective 3-space PG(3, 2)
Consider the vector space V(4, 2) of dimension 4 over GF(2) provided with the non-degenerate alternating bilinear form
B(x, y)
=
~1~2 + ~2~1 + ~3~~ + ~~~3' for X = ( ~ 1 I ~2 I ~ l I ~ 4) I Y :; ( ~ l I ~ 2 I ~ l I ~ ~) • We make the following observations:
1. The symplectic group Sp(4, 2) acts transitively on V ( 4 , 2 ) \ { 0} •
2. The 16 x 16 matrix (B(x, y)] yields (28- J],
x,yEV
which is a Hadamard matrix of order 16 (here J is the all one matrix of order 16).
3. Th~ columns of the matrix
(B J - B]
constitute a binary (32, 16, 8)-code.
4. Consider the graph whose vertices are the 15 vee-tors of V\{0}, in which x andy are adjacent if and only if B(x, y)
=
0. This symplectic graph is strongly regular, and even rank 3, under the action of Sp(4, 2). Obviously, Sp(4, 2) acts not 2-transitively on the graph!5. The extension of Sp(4, 2) by the translations of V(4, 2) does act 2-transitively on the vectors of V(4, 2). It maps the set of the triples {x, y,
z
}
withB(x, y) + B(y, z) + B(x, z) 0
onto itself. This set defines a regular two-graph. 6. Since Sp(4, 2) ~ Symm(6) the graph under 4 is easily drawn.
1.4. Equiangular lines
In a set of equiangular lines the angle ~ between each pair of distinct lines is the same. We consider two examples in IR3,where we have chosen a unit vector p. along each
l_
line (in either of the 2 directions), and their matrix C defined by
p = (<p., p .>),
l_ ]
1
C =cos ~(P-I].
The examples are the 4 diagonals of the cube and the 6 dia-gonals of the icosahedron:
1 3 l 1 I I 0
,..0.-r:
+ + + + +lfo
+:j
0 + +l:
0 + 0 + +c4
+ 0~j'
c6
l:
+ 0 + + + + 0 + + 0 We observe that (C 4 - I) (C ~ + 31) 0,c
2
= 6 51, $ 1 cos~ 1 cos =3'
=-/5
With these examples, we associate the graphs'
l.
;. ~·:
:
•' :or~ ':~ ' .t ,J 'f -~ 0 0 1 1 3 't 0 0 5If, in the second example, we take 6' instead of 6, then the graph is switched into
1
t,'
5
The effect on the matrix C6 of this new choice of 6' is C'
=
DCDwhere D is a diagonal matrix, whose diagonal elements are
) 1, 1, 1, 1, 1, -1. This exemplifies the relations between sets of equiangular lines, classes of graphs, and two-graphs. Indeed, the triples of lines are of two kinds: those with (ph, p.) (p., p.) (p., ph) > 0 such as (h, i, j)
l_ l_ ] J
=
(1, 4, 6), and those with (ph' p.)(p., p.)(p., ph) < 01. l_ ] ]
such as (h, i, j)
=
(1, 2, 3). In the second example, the automorphism group of the set of equiangular lines is the icosahedral group A5• It contains as subgroups the
dihe-dral groups 010 and 06, the automorphism groups of the two
graphs mentioned above, which are related by switching with
respect to vertex 6.
1.5. Complex equiangular lines
we rearrange the Hadamard matrix H~ of example 1.1 into
+ +
+ I + C.
+ +
+
The skew matrix C, multiplied by the complex number i , is
a hermitian matrix satisfying
(iC) 2
= 3I. j
Hence iC
+
II) is hermitian positive. semidefinite of rank2, and may be considered as the matrix of the hermitian
products of the vectors (
11,
0) , ( i,/2) , (
i, w/2) ,(i, w2
12)
where w is a primitive cube root of unity. Hence we have 4 equiangular lines in the complex 2-dimensionalplane
a:
2• In real space
m
~ these correspond to 4equi-isoclinic planes.
2. CODES
2.1. Definitions
Using a finite set S of cardinality s as an alphabet, we
may form sn distinct words of length n; that is, each
el-ement of the set
Sn = { x = x 1 x 2 • • • xn I xi f S, 1 s i s n}
is considered a word. The degree of a "spelling error"
220
might be measured by the number of positions in which a
given word differs from the intended word, so a useful
de-vice is the Hamming distance d(~,
l l
between two words in Sn:the number of positions in which x and
¥
differ.
n It is elementary to verify that d is a metric on S , and
we may occasionally refer to the closed ball
B[5:;
rl
= {~ ~: snI
d(5:, ~)s rl.
For instance, if S
=
{a, b}, then the two words x=
aabbb andl
=
abaab in S5 have Hamming distance 3, and B[aabbb;1]
consists of the words aabbb, babbb, abbbb, aaabb, aabab, and aabba.n
Often, we want only a subset of S to form a
vocabu-lary of meaningful words; such
a
subset is said to be a code, and its elements are called code words.A
code is called an (M, n, d)-code if the code consists of M codewords, each of length n, such that the Hamming distance
between any two code words is at least d. In
if c is a code word, then any element of
a[fi
this case, [d;l)] is
closer to c than
ing from c in at
to any other code word, so a word differ-d-1
most
(--2-] places might reasonably be as-sumed to be
f•
up to a spelling error; we say that thed . ld-1) .
of mutually dtsjoint closed balls covers all of sn, we say that the code is perfect.
2.2. Binary codes
For a binary code, one may easily take the alphabet
s
to be{1
, -1}.
Abbreviating1
and-1
to+ and- respectively, we may group corresponding positions for two words ~·l
in Sn as follows~ + + + + + + + +
-
- -
-y + + + + + + + +
No. of positions PI p2 p3
(where, of course, PI + p2 + p3 + p4 n) , so that Hamming distance is given by
d(~, y)
=
p2 + p3and the usual inner product < , > is given by
(2.A) <~,
l>
= P1 - P2 - P3 + P4 = n - 2·d(~,
y
)
.
We may use these notions to establishp4
the
THE PLOTKIN BOUND. If the parameters of a binary (M, n, d)-code satisfy n < 2d, then M $ 2d/(2d- n).
PROOF. The M code words may be used to form the rows of an M x n matrix P. According to (2.A), the Hamming distance
between distinct code words is at least d precisely when
fOr all dist1'nct ~~ V L E Sn. But t e 1nner products are h ' given as entries in the symmetric M x M matrix PPt.
There-:; ;
~
•.
fore, for an (M, n, d)-code,
(2. B) PPt $
element\lrise
ni
+
(n - 2d) (J - I)=
2di + (n - 2d)J.If jM is a column vector of M l's, then
(j~
P) (Pt jM) is the square of the ordinary euclidean norm ofj~P,
so (2.B) implies that0 < .t ppt.
- JM JM
~;
2dM - (2d - n)M2
•
If n < 2d, this establishes the Plotkin bound.
Now, a code is not essentially changed if we multiply each entry in a. column of P by -1. Thus, i t may be asswned that the first row of P consists entirely of l 's, so we may write
p
where Q is an (M - 1) x n matrix of l 's and -l 's. In the extremal case M = 2d/(2d- n), we find that
QQt 2di - (2d - n)J,
.t Q .t
JM-1 = -Jn'
Qjn (n- 2d)jM-l'
so Q may be regarded as the (+, -)-incidence matrix of a block design.
There are several interesting examples and applications
of binary codes. For instance, in the extreme case in which the Plotkin bound is valid (i.e. 2d - n
=
1 and M=
2d), the matrix [jM P) is found to be a Hadamard matrix oforder M. If we remove the restriction that led to the Plotkin bound, we may find codes with a larger vocabulary
(i.e. more code words). For instance, there is a class of binary (8t, 4t, 2t)-codes obtained from matrices
p
= [_: ]
where H is a Hadamard matrix of order 4t. A final example is the binary (16, 5, 2)-code whose code words consist of all vectors (x1, x
2, x3, x~, x5) with an even number of
-l 's .
For further details, we refer to [4] and to (15].
The general algebraic theory of codes is exposed in (17].
2.3. Ternary Golay codes
For ternary codes, the alphabet S consists of the elements
0, l, -1 of the finite field GF(3). In this way, the words of length n may be regarded as forming the vector space
V(n, 3). We are particularly interested in situations
where the code words form a linear subspace (such a code is called a linear code).
If l and -1 are abbreviated to + and - respectively, consider the 6 x 12 matrix
0 + + + + + + 0 + + + 0
+
+ + + 0 + + + + 0 ++
+ 0 224 + 0 0 0 0 0 0 + 0 0 0 0 + 0 0 0 0 + 0 0 0 0 0 0 0 0 0 0 + 0 0 0 0 0 0 + .I'I
(~here C6 is the same matrix as in §1.4). Since the column
rank is obviously 6, the rows form a basis for a subspace W of dimension 6 in V(l2, 3). Now, the Hamming distance between any pair of these basis vectors is congruent to 0
(mod 3), so any pair of distinct elements from W has Ham-ming distance 3, 6, 9, or 12. Moreover, i t can be shown that no two code words in W have Hamming distance exactly 3. It follows that the 36 words in W form a ternary (36
,
12, 6)-code, called the Extended Golay code. The automor-phism group of this code is the Mathieu group M12•
The ordinary Golay code is obtained by deleting any one column of the matrix [C6 I
6) . The shortened rows then
form a basis for a subspace U of dimension 6 in V(ll, 3). The Hamming distance between any pair of distinct words in
u
is at least 5, so the 36 words of length ll in U form aternary (36, 11, 5)-code. Moreover, each closed ball
B(~; 2] contains exactly
l + 2·11 +
2
2·(~
1)
= 35 words of V(ll, 3), so the family{B(~; 2]
I
~ E U}of disjoint closed balls covers 36·35 words of V(ll, 3).
Since there happen to be exactly 311 words in V(ll, 3), we
see that this Golay code is an example of a perfect code. A third ternary code is obtained as the complement of the Golay code. More precisely, if we take the 35 words
code words, we obtain a ternary (35, 11, 6)-code. The code
words of this last code have Hamming distances 6 and 9 only.
REMARK. There are also binary Golay codes with words be-longing to V(n, 2): a binary (212, 24, 8)-code which is
related to the Mathieu group M2~, and a perfect binary
(212
, 23, 7)-code which is related to M23 •
3. STRONGLY REGULAR GRAPHS
3.1. Definitions
Consider a graph G; a vertex P is said to be a neighbour of the (unordered) pair of vertices P1, P
2 if P i s adjacent
to both P1 and P
2•
LEMMA 3.1. Let G be a graph with n vertices such that
(i) each vertex is adjacent to at most k others;
(ii) each adjacent pair of vertices has at least
A
neighbours; and(iii) each non-adjacent pair of vertices has at least
IJ neighbours.
Then (n - k - l)IJ s k(k - 1 -
A)
.
PROOF. Choose a vertex P
0 of G. Define A(P0) to be the
set of vertices adjacent to P0, and NA(P0) to be the set of vertices not adjacent to P0• By assumption, IA(P
0) 1 s k
so INA(P0
ll
~ n - 1 - k. We now proceed to count the num-ber M of edges between A(P0) and NA(P0) in two ways. Since
. ·~
:;j
. ::
•.'
~;.
each vertex Q in NA(P
0) determines, with P0, a non-adjacent
pair, there are at least IJ vertices adjacent to both
P
0and Q; thus
Similarly, each vertex R in A(P
0) is adjacent to at most
k -
l vertices other than P0 and has at least
A
neighbourswith P
0, so there are at most k - 1 - A vertices adjacent
to R but not to P
0; hence
M s IA(P
0) j · ( k - l - A.) ~ k· (k - l - A) .
These two inequalities establish the lemma.
Examining the circumstances under which equality oc-curs in Lemma 3.1, we say that a graph is strongly regular
i f
(i) each vertex is adjacent to exactly k others;
(ii) each adjacent pair of vertices has exactly A. neigh-bours; and
(iii) each non-adjacent pair of vertices has exactly
IJ neighbours.
For instance, a pentagon is a strongly regular graph with
n
=
5, k=
2,A
=
0, and 1J=
1. The Petersen graphis a strongly regular graph with n
=
10, k=
3, X=
0, and~
=
1. Moreover, if T is a finite set of cardinality t,the
(~l
unordered pairs of distinct elements of T form thevertices of a strongly regular graph (called a triangular 1
graph T(t)) with n =
2
t ( t - 1), k = 2 ( t - 2), A= t - 2,and ~ = 4; two pairs are called adjacent precisely when
they have one element in common.
3.2. Graphs and matrices
The properties of a strongly regular graph are often studied
via the corresponding adjacency matrix. In particular, we
state without proof
THEOREM 3.2. Suppose A is a symmetric matrix with entries
0 and 1 such that all diagonal entries are zero. Then A
is the adjacency matrix of a strongly regular graph with
parameters (n, k,
X
,
~) if and only if there are numbersr and s such that
Aj = k·j I
n n (A- ri)(A- si) = ~J.
Moreover, r and s are the roots of x2 + (~ - X)x + ~ - k
=
0. In addition, k, r, and s are the eigenvalues of A,the eigenvalue k has multiplicity 1, and, if r + s ~ -1,
the eigenvalues r and s are both integers.
The adjacency matrix As of a pentagon satisfies A~ +
J so r + s =
-1;
in this instance, both r and shappen to be irrational. Similarly, the adjacency matrix
A10 of the Petersen graph satisfies Af0 + A10 - 2I = J so
228
r + s
=
-1 again; in this case, however, r ands
happen tobe integers.
The state of affairs for the special case in which
r + s
=
-1 while r and s are both integers is quiteinter-esting. If x and y are the multiplicities of the
eigen-values r and s respectively, we are able to derive
(2r + 1) 2 - 2(2r + 1) - l6(x- y)~(2r
+
1)=
16~2 1as a consequence of the proof of the above theorem. It
follows that 2r
+
1 16~2- 1. One class of such graphs
consists of the Moore graphs, which are restricted to ~
=
1 and A
=
0; since 2r + 1I
15 for this class, we find thatthe parameters are limited to r -1-2 7 s
=2
-3 -8 k -3-7 56 n---ro
50 3250The Petersen graph yields the first possibility, and i t is at present unknown whether the third possibility is
real-ized. The second possibility is realized by the
Hoffman-Singleton graph (7], which we now describe. The 50
ver-tices of the graph are given by the 15 points and 35 lines
of PG(3, 2). To be able to define adjacency, we require
the following fact (cf. [1]): "there is a 1-1
correspond-ence between the lines of PG(3, 2) and the unordered
tri-ples from {a, b, c, d, e, f, g} in such a way that two
lines intersect or are skew according as the
correspond-ing triples have an odd or an even number of elements in
con-sidered adjacent, a point and a line are concon-sidered adjacent
only if they are incident, and two lines are considered adjacent precisely when the corresponding triples are dis-joint. It is an elementary exercise to check that this graph does have the stipulated properties.
An alternative class of these graphs, cf. [2], is produced by the restriction ~
=
2 andA
=
1,
so that 2r+
163; the parameters in this instance are limited to
r k n 1 4 9 3 14 99 4 22 243 10 112 6273 31 994 49401
At present, i t is unknown whether the second, fourth, or fifth possibilities may be realized. The ·g vertices for the first possibility may be taken to be the ordered pairs from {a, b, c}; two ordered pairs are adjacent if they agree in one position. The 243 vertices for the third pos-sibility may be taken to be the elements of V(S, 3); two vectors ~ and
¥
are called adjacent if one of the differ-ences x -X
or ~ - x is a column of the 5 x 11 matrix of generators for the complement to the Golay code. In fact,if
f 1 , ~2, • • • , ~11
are the columns of this matrix, thenthere are 22 vectors ±c., 220 vectors ±c. ±c., and 1 vector
-1 -1 - )
o. Since
it
was shown in §2.3 that no four columns ofthis matrix are dependent, i t follows that these 243 vectors are pairwise distinct, so they exhaust V(5, 3) and may be taken as representatives of the vertices of this graph. Again, one may readily check that the graph does have the
.·.~ ·; ·--~ '.·
·'
'1 '• required properties. 3.3. Rank 3 graphsIf G is a permutation group on a set ~. we may define a relation ~ on ~ x ~ by
(a, 8) ~ (y, o) if and only if y g(a) and
o
g (8)for some g € G.
It is an elementary exercise to show that this relation is an equivalence relation, with equivalence classes of the form
{(g(a), g(8))
I
g € G}where a,
8
€ Q. Obviously, if 6 is an equivalence class, then so is'il
:= {(8, a) (a, 8) € 6}.If G is transitive on ~. the diagonal T of
n
x ~ is an equivalence class. Moreover, if G is 2-transitive, there are exactly two equivalence classes, namely T andn
x ~\T.The next most interesting situation, in which G is slightly
I
less than 2-transitive, may be defined as follows: "G is called a rank 3 group on ~ if G is transitive on ~. the order jGj is even, and there are exactly three equivalence classes T, 6, and
r."
Using the restriction to even order, one may then show that 6=
6 andr
=r,
so each unordered pair of distinct elements of ~ belongs either to 4 or tor,
but not both. We may use this observation to define two graphs with the elements of Q as vertices: two vertices a,(a,
a>
t 6, anq they are adjacent in the other graph if andonly if (a,
al
tr.
Therefore, two vertices are adjacent in one graph precisely when they are non-adjacent in the other. These two complementary graphs are called rank 3 graphs. Since G is transitive on both of the equivalence classes ~andr,
i t is transitive on the edges and on the non-edges of each of these graphs. Furthermore, G is also transitive on the vertices, so we find that any rank 3 graph is strongly regular.We give an example to illustrate these remarks. If T is a finite set of cardinality t (t ~ 4) and Q consists of the
(t]
l2 unordered pairs from T, i t is clear that the symmetric group ST on T is a rank 3 group on
n
,
since the three equivalence classes areT = { (w, w)
I
w. E Q}, { (w, w')r ,;{(w,w')
the unordered pairs w and w' have no}, element in common
the unordered pairs ~ and w' have} exactly one element 1n common · The strongly regular graph with edges taken from
r
is pre-cisely the triangular graph T(t) .4. QUADRATIC FORMS OVER GF(2)
4.
1.
BackgroundThere are many applications of bilinear and quadratic forms to codes and strongly regular graphs. In order to
demon-strate this, we reproduce some results about forms which may be fqund with detailed references in Dieudonne's
trea-tise La geometrie des groupes classiques.
An alternating bilinear form on V(2m, 2) is a bilinear map B: V x V ~ GF(2) such that B(~, ~) = 0 for every x E V; i t follows that B(~,
y)
B(y, ~) for all ~· y E V. IfB(~,
y)
=
0 for all y E V only when x_2
,
the form iscalled non-singular.
Since GF(2) contains only two elements 0 and 1, we may consider the entries of the square matrix
B := [B(~, y)J
~EV, ¥_EV
(of order 22m) as integers. With this interpretation, one may show that the matrix 2B - J, consisting of l 's and
{-1) 's, is a Hadamard matrix. On the other hand, we may obtain an example of a binary (22m+l, 22rn, 22m-l)-code from the columns of the 22m x 22m+l matrix [B J-B).
Using a non-singular alternating bilinear form B on V, we may define the symplectic graph: the vertices are the points of V\{9}, and two vertices~· yare called ad-jacent precisely when ~
t ¥.
and B(~, ¥_) = 0. We find that the symplectic graph is a rank 3 graph under the groupSp(2m, 2) = {a E GL(V) B(~ a I
y )
a = B(~, ¥_)for all ~·
¥.
E V}. Since W = {~ E VI
B(~, ~) o} is a linear subspace of V and~ is adjacent to every element of W\{~, ~}, i t followsthat the symplectic graph is a regular graph with k
=
22m-1
-
2 edges through each vertex. If B(~o' y ) _o=
0 and B(x I ¥1)=
1, then-
·
u
{z0 E
v
B(~, X - 0 ) 0 B(~,:l
0 ) }and u { z E
v
B(~, ~ 1) 0 B(~, ¥1) }j
are both the intersection of two hyperplanes in V(2m, 2),
so they each contain exactly 22m-2 elements; i t follows that the symplectic graph is strongly regular, with
>..
I
U o \ { Q,~
o ,r:
o}I
= 22m-2 - 3and iJ
A quadratic form on V(2m, 2) is a map Q: V + GF(2)
such that Q(~ + yl + Q(!i + Q(¥) is a bilinear form B(~, ¥) on V; i t follows that Q(Q) = 0 and that the associated bi-linear form is alternating. If this bilinear form is also non-singular, i t is known that there is a basis ~
1
, . . • ,~2
mof V with respect to which the quadratic form Q is given
by
s1(2 + (3(~ + . . . + s2m-ls2m
or
S
~
+ ( 1 ( 2 + (~
+ ( 3 ( 4 + · · · + t; 2m-1 t; 2mand that the number of zeros of Q is 22m-l + 2m-l or 22m-l m-1
- 2 respectively. Let
n
represent the set of all zeros of a non-singular Q.We may now define the orthogonal graph: the vertices are the elements of Q\{9} (alternatively, one may use the elements of V\ll), and two distinct vertices~· yare
ad-:·i jacent precisely when Q(~ + y) = 0. As before, we find
~·
that the orthogonal graph is a rank 3 graph under the or-thogonal group
{a E GL(V)
I
Q(~a) = Q(~) for all x E v}.·· We mention that the symplectic and orthogonal graphs may
:' ~·
be characterized in a more geometric manner (cf. (10),
[13)):
SHULT'S THEOREM. For a regular graph G which is not a com~
plete graph, the following are equivalent:
(i) G is isomorphic to a symplectic or an orthogonal graph;
(ii) For each pair of adjacent vertices a, b of G,
there is a vertex c adjacent to both a and b such that
each further vertex is adjacent to an odd number of vertices of the triangle {a, b, c).
4.2. Kerdock sets
We now investigate some related concepts following Kerdock, Patterson, and Goethals. Suppose 8 is the set of all al-ternating bilinear forms on V(2m, 2), allowing for singular forms and even the zero-form. Each form B(~, y) may be regarded as a matrix product
t
~ B¥
where B is a symmetric 2m x 2m matrix with entries in GF(2),
Since there are m(2m - 1) positions in such a matrix above the main diagonal, we find there are exactly 2m(2m-l) ele-ments in B.
Consider a subset S of B such that whenever B. and
1
B. E S the form B. +B. is non-singular. If the matrices
J 1 J
representing B. and B. have the same first row, then the
1 J
matrix representing Bi + BJ. has first row 0, and B. +B.
1 J
is singular. Therefore, the first rows for the matrices of the B. in S must all be distinct, so S contains at most
}_
22m-l forms. For this maximal case, we define: a subset K 2m-1
of 8 is called a Kerdock set if
IK
I
= 2 and, whenever Bi and Bj are distinct elements of K, the form Bi + Bj is non-singular. We give the following construction of a Kerdock set. GF(2), each a + ~ where element a E Tr(a) . . 2m-1 If V(2m, 2) 1s cons1dered to be GF(2 ) $element ~ in V may be written uniquely as ~
2m-1
a E GF(2 ) and ~ E GF(2). The trace of an
GF(22m-l) is defined by 2m-2
L
k=O and it is known that
Tr(a) € GF(2),
Tr(a +B)= Tr(a)
+
Tr(B) ,Tr (1) 1.
For each Y E GF(22m-l) , we define an alternating bilinear
form By on V(2m, 2) by
By(a,
Bl
= Tr(a8y2) + Tr(ay)·Tr(By),?11;
i
l : i tj
·~ 'l ., B (a, y t;) By(i;, €;) for all a,B
to V(2m, 2) . K=
{B y Tr (ay) , 0,E GF(22m-l) and t; E GF(2), extending linearly We shall show 'that
I
Y € GF(22m-1)}is a Kerdock set. It obviously has the right number of elements, so i t only needs to be shown that By + B
0 is non-singular when y
#
6. Suppose then that BY+
B6 is singular. Since the kernel of By
+
B6 must have even dimension, there 2m-1
exists a € GF(2 )\{0} which is also in the kernel; i t follows that
Tr(ay) + Tr(a6) = 0
and Tr(aByL) + Tr(a862) ~ Tr(ay) ·Tr(By) + Tr(a6)·Tr(86)
for all
B
€ GF(22m-1). If Tr(ay)=
Tr(a6) 0, thenTr(aB(y + 6) 2)
=
0 for all B € GF (22m-1) and we concludethat y
=
6. On the other hand, i f Tr(ay)=
Tr(a6)=
1,then Tr{ 13 (y + 6)(a[y + 6) + 1)}
=
0 for all 6 € GF (22m-1)and we again deduce that y
=
6. This establishes that Kis a Kerdock set.
4.3. Kerdock codes
Suppose that B is an alternating bilinear form on V(2m, 2)
and that Q and Q' are two quadratic forms associated with B as in §4.1; i t is then an easy exercise to show that Q
+
Q' is one of the 22m linear forms on V(2m, 2).Th~s,
there are 22m quadratic forms associated with each
alter-237
nating bilinear form on V(2m, 2). The count of alternat-ing bilinear forms in §4.2 shows that there are exactly
2m(2m+l)
=
2~m.2m(~m-l) elements in the set Q of allquad-ratic forms on V(2m, 2).
We may represent each Q
~
Q as a vector in V(22m, 2)whose
2~m
coordinates 0 and 1 are the valuesQ(~)
over all~ <:
v
12m, 2). From these binary "words" of length 2 2m , weform a Kerdock code by fixing a Kerdock set K and taking
as code words
{Q, l + Q
I
the bilinear form B(Q) E K}.Since
~here
are 22m-l bilinear forms in K, 22m quadraticforms for each bilinear form, and 2 code words for each
quadratic form, we have a total of 24m code words. Since
2m-1 m-1
the smallest number of l's in a combination is 2 - 2 ,
h d k d . f d b . ( 4m 2m
t e Ker oc co e 1s oun to e a b1nary 2 , 2 , 22m-1
m-1
- 2 ~)-code.
5. EQUIANGULAR LINES IN Rd
5.1. Background
If a collection of n lines through a point 0 in Rd has the
property that each pair of distinct lines determines the
same angle ~. the collection is called a set of equiangular
lines. A unit vector p. may be chosen along each line of
1
such a collection in either of two directions; i t follows that the inner product between distinct unit vectors is
given by 2.18 .,
)
···1 <p., p.> =±cos ~ -1 JThe matrix of inner products may thus be rewritten as
l ±cos ~
p I
+
C cos ~±cos ~ l
where
c
is a symmetric n x n matrix with 0 along the maindiagonal and ±l elsewhere. In order that the n lines span
Rd, we may assume that n > d. Since there are now more
directions than dimensions, the smallest eigenvalue of P
is 0, and its multiplicity is at least n d. Because C
is obtained from P in the above manner, we find that the
smallest eigenvalue of C is -1/cos ~, and its multiplicity
is at least n - d. Conversely, suppose that C n is a
sym-metric n x n matrix with 0 along the main diagonal and ±l
elsewhere, that y0 is the smallest eigenvalue of en' and
that y o has multiplicity n - d. It follows that C n - y I o
is a symmetric positive-semidefinite n x n matrix of rank
d, and thus that there is an n x d matrix U such that
(5 .A)
Then n rows of U may be considered vertices in Rd, and (5.A)
shows that en- y0I is the matrix of inner products for
these vectors. Since the non-diagonal elements of Cn - Y0I
are all ±1, we see that the n lines spanned by these vectors
are equiangular.
we look for large matrices C whose smallest eigenvalue has
n
a large multiplicity. For instance, there is a 16 x 16 symmetric Hadamard matrix H16 which may be written as
H16
=
C 16 - I where (C16 - 1)2=
161; i t follows that (C16 - 51) (C 16 + 31) = 0, so the eigenvalues of C16 are 5 and -3 (with multiplicities 6 and 10 respectively). Proceeding in the manner described above, we discover a set of 16 equiangular lines in R16-10(=R6 ).5.2. Graphs and equiangular lines
The matrix C associated above with a set of equiangular lines may be interpreted as the adjacency matrix A of a graph via the transition
C = J - I - 2A,
i.e. , 1 and -1 inC correspond to 0 and 1 in A respectively. For instance, the adjacency matrix A1 o of the Petersen graph was found in § 3. 2 to satisfy 2
+ A1 o 21 Aj
A1 o
-
= J, = 3jso the corresponding
c
1 0 satisfies 2c
1 0 91. Themulti-plicities of the eigenvalues 3 and -3 of C10 are both 5. Using the procedure of §S.l, we discover a set of 10 equi-angular lines in R1o-s (=Rs).
The correspondence between graphs and sets of equian-gular lines is not 1-1, due to the fact that we originally could have chosen the unit vectors along the lines in ei-ther of two ways. Examining how a choice of -pi instead of pi is reflected in the adjacency matrix
c,
we define a.i
"switching relation" on (±!)-adjacency matrices: C and C' are related if there is a diagonal matrix D whose diagonal elements are ±1 such that C'
=
DCD. This switching relation is easily seen to be an equivalence relation. rhe switching classes of graphs are now found to be in 1-1 correspondence with the sets of equiangular lines. For graphs with 3ver-tices, there are two switching classes, consisting of graphs with an even number of edges or graphs with an odd number of edges (the latter are called odd 3-graphs). Graphs with
4 vertices are distributed among switching classes according as the number of odd 3-subgraphs is 0, 2, or
4
.
5.3. Two-graphs
If Q(i) denotes the collection of i-subsets of the set
n,
a graph is occasionally defined as a pair (Q, f) where f:n<
2) + {1, -1} is an adjacency function. Similarly, we define a two-graph to be a pair (Q, g) where g:
n<
3l +{1, -1} is a map satisfying
g(a,
a, y)
·g(a,a, ol
·g(a,y,
o) ·g(f3,y, o)
= 1for all distinct a, (3, y, o E n. If a triple {a, (3, y} is
called coherent when g(a, (3, y) = -1, the above condition (essentially the cohomology condition og
=
1) allows that, whenever we know which triples containing a particular ver -tex a are coherent, we know the coherent triples for the whole two-graph.Given a graph (n, f), we may define a two-graph (0, g)
by g(a, 8, y)
:=.
f(a, 8) ·f((3, y) ·f(y, a); the coherent tri -ples of the two-graph are precisely the odd 3-subgraphs of the original graph. On the other hand, if we start with a two-graph (Q, g), we may associate with i t the switching class{ w,
f)I
f(a, f3)·f(f3, y)·f(y, a)= g(a, (3, y)} of ordinary graphs. In this way, we discoverTHEOREM 5.1. There is a 1-1 correspondence between
(i) two-graphs,
(ii) switching classes of graphs, and
(iii) sets of equiangular lines.
COROLLARY. If a graph (Q, f) belongs to the switching class of the two-graph (Q, g), then Aut(Q, f) is a subgroup of Aut(i1, g).
5.4. 2-transitive two-graphs
Shult employs this last observation to construct a two-graph with a 2-transitive automorphism group. Suppose that
(Q, f0 ) is a graph, X € Q, f = {vertices adjacent to X} X
and ~x = {vertices non-adjacent to x}. We suppose further-more that Aut
W,
f) is transitive onn,
and that there exist automorphisms h1 and h2 of the subgraphs fx and Axrespect-ively such that h1 (y) is adjacent to h2(z) precisely when
Y
and z are non-adjacent. Under these conditions, i t fol-lows that there is a two-graph (11 u {w}, g) with a2-tran-?II?
sitive automorphism group. In fact, the switching class of the graph (Q u
{w},
f) wherei f either y f (y, z) =
r
f 0(y, z) w or z=
w, otherwisemay be shown to contain an element interchanging w and x; since the original assumptions make i t clear that the sta-bilizer of w is transitive on the remaining vertices, we see that the switching class of (Q u {w}, f) has an auto -morphism group which is 2-transitive on Q u
{w}.
Theorem 5.1 now establishes the existence of the required two-graph.0 FIGURE A
Q
FIGURE BFor example, the pentagon in figure A has a transitive auto-morphism group. If h1 interchanges the two vertices of
f
0
=
{1, 4}
and h 2 is chosen to be the identity on ~ 0 ={2,
3},
then we find that our conditions are fulfilled. It follows that there is a two-graph with six vertices whose automorphism group is 2-transitive; this two-graph is as-sociated with the switching class of the graph representing the diameters of the icosahedron (figure B), and its auto-morphism group is A5• This procedure generalizes to producea two-graph with q + l vertices and an automorphism group PSL(2, q) which is 2-transitive.
Another example may be obtained from the symplectic graph on V(2m, 2)\{Q} introduced in §4.1. We saw that Sp(2m, 2) was transitive on the vertices of that graph. I f ~ is a vertex, then
r
{:f ~ 0I
B(~, y) = 0} and b.~ ~
{~ B(~, z) = 1}. I f hi is the mapping
hi:
r
~r
xi y 1-+ X + y ~
and h2 is chosen to be the identity on 6 , then our
condi-~
tions are fulfilled and there exists a two-graph on 22m vertices with a 2-transitive automorphism group Sp(2m, 2)· V(2m, 2). If the vertices are the elements of V(2m, 2), we find that ~· ~· ~ are coherent exactly when B(~, ~) +
B(~, ~) + B(~, ~) = 0.
As a final example, we consider the orthogonal graph on the quadric
{~ E V(2m, 2)\{Q}
I
Q£(~) = O}where £ ±1, as introduced in §4.1. The graph has 0£(2m,
2) as a transitive automorphism group, and a construction
as above yields a two-graph with 2m-l(2m + £) vertices and a 2-transitive automorphism group Sp(2m, 2).
For further examples such as those involving Conway's group ·3, we refer to [13].
6. REGULAR TWO-GRAPHS
6.1. Definitions and properties
We proceed with notions due to G. Higman [11] I (16]. A
two-graph (Q, g) is called regular if each pair a,
B
of-~
-vertices is contained in exactly r coherent triples. In
terms of the ideas of §5, i t may be shown that this is equivalent to insisting that any matrix C (of size n x n) in the switching class associated with (Q, g) has exactly two eigenvalues, say p1 > p 2 , and (C- p1I)·(C- p2I)
=
0.We shall set forth several consequences of this latter
de-finition. The multiplicities of the eigenvalues p
1 and
p
2 will be denoted by ~
1
and ~2
respectively, so~~ + ~2 = (6 .A) n 1
-
plp2, ~lpl + ~2.p2 0. It follows that (6.B) IfP
1 + p2 ~ 0, then p1 and p2 are odd integers. (6.C) If p
1 + p2
=
0, then p1=
-p2=
;n-=-I andc
2
=
(n- l)I; furthermore, i t may be shown that n=
2 (mod 4)and that n - 1 is the sum of two squares. Examples have
k
been constructed when n - 1
=
p=
1 (mod 4), and also for n = 226; the first unknown possibility occurs when n = 46.Solving (6. A) for ~2.' we find
p3
-PI p2 1 ~2=
1-
pl-
Pz and we deduce (6.D) If p 1=
3, then n=
10, 16, or 28. If p1=
5, then n = 16, 26, 36, 76, 96, 126, 176, or 276. Moreover, exam-245ples of each possibility are known, except for
P
1 5 andn
= 76 or 96.Furthermore
(6. E)
This is a consequence of the following (cf. also (21])
THEOREM 6.1. The number of equiangular lines in Rd is at most ¥<d + 1).
PROOF. Suppose p , ... , p are unit vectors along n
equi-1 n
angular lines in Rd, so <p., p.> =±a if i
~
j. Letl. ]
Pi: Rd + Rd,
X~
<~,
pi>pibe the projection onto the vector pi; then Pi is a symmetric linear map (1 5 i 5 n). Moreover, if i ~ j,
Tr (P. P.) l. ] k E<P.P.ek, l. J- ~k> l:<p., ek>·<p., p.>·<p., ~k> k l. - l. ] ] <p., p.>2 = a2. l. ]
Since an inner product of two symmetric linear maps Q and R is given by tr(QR), we may deduce from this that P1, • • • ,
p are linearly independent. Since the space of symmetric n
linear maps on Rd has dimension
~(d
+ 1), i t follows thatn 5 ict<d + 1), as required.
6.2. Unitary two-graphs
Let q be an odd prime power, PG(2, q 2 ) the projective plane
246
<j
obtained from V(3, q2) , H a non-degenerate hermitian form
on V(3, q 2), and
n :
= { < ~ > £ PG < 2 , q 2 >1
u
< ~ , ~ > =o }
the quadric associated with H. It is known that
1111
= q3+ 1 and that PfU(3, q 2 ) acts 2-transitively on
n.
Suppose 6 is the set of distinct(!), (~) , (!) inn such that{ a square in GF(q2) if q : -1 (mod 4), H(!, x>·H(~, !)•H(!, ~)'"' . a non-square l.n GF(q2) if q - 1 (mod 4).
Then (1!, 6) is a regular two-graph, with n = q3 + 1, ~ = q2 - q
+
1,1
(C- q2I)·(C
+
qi) = 0~2 = g(q2- q + 1),
(it is a slightly complicated process to establish these results about the parameters). Theorem 5.1 now implies that there is a set of q3 + 1 equiangular lines in Rq2-q+l
with cos ~
=
1/q. Thus, the maximum number of equiangular lines in Rd is at least d/d.6.3. Equiangular lines
We now tabulate the maximum number n(d) of equiangular lines in Rd as follows: d n(d) reference 2 3 3-4 6 5 10 5.2 6 16 5.1 7-14 28 5.4, 6.1 15 36 5.4 16 40 17-18 48 ?117
d n(d) reference 19 72 20 90 21 126
u
(3, 52) 22 176 Higman-Sims 23-42 276 constructed below 43 344 U(3, 72) 6.4. The 276-two-graphA non-trivial regular two-graph on 276 vertices may be con-structed as follows [5]. From (6.A), (6.B), and (6.C) , we deduce that the only possible eigenvalues p1,
P
2 of Care55 and -5. Now the complementary Golay code consists of the 243 vectors, having Hamming distances 6 and 9, in a 5-dimensional subspace of V(l l, 3). Let
PR.: V(ll, 3) -+ GF(3) I
~
be the projection onto the ith coordinate (1 ~ i ~ 11); there are 33 pairs (PRi' x) where PRi is one of these 11 projections and x E GF(3). Let the 276 vertices of a graph consist of these 33 pairs and the 243 elements of the com-plementary Golay code. Adjacency is defined by:
(PR. I X) is adjacent to (PR. I y) <=> X
=
y ~ J (PR. I X) is adjacent to!::
= (y 1 I • • • I y 1 I) ~ <=> X=
PR. (y) 1 - yi ~ (x 1, • • • , x11) is adjacent t o y = (y1 , • • • , y11 )<=> their Hamming distance is 9.
we find that the adjacency matrix
c
for this graph satisfies (C + 5I)·(C- 55!)= 0so the two-graph associated with the switching class of
'lAO
\
this graph is regular. This two-graph is the only non-trivial regular two-graph on 276 vertices, since the above process may be reversed. Another way to construct 276 equi
-angular lines in R23 is to consider in the Leech lattice in R24 the 276 rhombi having a fixed diagonal at the second
smallest distance occurring in the latti~e.
REFERENCES
1. F.C. Bussemaker and J.J. Seidel. Symmetric Hadamard matrices of order 36. T.H. Report 70-WSK-02, Tech. Univ. Eindhoven, Netherlands (1970).
2. E.R. Berlekamp, J.H. van Lint, and J.J. Seidel. A strongly regular graph derived from the perfect ternary Golay code. A Survey of Combinatorial Theory 1973
(J.N. Shrivastava, editor), 25-30. North Holland Pub. Co.
3. P.J. Cameron and J.J. Seidel. Quadratic forms over GF(2). Proc. Kon. Ned. Akad. Wet. Ser. A, 76 (1973):
1-B.
4. J.M. Goethals. Some combinatorial aspects of coding theory. A Survey of Combinatorial Theory 1973 (J.N. Shrivastava, editor), 189-208. North Holland Pub. Co. 5. J.M. Goethals and J.J. Seidel. The regular two-graph
on 276 vertices. Discrete Math. 12 (1975): 143-158. 6. M. Hall. Combinatorial Theory. Waltham, Mass. (1967) . 7. A.J. Hoffman and R.R. Singleton. On Moore graphs with
diameters 2 and 3. IBM J. Research Develop. 4 (1960): 497-504. B. 9. 10.
11.
12. 13.P.W.H. Lemmens and J.J. Seidel. Equiangular lines. J. Algebra 24 (1973): 494-512.
J.J. Seidel. Strongly regular graphs. Progress in Combinatorics 1969 (W.T. Tutte, editor), 185-197. Academic Press.
On two-graphs and Shult's characterization of symplectic and orthogonal geometries over GF(2). T. H. Report 73-WSK-02, Tech. Univ. Eindhoven, Nether-lands (1973).
A survey of two-graphs. Proc. Int. Call. Teorie Combinatorie, Ace. Naz. Lincei, Rome (to appear).
Graphs and two-graphs. Fifth Southeastern Conf. on Combinatorics, Graph Theory
&
Computing. Utilitas Math. Publ. Inc., Winnipeg (1974): 125-143.E.E. Shult. Characterizations of certain classes of graphs. J. Comb. Theory (B), 13 (1972): 1-26.
14. The graph-extension theorem. Proc. Amer. Math. Soc. 33 (!972): 278-284.
15. N.J. Sloane and J.J. Seidel. A new family of
non-linear codes obtained from conference matrices. Int.
Conf. on Combinatorial Math. 175 (1970): 363-365.
16. D.E. 'l'aylor. Regular two-graphs. Proc. London Math.
Soc. (to appear).
17. J.H. van Lint. Coding theory. Springer Lecture
Notes 201.
18. Combinatorial theory seminar. Springer Lecture
Notes 382.
19. J.H. van Lint and J.J. Seidel. Equilateral point sets
in elliptic geometry. Proc. Kon. Ned. Akad. Wet. Ser.
A, 69 (1966): 335-348.
20. J.S. Wallis. Hadamard matrices. Springer Lecture
Notes 292.
21. Ph. Delsarte, J.M. Goethals, J.J. Seidel. Bounds for
systems of lines, and Jacobi polynomials. Philips