You are on page 1of 155

LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES

Managing Editor: PROFESSOR I. M. JAMES, Mathematical Institute,
24-29 St Giles, Oxford
Already published in this series
1. General cohomology theory and K-theory, PETER HILTON.
4. Algebraic topology: A student's guide, J. F. ADAMS.
5. Commutative algebra, J. T. KNIGHT.
7. Introduction to combinatory logic, J. R. HINDLEY, B. LERCHER,
and J. P. SELDIN.
8. Integration and harmonic analysis on compact groups, R. E.
EDWARDS.
9. Elliptic functions and elliptic curves, PATRICK DU VAL.
10. Numerical ranges II, F. F. BONSALL and J. DUNCAN.
11. New developments in topology, G. SEGAL (ed.).
12. Symposium on complex analysis Canterbury, 1973, J. CLUNIE and
W. K. HAYMAN (eds.).
13. Combinatorics, Proceedings of the British combinatorial con-
ference 1973, T. P. McDONOUGH and V. C. MAVRON (eds.).
14. Analytic theory of abelian varieties, H. P. F. SWINNERTON-
DYER.
15. An introduction to topological groups, P. J. HIGGINS.
16. Topics in finite groups, TERENCE M. GAGEN.
17. Differentiable germs and catastrophes, THEODOR BROCKER and
L. LANDER.
18. A geometric approach to homology theory, S. BUONCRISTIANO,
C. P. ROURKE and B. J. SANDERSON.
19. Graph theory, coding theory and block designs, P. J. CAMERON
and J. H. VAN LINT.
20. Sheaf theory, B. R. TENNISON.
21. Automatic continuity of linear operators, ALLAN M. SINCLAIR.
22. Presentations of groups, D. L. JOHNSON.
23. Parallelisms of complete designs, PETER J. CAMERON.
24. The topology of Stiefel manifolds, I. M. JAMES.
25. Lie groups and compact groups, J. F. PRICE.
26. Transformation groups: Proceedings of the conference in the
University of Newcastle upon Tyne, August 1976, CZES
KOSNIOWSKI.
27. Skew field constructions, P. M. COHN.
28. Brownian motion, Hardy spaces and bounded mean oscillation,
K. E. PETERSEN.
29. Pontryagin duality and the structure of locally compact abelian
groups, SIDNEY A. MORRIS.
30. Interaction models, N. L. BIGGS.
31. Continuous crossed products and type III von Neumann algebras,
A. VAN DAELF.
32. Uniform algebras and Jensen measures, T. W. GAMELIN.
33. Permutation groups and combinatorial structures, N. L. BIGGS
and A. T. WHITE.
34. Representation theory of Lie groups, M. F. ATIYAH.
35. Trace ideals and their applications, BARRY SIMON.

36. Homological group theory, edited by C. T. C. WALL.
37. Partially ordered rings and semi-algebraic geometry, GREGORY
W. BRUMFIEL.
38. Surveys in combinatorics, edited by B. Bollobas.
39. Affine sets and affine groups, D. G. NORTHCOTT.
40. Introduction to Hp spaces, PAUL KOOSIS.
41. Theory and applications of Hopf bifurcation, B. D. HASSARD,
N. D. KAZARINOFF and Y. H. WAN.
42. Topics in the theory of group presentations, D. L. JOHNSON.

London Mathematical Society Lecture Note Series. 43

Graphs, Codes and Designs

P. J. Cameron
Fellow and Tutor,
Merton College, Oxford

J. H. van Lint
Professor of Mathematics,
Technological University, Eindhoven, The Netherlands

CAMBRIDGE UNIVERSITY PRESS
CAMBRIDGE
LONDON NEW YORK NEW ROCHELLE
MELBOURNE SYDNEY

Subject to statutory exception and to the provisions of relevant collective licensing agreements. Melbourne. Cambridge CB2 8RU.cambridge.org Information on this title: www. UK Published in the United States of America by Cambridge University Press. New York www.cambridge. no reproduction of any part may take place without the written permission of Cambridge University Press.org/9780521231411 © Cambridge University Press 1980 This publication is in copyright. Singapore. Cape Town. First published 1980 Re-issued in this digitally printed version 2008 A catalogue record for this publication is available from the British Library ISBN 978-0-521-23141-1 paperback . Madrid. New York.CAMBRIDGE UNIVERSITY PRESS Cambridge. Sao Paulo Cambridge University Press The Edinburgh Building.

1-factorisations of Kb 56 9. designs and projective planes 84 14. Strongly regular graphs with no triangles 37 6. Self-orthogonal codes. Symmetry codes over GF(3) 107 16. Extensions of graphs 49 8. Finite geometries and codes 76 13.Contents Preface vii 1. Strongly regular graphs 16 3. Nearly perfect binary codes and uniformly packed codes 113 17. Partial geometries 32 5. Quasi-symmetric designs 25 4. Association schemes 124 References 136 Index 145 . Codes 61 10. Polarities of designs 45 7. A brief introduction to design theory 1 2. Quadratic residue codes 95 15. Threshold decoding 74 12. Cyclic codes 68 11.

.

New material includes ovals in symmetric designs (Chapters 1 and 13). partial geometries. and the Krein bound (Chapter 17). planes and biplanes (Chapter 13). The audience for the lectures consisted mainly of design theorists. For the present volume. two-weight projective codes (Chapter 16). the inequalities of Ray- Chaudhuri and Wilson (Chapter 1). the format has been kept. An introductory chapter on designs was added. for the benefit of readers without the background of the Westfield audience. London. but extensive revisions and updatings have been made. and the aim was to present developments in graph theory and coding theory having a bearing on design theory. with the Hoffman- Chang and Hall-Connor theorems (Chapter 4).Preface The predecessor of this book (London Mathematical Society Lecture Note Series 19) had its origin in several short courses of lectures given by the authors at Westfield College. equidistant codes (Chapter 12). in 1973. 1-factorisations of K6 (Chapter 8). generalised quadratic residue codes and inversive planes (Chapter 14). vii .

.

where they are permitted. simply omit the condition that any k points are incident with at most one block.(12. and the same subset of S may occur more than once as a block. 1) designs and their contractions. 1) and 5 . Four of these. k. We assume that S and 1) are non-empty. that is. a) is a family rather than a set. the 5 . Only finitely many Steiner systems with t > 3 are known. 8. 6. The others were 1 . A t-design with parameters (v.1 A brief introduction to design theory These lectures were given to an audience of design theorists. and that v ? k ? t (so A > 0). are particularly important.) In these notes. we normally do not allow repeated blocks. even 5-designs are sufficiently rare that new constructions are interesting. with a relation called incidence between points and blocks. A) (or a t . such that every member of 2 contains k points. the introductory chapter describes some of the concepts of design theory and examples of designs to which we shall refer. Alternatively. and will be described later. A t-design with A = 1 is called a Steiner system. and any set of t points is contained in exactly d members of 5). Of course. There are 'non- trivial' t-designs with repeated blocks for every value of t. Sometimes a t-design is defined so as to allow 'repeated blocks'. for Steiner systems the question of repeated blocks does not arise. a t-design may be defined to consist of a set of points and a set of blocks. (This is more natural with the 'rela- tion' definition.(24. The existence of non-trivial t-designs with t > 5 is the most important unsolved problem in the area. Various conditions are usually appended to this definition to exclude degenerate cases. for those outside this class. satisfying the appropriate conditions. k. but the only known examples without repeated blocks with t > 5 are those in which every set of k points is a block.(v. A) design) is a connection 0 of subsets (called blocks) of a set S of v points. we shall say so.

) The conditions that any block contains k points. The present convention is adopted because we shall wish to regard the characteristic functions of blocks. we obtain i)A. and consider the subspace they span. that is. in any 1-design.1) = (v .i further points and a block containing all t distin- guished points. with the result that our incidence matrix is the transpose of that appearing in these books. as row vectors. let Xi denote the number of blocks containing a given set of i points. and any pair of points lies in x blocks. (1. A 0-design is simply a collection of k-element subsets of a set. Remark. Counting in two ways the number of choices of t . With t = 1.1) X. In a t-design. (1. with 0 < i s t. 3) r(k . The parameters A0 (the total number of blocks) and Al (the number of blocks containing a given point) are usually denoted by b and r respectively.constructed by Denniston [45]. the entry in row B and column p being 1 if p e B. i = 0. or simply a design. abbreviated to BIBD. a t-design is also an i-design for 0 < i 5 t. or rows of M.i) _ t - It follows that Xi is independent of the i points originally chosen. 0 otherwise. can be expressed in 2 . for example in the books by Dembowski [44] and Hall [60]. l) i(t . A 2-design is often called a block design. any point lies in r blocks. In a 2-design we have (1. (1. 2) bk = vr. 1) shows that. in the literature the term 'balanced incomplete block design'. is used in the case where not every k-subset is a block. (The reader is warned that a different convention is often used. An incidence matrix of a design is a matrix M whose rows and columns are indexed by the blocks and points of the design respectively.

X)I + AJ)M 1 = MT. we have b ? v. Its dual is obtained by reversing the roles of points and blocks. the matrix MTM is non-singular. the dual is a symmetric 2- design with the same parameters. 5) Theorem. 6) Theorem.)v-1. and J the matrix with every entry 1. MTM = (r . as throughout this book. A 2-design satisfying the conditions of (1.x)I + xJ) = rk(r .1. having incidence matrix MT. (Here. so. of the appropriate size. from which follows Fisher's inequality: (1.terms of M: (1. the following are equiva- lent: (i) b = v. 6) is called symmetric. Or 3 . then MJ = JM. if b = v. (ii) r = k.A)I + AJ. (iii) any two blocks have A common points. (1. In a 2-design with k < v .) It is not difficult to show that det((r . So MMT = (r . for any point p and block B. A point p (resp. p E B if and only if Ba E po. JM = rJ. thus M commutes with (r . that is. identifying a point with the set of blocks containing it.1. a one-to-one correspondence v between the points and blocks of D such that. Furthermore. if r > A. I is the identity matrix..A)I + AJ. from which we see that any two blocks have A common points. In a 2-design with k < v . 4) MJ = kJ. and so with ((r .A)I + AJ. a block B) is absolute with respect to the polarity if p E po (resp. Ba E B). A polarity of a symmetric design 0 is a self-inverse isomorphism between 2 and its dual.

Let d be the number of blocks different from but not disjoint from B. (ii) if v is odd. we obtain (with summations over i running from 1 to k): nd. if b = v. 7) Theorem. The Bruck-Ryser-Chowla theorem gives necessary conditions for the existence of symmetric designs with given parameter sets (v. k. 2) and (1.1).(v. i(i .1). In a 2-design. ini = k(r . 5) follows from b-1 >_ k(r-1)2/((k-1)(A-1)+(r-l)) on applying (1. Counting in two ways the number of choices of j points of B and another block incident with these j. 8) Theorem. 6) follow from a more general result. Then (i) if v is even then n is a square.2k(r-1)x + k((k-1)(A-1) + (r-1)).1)(x . (1. 3).1)A = k(k . points. So E (i-x)2ni = dx2 . and 1 + (k-1)(A-1)/(r-l) = X. for j = 0. It vanishes only if d = k(r-1)2/((k-1)(A-1) + (r-1)) and ni = 0 for all i # 1 + Remark.1)ni = k(k . k. Now (1. This quadratic form in x must be positive semi-definite. Put n = k . proving the inequality. A) design. the constant is 1 + (k-1)(x-1)/(r-1). 1. Proof. (1. Also. then r = k. 2.1). the number of blocks not disjoint from a given block B is at least k(r-1)2/((k-1)(A-1) + (r-1)). A) satisfying (v . then the equation Z2 = nx2 + (-1)(v-1)/2Xy2 4 . Suppose there exists a symmetric 2 .X. if this occurs. (1. Equality holds if and only if any block not disjoint from B meets it in a constant number of points. suppose ni of these blocks meet B in i points. which we will need in chapter 5. 5) and (1.

1) in chapter 11. we obtain a 3 . i n.(n-1. A Hadamard matrix is an n x n matrix H with entries ±1 satis- fying HHT =HT H = nI. The Hasse-Minkowski theorem guarantees the existence of a rational matrix M satisfying the incidence equation when- ever the conditions of (1. so we may assume that all entries in the first row and column are +1. However. Such a design is called a Hadamard 2-design. = q a prime power (q = 3 (mod 4)). Any row other than the first has Zn entries +1 and i n entries -1. (1.has a solution in integers (x.) If we take columns as points. where Q is the set of nonzero squares in GF(q). Any design with these parameters 4 arises from a Hadamard matrix in this way. but this does not mean that a design exists. (This partition is unaffected by changing the sign of the row. 5 . We shall have more to say about the case (v. a Hadamard matrix can be modified by permuting rows and columns. n-1) design called a Hadamard 3-design. with n . and the blocks are the sets Q + a (a E GF(q)). 8) can then be verified by applying the Hasse-Minkowski theorem to them. The points of the design are the elements of GF(q).J shows that the matrices I and nI + xJ are rationally cogredient. (It is so called because its determinant attains a bound due to Hadamard. not all zero. A) _ (111. so from 'equivalent' Hadamard matrices it is possible to obtain different Hadamard 2-designs. in which every entry in the first row is +1. and the sets determined in this way as blocks. n-1) design. though more elementary proofs are available.) Changing the signs of rows and columns leaves the defining property unaltered. in-1. z). we can recover a Hadamard matrix by reversing the con- struction. y. 11.(n. It should be pointed out that Hadamard matrices are very plentiful. k. Let H be a Hadamard matrix with n > 4. The incidence equation MTM = nI + A. thus determining two sets of i n columns. Examples of Hadamard 2-designs include the Paley designs. and indeed it is not known whether these conditions are sufficient for the existence of a design. we obtain a matrix M which (if n > 4) is the incidence matrix of a symmetric 2 . 8) are satisfied. From a design 4 with these parameters.1. If we then delete this row and column and replace -1 by 0 throughout.

q) is unambiguous for m > 2. A collection of subsets (called 'lines') of a finite set of points is the set of lines of a projective geometry or projective plane if the following conditions hold: 6 . familiar geometric statements hold: two points lie in a unique line.Examples are known for many orders n divisible by 4 (the smallest unsettled case at present is n = 268). where m is the dimension and q = I F . we mention one due to Veblen and Young [117]: (1. These facts can be verified by counting arguments. loosely speaking. its parameters are ((qm+1 . the blocks of a projective plane are called lines.1)). 2) is a Hadamard 2-design for all m . in that PG(2. the dimension of the geometry is that of the whole space. the subspaces of given positive dimension are the blocks of a 2-design. A projective geo- metry is often regarded as a lattice. A projective geometry over a (skew) field F is. q) is a projective plane. in which points are atoms and every element is a join of atoms.1).1)/(q-1). or by using the transi- tivity properties of the general linear group. The dimension of a subspace is one less than its vector-space rank (so points have dimension 0). Thus PG(2. regarded as a subset of the point set. 9) Theorem. If F is finite. etc. A projective plane is a symmetric 2-design with A = 1. however. the collection of subspaces of a vector space of finite rank over F. and for moderately small n there are many inequivalent matrices.1)/(q.1)/(q.2. Thus. q is not even restricted to be a prime power (though it is in all known examples). q) to represent any projective plane with k = q + 1. Note that PG(m. a non-incident point-line pair lies in a unique plane. the symbol PG(m. hyperplanes are subspaces of codimension I. q) may denote several different designs. This agrees with the previous terminology. This is motivated by various characterisations. Lines and planes are subspaces of dimension I and 2 respectively. We will modify our notation and use PG(2. The hyperplanes form a symmetric 2-design. (qm. (qm-1 . by analogy. The points of the geometry are the subspaces of rank 1. We shall identify a subspace with the set of points it contains. q). which we shall denote by PG(m.

) design. If F is finite. points are just vectors (or cosets of the zero subspace). Projective planes over fields are called Desarguesian. in terms of the concept of 'parallelism'. Hall [59] has given a counterexample with three points on any line. while the subspaces of given dimension d > 1 form a 3-design. forms a projective plane. affine planes over fields are characterised by Desargues' theorem. the geometric dimension of a coset is equal to the vector-space dimension of the underlying subspace. 7 . (It is also possible to give an axiomatic definition of an affine plane. An affine plane AG(2. Projective planes can be axiomatised in a way which permits the extension of the definition (and of (1. then any line has two points (and any set of two points is a line).) Again our notation is ambiguous. Again. and there exist four points of which no three are collinear. any two lines have a unique common point. and no line contains every point. 9)) to infinite planes. We note that an affine geometry can be obtained from the corres- ponding projective geometry by deleting a hyperplane (the 'hyperplane at infinity') together with all of its subspaces. the require- ments are that any two points lie on a unique line. 2) is a Hadamard 3-design. An affine (or Fuclidean) geometry of dimension m is the collection of cosets of subspaces of a vector space of rank m over a field F. (ii) any two points lie on a unique line. q). we identify a subspace with the set of points it contains. If IF I = 2. q. Here. The same is true for pro- jective and affine planes. where q = IF 1. (iii) any three non-collinear points lie in a subset which. the subspaces of given positive dimension form a 2-design. q) can be defined to be a 2 . The exact analogue of the Veblen-Young theorem has been proved by Buekenhout [23] under the restriction that each line has at least four points. 1. AG(m. Again. (i) any line contains at least three points. together with the lines it contains. since they are characterised by the theorem of Desargues.(q2. We denote the design of points and hyperplanes by AG(m. while the projective geometry can be recovered from the affine geometry.

k. An extension may not exist. K-A. we must find a suitable design Zp. the same assertion is true. We will prove this theorem. If we put v = V-K. A) design. The question is: which quasi-residual designs are residual? A quasi-residual design with A = 1. which is very important in design theory and permutation groups. Let 5) be a symmetric 2 . is a 2 . k. The residual design Op with respect to p has the same point set as 1) p.{p ) for each block B of 2 which contains p. and the blocks of the design are the cosets of those subspaces of rank 2 which are totally isotropic with respect to the form. Given a t-design 0.1)/x. Here the vector space carries a symplectic bilinear form. that is. There are quasi-residual designs which are not residual: see van Lint [82]. an affine plane. 2) to the extension. (To construct the extension. and B a block of 0. Shrikhande and Singhi [19] concerning quasi-residual designs with A > 2. is it isomorphic to 7P for some (t + 1)-design D? a) is called an extension of the given design. in Chapter 4.) By applying (i. by a theorem of Hall and Connor [61]. the derived design 5) p with respect to a point p is the (t-l)-design whose points are the points of 0 different from p. We call 0B the residual design of 0 relative to the block B. we obtain a simple necessary condition for extendability: 8 . A) design with v = k(k + x + 1)/A is called quasi-residual.(v. If B and all its points are deleted from D. and state a theorem of Bose.(k2. We shall consider briefly in chapter 5 the affine symplectic geo- metry of dimension 4 over GF(2). A converse question. and whose blocks are the sets B . the resulting structure 0B is a 2 . An arbitrary 2 . K. A) design. For A = 2. but its blocks are the blocks of D not containing p.(V-K. k = K-A. A similar question can be asked about more general designs. is that of extendability: given a t-design. it is also a (t-l)-design.(V. The familiar procedure of adjoining a line at infinity to an affine plane shows that any such design is residual. then we have v = k(k + A . 1) design. or there may be more than one.

then k + I divides b(v + 1). then q = 2. PG(2. 1.k=q+1. 4) at most three times. The exis- tence of PG(2. 10) Proposition. The 2-design PG(2. then there are kA/A + 1) blocks containing p and q and meeting B in A + 1 points. k. A = 3. 2) is unique. the number of blocks of 0 o is 9 . q e B. q) is extendable. 3) (or by (1. we obtain a result of Hughes [70]: (1. (k-A-1)/(A+1)) design.2). This shows that the incidence structure D O. By (1. 2) and (1. Proof. and has a unique extension. A) designs with any given value of A. Applying (1.12) Theorem.2)A = (k . 4. k = 40. 2).(v-k. PG(2.0). k. A) design with these properties has r = v . A) design with b blocks is extend- able. First we note that a 3-design D is an extension of a symmetric 2-design if and only if any two blocks of a) meet in 0 or A + 1 points. k. 10) can be extended at most once. and (v . (iii) v = 112. k. each successive extension being unique (up to isomorphism). k = 12. 4) is also unique. and PG(2. 1. Much effort has been devoted to these projective planes and their extensions.1)(k .1. 2) and PG(2.1)/(A + 1) blocks disjoint from B.(v.A . Further application of (1. 10) is still undecided (we will discuss this further in chapter 11) and its extendability appears rather remote at present. If PG(2. 11) Theorem. or 10. where points are those outside B and whose blocks are those disjoint from B. then one of the following occurs: (i) 5) is a Hadamard 3-design.1 . q) has v=q2+q+l=b. If a 3-(v. (1. (ii) v = (A + 1)(x2 + 5A + 5). k = (A + 1)(A + 2). A) design 5 is an extension of a sym- metric 2-design. The strong- est result in this direction is due to Cameron [25]. In fact.(v. 7) applied to O)). namely AG(3. A 3 .(1. (iv) v = 496. If p. x2 = k . k.(v. Let B be a block of 5).0) shows that PG(2. and can be extended three times. A = 1. Hughes showed further that there are only finitely many extendable symmetric 2 . and hence (k . If a t . is a 2 .

3k + 2A + 2)(k2 . The simplest example is constructed as follows: the point set is { -} u GF(g2). Any derived design of this inversive plane is isomorphic to the Desarguesian affine plane of order q.1)/k = (k2 .3k + A + 2)/kX2.1)(k. Dembowski [42]. only one design is known which satisfies the hypotheses of (1. we may apply Fisher's inequality to 13 O yielding (v . d e GF(g2).be # 0 }. 12) then shows that the complements of all these blocks must be included as blocks. b = v(v . then v = 2k. 12) that the projective plane of order 4 is the only twice-extendable symmetric design. 1) design) is called an inversive plane or Mobius plane of order q. Also. 12). [43] has shown that any inversive plane of even order q can be represented similarly as an 'ovoid' in PG(3. Apart from these. the situation is a little different.(A+1)(A+2))= 0. a 2 . 10) is always satisfied. q + 1. For affine planes. b. An extension of an affine plane (that is. ad . Hadamard designs exist for many values of A. and these are all the blocks. q) 10 . (k . the new point p must be added to all existing blocks. namely the 3 . since the necessary condition (1. Another description of this design is as the point set of an elliptic quadric in PG(3. As previously noted. It is an easy deduction from (1.k .(22.A . if k = 2(A + 1)(A + 2). 1) design (case (ii) with A = 1): this will be discussed later. The same expression shows that. A+1 If a) o is degenerate (having but a single block).1)(k . k = 2(A + 1). For.(q2 + 1. q). blocks are all images of { 00 } u GF(q) under the linear fractional group PGL(2. so k divides 2(A + 1)(A + 2). in forming the extension. the projective line over GF(g2). the proof of (1. whence v = 4(A + 1). giving cases (iii) or (iv) of the Theorem. 6. and 0 is a Hadamard 3-design. c. and so k j (A + 1)(A + 2).1)2! k(k-1)(A+1). However. Otherwise. the blocks being the non-trivial plane sections. If k = (A + 1)(A + 2) then we have case (ii). any Hadamard 2-design is uniquely extendable. (v-k)(v-k-1)(k-X-1) k k. q2) _ {z ~' cz + d la. then A = 1 or 3.

Note that ovals can exist only if A Ik . A) design with k .1. An n-arc is a set of n points of a). 10). then q = 3 or 13. and so any point lies on at least one tangent to S. and p E S. Kantor [74] showed that if AG(2. 13) Proposition. Let 2 be a sym- metric 2 .1)/A). A and n are all odd. Proof. and p a point outside S. If a symmetric 2 . B) where B is a secant containing p and q. n = 1 + k/A). In fact. a block B is called a secant. A) design. or whether any non-Desarguesian affine plane is extendable. // (1. tangent or passant to S according as I B A S I = 2. k. AG(2. k. Let S be the oval. Any point of an n-arc in a symmetric 2 . A) design has an oval of type II. Proof. Proof. Then any point is on either one or all tangents to S. The number of secants containing p is i nA = i (k + A). and an oval of type II if it has no tangents (that is. n:5 1+k/A.(v. Let S be a type I oval in a symmetric 2 .1 or A. 14) Proposition. k.Ik respectively. q E S. 15) Proposition. Let ni be the number of points which 11 . 4). Using this and a further embedding theorem in conjunction with (1. A) design lies on (n . the extensions are all 'embedded' in pro- jective spaces over GF(3) in a similar way. In particular. and are closely related to the extensions of PG(2. 1 or 0. then k .with its non-trivial plane sections.(v. q) (with q > 2) is twice extendable. // (1. 13) is twice extendable. 3) is three times extendable.)A secants and k .A is even.(v. (1. n = I + (k .(v. Given an n-arc S.(n . It is not known whether any AG(2. Our next topic in this chapter concerns ovals.1)A tangents. Observe that k. An n-arc is called an oval of type I if each point lies on a unique tangent (that is. Let S be an n-arc. k. Count pairs (q. no three of which are contained in a block.A even.

5. 2) Paley design a). // It follows from (1.1) (i .(22. i. and B a block not containing p. then B is a type II oval in the projective plane 5) D. whence every point is on one or all the tangents. an extension of PG(2. We define a new design Z' as follows. the knot (or nucleus) of S. e. (ii) a block of 2 and the five points of 5) incident with it. Assmus.A). If 0 is a 3-design which is an extension of a projective plane. p a point of D.are on i tangents. It is now easy to verify that D' is a 3 . Z (2)ni = (z).(11. Mezzaroba and Salwach [8]). 6. 1) design. Therefore 1. Then ni = v. l) design.X) /2X. and by Ray-Chaudhuri and Wilson [98] in general.n)ni = 0. 12 . Consider the 2 . (i . For a further reference on ovals see Assmus and van Lint [2]. all the tangents to a type I oval S pass through a point p. Points of V are the points and blocks of 0. Given a type II oval S. Blocks of D' are sets of size 6 of the following types: (i) a point of 0 and the five blocks of 5)incident with it. (iii) an oval of D and its three tangents. 4) (cf. This result is as follows. Any three points either lie in a block or form an oval. and {p } u S is a type II oval. This was proved by Petrenjuk [93] for t = 4. 2 (k . We turn next to a generalisation of Fisher's inequality for designs with larger values of t. the passants to S and points outside S (with the reverse of the given incidence relation) form a 2 = ((k . 15) that. Every triple of points not contained in a block is a type I oval. in a projective plane of even order. Hence any two points of 0 are on two distinct blocks of 0 and three ovals.2) (k . ini = nk.

and Ys the vector with 1 in the column with label S. and are very plentiful.s. Let a) be a t . however. x) design. This theorem will be discussed further in chapters 3 and 17. Then b ? (s). xs = YS is a linear combination of y's. Ito [71]. Fnomoto. In particular. j=0 I S' nS I =j B?S' IBnSI=i i s - j=0i . Up to complementation. k. 13 . = F YS . we have a system of s + 1 linear equations for the S'nSI=j x's in terms of the y's. with non- zero diagonal entries s. the position is different. For 0 < i s and a fixed s-set S. We shall prove the inequality b (S). Ito and Noda [91]. So the equations have a unique solution. put y E P B 1 IBnSI=i i E E Ys. Accordingly. 17) Theorem. let pB be the row of M with label B.k < v . S) entry I if S S B.j)As+i-j. with (B. with t = 2s and t k s v .s).(23. 0 otherwise. 16) is called tight.s-iIS S YS I=j where am n denotes the number of blocks containing a given m-set M and disjoint from a given n-set N (where M n N = 0). the only tight 4 . Thus tight 2-designs are nothing but symmetric 2-designs. For t ) 2. Let M be the 'incidence matrix' with rows indexed by blocks and columns by s-subsets of the point set. 1) design. proving the assertion. Putting x. Equality holds if and only if the cardinality of the intersection of two blocks takes precisely s distinct values.(v. this number is easily seen to depend only on m and n if m + n t. 0 elsewhere. A) design with 2 < k < v-2 is the 4 . Proof.(v. A t-design attaining the bound in (1. and so is in the row space of M.(l. s-i (since s -. 16) Theorem. and Bremner [20] have shown: (1. k. 7. It is v enough to show that the rows of M span R(s). The coefficient matrix is triangular.

when the geometric configurations are Baer subplanes and symmetric differences of pairs of lines respectively. 6. 4). and adjoin each pair to all ovals in the corresponding class. 4). When i = 2. They occur also in coding and group theory. Similar arguments deal with the cases i = 1 and i = 0. and Bannai [10] has shown that. A similar method can be applied to obtain the 5 . We post- pone discussion of this until chapters 13 and 17. and showed group-theoretically that they are groups of automorphisms of designs with the appropriate parameters. 3). 8.(12.21]. it can be located as having point set {p. and take as blocks all sets {p. The simplest constructions use the ideas of coding theory.S2 (where S1 and S2 are ovals of II) if Is n S2 I is even. a block containing i of the new points contains also 8 . since the two designs are connected with the two perfect Golay codes. Choose a bijection between equivalence classes and pairs of new points. Luneburg [85] gave a construction which is similar in spirit (in that it involves extending known structures) but combinatorial in nature. sufficient to convince the reader of their importance in design theory.i points of H. r } u L. and counting arguments show that 168 blocks of this form are required. 15)).(24. r I U U. q. r. this configura- tion is a Type II oval (see the remark following (1. Alternatively. and consider the problem of extending 1I to a 5 . where L is a line of fI. It can be shown that this is an equivalence relation. q. 1) design from AG(2. for given even t > 10. forming some geometric configuration in H. 3) several times in this chapter. q. We must add three new points p. where U is a unital of PG(2. Witt [120] gave an indirect construction of the designs. In this way the design can be constructed and proved unique. and as blocks all 14 . 4) and AG(2. We con- clude this chapter with a brief survey of some constructions of them. There are 168 such ovals in II. there are only finitely many tight t-designs. Let H denote PG(2. We have touched on extensions of PG(2. He also proved the uniqueness of the designs [1. 1 with three equivalence classes of size 56. 1) design. Write S1 . There are also nonexistence results for larger values of t. For i < 3. He con- structed the two 5-fold transitive Mathieu groups M24 and M12 as extensions of known permutation groups.

those blocks of the 5 .) design which meet this set in six points. 15 . 1. Yet another construction is outlined in chapter 8. 8.(24.

This generality means that usually either the questions answered are too particular. The unifying theme is provided by a class of graphs. with the properties that any edge is incident with two vertices and any two vertices with at most one edge. the 'strongly regular graphs'. in the next five chapters. without loops or multiple edges.) As with designs. where each edge is a subset of the vertex set of cardinality 2. Given a subset S of the vertex set. First. By contrast. (In classical terms. introduced by Bose [16]. several of these areas will be considered. whose definition reflects the symmetry inherent in t-designs. the large and amorphous area called 'graph theory' is mainly concerned with questions about 'general' relations on a set. to have useful consequences for design theory. or the results obtained are not powerful enough. there is an alternative definition: a graph con- sists of a finite set of vertices and a set of edges. we shall mention an example of the kind of situation we shall not be discussing. A graph is complete if every pair of vertices is adjacent.2 Strongly regular graphs The theory of designs concerns itself with questions about subsets of a set (or relations between two sets) possessing a high degree of sym- metry. The complement of the graph r is the graph T' whose edge set is the complement of the edge set of r (in the set of all 2-element subsets of the vertex set). however. Still another definition: a graph consists of a finite set of vertices to- gether with a symmetric irreflexive binary relation (called adjacency) on the vertex set. with an 'incidence' relation between vertices and edges. There are some places where the two theories have interacted fruitfully. In any graph r. and null if it has no edges at all. r(p) will denote the set of vertices adjacent to the vertex p. A graph consists of a finite set of vertices together with a set of edges. r IS denotes the graph with vertex set S whose 16 . our graphs are undirected.

. ) As in design theory.. a regular graph is a 1-design. r is regular if and only if A(r)J = aJ for some constant a (which is then the valency of r). A more useful matrix is the adjacency matrix A(r). where n is the number of vertices and a the valency. so M(r)TM(r) = A(r) + aI if r is regular with valency a. p) entry 1 if p e e. n }. c.edges are those edges of r which join pairs of vertices of S. the matrix of the adjacency relation (with rows and columns indexed by vertices. Only the complete graph is a 2-design. and the common valency is the valency of the graph. the graph is called regular. For example. If every vertex has the same valency. Now it is well-known that there exists a (projective or affine) plane of order n if and only if rn contains the complete graph on n .. and let V be the set of all n x n latin squares with entries { 1. Form a graph rn with vertex set V by saying that two latin squares are adjacent if and only if they are orthogonal. let n be a natural number. where (as always) J denotes a matrix with every entry 1. with (e.1 vertices as a subgraph. . Note also that A(r) = J . a graph is a 0-design with k = 2. these are far too weak to force the existence of a plane of given order except in trivial cases.I . sometimes called the pair design. 17 . Note that M(r) M(r) is the sum of the symmetric A(r) and a diagonal matrix whose (p. a. with (p1. the number I r(p) I of vertices adjacent to p). This simple approach seems unlikely to produce any useful results for the theory of finite planes. and has the property that the number of vertices adjacent to p1 and p2 (p1 # p2) depends only on whether or not p1 and p2 are adjacent. but not com- plete or null. 0 other- wise). p) entry is the valency of p. the valency of p is the number of edges containing p (equivalently. and the number of vertices adjacent to p1 and p2 is c or d according as p1 and p2 are adjacent or non-adjacent. If p is a vertex of a graph r. (Thus. p2 1 is an edge. we can define the incidence matrix m(r) of a graph r to be the matrix of the incidence relation (with rows indexed by edges and columns by vertices. A strongly regular graph is a graph which is regular. p ) entry 1 if {p1. Its parameters are (n. not surprisingly. Graph theory provides us with lower bounds for the size of complete subgraphs of graphs. 0 otherwise).A(r). d).

(Be warned: this notation is not standard.c . of the two main contenders. one is impossible here because it uses the symbols k and A. 2. and the other is a special case of notation for association schemes and so is too cumbersome. 0 0 (i) L ::4 Fig. 1 1.1) are adjacent to neither p1 nor p2. Similar remarks hold if p1 and p2 are non-adjacent.(a . to be abandoned for the design theory notation whenever a design appears.c . ) If IF is a strongly regular graph and p1 and p2 are adjacent vertices. and so (n .a .8 .1 vertices are adjacent to p1 but not to p2. The notation used here is only temporary. then a . So the complement r is also strongly regular. We use it because a standard notation does not exist.1) .

1. 2. d = 2.1). 1 are P(5) and P(9). d = 2.. where 6) is the set of Sylow 3-subgroups of the alter- nating group A6. T(m) is strongly regular with parameters n = z m(m . The Gewirtz graph on 56 vertices. Larger graphs must be described verbally. two distinct vertices are adjacent if and only if they are not disjoint. the triangular graph T(m) (m > 4) has as vertices the 2-element subsets of a set of cardinality m. with two vertices adjacent if and only if their difference is a nonzero square. 2. In this graph. In drawings such as Fig. The second and third graphs of Fig. It is strongly regular. the set Y(p) of vertices not adjacent to a vertex p carries the Petersen graph. 2.1. join 19 . indeed. . . c = m . c = m . d = 4 (q . and is iso- morphic with its complement.5). c=m-2.). 1. The Clebsch graph has as vertices all subsets of 11. with a = 10. r(k. a small circle represents a vertex. Thus. If q is a prime power with q = 1. (mod 4). the Paley graph P(q) has as vertex set the elements of GF(q). any strongly regular graph with d = 0 is of this form. a = 2(m . join P E 3) to q E Q whenever qPq = P.to all vertices in P. The fourth graph in Fig. but two arcs may 'cross' without requiring a vertex at the crossing-point. which is called the Petersen graph. The complement of r(k. with parameters n = q. 2. and 2 is the set of involutions in A6. 2) is called a ladder graph. 2).1. is the complement of T(5). m) with parameters n = mk. c = 0. a=2(m-2). 1. Four strongly regular graphs are displayed in Fig. with parameters n = m2. c = 4(q .2. 2. m > 1) is also a strongly regular graph r(k. The disjoint union of k complete graphs on m vertices each (k. where S is a set of cardinality m. a = i (q . L2(m) is strongly regular.. The lattice graph L2(m) (m > 2) has as vertex set S X S. The vertex set is { ° } U 5 U Q. two distinct vertices are adjacent if and only if they have a common coordinate. d = 0. 1 is the complete bipartite graph r(2. can be defined by the following construction due to Sims.1 are L2(s) and L2(3). The first and third graphs in Fig. Only a few graphs are small enough to be drawn. Join . The first graph in Fig. two vertices are adjacent whenever (as subsets) their symmetric difference has cardinality 4. 2. d=4. 5 1 of even cardinality.1). a = m .1). and an arc an edge. m) is called a complete k-partite graph.2.

. Gewirtz [51. c. q2 E Z whenever g1g2 has order 4. whose edges are the unordered pairs corresponding to ordered pairs in 0. d). The Paley graphs are not in general determined by their parameters. say p and q. counterexamples are given at the end of this chapter. 2. Clebsch and Gewirtz graphs are determined up to isomorphism by their parameters. Form a graph r with vertex set V. The (p1. respectively. . Let A be the adjacency matrix of the strongly regular graph r with parameters (n.q1. The complete multipartite graphs and the Petersen. since G is tran- sitive on adjacent pairs and on non-adjacent pairs of vertices. If the rank 3 group G has even order. 4) (see the remarks on Luneburg's construction in chapter 1). such graphs are sometimes called rank 3 graphs. If G is a permutation group on a set V.. 31. (For the last three. We say G has rank 3. 14. Another construction of the Gewirtz graph uses as vertices one class of ovals in PG(2. G is a group of automorphisms of T. and the other orbit is also. Combinatorially. then G has a natural component-wise action on V X V. The fact that this graph is strongly regular will follow from (3. Since G is transitive on vertices. and lattice graphs L2 (n) with n # 4. F is regular. 6)) to (12)(45). Then typical edges of the second and third types join {{1. it contains an involution interchanging two points. two ovals joined whenever they are disjoint. see Seidel [101]. p2) entry of A2 is the number 20 . r is strongly regular. . if it is tran- sitive on V and has just three orbits on V X V. thus the orbit containing (p. 0'. Hoffman [68] and Shrikhande [107]) will be described in chapter 4.) The same holds for the triangular graphs T(n) with n 8. and (1. 6 1 (its orbits). including all those already mentioned. 5. p) I P E V) and two others 0. 2). a. q) is symmetric. the diagonal i (p. The many known families of rank 3 groups of even order yield large numbers of strongly regular graphs.].2)(34) to (23)(56). these results (due to Chang [32]. we may identi- fy P E 6) with a pair of disjoint 3-subsets of 11.

1) and (2. or non-adjacent. So (2. so the other eigen- values p1.a .1) = (n . A span a real algebra of dimension 3 which is commutative and consists of symmetric matrices. If p1 and p2 have multiplicities f1 and f2.d ± V((c . adjacent. 0=Tr(A)=a+f1p1 +f2p2. Also. Indeed we find without difficulty that (2.1) A2 = al + cA + d (J . Any other eigenvalue of J is zero. Thus the matrices I. A has the eigenvalue a with multiplicity 1. 3) a(a . 2 whence pi. p3 1 with p2 adjacent and p3 non- adjacent to p1. There is an orthogonal matrix which simultaneously diagonalises I. These equations determine f1 and f2 (since c = d = a is not possible). p2 of A satisfy the equation p =(a-d)+(c-d)p. or d according as p1 and p2 are equal. Indeed.1)d.d))]. an equation which can also be verified by selecting a vertex p1 and counting the number of edges {p2. a strongly regular graph could be defined as a graph whose adjacency matrix satisfies (2. 4) fi.2ad)) + - 21 . this number is a. corresponding to the eigenvalue n of J.1 t (n . f2 = z [n .c .1)(d2 C)2 c) . J. c. thus (2.A).of vertices adjacent to p1 and p2. p2 i [c . (2. J and A.d) 2 + 4(a .I . 2). then n = f1 + f2 + 1.2) AJ=JA=aJ.

c) . as the 22 . and the quotient is congruent to n . d). and A(t) is precisely the centraliser algebra of the set of matrices in the permutation representation of the group G = Aut(T).p1I)(A .c)2 . some will appear later. d = a +P 1P 2 * However. the graph L2(3) = P(9) is an example of this.c) = 2a. f1 and f2 are non-negative integers. then the algebra generated by I.2a. A few other conditions are known. We may distinguish two types of parameter sets for which f1 and f2 are integers. J. and the multiplicities of the eigenvalues of A(t) are the degrees of the irreducible constituents of the permutation character of G. jl is A-invariant. Sometimes it does give strong partial information. Thus d-c=1 andwefind c=d-1. plp2 = .1 (mod 2).4(a . then (A . a=2d. Type II. whence c = a + p1 + p2 + p1p2. (a. (We have already seen the 'only-if' part of this. This observation pro- vides quite strong conditions on the parameters. A and r is strongly regular. They include most of the known non-existence criteria for strongly regular graphs. whence A2 E (I. Given a graph r with adjacency matrix A. Then F is strongly regular if and only if A I j1 has just two distinct eigenvalues. ) We have seen that the parameters of the strongly regular graph r determine the eigenvalues of A and their multiplicities.1)(d .d)p.p1I)(A . when (d . so 0< d-c< 2.d) is the square of an integer u. with (n . knowledge of the multiplicities alone does not in general determine the parameters.Obviously.p2I) 1 j1 = 0. the eigenvalues determine their multiplicities. Note that type I graphs are also of type II if. the so-called 'rationality conditions'. so we have p1 + p2 = c . n=4d+1. For p1 and p2 are roots of the quadratic p2 = (a . The Paley graphs are of type I.1)(d . the corresponding eigenvalue is the valency. n is a square. Here n = 1 + d 2ac > 1 + a. u divides (n .p2I) = ceJ for some a. Since A is symmetric. If I is a rank 3 graph. and only if. if ( A .d. F is regular if and only if the all-1 vector j is an eigenvector of A.d) + (c . J. It can also be shown that a Bruck-Ryser like condition holds: n must be a sum of two squares of integers. Type I. Conversely. Conversely.

1) vertices contained in these cosets. a = s(2s + 1). [55]. This graph is strongly regular. due to Wielandt [119].following result. We have a + (m . We conclude this chapter with an example. and add new edges (x. y g X. Suppose r is a strongly regular graph on n = 2m vertices. and the value of c follows from the third equation. Next.prime power.s.1. and select any 12-(q . and (i) holds. (2. so either p1 = a or p1 = a . p2 = -a.s. c = s2 . add a new vertex 23 . by replacing the graph by its complement if necessary. c = 4(q2 5).1) of its cosets. We may assume a < m.-dimensional subspaces of V into two disjoint sets P and N of equal size 2(q + 1).1 . due to Delsarte and Goethals and to Turyn.m. whose eigenvalues have multiplicities 1. the second equation then yields m = 2s2 + 2s + 1. choose a member of P. let V = V(2. m . where q is an odd prime power. or (ii) r or its complement has parameters n = 4s2 + 4s + 2. for some positive integer s. Then either (1) r or its complement is a ladder graph. y) for x E X. p1 = -1 . 2ma2 = 2ma.1. Finally. q). a = 1. // The only graph with the parameters (ii) of (2. shows. They are related to the so-called 'symmetric conference matrices' [40]. 5) with s = 1 is the Petersen graph. Par- tition the l. 5) Theorem. d = 4(q2 1) (the same parameters as the Paley graph P(q2)). a = 2 (q2 1). We find a = m .y) E P. by the Perron-Frobenius theorem.1)p3 + mpz = 2mac and Ip1 I < a. Proof. with n = q2.1)p 1 + mp 2 = 0 a2+(m-1)p2+mp2=2ma a3 + (m . Delete each edge {x. y ) for all previously non-joined pairs of this form. First. m. d = s2. Form the graph with vertex set V in which x and y are joined whenever (x . let X be the set of 2q(q . The first equation shows a ° p1 (mod m). which shows that these graphs exist whenever 2s + 1 is a . In the first case. In the second case put p2 = s.

S is a subset of the unit sphere SZ. We must show that such a set can have cardinality at most z f(f + 3). The adjacency matrix A = A(F) has three distinct eigenspaces.I . The above 'switching' construction will be considered further in chapter 7. and so it is the Gram matrix of inner products of a set S of vectors in Rf.]. We conclude this chapter with an inequality connecting the para- meters of a strongly regular graph. and J .R be the function defined by fv(x) = ((v.I .A). e. Proof. fv(w) = 0 for v.joined to every vertex in X. and we may ignore the constants. and the functions fv (v E S) are linearly indepen- dent. Thus n = ISI :!s f+Zf(f+1). 24 .f.y)/(1 . Since F = al + (3A + y(J . as claimed. Then F is positive semidefinite symmetric. let fv : 11 . In particular. We may normalise to assume a = 1. 1 with multiplicities n .1)((v. v # w. there is such a linear combination F having the eigenvalues 0. and that the adjacency matrix of IF has an eigenvalue of multiplicity f (f > 1). Goethals and Seidel [41. For v E S. and two vectors in S have angle cos-1 ((3/a) or cos-1(y/a). and any matrix with these eigenspaces is a linear com- bination of I. since fv(v) = 1. Let t be a strongly regular graph on n vertices.. since xi + . any vector in S has length a2 . 6) Theorem. having the properties that r and f are both connected. (2. and has the parameters of (2. w c S. + xf = 1 on S2.y).5ii) with q = 2s + 1. A. But the spaces of homogeneous linear and quadratic functions on SZ have dimensions f and 2f(f + 1) respectively. f respectively. due to Delsarte. Now fv is a polynomial function of degree 2. Then n s Zf(f + 3). i.0)(1 .A.. The reader is invited to check that the resulting graph is strongly regular. x) . x) ..

say x and y.(v. Bose [16] studied the line graph of such a design. Indeed. clearly. Notice that the class of strongly regular graphs obtained in this way includes the triangular graphs: they are the line graphs of the 'pair designs' with k = 2. The blocks of such a design are often called 'lines'. easy counting arguments give us the parameters: n-b . with equality if and only if it is symmetric. The graph is not the complete graph. y } = 10. 1 }. The line graph of a 2 .1) = c=(r-2)+(k-1)2=v----. A 2-design is called quasi-symmetric if the cardinality of the intersection of two blocks takes just two distinct values. He observed that (3. so { x.. with two vertices adjacent whenever their inter- section is non-empty.3 Quasi-symmetric designs Recall that a 2-design is symmetric if the cardinality of the inter- section of two blocks is constant. since r > 1.2k i 1+(k-1)2 d=k2. k.kk-11)-v(v- kkv lk) a = k(r . It will come as no surprise to design theorists to learn that the rationality condition gives no extra information in this 25 . and not the null graph. 1) Theorem. two blocks have at most one common point. the graph whose vertices are the lines. b > v. since b > v. here. Thus. One simple class of quasi- symmetric designs is the class of 2-designs with A = 1 which are not projective planes. 1) design with b > v is strongly regular. in a quasi-symmetric design. Any 2-design has at least as many blocks as points.

We prove this by matrix methods resembling those used for Fisher's inequality (1. but the present proof has the advantage of giving us the multiplicities. J. 5). and A the adjacency matrix of its block graph. We have NTN = (r . // 26 .case. and NTN I j1 has only the one eigenvalue r . and A J j1 has just two eigenvalues.X)I + XJ. due to Goethals and Seidel [56]. Proof. j is an eigenvector of NTN with eigenvalue r. Since x # y. Let N be the incidence matrix of the design.v = rk. (3.A. 2). The block graph of a quasi-symmetric design is strongly regular. NNT=kI+yA+x(J-I-A). and NNT. Dr. 1) was generalised by Bose to 'partial geometries'. (3. and so j is an eigenvector of A. Here we pursue a different generalisation. We shall soon see the real reason for this. A is a linear combination of I. this is why the rationality conditions are empty. However. Notice that the multiplicities have automatically turned out to be positive integers.v). Norman observed that a simple counting argument will suffice to prove (3.A + X. Its parameters can be computed from the eigenvalues and their multi- plicities. C. From the first equation. in the last section. In an arbitrary quasi-symmetric design with intersection numbers x and y (x < y). which we discuss in the next chapter.1) and 0 (with multiplicity b . with two vertices adjacent if and only if their intersection has cardinality y. as described. Since jNT = kj (with a different j!) it follows that j is an eigenvector of NNT with eigenvalue rk. Thus the graph is strongly regular. and that NNT 1j1 has eigenvalues r . define the block graph to be the graph whose vertices are the blocks. the expressions for the parameters and eigenvalues of the block graph do give non-trivial divisi- bility conditions on the design parameters.A (with multiplicity v . 2) Theorem.

A).y)a+x(k. but from expressions for the eigen- values of r we have relations connecting them.a) = 0. (ii) If r is the complement of a ladder graph then 5) is a Hadamard 3-design. and is clearly not j. (3. 2) is (3. e2) be a matrix with rows indexed by pairs of points and columns by blocks. 1) and (1. 1)M(c0. el. Suppose b av(v . 4) Proposition.x)(-1 . We find that MT(0. A result closely related to (3. p2 } [1 B = i. 0) _ (y(k-y)c1 + (k-y) co) A + (x(k-x)81 + (k2x) eo) (J .a)= 0 (k2y)a + 2x) (-1 . In particular. Goethals and Seidel went on to examine some special strongly regular graphs. 2. 27 . and these are sufficient in some cases. Clearly we have the parameters v and b of 2. we have with (ce. 1. The other para- meters of 7) are not determined. c1) _ (0.I .1). (iii) r is never a lattice graph or the complement of one. p2 }. for i = 0. In a quasi-symmetric design (without repeated blocks). We shall prove (ii) later. (i) If r is a ladder graph then ID is obtained from a symmetric design by counting each block twice. B) entry ci if I {p1.1). and we know r. using only the knowledge of b and v. Then this matrix product is singular. Let M(co. to see whether they could be the block graphs of quasi- symmetric designs. for all choices of e0 and c1. So if a is the corres- ponding eigenvalue of A. b < zv(v . An eigenvector with zero eigenvalue is also an eigenvector for J and A. 0) respectively. with ({p1. Suppose r is the block graph of 2. 3)Proposition. E: . 0. Proof. y(k.

For the first suppose x = 0 and consider the incidence structure whose 'points' are the blocks of n containing p and whose 'blocks' are the points of a) different from p. there is up to complementation a unique design satis- fying (i)-(iii).1). 5) Proposition. // By the theorem of Petrenjuk (1. 1) design. (iii) b = Zv(v . In fact. so this structure is a 2-design. (ii) x = 0. (ii) 0 is a 4-design. so the determinant of these equations must be zero. which gives the result.1).4. 6) Proposition. Proof. and hence of strongly regular graphs. and any 'block' has A 'points'. a 4-design satisfies b ? Zv(v .1). y = 1 implies b = v(v . Applying Fisher's inequality to it.1 'blocks'. For a 2-design a) with 4 < k:5 v .We have already seen the second part. In a quasi-symmetric design 5) with intersection numbers x and y. we have the following result. Recall the 28 . We shall see a much more general result in chapter 17.1)/k(k . Stronger results hold in special cases. 7. with incidence the dual of that in 5). 2) has been used here. so k = x or k = y or k = 1 or x = y. while the last two are impossible. Thus (3.a is impossible. (i) x = 0 implies b < v(v . Any two 'points' lie in y . Thus. Note how little actual information from (3. 17). // Remark. any two of the following imply the third: (i) 0 is quasi-symmetric. 16). The first two imply that the design has repeated blocks. This 4-design provides us with several more examples of quasi- symmetric designs.1 >_ r. v .1). The pair designs satisfy (i) and (ii) but not (iii) (except in a trivial sense).1)/k. and is tight. a quasi-symmetric 4-design satisfies b = Zv(v . By (1. namely the 4 .(23. (3.a = 0 = -1 . But this determinant is 2 (k-x)(k-y)(k-1)(x-y).

Op. (2) This graph. Dpq) form an equivalence class of ovals (resp. By contrast. 4). (3. then Dp. 1 1 3 5 1 4 A 21 5 16 4 12 Al 77 21 56 16 40 A = n 253 77 176 56 120 0 a 112 16 70 10 42 c 36 0 18 0 8 d 60 6 34 2 18 Note (1) (2) (3) (4) Table 1 Notes: (1) These graphs are the complements of the block graphs as previously defined. and p and q are distinct points of D. 7) below describes a better- behaved situation.definitions of Dp. 29 . 4). shows that the multiplicities of the eigenvalues of a strongly regular graph do not determine the para- meters of the graph. The blocks of Dp (resp. together with T(23). 1) design. as described in Luneburg's construction in chapter 1. (Dpq is in fact symmetric. 7. Baer subplanes) in PG(2. Dp and Dpq are all quasi-symmetric.(23. Design a) Dp Dp Dqp Dpq v 23 22 22 21 21 k 7 6 7 6 7 X. 5)(i). Dp from chapter 1. The parameters of these designs and graphs are given in the following table.) Dp achieves the bound of (3. even when it is known to be the block graph of a quasi-symmetric design. being the projective plane PG(2. If D denotes the 4 .

Otherwise. We will examine this construction in a wider context in the next chapter. r = 2k . since x :s k we have x = 0. as follows. Proof. v . 7) Theorem. d = 6. Higman and Sims [67] constructed a strongly regular graph with n = 100. 5) applies.1)y + x = 2k(k . which implies s = 1.a)x = 2k(k . (3.v = v . Thus 2(k . join two blocks whenever they are disjoint. and so s(2s + 1)y + (s + 1)(2s +i)s=2s(s+l)(s2 + s + 1). count the number of choices of a point in that block and another block containing that point: ay + (4k . It is easy to show that there is a unique design with these parameters.A quasi-symmetric design with b = 2v .2. (4) This is the Gewirtz graph defined in chapter 2. b . we have v . a=s(2s+ 1).to all points. c = 0. (3. 5). 5) asserts that a symmetric 2- design cannot be a 3-design.1. and so a Hadamard 3-design. say x* and y*. 2) design.2 is either a Hadamard 3-design or the unique 2 . a = 22.1).1) divides x.2. // A result of Dembowski ([44] p. So (2. Thus 2s + 1 divides 2s(s + 1)(s2 + s + 1). D is a 3-design. 3.(6.1). k=s2+s+1. where 6) and G are the point and block sets of 5) D Join . b = 4k .1.1 = 2s 2 + 2s + 1. Note also that v = 2k. (3) From this design. and so 4(k . This class contains all symmetric 30 . Suppose we consider the class of symmetric designs in which the number of blocks containing three points takes just two distinct values. The eigenvalues of the block graph have multiplicities 1. The vertex set is { -) u P u B. By the remarks after (2. If the graph or its complement is a ladder graph. join a point and a block whenever P* they are incident. Given a block. 3)(ii) is a consequence of this result. we can assume a = n . y = Zk.3 .2.

all the projective geometries. so a complete determination seems difficult. // Note also that if x* = 0 then 0* achieves the bound of (3. 8) Proposition. we can obtain strong information about the parameters of such a design. If 5) is a symmetric design in which any 3 points lie in 0 or y* blocks.(11. If 0 is the complement of the given design. From this we can show (3. 5)(i). b* = v . 12). then the same is true of the dual of 5).designs with A s 2.1. But some progress can be made. 7. It is not known whether this remains true when 0 is replaced by an arbitrary x*. then D* satisfies the hypotheses of 3. 2) design. if a)* is a Hadamard 3-design for any point p it is easy to show that the original design is a projective geo- metry. 31 . for any point p. It can also be shown that (3. Using result (1. and the complement of each of its members. large numbers of designs with v = 22m. Proof. the blocks containing p and points differ- ent from p form a quasi-symmetric design n* with parameters v* = k. etc. 9) Proposition. 5. A Hadamard 2-design in which any three points lie in x* or y* blocks is either a projective geometry over GF(2) or the unique 2 . Thus.

we prove the theorem of Chang [32] and Hoffman [68] mentioned in chapter 2. choose such a circuit C = (xo. 32 . for such an edge would divide C into two smaller circuits. k * 3. one of which would have odd length. From the preceding para- graph. Then there are no edges { xi. there are at least n .2 . and let m vertices of r(x) be adjacent to y and z. If m = 3. Similarly there are n .. so m ? 2.2 adjacent to neither. a contradiction. xj I of o with i . 1) Theorem.2 3 + 3.. where n _> 8. n . and m . Suppose that it contains a circuit of odd length. We must have k = 5.. these must be nonadjacent to both x1 and x4. . contrary to assumption. Then every vertex of r(x) adjacent to w is adjacent to y or to z.2 .m vertices of r(x) adjacent to y but not z.4 0 common neighbours x3. let w be the vertex adjacent to neither y nor z. By the preceding paragraph.6 vertices outside C nonadjacent to both x3 and x4. (4. Consider the complementary graph W.2 on 2(n . whence n . Let y.4 Partial geometries To introduce this chapter. A clique in a graph is defined to be a maximal complete subgraph. . Since x1 is non- adjacent to n . Since c = 4.j ±1 (mod k). Let r be a strongly regular graph with the same parameters as T(n). . x1. Also. Proof. Then r is isomorphic to T(n). and x is adjacent to y and z. There are n ..2) certices.m to z but not y. n-6 1. so k 6. For any vertex x of r. . x and x1 are nonadjacent in o and have k . xk = xo) with k minimal. r(x) carries a regular graph A of valency n .4 vertices outside C.5 vertices nonadjacent to x1 and x3. So m = 2. we have m < 3. There are n .5 vertices adjacent to x0 in Z (other than x1 and x4). xk-2 there. z E P(x) be non- adjacent.

2) Definition. discovered by Chang [33].1. and any edge in a unique grand clique. Equivalently. two vertices adjacent whenever they are collinear. 3) Definition. Bose [16] made the following definition. d=rt. with a = r(k . This means that A contains two disjoint cliques of size n . For n = 8.2) + (r . (4. So X contains no odd circuits. Thus the grand cliques and vertices are the points and blocks of a 2-(n. and it is bipartite. t).2. (4. In the proof of (4. we showed that the points and grand cliques formed a partial geometry with parameters (2. k. k. 1) design (the pair design on n points).1)n. r. To study large cliques in more general strongly regular graphs. any two having at most (and hence exactly) one vertex in common. There are 2 (n . The cases n < 8 can be dealt with individually. so r = T(n).1). and any line with k points. 2/(n . t) is an incidence structure of points and lines satisfying the following conditions: (i) any point is incident with r lines. t) is a partial geometry with parameters (k.1)(t .1. there are just three exceptional graphs. n . // The line graph of a partial geometry is defined dually. // The theorem is in fact true under the weaker assumption n # 8. (4. (ii) two points are incident with at most one line. and r is its block graph. then there are exactly t points incident with L and collinear with p. The proof of 33 .1) = n grand cliques. 1).The point graph of a partial geometry has as vertices the points of the geometry. Note that the dual of a partial geometry with parameters (r. Let r be the point graph of a partial geometry with parameters (r. Then r is strongly regular.1). 2). 4) Proposition. any vertex in r lies in two 'grand cliques' of size n . It is also strongly regular. (iii)if a point p is not incident with a line L. c = (k . k. 2. t). A partial geometry with parameters (r.

for n > 4. A strongly regular graph with the same parameters as L2 (n). referring interested readers to [16] or [113]. // This theorem is true for n * 4. n. 7) Theorem (Shrikhande [107]). It is geometric if it is the point graph of a partial geometry.By (4. Theorem (Bose [16]). is isomorphic to L2 (n). 1) is necessarily an n x n square lattice. Another important special case is the following.1).(4. Instead. 2. 2. 6) geometric if k > a [r(r-1) + t(r+1)(r2-2r+2)]. t) is (4. But a partial geometry (n. 2) is necessarily the pair design on n points.1. Fig.2 + (r . A pseudo-geometric graph (r.1 We will not go further into the theory of partial geometries.1)(t . 6). 4. k. 34 . 5) Definition. k. 6). (4. Shrikhande found that there is a unique exception. In the case n = 4. (4. with opposite sides identified as shown (so it should be regarded as drawn on a torus). 1) is a complete bipartite graph. It is the graph drawn below. 1) may be regarded as a special case of (4. (4. and so its dual (2. we describe four connections with the theory of designs. t) if its parameters are given by a = r(k . A strongly regular graph is pseudo-geometric (r. Proof. c = k .1). 6) below by Bose [16] uses the same idea. // Since a partial geometry with parameters (n . d = rt. such a graph is geometric.

Let 11 be a projective plane of even order n. where n is a power of 2.1. First. Counting as in (1. 7). (A partial geometry with t = 1 is called a generalized quadrangle. we remark on two constructions of partial geometries from ovals. while k = 6 requires separate treatment). We suppose k > 6 (the cases k < 6 follow from the strengthened version of (4. Then the points outside 0 and secants to 0 form a partial geometry with parameters r = (n + 2). Second. (4.1) = r(k . we may regard 0 as a collection of 1- II dimensional subspaces of the vector space V = V(3. Let 0 be a 2 . and 0 a type II oval in fl. z a If is a Desarguesian. That is.1)). t = 1. we show how the Hall-Connor theorem can be deduced from (4. and 2 is quasi-symmetric. Proof. we can label 35 . t = (n .) Third.(v. its line graph is the block graph of the design as defined in chapter 3. Let B be a block of 2.(v. and the cosets of members of 0 as lines. Thus (i-1)(i-2)ni = 0. 8) Theorem (Hall-Connor [61]). A quasi-residual design with A = 2 is residual. n). 1). By (4. we obtain a partial geometry with r = n + 2.2). and let ni denote the number of blocks which inter- sect B in i points (i >_ 0). k = n. whence ni = 0 unless i = 1 or 2. we find that ni = Z k(k + 3) ini = k(k + 1) i(i-1)ni = k(k . k. Let r be the block graph of 0) (where adjacency corresponds to intersection of size 1). So any two blocks of 0 meet in 1 or 2 points.1). 1). k. Taking the elements of V as points. Then r is strongly regular and has the same parameters as T(k + 2). a triviality. k = n . 1) design (with (v . A partial geometry with t = k is the same thing as a 2 . 1) in the same way. 2) design with v = Zk(k + 1). r = T(k + 2).

n. and let X be a new block. There is a function g defined on the natural numbers with the property that a quasi-residual design with k > g(A) is residual. 36 .i points for i = 1. Then the cosets of the members of S are the lines of a net on the point set V. 9) Theorem (Bose-Shrikhande-Singhi [19]). where M = (A . Bruck [21] proved a special case of (4.3A + 3). they gave 76 for A = 3 2(A-1)(A4-2A2+A+2) for 4!5A:59 g(A)= 2(A-1)(A-2)(M+VM2+4(A-1))-(A-1) for A?10. p. (This is consistent with our notation L2(n). The same argument shows that. Moore geometries [75].1 mutually orthogonal Latin squares of order n (see [44].. Related classes of geometries include generalized polygons [50]. and its point graph is said to be of Latin square type Lr(n). n) be a 2-dimensional vector space.. . Such a partial geometry is called a net. // Specifically. semi-partial geometries [36]. 2. Now adjoin X to the point set of 5). // Using an elaboration of these ideas. k+2 I in such a way that two blocks meet in i points if and only if their labels meet in 2 . r-1) is equivalent to that of r . let V =V(2.142 ). more generally. We shall return to partial geometries in chapter 12.the blocks with the 2-subsets of X = 11. adjoin to each block its label. Finally.1 such squares. a connection with Latin squares. and partial- geometric designs [19]. We obtain a symmetric design with A = 2 whose residual (with respect to X) is 5).) For example. and S any set of 1-dimensional sub- spaces.2 orthogonal Latin squares of order n. (4. the existence of a partial geometry with parameters (r. . It is known that the existence of an affine plane of order n is equivalent to that of a system of n .1)(A2 . 7) and used it to show that a set of 'sufficiently many' mutually orthogonal Latin squares of order n can be completed to a set of n . Shrikhande and Singhi proved the following result. Bose.

a path is a sequence of vertices in which consecutive vertices are adjacent. so u = 3. Indeed it shows that [a2 ± a(a . Equality holds in either case if and only if the graph is strongly regular with c = 0. The function d defined by d(p. here the pentagon is the unique graph.2) ] 4a 3 are integers. It is worth noting that analogous bounds exist for larger values of diameter and girth. 7 or 57. d = 1. is a metric in a connected graph. d = 1 is so strong that the rationality con- dition can be expected to give useful information. 4 if c = 0 and d> 1. Such a graph is called a Moore graph with diameter 2. so u divides 15.5 Strongly regular graphs with no triangles In a graph. r is connected with diameter 2 if d > 0. In type II we have a = 4(u2 + 3) for some integer u which divides -16(u2 + 3)(u2 . 37 . The girth of a graph is the length of the shortest circuit which contains no repeated edges. r has a girth pro- vided a> 1. For strongly regular graphs. non-trivial examples do exist. and 5 if c = 0. The condition c = 0. diameter and girth are simply determined by the parameters. the girth is 3 if c> 0. n = 10. the diameter of the graph is the largest value assumed by d.5). provided such a circuit exists. 50 or 3250. and a graph with girth 5 and minimal valency a has at least a2 + 1 vertices. In the diameter 2 case. and a circuit is a path with initial and terminal point equal. d = 1. and is disconnected if d = 0. a = 3. A graph is connected if any two vertices lie in a path. the connectedness. It is easy to see that a graph with diameter 2 and maximal valency a has at most a2 + 1 vertices. however. u = 1 is not possible. q) = length of the shortest path containing p and q. 5 or 15. but recently Bannai and Ito[11]and Damerell [35] have shown that they are attained only by graphs consisting of a single circuit. Type I occurs precisely if a = 2.

If n = 10. though it is not as strong as in the Moore graph case. it is easy to see that the Petersen graph is the unique example. Clearly any block is incident with d points. with a point and block incident if and only if they are adjacent in r. p) denote the incidence structure with point set r(p) and block set ?(p). and let D(I'. Proof. which are uninteresting. (5. 1 < d < a. // For the remainder of this chapter.4k + 4v) are integers.1. p) for some strongly regular graph r of girth 4? The rationality conditions give us some information.1 are expressible as D(r. 1 < d < a. we shall exclude it and con- sider strongly regular graphs of girth 4. This was first proved by Hoffman and Singleton [69]. They are both adjacent to p. b = v(v . and so are not adjacent to each other. we will allow designs to have repeated blocks. and d . all of which must lie in T(p). Take two points of D(r. Let r be a strongly regular graph with c = 0. Then 38 . so we ex- clude them as well and assume c = 0. so k2 . 1) Theorem.3) ] (k2 . The other parameters follow. possibly having repeated blocks.4k + 4v is a square. and will use the design parameters v and k in place of a and d. d = a).1 more vertices are adjacent to both. Hoffman and Singleton were able to show the existence and uniqueness of a graph with n = 50.4k + 4v ° k2 (mod 2)). Type I cannot occur (d * c + 1).) The final case is too large for any complete analysis so far. Now we can ask: which 2-designs with A = k . who went on to analyse the first two cases. Then D(r. (We outline a construction and uniqueness proof in chapter 8. These include the complete bipartite graphs (with c = 0. a = 7. k = d.1. r = v . p). say k2-4k+4v=(k+2s)2 where s is an integer (since k2 . s > 0. a = 3.1)/k. A = k . Let p be any vertex. p) is a 2- design with v = a. They show that [n-lt v(v + k . Since this case is fairly well settled.

and ((s+l)k+s2)((s+2)k+s2-1) t ((s+l)k+s2)((s+2)k+s2-3) k k+2s ] are integers. if k = 2. 39 .6). D(r. the two terms inside the square brackets are both integers.x blocks must be adjacent to both. The first of these is the Gewirtz graph. In fact. if two blocks are non- adjacent and have x common points. two blocks are adjacent if and only if they are disjoint. then v = (s + 1)2 + 1.k others. in which n(I'. The graphs obtained are the complement of the Clebsch graph (n = 16) and the Higman-Sims graph (n = 100) respectively. and adjacent blocks must be disjoint (since the graph contains no triangles).4)(k .) Any block must be adjacent to v . In particular.2)(k .1 as a potential D(r. These are the graphs on 56 and 77 vertices constructed in the last chapter. For example.k blocks are disjoint from any given block. and s + 1 is not divisible by 4. Having chosen a 2-design with i = k . then k . Two examples of this. Sometimes it happens that exactly v . p). 6. p) is the pair design on 10 points. If p and q are adjacent vertices in the Higman-Sims graph I'. 1) design. p) is the 4-dimensional affine symplectic geometry over GF(2) (see chapter 1). are the pair design on five points. 4 or 6. In the second case. there are only finitely many possible values of v. The last condition implies that k + 2s divides k(k .(5. furthermore. where the resulting graph is strongly regular. 4 or 6. (Thus the problem resembles that of extending a given design.(22. 2) Proposition. v = (s + 1)k + s2. the rationality condition allows infinitely many values of v. Then there is no choice in the construction. whence k divides s2 (s2-1) and k+2s divides s(s+l)(s+2)(s+3). For k = 2. two blocks are adjacent whenever they are disjoint but not parallel. This was observed by Biggs [13]. and the unique 3 . we know all the edges of r except those joining distinct blocks. only two further examples of strongly regular graphs of girth 4 (other than complete bipartite graphs) are known. so for any value of k different from 2.

k) = (9. and 1 < k < v. for example (v. d = k.then r I T(p) and r i (r(p) n I'(q)) are the graphs on 77 and 56 vertices just described. 5. 2). others. which gives the result. Instead of requiring that the strongly regular graph contains no triangles. the Petersen graph can be partitioned into two pentagons. however. the number of blocks disjoint from a given block is at most v(v-1) k .1). ) We shall see later the reason for some of these facts.) In addition. // (5. 1). So k(k2 .k(v-2)2 _(v-k)2(v-k-1) .3k+v)<(v-k)(v-k. but this number is at - k2-3k+v k(k2-3k+v) least v . By (1. the observation that adjacent blocks are disjoint gives the following result. 7). and the Higman-Sims graph into two Hoffman-Singleton graphs. then rIr(p) is the Petersen graph. (5. remain slightly mysterious. Let IF be a strongly regular graph with a = v. (Similarly. 48). it is enough to assume that it contains no configurations of shape as in Fig. Then v Z[3k+1+ (k. Fig. and r I (r(p) n T(q)) is a ladder graph.k. If v=(s+1)k+s2 (as in (5. Let D(r. (The last fact was proved by Sims [109]. c = 0. Proof.1 . if p and q are adjacent vertices in the complement of the Clebsch graph. 3) eliminates some parameter sets satisfying the rationality condition (5.1) V (4k+1)]. 1 40 . 4) or (256. 2)) then k<_s(s+1). 1. It is worth noting that a more general result can be proved by the method. 3) Theorem. In the situation of a strongly regular graph with girth 4. p) be the 2-design of (5. 5.

This gives an inequality for a class of strongly regular graphs which in- clude all those with c 1. suppose k > 2. (iii) r I r(p) is strongly regular. Proof. 12) hold). 6). The three statements are equivalent respectively to the following three statements about a) p (i) 5) p is a 2-design. If either holds. 4) Lemma. Then the following are equivalent: (i) D(T. 3). Higman [65]. the point and line graphs of generalised quadrangles. as the Gewirtz graph on 56 vertices shows. (iii) b = v(v . Before proving this.1) . 4). we note that the condition k 2 is essential. p). (5. Any one of these implies that v = s(s2 + 3s + 1). p) is a 3-design (ii) D(r. With the hypotheses of (5./(4k + 1)]. any two together imply the third. (i) and (ii) are equivalent (since b = v(v. (iv) v = 21[3k + 1 + (k . 5) Theorem. then (1. Remark. which should be compared with (3. (iii) 2) p has equally many points and blocks. G. Proof. (ii)) is a quasisymmetric 2-design with x = 0. By (5. p) is quasisymmetric. (5.1)/k. (ii) the dual of 2p is a 2-design. and that TIr(p) is the complement of the block graph of 5)(I'. and many other interesting graphs. (i) l is a 3-design. What about the case of equality? We need first a simple lemma on designs. 12) gives us rather limited possibilities for the parameters of E)(P.1) /k). Of the following three statements about a design 5) with 2<k<v-1. For generalised quad- rangles it gives a simple proof of an inequality due to D. and imply also that a) is an extension of a symmetric 2-design (so that the conclusions of (1. k = s(s + 1). p): 41 .

Note also that if p and r are non-adjacent. adjacent blocks are disjoint. if p and q are adjacent vertices. (iv) v = 496. while non-disjoint pairs of blocks have equally many common points. and the proof of (1. Then any two non-adjacent blocks are both adjacent to t further blocks. s).1). Thus o (r. We are left with v = y(y2 + 3y + 1). and so two blocks are adjacent if and only if they are disjoint. 2) + 1) (Hadamard 2-design). contain) k . r I r(p) and r I (I'(p) (' P(q)) are strongly regular graphs containing no triangles. Let us exhibit one corollary of (5. 3). y = 2. 12) shows that the parameters of an extendable symmetric design are of one of the forms (i) (4a + 3. s2. for some integer t. Much of the proof of (1. So (ii) and (iii) hold. with incidence = adjacency in r. k = y(y + 1). 5). the number of blocks disjoint from a given block is at least v .k.1) and 2 . (i) v = 4y. p) is quasi-symmetric and r IT(P) is the complement of its block graph. except in the Hadamard case. for which (iv) is easily verified.1). the first is excluded by (5. and the third and fourth fail the rationality condition (5. s2 . (iii) v = 112. ) If F satisfies the hypotheses of (5. // Remark. s2 . A). The parameters of the corresponding designs are 2 . In the situation of (5. 12) required for (5. Of these. s(s + 1).s . (ii) v = Y(Y2 + 3y + 1). the special case of (1.1). 2). (Indeed. If (iv) holds then we have equality in (5. + 3X + 1. 42 . then D(r. 1). p) is a quasi- symmetric 3-design. 3). 11. 12) is concerned with estab- lishing that.t points. k = Y(Y + 1). and so are adjacent to (that is. Finally. 5) was proved first. k = 12. (r(p) n f (r). suppose (iii) holds. y = 4. s(s . 5). this is automatically true. p)q is a quasisymmetric 2-design for any point q. 5). concerning extendability of designs. Thus. and so D(r. k = 40.(s(s2 + s . x2 (ii) ((A + 2)(A2 + 4x + 2). F(p) n r(r)) is a symmetric 2-design with parameters (s 2 (s + 2). k = 2y. 12) can be considerably shortened. (iii) (111. (1.(s2(s + 2).

However. so they are adjacent to c .((A+2)(A2+4A+2). Mezzaroba and Salwach [8] and Denniston (personal communication).c . the case of equality is characterised by a result resembling (5. which includes (5. exactly v .A . r=a. 4). With A = 2. Let 2 be a 3-design and q a point of 5) for which 5) q is isomorphic to the given symmetric design. p) is a 1-design. Let us consider case (ii). there is a unique design. q)p is the dual of the given symmetric design. a weaker result is true: (5. p) = a). whence 43 . Thus r I r(p) is strongly regular. namely PG(2. 3). we have cA = (a-c-1)(d-A-1). If IF is an arbitrary strongly regular graph. The question has not to our knowledge been settled for the others. 6) Theorem. b=n-a-1. and there is a strongly regular graph r of girth 4 with a vertex p such that a) (r. q) is also a 3-design. A2+3A+1. With A = 1. Assmus. but it is not in general a 2-design. whether or not they are adjacent.a . This is not true for the other parameter sets. due to Gewirtz [51] (see p.(56. 11. then 2)(r.1 points of r(p). There is an inequality (the 'Krein bound') for arbitrary strongly regular graphs.1. Baumert and Hall (personal com- munication) have shown that the first of these is not extendable.c . 5). there are four known 2 .1) = (n . p) is a 2-design. It is known that any Hadamard design is uniquely extendable. (provided only that d > 0). In a). A) design is extendable. Proof. it follows that D(r. 19).) Sup- pose c > 0 and D(r. 2) designs.1 or d .1)d' to it. according as they are adjacent to each other or not. These matters will be discussed in chapter 17. Hall. k=d. Lane and Wales [62]. and it is easy to see that a) (r. If a symmetric 2 . // The results of this section can be generalised. (iv) (495.k blocks are disjoint from any block. 3). and it possesses three different (though isomorphic) extensions. (The parameters are v=a. then so is its dual. Since (iv) of (5. Then two points of r(p) are adjacent to A points of r'(p). Applying 'a(a . 39.A . 5) is independent of the vertex p.

with parameters (n. 10. p) is a 2-design if and only if rl r(p) is strongly regular with parameters c(d . d).(10. J. Paulus found six non- isomorphic graphs with these properties and parameters.1) c(d . a. 1). personal communication). 3.d + a-1 ' a-1 An example of such a graph r was found by Paulus and Seidel (J. Indeed. with parameters (10.1) (a. where c > 0 and d > 0. c. Then ow. and 0(r. Seidel. 3. 44 . 4). 2) design. r ( r(p) is a Petersen graph. 7) Theorem. c . =(a-c-1)(d-1) a-1 Thus (5. c. 4. Let r be strongly regular. 0. p) is a 2 . r has parameters (26.

) 45 . ([44] p. and a point and block are incident if and only if their labels are adjacent. Not surprisingly. k. k. we see that u divides k = J + u2. The fact that the triangle is the only (v. 1) graph has appeared in different guises. Renyi and S6s [49].6 Polarities of designs Let r be a strongly regular graph in which n = v. A) design can be constructed from it as follows: the points and blocks are both indexed by the vertices of r. It is also the fact. and so u divides X. Then any two vertices are adjacent to exactly A others.A = u2. a = k. the existence of such a polarity is a very strong restriction on the design. and the point graph of the 27 lines in a general cubic surface with A = 3. indeed. that a polarity of a finite projective plane has absolute points. Clearly a symmetric 2 . for which several elementary proofs (avoiding use of the ration- ality conditions) have recently been given. The smallest examples are the triangle with A = 1. c = d = A. The rationality condition shows that (v. A)-graphs with a given value of A. Such a graph is called a (v. k. This design has a further remarkable property: the correspondence between points and blocks indexed by the same vertices is a polarity with no absolute points. Ahrens and Szekeres [1] showed the existence of a graph with these parameters for each prime power value of A.1)t kk ] are integers. In the extremal case with A = u we have k = A(A + 1). Thus there exist only finitely many (v. It is the Friendship Theorem of Frdos. 152. Putting k . X)-graph (among other things). v = A2 (A. k. due to Baer. + 2). their graphs are line graphs of generalised quadrangles. L2 (4) and the Shrikhande graph (chapter 4) with X = 2.(v.

and p and q are adjacent if q e p°.) So the two concepts are equivalent. 6. Construct an incidence structure as follows: points and blocks are indexed by vertices. p and q are adjacent if q E p° but q * p. A) graph or one of the 'windmills' of Fig. k. d = A. k. (Vertices are points of 5).) This time the rationality condition shows that 46 . c = A . k. we can recover the (v. and irreflexive since a has no absolute points. A) design. A) design has a polarity a with no absolute points. This incidence structure is a symmetric 2 . A) graph as follows: the vertices are the points. and the correspondence between points and blocks with the same label is again a polarity. and a point and block are incident if and only if their labels are equal or adjacent. (Such a polarity is called null. k. Now. 1 If a symmetric 2 . let r be a strongly regular graph in which n = v. but this time every point is absolute. A graph in which any two vertices are adjacent to A vertices is either a (v.2.) Again the graph can be recovered from the design 2 and the polarity a. 1 with A=1. (This relation is symmetric since a is an involution.(v.(v. 6. the other conditions follow from the design properties. a = k . Fig.1.

A null polarity of a design induces a polarity of the complementary design with no point absolute. since d = A . q E B . This fact. 47 . and the graphs are among those discussed in the preceding chapter. 2) design. so the two cases are really the same. and a a null polarity of 0. a is uniquely determined by the image of a point or block. a3 are three distinct null polarities. Let 0 be a symmetric 2-design with A = 2.(56. 1) Theorem.(16.2 is non- negative. Examples with u = 2 and u = 3 are the Clebsch and Gewirtz graphs.(16. The first of these is the same one that also has polarities with no absolute points! The next result is easy to verify. (i) For any block B and any two points. 3. clearly not peculiar to the finite case. The case A = 1 cannot occur here. Thus. Here c = A . z[v -lf k ] are integers. 6. (ii) (iii) Two distinct null polarities commute. An interesting case is that with A = 2. Each is characterised by its parameters. and their product is a fixed-point-free involution.(4.2 = 0. It follows from a previous remark that the only non-trivial 2- design with A = 2 admitting three distinct null polarities is the 2 . 2) and 2 . p. (iv) If al.{ Ba) . is also a consequence of the fact that an absolute line in a projective plane carries a unique absolute point. This does not give a bound for v in terms of X. 2 . (6. However. 2) designs with null polarities exist and are unique. a2. 11. where we saw that k = u2 + 2 for some integer u not divisible by 4. q and Ba 'generate' the trivial 2 . 2) design (which has 6 such polarities). Null polarities have received less attention than polarities with no absolute points. then a1 a2 03 is a polarity with no absolute points. p. it is often convenient to separate them. 6. Next we give some further examples of graphs with the above properties.

Suppose H is a Hadamard matrix. We may interpret H as an
incidence matrix (with, say, -1 for incidence and +1 for non-incidence).
If H is symmetric and has a constant diagonal, we can change signs to
make the first row and column consist of -1's without affecting these
properties; deleting them we obtain the incidence matrix of a Hadamard
2-design with parameters 2 - (4u2 - 1, 2u2 - 1, u2 - 1) with a null
polarity. If in addition the original H has constant row sums, then it
or its negative is the incidence matrix of a 2-design with parameters
2 - (4u2, 2u2 ± u, u2 ± u) with a null polarity; the sign is + or -
according as the signs of the diagonal and the row sums agree or not.
Such Hadamard matrices are quite common. Thus, one exists realising
each possibility for the signs if a Hadamard matrix of order 2u exists.
They exist for other orders, such as 36, as well. (See [55] and [24]. )
Other examples of null polarities are provided by the symplectic
polarities of projective (2m - 1)-space over GF(q), m > 2. Also, if a
is an orthogonal polarity of projective 2m-space over GF(q), with m ? 2
and q is odd, then the absolute points and hyperplanes form a symmetric
design, and a induces a null polarity of this design. The design has the
same parameters as PG(2m - 1, q), but is not isomorphic with
PG(2m - 1, q).
Further examples when q = 3 arise as follows. Let Q be the
quadratic form associated with a. Since Q(x) = Q(-x), Q induces a
function from projective points to GF(3); denote it by Q as well. If
P.i is the set of points on which Q takes the value i (i = ±1), then
(Pi, P°) is a symmetric design, and or induces a polarity of it with no
absolute points. The parameters of this design are

2 - (z 3m(3m + i), z 3m-1
(3m - i), 2 3m-1(3m-1 - i)), i = ±1,
provided the sign of Q is chosen correctly.

48

7 Extension of graphs

In this section we shall be considering graphs from a slightly
different point of view, as incidence structures in their own right (as
remarked in Chapter 2). In particular, a regular graph is a 1-design
with k = 2, and conversely (provided we forbid repeated blocks). An
interesting question is: when do such 1-designs have extensions? Since
2-designs with k = 3 are so common, this problem is too general, and
we shall usually impose extra conditions on the extension. Extensions of
designs were originally used by Witt, Hughes and Dembowski for studying
transitive extensions of permutation groups, and it might be expected
that extensions of regular graphs would be useful in the study of doubly
transitive groups. This is indeed the case.
If we are extending a design by adding an extra point p, we know
all the blocks of the extension containing p - these are obtained simply
by adjoining p to all the old blocks - but it is in the blocks not containing
p that possible ambiguities arise. Let us then suppose, as a first possi-
bility, that we have a set o of triples, called blocks, in which the blocks
not containing a point p are uniquely determined in a natural way by those
containing p; in particular let us suppose that, when we know which of
{pqr }, {pqs } and {prs } are blocks, we can decide whether {qrs }
is a block. It is easy to see that the rule must be that, given a 4-set
{pgrs }, the number of its 3-subsets which are blocks belongs to a sub-
set S of 10, 1, 2, 3, 4 } with the property that i, j e S, i * j= I i-j I 2:2.
Further analysis shows that S is a subset of one of { 0, 2, 4 }, 11, 4 },
or { 0, 3 1, and that in the second or third case either A or its comple-
ment is the set of all triples containing a particular point, an uninteresting
situation.
A set A of triples, with the property that an even number of the
3-subsets of any 4-set belong to A, was called a two-graph by G. Higman,

49

who first studied such sets. Two-graphs turn out to be fascinating ob-
jects, closely connected with graph theory, finite geometry, real geom-
etry (sets of equiangular lines in Euclidean space), etc. In addition, a
large number of the known finite doubly transitive groups act on two-
graphs. (These include PSL(2, q) for q = 1(mod 4), PSU(3, q) for q
odd, 2G2(q), Sp(2n, 2) in both doubly transitive representations,
2), HS, and . 3.) Two-graphs were studied extensively by
V22nSp(2n,

D. Taylor and J. J. Seidel, and many others contributed to the theory.
For surveys, see [104], [105], [112].

Clearly a two-graph can be regarded as a map from the triples
on a set into Z2 with vanishing coboundary, that is, a 2-cocycle. Since
all relevant cohomology is trivial, it is thus the coboundary of a 1-cochain,
the latter being a function from pairs into Z2, that is, a graph; the
triples of the two-graph are those carrying an odd number of edges of the
graph. Two 1-cochains have the same coboundary if and only if they
differ by a 1-cocycle, that is the coboundary of a 0-cochain (a function
from the point set to Z2); this means there is a partition of the point
set as X U Y, and the two graphs agree within X and within Y and are
complementary between these sets. In Seidel's terminology, the graphs
are related by switching with respect to the partition X U Y. Thus a
two-graph is equivalent to a 'switching class' of graphs. Seidel uses
(0, -, +) adjacency matrices for graphs (these have 0 on the diagonal,
-1 for adjacency, and +1 for nonadjacency). The adjacency matrices
A, A' of switching-equivalent graphs are related by A' = DAD, where
D is a diagonal matrix with diagonal entries t1. The null two-graph
corresponds to the switching class of complete bipartite graphs, and the
complete two-graph to that of disjoint unions of two complete graphs.
These will be excluded in future.
If A is any set of triples on P and p is any point of P, let
o(p) be the graph with vertex set P - {p I and edge set
{{q, r I {p, q, r l e z ). It is easy to see that, if 0 is a two-graph,
I

then the switching class corresponding to 0 contains the disjoint union
of the isolated vertex p and o(p). (Take any graph in the class and

50

(Thus. A(p) is strongly regular with d = i a. Now suppose that a graph in the switching class of a regular two-graph is itself regular. 1). in the above construction. then it holds for every point. 1) Theorem. The precise answer is: (7. if A(p) is the Paley graph P(q) then A admits the doubly transitive group PSL(2. it is enough to prove that the graph is regular. 1). 2) Theorem. it yields a regular two-graph A. 51 . The proof that this graph is strongly regular can be simplified by using the following result of Seidel: (7. when is A an extension of A(p) in the design sense. Next. Note that the equation (A . depending on which of the two eigenvalues is associated with the all-1 vector j. and A has only two eigenvalues on j1.switch with respect to the neighbours of p. Then the all-1 vector j is an eigenvector of A. Using this and the rationality conditions. In the first step. A two-graph A is regular if and only if.p2I) = 0 is invariant under switching. q). 5)(ii) in chapter 2 can be interpreted in terms of two-graphs. that is. A(p) determines A uniquely. The construction of strongly regular graphs with the parameters of (2.) Suppose A(p) is regular. it is shown that the switching class corresponding to A contains a strongly regular graph. a strongly regular graph LZ(q+1)(q) with d = Za is constructed: by (7. We give two examples. -. so the graph is in fact strongly regular.p 1 I)(A . for some point p. when is A a 2-design? A two-graph which is also a 2-design is called a regular two-graph. These two possibilities may or may not occur. If this holds for one point. one sees that the para- meters of regular two-graphs are fairly restricted (though still general enough to include all the examples with doubly transitive automorphism groups mentioned earlier). +) adjacency matrix of some (and hence every) graph in its switching class has just two eigenvalues. Any strongly regular graph of Type I satis- fies the condition of (7. A two-graph is regular if and only if the (0. ) There are two possibilities for the parameters of a strongly regular graph in the switching class of a regular two-graph.

4) and (28. (Problem: Find a switching partition. The second parameter set is realised by T(8) and the three exceptional Chang graphs (see chapter 4). there is a unique regular two-graph with design parameters 2 . {7. for some vertex p of r. A strongly regular graph on 9 vertices with a = 4. 15. 3. r) E I = {qh1. {3. the Chang graphs are most easily defined by giving switching sets of vertices in T(8). 6). Indeed. {3. 1. r E T(p). 6). ) 2. the set of triples of lines for which the product of cosines is negative forms a two-graph. with any two at an angle with cosine ± a(a # 0). 9. 3).(28. the connection with equiangular lines is as follows. rh2) c r. Thus. for q e r(p). the theory of regular two-graphs also gives upper bounds for the number of lines. Given a set of lines through a point of Euclidean space.(10. 5). Let r be a graph with a vertex-transitive automor- phism group G. 15. 3). 17. r) % r= {qh1. 7). 15. The possible parameters of a strongly regular graph in its switching class are (28. 8). 6. 10). 0. c = 1. Suppose that. 3). 52 . Shult [108] proved the following result on transitive extensions: (7. 14. {7. 3. 16. 4). 5).1). 12. The first is impossible by (5. 4). they must be equivalent under switching. so there is a unique regular two-graph with design parameters 2 . all these graphs are switching- equivalent. 8). {q. 1. ) F. The appropriate sets are 11. 4). 2). 7). 18. there exist elements h1 E Aut(I jr(p)) and h2 E Aut(r J t (p)) such that. Since both the Petersen graph and its complement lie in the switching class of such a two-graph. 10. 10). 14. 16. 12. 15. {1. 8). {8. (See Lemmens and Seidel [77]. d = 2 is necessarily isomorphic to L2 (3). conversely any two-graph can be represented in this way. 3) Theorem. 6). It can be shown that there is a unique strongly regular graph with parameters (27. rh2) r and {q. thus. Then G has a transitive extension. 11. The 'best' known sets of equiangular lines (with most lines in space of a given dimension) often represent regular two-graphs. 4). 1). 2). 2). Briefly. 0. 5) (see [102]).

Again. Proof. all of whose triples belong to o.(12. 2). apart from D. If A is the two-graph for which a(p0) = r for some vertex p0. {pqs }. t + 1.1. however. little machinery for studying regular t-graphs has been developed. The blocks are the sets S with IS I = k. then it is easy to see that the permutation which interchanges pp and p and acts as h1 on r(p) and h2 on r(p) is an automorphism of A.(2t + 2. It seems that if 0 only (and not its complement) is assumed to satisfy (*). A two-graph could be defined as a set A of triples such that 0 and its complement both satisfy the following condition: (*) If {pqr }. Higman's generalisation of (7. one could define a t-graph to be a t-cocycle over Z2. 1) design is a regular t-graph. one could generalise by considering cocycles over other abelian groups. Let r be a graph on v .{ p } is a (k-1)- 53 . and any edge lies in at least one of them. Indeed it is trivial that any t . Suppose 0 is a set of triples satisfying (*). {prs } E A then {qrs) E A. but only two such designs are known. G. Also. but the structure of A is sufficiently restricted that some progress is possible. 6. x) design D. 4) Lemma. Then there is a 2 . k.(p) = r for any point p. and a regular t-graph to be a t-graph which is also a t-design. G.(v. any vertex lies in A of them. such that A is the set of triples which are contained in some block of D. the methods used for two-graphs fail. but here too examples are fairly scarce and general theory is lacking. these are the sets S for which I S I . and the latter a double extension of the Petersen graph with automorphism group M11 discovered by D.1 vertices in which the largest complete subgraphs (cliques) have size k . A useful remark is: (7. For t > 2. 1) design). // More generally. Higman) and a regular 5-graph on 12 points (the 5 . 3) to t-graphs [66]. examples seem to be much more limited. with . The only ones we know are regular 3-graphs on 8 and 12 points (the former is AG(3.

the second to point sets of 2 .(v. Let H be a permutation group of degree k inducing 54 . 5) Theorem.) The converse is also true. . // b = m+1 The only known design with k > 3 satisfying the axiom of (7. {13 1. 2) design and B a If 1) block of 0 such that any four points of B generate 016. 1) design.(16. then A is represented by a 2 . is a symmetric 2 . then m = 3 and 0 is the unique regular two-graph on 10 points. 2) designs. 7) Theorem.. Clearly two points lie in A such blocks. so r = 2m. and conversely. 6) Theorem. { lk } } and 1(12 } .(4. Note that r satisfies the hypothesis of (7. and uniqueness is easy.1 ver- tices. then the 'second blocks' through pq.Suppose T is a graph. and A a set of triples satis- fying (*) with A(p) = r for any point p. {23 }} . k. (ii) If r is a triangular graph T(k). A few applications follow. Call this design a) 16 for the moment. (iii) If r is a lattice graph L2 (m). 2) design.(4. The first extend to blocks. 2m(m2 + 1) whence m = 3.(v. 5) (ii) is a 2 . closely related to extensions of triangular graphs: (7. 4) if it admits a vertex. (i) If r is a disjoint union of complete graphs on k . 6. k. triangular graph T(k) has two kinds of cliques. It 12 1. (7. .) For (iii) we have v = m2 + 1. In (ii) note that for k > 4 the Proof.(v. qr and rp are concurrent. 3. r e B. 3. A = 2. A is a regular two-graph! Characterising it by this axiom seems a difficult problem. 2) design which we met in the last section. (The case k = 4 can be done by hand. for this design. (Thus p. r generates the 2 . The proof of (i) is trivial. and a triple is contained in a block if and only if it is in A. (7. We shall say the design represents the set of triples.clique in o(p) for some (and hence every) p E S. 2) design satisfying the following axiom: if B is a block and p. q.and edge- transitive automorphism group. {13 } . then 16' This has as a consequence a result on transitive extensions. viz. k. q. then 0 is represented by a symmetric 2 . We mention a weaker characterisation.. k = m + 1.

Let H be a rank 3 group of automorphisms of L2(k). and G a transitive extension of H. G=V16. Then k = 3. H = A5. k.S6. 55 . A similar result holds for lattice graphs: (7. 9) or PEL(2. Then one of the following holds: (i) k = 4. and G a transitive extension of this rank 3 group.A6 or V16. and G = PSL(2. 9) acts on the regular two-graph on 10 points. See Cameron [26]. 8) Theorem.(v. 7) (ii) k = 5. G = PSL(2. In each case G acts on a symmetric 2 . [27]. 11) (iii) k=6. 2) design.a rank 3 automorphism group of T(k). H=A6 or S6. H = S4. G = PSL(2.

1) Proposition. In place of edges. f } be a set of size 6. .factorisations of K6 In this chapter we explore a remarkable combinatorial property of the number 6. c. we shall speak of edges.. b. 1 56 . 1-factorisations. hence a hexagon. any 1-factorisation contains 5. Two disjoint 1-factors together form a union of circuits of even length. Let A = {a. of the complete graph on A. 8. and the 1-factorisation is uniquely determined (Fig.. Two disjoint 1-factors are contained in a unique 1- factorisation. We regard A as the vertex set of a complete graph K6. d. etc. This material is loosely based on lectures by G. 1-factors. 8 choices of two disjoint 1-factors. which can be exploited to give constructions and unique- ness proofs of the projective plane of order 4. 4 such pairs. and the 5 . of A. There are 15. 8. A 1-factor is a set of three mutually disjoint edges (covering A).8. Proof. 6.(12. any two isomorphic. A further 1-factor must include one or all of the long diagonals of the hexagon. each 1-factor must include one long diagonal. Since there are three long diagonals. and a 1-factorisation is a partition of the set of fifteen edges of A into five 1-factors. . the Moore graph of valency 7. e. (8.1. There are six different 1-factorisations of A. 1) Steiner system. 1). Higman. Fig.

e. It can be shown that a projective plane of order 4 necessarily contains a type II oval A. By (8. edges of X. Any edge of A defines a line. // (8. points. 57 . whereas a 1-factorisation contains five 1-factors. The projective plane of order 4 is unique (up to iso- morphism).1-factorisation. the five 1-factors defined by its points form a 1-factorisation. by (8. a unique edge of the 1. On edge of A lies in three 1-factors. or between 1-factors and 1 -factorisations.-factor contains the point. their labels are disjoint. 2 is a projective plane of order 4. edge . 3) Corollary. 1). Proof. The five 1-factors of X corresponding in this way to the five edges through a vertex of A form a 1-factorisation. // Let X be the set of 1-factorisations of A. Proof. Given a line disjoint from A. There are 21. (8. and so form a 1-factor of X. the blocks are the edges and 1-factorisations of A. Proof. the three lines on x meeting A define a 1-factor.1-factor. (ii) given a point and a 1-factor. Now consider the following structure 0: the points are the vertices (elements of A) and 1-factors of A. k = 5 holds since an edge contains two vertices and is contained in three 1-factors. is defined by inclusion. while incidence between edges and 1-factors is reverse inclusion. Given a point x outside A. Two blocks in a projective plane meet in a unique point. 3). Incidence between vertices and edges. 2) Proposition. // (8. A = 1 follows from a consideration of cases: (i) two points define a unique edge. Any two 1-factorisations share a unique 1-factor. 1-factors of A can be labelled with pairs of elements of X. i. Note that A is a type II oval in the plane Z. Thus we have a 'duality' between A and X: vertex . (iii) two 1-factors either share a unique edge or lie together in a unique 1-factorisation. Thus the given plane is identified with 2. 4)Corollary.

and X with the set of six 1-factorisations of A. x) ). (Of course. y) are joined to (c. then { (a. It remains to describe the edges within A X X.xy if this occurs. b. given a.ab similarly. y)) and { (a.(This can also be verified by use of the dual design of 0. (ii) the nonexistence of M(4): for if A = {a. x) and (b. y}. y. (b. y). Any further vertex is joined to a unique a E A and to a unique x E X. x). We immediately have (i) the uniqueness of M(3): for if A = {a. Then A and X are disjoint. x). (8. ) Now the construction of M(7) is as follows. b E A. then both (a. x).xy. x). This definition is self-dual (i. there is a unique y E X such that (a. We turn now to the construction of the Hoffman-Singleton graph. (a. (a.xy if ab is an edge in the I-factor with label xy. b. z). e. x) E A x X. whenever { (a. (b. 0. then b c and y * z.) Define ab . The uniqueness proof for M(7) depends on the following lemma. x) is joined to (b. we have xy . 4). z } . Then it contains the fifteenth edge also. y). y). (b. 5) Lemma. Let { -. With the above notation. this ensures that. and if I (a. this is also a consequence of (2. 0 } be any edge. y) I whenever ab .xy). (b.ab if and only if ab . a. (b. A the set of vertices (other than 0) joined to -. y. (b. y)) is an edge. (b. defining xy . y). and so can be labelled with the ordered pair (a. Suppose a set of ten vertices of M(7) contains fourteen of the fifteen edges of M(3). (b. z). Suppose first that r = M(n + 1) is a Moore graph of valency n + 1 and diameter 2. x). It is straightforward to check that this relation arises from the identifica- 58 . and X the set of vertices (other than -) joined to 0. whose proof we omit. x) ) are edges. For any such edge { (a. y). x. y) and (c. Identify A with our original set of six vertices. moreover. Thus. y) I is an edge. yielding a triangle. (b. X E X. (Consider the ten vertices -. x). c) and X x. Introduce edges { (a. x) is joined to (b. ) It is convenient to define a relation between edges of A and edges of X as follows: ab . Considera- tion of a number of cases shows that this graph is strongly regular. x). y) 1. b } and X = {x. as is I (a. if (a. we must have a * b and x * y.

Its uniqueness can be proved in a similar way. c I U { x.(12.(24. yz. b) whenever ab.-xy. and define a graph t on 12 with edges {{ B. (Note that two blocks of the 5 . X. Let 62 be the set of 66 pairs of disjoint blocks. 6.(12. and blocks BU Ix. r = T(12). 1) on {a. 8.(12. we have a 'duality' between A* and X*. a little more notation. y) U {a. Choose three elements x. y.xyzw. 1) design. b. It is possible to define a 5 . 6. Y' } . zx of A are pairwise disjoint but not contained in a I -factorisa- tion.(12. { a. y. The three 1-factors xy. This relation is also self-dual. Write abc . z E X. First. where IX* I = 12.xyz if abc is one of these triangles. 3 or 4 points. 1) design. z I whenever abc .(12. 6. B') '-' { x.xyz. z. 6. b. c. b I { a. 6. y}. X\{x. b I where { Y. and it is easily checked that any five points lie in at least one block. b. with the same parameters as T(12). We turn now to the 5 . Let A* be the point set of the 5 . Thus.tion of X with the set of 7 -factorisations of A as previously defined. y. c. The complement of any block in this design is a block. that is.) Counting arguments show r is strongly regular. C' ) } whenever I B A C I = 3. A\{a. { C. for a certain relation on the quadruples of points of A* and X*. w I whenever abcd . Now take the incidence structure with point set A U X and blocks of the following types: A. f can be identified with the set of 2-element subsets of X*. y ) Y U { a. d I u { x. So we have the required design.-. There are 132 = (5)/(5) blocks.). Now define a design with point set A* u X*. By (4. b) u {x. d I correspond to four 59 . y } where { B. 1) design. 1. 1) design meet in 2. It is possible to modify this construction to obtain the 5 . { a. so the remaining six edges must form the union of two triangles. (The four blocks of the 5 . 1) design on the point set X* in such a way that a repetition of this construction leads back to A*. B' 1.

y) is a cycle of the permutation 0(a. 1) design. where S6(A) is the symmetric group on A. 8. It is known that 6 is the only cardinal number n for which Sn possesses non-inner automorphisms. We conclude with a group-theoretic interpretation of the duality between A and X. Our constructions show that S6 is a subgroup of the automor- phism groups of each of the three constructed objects. and that an automorphism pre- serving the conjugacy class of point stabilisers is an inner automorphism. the duality between A* and X* yields the existence of an outer automorphism of the Mathieu group M12. b). since the stabiliser of a point and the automorphism group of a I -factorisation are not conjugate in S6(A). So IAut S6I = 2IS6I. y. Identifying A and X by means of any bijection.disjoint 2-subsets of X*. Since M24 has no outer automorphisms. w). 1) design.S6(X).(24. furthermore. These facts enable some simplifications to be made in the existence and uniqueness proofs. Thus we have an isomorphism 0 : S6(A) . it can be shown that any subgroup of index 6 in S6 is of one of these types. 6. the uniqueness proofs enable easy calculation of the orders of their full automorphism groups.xy if and only if (x. This is not an inner automorphism. etc. the process cannot be continued! 60 . z. However. and that abc . ) This yields a 5 . It can be shown that ab . the auto- morphism group of the (unique) 5 . Any permutation of A induces naturally a permuta- tion of the 1-factorisations of A.(12.xyz if and only if (xyz) or (xzy) is a cycle of 0(abc). this gives an automor- phism 8 of S6(A). whose complement is {x. In similar fashion.

and calls these n-tuples words and n the word length. 2) Definition. where t < e (i. Consider a subset C of (R (n) with the property that any two dis- tinct words of C have distance at least 2e + 1. (As usual. If we take any x in C and change t coordinates. y) P }. 1) For x E 61(n) and y e (R(n) we denote by d(x. e. In (R (n) we introduce a distance function d (called Hamming distance) which is the natural distance function to use when one is inter- ested in the number of errors in a word that is spelled incorrectly.9 Codes In coding theory one considers a set F of q distinct symbols which is called the alphabet. 0 denotes the zero-vector in (R(n). For x c (R (n) we define the weight w(x) of x by w(x) := d(x. (9. i. In practice q is generally 2 and F = (0. If F = GF(q) we shall denote the set of all words by (R (n) and interpret this as n-dimensional vector space over GF(q). y) the Definition. e. we make t errors). 1 ). In most of the theory one takes q = pr (p prime) and F = GF(q). (9.(n) we define the sphere of radius p with centre at x by S(x. P) := {y E GR(n) I d(x. . then the resulting word still resembles the original word more than any of the 61 . number of coordinate places in which x and y differ. 0). Using the symbols of F one forms all n-tuples. (9. The following two definitions are directly connected with d and the language of metric spaces. Fn. 3) For p > 0 and x e Definition.

it has a smaller distance to x than to other words).others (i. y) = d(x-y.correcting code C is a subset of (R(n) with the property VXECVyEC[x *y=t. d(x. VXEcVYEc[x * S(x. e) _ (R. g. coding.error. Proof. For an arbitrary code consisting of K code words there are (K) comparisons to make to find this minimum distance.(n) X EC One of the more interesting types of codes is the linear code: (9. The error-correcting capacity of a code is determined by the minimum distance between all pairs of distinct code words. (9. we can correct the t errors. decoding. The purpose of coding theory is to study such e-error-correcting codes (e. There- fore. If X E C. 4) Definition. i. 6) A k-dimensional linear subspace C of Definition. k)-code over GF(q). y E C. 5) Definition. 0) _ w(x-y). 7)Theorem. A special type of code which is not too important in practice but of great interest to combinatorialists and group theorists (because of inter- esting automorphism groups) is a perfect code: (n) (9. An e-error-correcting code C in (R is called perfect if U S(x. An e. In a linear code the minimum distance is equal to the minimal weight among all non-zero code words. then x . e.y E C and d(x. e) n S(y. R(n) is called a linear code or (n. implementation). e) = gl] . e. y) ? 2e + 1]. For linear codes we have the following advantage: (9. if we know C. In these lectures we shall be interested in certain connections to combinatorial designs. // 62 .

y) denotes the normal inner product) is called the dual code of C. 10) C = {x E R(n) IxHT = 0 ) . 8) an (n. If H is a generator matrix for Cl then H is called a parity check matrix for C. We shall now look at two ways of describing a linear code C. Therefore a code C is defined by such a parity check matrix H as follows: (9. The first is given by a generator matrix G which has as its rows a set of basis vectors of the linear subspace C. 9) Definition. A code C is called systematic if there is a k-subset of the coordinate places such that to each possible k-tuple of entries in these positions there corresponds exactly one code word. g. With this in mind we see that for every linear code there is an equivalent code which has a generator matrix of the form (9. n-k)-code.k matrix. If C is a linear code of dimension k then (n) C1 := {x E (R IVVEC[(x. Note that by (7. 63 . In general a parity check for the code C is a vector x which is orthogonal to all code words of C and we shall call any matrix H a parity check matrix if the rows of H generate the subspace C'. The code C1 is an (n. 8) G = (Ik P) where Ik is the k by k identity matrix and P is a k by n . y) = 0] } (where (x. Since the property we are most interested in is possible error-correction and this property is not changed if in all code words we interchange two symbols (e. We call this the standard form of G. the first and second letter of each code word) we shall call two codes equivalent if one can be obtained by applying a fixed permutation of symbols to the words of the other code. (9. If one is not interested in linearity one can extend the concept of equivalence by also allowing a permutation of the symbols of the alphabet. We now come to the second description of a linear code C. k)-code is systematic.

k)-code. then obviously the minimum distance of C. xEC Let C be an (n. then the extension is trivial because co = 0 for all words of C. If x E C then I S(x. this code has dimension n . C. Proof. It is called the (n. Hamming codes are perfect codes. This is called an overall parity check.1)). is d + 1. e. Form the matrix H (m rows. In this way we obtain a new linear code C which is called the extended code corresponding to C. (9. Clearly. Remark. The reader is warned that often two codes. q) and along each of these choose a vector xi (i = 1.1). 64 . n-m)-Hamming code over GF(q). c2.m. . However. Let C be a (n.... + cn = 0. Let us consider an important example of linear codes. 11) Theorem. cn) of C we add an extra letter c 0 (say in front) such that c 0 + c 1 +.1)/(q . n :_ (qm . . If H is a parity check matrix for C then H= H J is a parity check matrix for Of course if the all-one vector (which we denote by 1) is a parity check vector for H. To every word (c1.. i. Then n = (qm . Consider the lines through the origin in AG(m. 10)). Since no two columns of H are linearly dependent we see that H is a parity check matrix of a code C in which every non-zero code word x has w(x) > 3 (by (9.. n-m)-Hamming code over GF(q). n columns) with the xi as column vectors.1) = qm. 1) = qn.. C* are called duals if C* is equivalent to C This often leads to con- fusion but it also has certain advantages. 1) I = 1 + n(q .1)/(q . . The previous remark shows that this is a single-error-correcting code. 2.. C is perfect. Since IC I = qn-m and C is single-error-correcting we find that U S(x. if C is a binary code with an odd minimum distance d.

For our first connection to the theory of designs we consider as
an example the extended (8, 4)-binary-Hamming code. We have seen
above that

1 1 1 1 1 1 1 1
10 0 0 0 1 1 1 1
H:= 0 0 1 1 0 0 1 1
0 1 0 1 0 1 0 1

is a parity check matrix for this code. It is easily seen that the code is
equivalent to a code with generator G = (14 J4 - I4). If we make a list
of the 16 words of this code we find 0, 1 and 14 words of weight 4. Now,
since two words of weight 4 have distance ? 4, they have at most two
1's in common. It follows that no word of weight 3 is a 'subword' of
8
more than one code word. There are (3) = 56 words of weight 3 and
each code word of weight 4 has four subwords of weight 3. We have
therefore proved:

(9. 12) Theorem. The 14 words of weight 4 in the extended (8, 4)-binary
Hamming code form a 3-design with parameters v = 8, k = 4, A = 1.
We leave it as an exercise for the reader to check that this design
is the extension of PG(2, 2), i. e. AG(3, 2).
Besides as an example, the discussion above has served a second
purpose, namely to show that it can be useful to know how many words of
some fixed weight i are contained in a code C. A simple way to des-
cribe such information is given in the following definition.

(9. 13) Definition. Let C be a code of length n and let A.
(i = 0, 1, ... , n) denote the number of code words of weight i. Then
n
A(, 77) := E Aid i n-i 77
i=0
is called the weight enumerator of C.
The weight enumerators of a code C and its dual C1 are related.
The relation is due to F. J. MacWilliams.

65

(9. 14) Theorem. 77) be the weight enumerator of an (n, k)-
Let
code over GF(q) and let A( , r)) be the weight enumerator of the dual
code. Then

71) = q-kA(71 - , 71 + (q - 1) ).

The proof of this theorem depends on the following lemma. (We
write (R instead of cR(n). )

(9. 15) Lemma. Let x be a non-trivial character on the group
(GF(q), +) and for any v E CR define Xv : (R - CX by

VUE(R[x (u) := X((u, v))].

If A is a vector space over C, f : CR - A, and if g : (R -A is defined
by

VuE(R[g(u) := I f(v)Xv(u)],
VE(R

then for any linear subspace V C (R and the dual subspace V1 we have

E g(u) = I v I; f (v).
UEV VEV1

Proof. Z g(u) = I I f(v)xV(u) _
uEV uEV vEC

I f(v) I X((u, v)) _
vE(R uEV

_IVI Z f (v) + f (v) I x ((U' 0.
VEV1 v%V1 uEV

In the inner sum of the second term (u, v) takes on every value in
GF(q) the same number of times and since I x(a) = 0 for every
a EGF(q)
non-trivial character, this proves the lemma. //

In Lemma (9. 15) let A be the space of all poly-
Proof of (9. 14).
nomials in 2 variables , 71 with coefficients in C and let
f(v) :_ w(v)77n-w(v). If a E GF(q) write w(a) := 1 if a * 0, w(0) := 0.

66

Then we have

g(u) = E .. I
vl EGF(q) vnEGF(q)
w(v )+, , ,+w(v ) (1-w(v ))+, , ,+(1-w(v n))
1 n 1

n
711-w(v)
= II { 1 w(v) X(uiv)
i=1 vEGF(q)
Since the inner sum is 71 + (q - 1) 4 if ui = 0 and
r+ ( F X(a)) = rl - if ui # 0, we find
a EGF(q) \ { 0 }
)n-w(U)
g(u) = (17 + (q - 1) (n - )w(u).

Now by the result of Lemma (9. 15) we have (taking V = C):

1, f (v) = 71) = q-k I g(u)
vECI- uEC

= q-kA(n] - , 77 + (q - 1) ). //

In Chapter 13 we shall show how this equation played a role in
recent investigations on the existence of a projective plane of order 10.
For the sake of completeness we mention that generalisations of
Theorem (9. 14) to nonlinear codes are known (cf. [39], [87]).
We give one example of a nonlinear code constructed by using
design theory. It is not difficult to show that a binary code of length 10
with minimum distance 5 can have at most 12 code words. To construct
a code meeting this bound we take the rows of the incidence matrix of
the 2 - (11, 5, 2) Paley design and add the all-one word 1 to this set.
These twelve binary words then clearly have mutual distance 6. By
deleting a fixed coordinate we obtain a code with the required properties.
The reader interested in nonlinear codes is referred to [86].

67

.... From now on we make the restriction (n. k)-code C over GF(q) is called cyclic if V(co... . r 0 0 . 0 0 0 (10. We shall call g(x) the generator (-polynomial) of the cyclic code.(n)) is generated by a polynomial g(x) which divides xn . .1. c1.. . .. cl.. an-2). +hdx d... co.a0 + a x + . e. . The words g(x). i..1). a1. a0. . gn-d 0 .. . .. gn-d 0 g g1 . ... . 0 go .(n). An (n.2) G := 0 0.1. .' an-1) ~ (an-1.. .. . . cn-2) E C]. q) = 1. . + since it is obvious an-1xn-1 1 that the polynomials of degree < n form a set of representatives for R mod S.1) Definition. . . . Every ideal in R mod S (i. xd-1g(x) form a basis of the code C.. From now on we do not distinguish between words and poly- nomials of degree < n (mod xn . .. .. . xg(x).. every cyclic code in . Clearly the residue class ring R mod S (considered as an additive group) is isomorphic to . Let R be the ring of all polynomials with coefficients in GF(q) and let S be the ideal generated by xn .10 Cyclic codes Many of the most interesting known codes are cyclic. . an-1) t. +gn. . 0 go g1 g2 . . . Let g(x) = go + glx + . .dx n-d be the generator of a cyclic code C and let h(x) := (xn . a1. .. We define these as follows (10.1)/g(x) = ho + h1x + . e. . . An isomorphism is given by (a o. .. gn-d L J 68 . . . . From this it follows that a cyclic code C corresponds to an ideal (which we also denote by C) in R mod S. Note that multiplication by x in R mod S amounts to the cyclic shift (ao' a1.. ..

Let q = 2. .. ... The code generated by h(x) is the dual code of C. Construct a matrix H for which the j-th column m-1 is (c O. 2m-2). h1 ha hd . 2. Very often cyclic codes are described. not by giving the generator g(x). Clearly this is an equivalent description be- cause the requirement that all code words are multiples of g(x) means that they are 0 in the zeros of g(x). 1. .... ho 0 (10. .... an-1)..1 and let a be a primitive element of (GF(2m).. 0 hd hd-1 . em-1)T if al = Z 6ia1 (j = 0. a(x) := a 0 + a1 x+. Let m1(x) _ (x-a)(x-a2). but by prescribing certain zeros of all code words (in a suitable extension field of GF(q)). .. Since the columns of H are a permutation of the binary representations of 1. but with the symbols in reversed order (see remark following 9.... . +a n-1xn-1 then the vector allT corresponds to the field element a(a).. We call h(x) the check polynomial of the cyclic code C.. 10). n = 2m .. Therefore aHT = 0 means the same thing as m1 (x) I a(x). El.. To show how simple this description can be we give an example. .3) H:= 0 ha ho 0. Note that the dimension of the cyclic code is equal to the degree of the check polynomial. .. Since g(x)h(x) = 0 in the ring R mod S we see that 0 0 ... (x-a2m ) be the minimal poly- nomial of a. 2m-2)... a1. . If i=0 a := (a0. m-1 Every element of GF(2m) can be expressed uniquely as eia' where i=0 ei E GF(2).. 1... We wish to consider the cyclic code generated by m1(x). . 2m-1 we have proved: 69 . This gives a translation into the termi- nology of Chapter 9... The non-zero elements of GF(2m) are the powers a] (j = 0.. 0 is a parity check matrix for C..is a generator matrix for C..

an-1). Example. then n = 2m . It is equivalent to the (7.1 points (as we saw above for m = 3).. [4])? One easily sees that the binary Hamming codes as defined in chapter 9 have the property that the words of weight 3 represent a Steiner triple system on 2m .. PG(2.1. 4) Theorem. where 70 . in GF(q).1 and C is the Hamming code and the Steiner system is PG(m . 3). In our derivation of Theorem (10. This example is connected to an interesting question linking design theory with coding theory.I for which the generator is the minimal polynomial of a primitive element of GF(2m) is equivalent to the (n. 0. The binary cyclic code of length n = 2m . Suppose a Steiner triple system S (like e. In this case the cyclic code C generated by g(x) has dimension 4. 6) linear code over GF(q) such that the nonzero positions of code words of weight 3 represent AG(2. Take q = 2. g(x) = 1 + x + x3. We refer to [48] for a proof that the two nonisomorphic Steiner triple systems on 13 points do not arise from linear codes in this way. Let q = 1 (mod 3) and let a be a primitive cube root of 1. 4) we have demonstrated that the requirement that all the code words have a as a zero corresponds to prescribing the parity check matrix H :_ (1 a a2 . Define 0 0 0 -1 -a -a 2 1 1 1 H= 1 1 1 0 0 0 -1 -a -a2 1 -a -a2 1 1 1 0 0 0 Then H is the parity check matrix of a (9. n-m) binary Hamming code. Is there a linear code C over some field GF(q) such that for ecery code word of weight 3 its nonzero posi- tions form a block of S and all blocks of S are of this form (cf. 2) for this example form the incidence matrix of PG(2. [48]). g. 4) Hamming code which corresponds to our example following (9. 2) in our example) is given. 2).(10. 2). n = 7. In fact if a Steiner triple system on n points corresponds to the set of words of weight 3 of a binary linear code C. 13 points the problem has been solved (cf. We remark that the 7 cyclic shifts of the first row of (10. 11). For the Steiner triple systems on 9 resp. and the 8 words obtained by interchanging 0 and 1. The code consists of these 7 words.

.. with the restriction that no factor is taken more than once.we have to interpret al as (e ei ' ' ' sm -1) T if al = nZ-1 c1ai. . The determinant of the submatrix of H obtained by taking the 71 .. The minimum distance of a BCH-code of designed distance d is at least d.. D... If c had weight < d.... d-1. .... 2. They were discovered by R.. Let 0 be a primitive n-th root of unity in GF(gm).. We use the abbreviated notation introduced above and define the m(d . The requirement that 1 is a zero of all the code words means that they all have even weight. ad-1 as zeros is called the BCH-code of designed distance d. the code is called primitive.. 5) Definition..1). i. In terms of parity check matrices we can say that the all-one row has been added as a parity check.... . n-m) Hamming code. (32.... then there would be d . C. n-m-1)-code consisting of all the words of even weight in the (n. . K. 6) Theorem... e.... 1 /3d-1 02(d-1) 13(n-1)(d-1) A word c is in the BCH-code iff cHT = 0. Bose. The terminology 'designed distance' is based on the following theorem: (10.... i=0 A second useful example is obtained by replacing m1 (x) by (x + 1)m1(x). We now introduce a very important class of codes known as BCH- codes. 5) is the product of the minimal polynomials mi(x) of 0i where i = 1.1) by n matrix H over GF(q) by 1 1 (3 a2 R2 34 . The generator g(x) of the code defined in (10..1 columns of H which are linearly dependent over GF(q). If R is a primitive element of GF(qm) (n = qm . 13n-1 132n-2 H:= .. (10. Hocquenghem. The cyclic code in (R (n) consisting of all the words which have /3.. we now have the (n. Ray-Chaudhuri and A. Proof.. Clearly the new code is a subcode of the Hamming code.

cam). d-1 .. This group. Therefore Pa 0 transforms every extended cyclic code into itself. We shall now show that the extended code is invariant under the permutations of the affine permutation group on GF(qm) applied to the positions of the code... We represent the code words as cx00 co + clx + . We extend the code by adding an overall parity check. n-1) where Xi := a1. We now denote the positions of the sym- bols in the code words by Xi (i = 0. The proof is the same as above.. 5) we also require 1 to be a zero of all the code words then the code has minimum distance at least d + 1. Remark. . Now let (c0.. Xn-1 and that it leaves X. qm).II 0. Let us now take a look at BCH-codes from the point of view of the group theorist. X1.. . 1 i]j d-1 d-1 d-1 1 '2 ' ' ' d-1 This shows that all code words c * 0 indeed have weight ? d. We consider a primitive BCH-code of length n = qm . and make the conventions 10 := 1.. :_ (32 := 0 d-1 is 1 2 ' d-1 2 2 2 2 _ 2 . . consists of the permutations Pu v(X) := uX + v (u E GF(qm). c1..columns headed by :_ (31... v E GF(qm). a2 ad-1 (where a is a primitive element in GF(qm)). u * 0). 1. invariant.) be any code word in the extended BCH-code and apply Pu v to the positions to obtain (co. d-1 we have 72 ... .... We first observe that Pa 0 is the cyclic shift on the positions X0. Then for k = 0.. . . 1.1 over GF(q) with prescribed zeros a. It is a doubly transitive group. c. ci. .. We denote the addition- al position by X. If in (10. .. + cn-lxn-1 + c . (al)00 := 0 for i 0 (mod n). which is denoted by AGL(1.

8) holds for any binary code for which the extended code is invariant under a transitive permutation group. 73 . Every extended primitive BCH-code of length n + 1 = qm over GF(q) is invariant under the affine permutation group on GF(qm). i. We have proved: (10. Remark.. We have i= O A2i = a2i-1 + a2i and A2i-1 = 0. 1. Notice that the result of Theorem (10.. The total weight of the code words of weight 2i in C is Since the extended code is invariant under 2iA2i. Therefore the per- muted word is also in the extended code. cialk = (Z )ul ail v -Z = ci(ual + v)k = l ci E l=0 k I (L)uly -l I ci(al)i = 0 1=0 i because the inner sum is 0 for Z = 0. a transitive permutation group the weight is equally divided among all the positions. . The minimum weight of a primitive binary BCH-code is odd. As an application consider a primitive binary BCH-code C of n i77 n-i length n.+ 1)a2i-1 = 2iA2i A consequence is: (10. (n. 7) Theorem. In C there are a2i-1 words of weight 2i with a 1 in position X. be the weight enumerator of C and let n+1 i=0 1 F Al be the weight enumerator of C. d-1. 8) Theorem. Let E a ...e.

followed by a cyclic shift of the received word etc. . if there are t < [ r +2xA . to correct position i. To show how decoding works let i = 1 and let a be a word with t [2 r] errors.. such that (i) Yiv) = 1 (v = 1. In the case of GF(2) the error can immediately be corrected. r). The procedure is applied to every position of a code. . y(r).1).1] errors in a. Then. (a. . Here we see one advantage of cyclic codes because the procedure has to be applied to one fixed position only.11. we can use the set of parity checks.(t . 2. (11..Threshold decoding There are several codes which allow a very simple error correc- tion procedure of which the simplest example is majority decoding. say y(1 .. Take an orthogonal check set y(l). It is not difficult to generalise (11. Of course the reader has seen the connection to combinatorial designs.a then position i is in error. 1.X(t . We consider a linear code C. v (ii) if j # i then yj : 0 for at most one value of v.) Definition. In the same way as above it is easily checked that if the number of parity checks which do not yield 0 exceeds the threshold r . Let A be the incidence matrix of a BIBD with parameters 74 .1) . Suppose that in (ii) we re- place 'at most one' by 'at most V.(t-1) values of v if al is in error. Y 2). Since r . Then t values of v if the symbol a is correct.1) t. y(r) on position 1. now called a A-check set of order r. An orthogonal check set of order r on a position i of C is a set of r words of Cl. . the majority of the values of (a.. y(" )) (zero or not zero) decides whether al is correct or not. Y(v)) # 0 for 1 r .

then C consists of 0 only which is hardly interesting! We shall return to this problem in chapter 13. Here we encounter a problem in coding theory which has been solved for a few cases only. in reversed order.(b. v. 1 1 0 1 0 0 0 0 1 1 0 1 0 0 0 0 1 1 0 1 0 A:= 0 0 0 1 1 0 1 1 0 0 0 1 1 0 0 1 0 0 0 1 1 1 0 1 0 0 0 1 In the example following (10. k. Instead of using a number of parity checks to check one position we can use the same procedure to determine whether a certain subset of the positions of a word contains an error. For every position of this (cyclic!) code there are 3 rows of A which form an orthogonal check set of order 3 on that position (correcting 1 error). We consider the rows of A as basis vectors of a linear code C over GF(q). is generated by (1 + x)(1 + x2 + x3). Once this has been decided it may be possible to use other parity checks (in the same fashion) to localise the error. [63]). Of course. We refer the reader who is interested in learning more about the topic of this chapter to Goethals [54] and Massey [90]. We consider one example now with q = 2. A). This is called multistep majority decoding (see Chapter 12). Then the dual C has the rows of Cl as parity checks and for every position i of C we can find a A-check set of order r on position i (by the definition of BIBD). there is a much simpler error correction procedure for Hamming codes but this example serves as an introduction to the discussion of codes generated by projective planes. From our second example in Chapter l0we know that the code C consists of the words of even weight in the (7. 4) we learned that the first 4 rows of A generate the cyclic code C1 over GF(2) which has g(x) = 1 + x + x 3 as generator polynomial. namely the problem of determining the dimension of C in the situation described above. 4) Hamming code. If this rank is v.) 75 . r. (Also see chapter 13. A further generalisation should be mentioned here. 2). [47]. Let A be the incidence matrix of PG(2. This implies that C. This amounts to finding the rank of A over GF(q) (cf.

I an-1) and b = (b0. m If j = z i. Reed. 2). n := 2m columns). 2). I an-1bn-1). v7. 3) Definition... and v0 := 1. (12. (Here the rows of F are considered as words in 61(n). (12. e. n A. . 1. If a = (a0... . If the indices 11 12 i1. We consider AG(m. 3) the product vi11 v12 . a1b1. a1. We make a list of all the points of AG(m.. . 2) I ij = 1 }. .' n-1) then ab := (a0b0.. := lxj e AG(m. n A. 2). To be able to introduce them we first need a lot of notation. The order of the columns is given as follows: column j is the binary repre- sentation of the integer j (j = 0. . 2) Definition. 2. b1) . . The standard basis vectors will be called u1... . For i = 1 . A. Note that A i is an (m-1)-dimensional affine subspace (an (m-1)- flat). Note that by (12.. the characteristic function of Ai. .. 2m-1). K is the characteristic function of A.. m vi := i of F..12 Finite geometries and codes The first class of codes connected to finite geometries which we shall discuss was first treated by D.1) Definition.. 2) as columns of a matrix F (m rows. (12. . then i=1 m the j-th column of F is xj ui.. ... um (considered as column vectors). F..lk. We need the following definition: i=1 (12. the characteristic function of AG(m.2 i-1 (a notation we use throughout this section). n . S. 1). and (12. Muller and I. ik are distinct this is an (m-k)-dimensional affine subspace of 76 .. i.

Then v0=(1111111111111111) v1=(0101010101010101) V2(0011001100110011) v3=(0000111100001111) v4=(0000000011111111) v1v2=(0001000100010001) v1V3=(00000101 00000101) V1V4=(0000000001010101) v2v3=(0000001100000011) v2v4=(0000000000110011) v3v4=(0000000000001111) 77 . . i. (12.. i=1 Therefore each of the 2m basisvectors of (R (n) can be written as a polynomial of degree < m in v1. vm. . + (m) In = 2m vectors.. . . vk) = 2m-k. V1. V2. 6) e. The vectors v0 . 0. v v . v1 . .. = H {vi + (1 + ij)v0 1. v v ... . i=1 The standard basis vector ej = (0. i.... . If S C 11. We have m (12. e. Therefore (R(n) is generated by the set {v0.. n = 16. we have proved: (12.. 2). This set contains 1 + (m 1 + . Vm 1. 2m-1 1 is defined by m ij21-1 C(S) {j Ii g ij = 01. . . .. 1. w(V1v2 .... m = 4..... . . .4) Theorem. . . vl. v m 1 2 1 2 form a basis of (R(n) Example. 0. V1V2 .. 0) e(R(n)I (the 1 is in position j).AG(m. 7) Theorem. 1. Therefore we have (12.. .. . 5) Definition. v m... e. it is a basis of (R(n). is the characteristic function of {xj 1. Vm.. 0... V2.. 2. m 1 then the subset C(S) of { 0... 0.

Then ab is a basis vector of the (m-1)-st order RM code and by (12. Con- sider all polynomials of degree r in m variables x1. Many authors prefer the following description of RM codes.. The dual of the r-th order RM code of length 2m is the (m-r-1)-st order RM code of length 2m (12. v1.. 2). Since the dimension of the (m-r-1)-st order RM code is 1 + (m) + . v. The dimension of the r-th order RM code is 1 + (m) + . .. 4) it therefore has even weight. Consider a fixed order of the points in AG(m. 2) is a codeword of C. (n) The linear subspace of (R which has as basis v0. 11) Theorem. . b) = 0.9) Theorem. The following theorem establishes a connection with the previous paragraph. v1 v2v3=(0000000100000001) v1v2v4=(0000000000010001) V1v3v4=(0000000000000101) v2v3v4=(0000000000000011) v1v2v3v4=(0000000000000001) (12... x2. . + ( m rl 1) = r1+1 + . ..l be a basis vector of the (m-r-1)-st order RM J1 code.. v ik be a basis vector of the r-th order RM code and let b = v.. The (m-2)-nd order RM code of length n = 2m is the (n. vnl and all products vi . + (m) we have proved: (12. Let a= v1. (12. 1 .. Let C be the (m-l)-th order RM code of length 2m. n-m-1) extended Hamming code.. 8) Definition. 78 . 10) Corollary.. Then the characteristic function of any 1-dimensional affine subspace of AG(m. xm which can take values in GF(2). This implies that (a.. The 0-th order RM code has vo as a basis. with k r is called the 1 r-th order Reed-Muller code (RM-code) of length n = 2m. + (m).. . v l. It is a repetition code of length 2m.... The sequence of values of a polynomial in these points is a code word of length 2m in the RM code of order r.

. 9) and (12.. ik) jEC(il. Suppose f = . 2).. These (t+l)-flats form a BIBD with parameters: 2m b _ 2 m-t-1 (2m-1)(2m-1-1)... (2m-t-1) r . 11 f. 2) and let f E 6R(n) be n-1 the characteristic function of A. e. There are 2m-t . Let A be an 1-flat in AG(m. Now I f. ik) 12 lk because by (12..12)f = I I { I f. ik) }. .. We now know that every (t+l)-flat in AG(m.i. i.. (2m-t-1) k _ 2 t+1 v = ' (2t+1-1)(2t-1). . } v .=0 if k> m-l. .)J A which are also in the k-dimensional linear subspace L:={x... using (12. By the procedure described in Chapter 11 we can let a majority vote provide a parity check on the positions of T if there are 2m-t-1 less than errors.. i2.. I 0 fjej. 11). . is the number of points of jEC(11. ik}. (2m-t-1) _ (2m-1-1).. . i2.. v. . Then m (12.. 6) vi . i2.. 2) j jEC(i1.. i2. If k> m-i then L nA is empty or an affine subspace of dimension > 0. In both cases L A A has an even number of points.. Proof. This proves that f is in the (m-1)-th jEC(i1. 13) Theorem..(2t+1-1)(2t-1).. The minimum weight in the t-th order RM code of 2m-t length 2m is 79 . Let C be the t-th order RM code of length 2m.. . Since the minimum weight in the t-th order RM code is even we have... .. occurs in the expansion of ej only if ij = 0 1 for all i %`{il. v.. v. ik) order RM code. Every point outside T is in exactly one of these (t + 1) flats. // Now we combine theorems (12. By induction one sees that in t + 1 steps of 2m-t-1 majority decoding up to . (12. 2) provides us with a parity check equation. 4). (2-1) (2t-1)..1 distinct (t+l)-flats containing T.. (2-1) (2m-1)(2m-1-1).1 errors can be corrected. 11 .EAG(m. . k=0 (il. . (2-1) ' Consider any t-flat T in AG(m....

2). Now we take another look at groups.2] with w(j) < m . a triply transitive group of order m-1 2II m (2m . By Theorem (12. By definition every code word in the r-th order RM code of length 2m is a sum of characteristic functions of affine subspaces of dimension ? m . We do not prove this theorem here. 6) and (12. then Zs is equivalent to the r-th order RM code. It follows that every permutation of AG(m. the group of affine transformations of AG(m. Fvery RM code of length 2m is invariant under all affine transformations (acting on the positions interpreted as points in AG(m. 2)). 2m . For the sake of com- pleteness we must mention that it was discovered later that RM codes are extended cyclic codes! To give the cyclic definition we introduce the notation w(j) for the weight of the binary representation of j (interpreted as vector over GF(2)). 13) can easily be derived from (10. i i= O (12. Remark.r.r. 15).r is a code word in the r-th order RM code. (12. One can show that RM codes are equivalent to extended cyclic codes by considering AG(m. The proof is not difficult.15) Theorem. 14) Theorem. This is AGL(m. 13) can be proved directly by induction on m and t. Let a be a primitive element in GF(2m).1 generated by g(x). 15) is that it can be generalised to other fields GF(q).2). 11) a sum of characteristic functions of affine subspaces of dimension m . Since the original definition of RM codes is more combinatorial we have chosen this way to present them here. The advantage of the definition (12. Remark. If C is the cyclic code of length 2m . 2) as the field GF(2m) and then looking at the mapping xh ax where a is a primitive element of 80 . Let g(x) := ll*(x . We remark that (12. 2) which also permutes the affine subspaces leaves all RM codes invariant. The reader interested in a proof is referred to [79].al) where II* is the product over all integers j in (0. 2). We leave it as an exercise for the reader to show that Theorem (12.

By (12. Also by (12. To conclude this paragraph we give one more example linking RM-codes to a familiar part of combinatorial theory. By straight- forward calculation one can show that C1 has exactly 21 words of weight 5. g. In other words: A can be retrieved from its linear span! Once again we have an example of a linear code for which the vectors of minimum weight (* 0) form a 2-design. 14). 2m columns). If we re- place 0's by -1's in this matrix the result is a Hadamard matrix! We-now consider so-called equidistant codes. and let A be the incidence matrix of points and hyperplanes of PG(m. The code has dimension m + 1. Closely related to these codes are the Projective Geometry Codes. The dual code C is called a pro- jective geometry code. q = pa. 10) this is the dual of the extended Hamming code. From each such pair we choose one word and with these we form a matrix (2m rows. The code C1 has dimension 10. We treat only the binary case. (a. (Readers familiar with the proof of Singer's theorem will see the connection. Consider the 1st-order RM code of length 2m. [54]).GF(2m). An example connected to the topic of our next p-1 chapter is the case p = a = m = 2. The code treated at the end of chapter 9 is an example with d = 6. 4). e. If d is odd the code clearly can have 81 . e. The rows of A generate a code C1 over the alphabet GF(p). 4). 9) and (12. [54]) that C1 has dimension 1 + (m + p . b) = 0 for every pair of code words. For each code word a in this code 1 + a is also a code word. These are codes with the property that any two distinct code words have the same distance d. Let p be a prime. We shall give only one example. In this case A is the incidence matrix of PG(2. It can be shown (cf.1)a.1 and from Theorem (12. (12. ) The RM codes treated in this paragraph are examples of a larger class of codes known as Fuclidean Geometry Codes (cf. The result follows from the fact that this is a linear mapping of order 2m . This implies that a code word of C1 has weight 5 if and only if it is a line of PG(2. 9) this code is self-orthogonal. 16) Definition. i. q).

2).e. It has parameters r = k = 6. 2. 2). 2. To construct the partial geometry one has to pick certain 6-cliques and take these to be the lines. y = c + cj. c3. 30. a code is called trivial if the m by n matrix with the code words as rows has the property that every column has at least m . It is much more difficult to show that the converse is true. 9. c4 = (2. e. [58]). Therefore all differences ci . Here. We take S := {co. 2). c + S. 2.only two words. chapter 4). 2. 1. The resulting matrix has as its rows the code words of an equidistant code with k2 + k + 2 words with mutual distance d = 2k.1 equal entries (cf. Two points x and y are collinear if there is a c and i and j such that x = c + ci. To A we add k . Define co = 0.y = ci . It follows that two points are on at most 82 . 2. d = 12. c5 } as a line and define the remaining lines by translation. where c E C. c2 = (2. eigenvalues 3 and -6 with multiplicities 50 and 30. c2. 2. Now. 2. 12). If d = 2k and the code has m words of length n (where n is sufficiently large) then it is not difficult to show that n k2 + k + 2 or that the code is trivial. x . 1) = 0 and to join two vertices by an edge iff their Hamming distance is 2 or 5. We now discuss a connection between coding theory and partial geometries (cf. We remark that it has been shown that in the case that k = 6. i. 1. i. 2. i. namely that if an equidistant code with k2 + k + 2 words and mutual distance 2k exists. k). At present only one partial geometry with t = 2 and min { n.cj (i * j) are different. The essential observation in the proof that we now have a partial geometry is the fact 5 that c.1 columns of ones and then we add a row of zeros. [81 ]). e. We do this as follows. c3 = (2. c4. These six code words of C form a clique in the graph. then a PG(2.cj. k } > 3 is known. 2. cl = (1. the maximal number of words is 32 (cf. The easiest way to describe the corresponding strongly regular graph is to take as its 81 vertices the code words of the (5. 4) linear code C over GF(3) defined by c c C (c. 2. [46]). k) exists (cf. 2. 1). c5 = (2. 1. = 0 is the only nontrivial linear relation involving c i to c 5 . 2). 2. cl. It is not difficult to check that this graph is indeed strongly regular with parameters (81. let A be the incidence matrix of PG(2.

each line has 6 points and each point is on 6 lines. j. If k # i then we see that x is also collinear with ci. 83 . and k all different.cj. We saw above that x is collinear with ck if x = ck + ci . L) the average number of lines through x and meeting L is 2..one line. and x is collinear with c. By the observation made above these 75 linear combinations are all different! Since there are 75 points x S the proof is complete. By a simple counting argument we see that for a nonincident point-line pair (x. For more information on this partial geometry and connections to association schemes we refer to [84].c. . It is now sufficient to take x S and to show that there are two points in S collinear with x. There are 15 1 J J combinations ck + ci . If k = i then x=-c.cj with k = i j and there are 60 such com- binations with i.

10)) which are self-dual by Theorem (12. We now give an example of such a code over GF(3). a6) then 84 . a3. 9). designs and projective planes The (8. Let SS be the circulant with first row (0 1 -1 -1 1). Define G by 1 0 0 0 0 0 1 1 1 1 1 0 S5 (13. 6) code C over GF(3) and H is a parity check matrix for C. We remark that 0 0 0 0 0 0 0 (13. a5. 1) G 0 15 0 0 and H by 11 S I H := 1 5 5 1 1 Then GHT = 0.13 . 4)-extended binary Hamming code discussed in Chapter 9 coincides with its dual. Clearly the rows of G are independent and hence G is the generator matrix of an (11.2) GGT= 0 -Js 0 0 This implies that if a = (al. a4.Self-orthogonal codes. This is an example of the class of r-th order RM codes of length 22r+1 (see (12. a2.

b has weight 12 . i. If we had started with instead of C..k . Now consider the extended code C. the code C is self-dual. a (12.[k2 1]. 2) is 1 + 2(11) + 4(12) = 35. 1). Let there be k positions where a and b both have non-zero coordinates. i. both with weight 6.. a Steiner system.. . Consider two code words a and b in C. We now have the following theorem. all words in C have weight I (mod 3). 3) 77 + 132k5776 + 1326775 + 330ksr73 + 110491]2 + 24k11. Let Z be the collection of 6-subsets of the set S:= { 1. This implies that a linear combination of 3 rows has weight > 3. . Then either a + b or a . the code is perfect! This code is known as the ternary Golay code. This implies that either a = ±b or k 4. 2. e. e. The weight enumerator is easily found to be 11 (13. 6 a GGTaT = -(I ai) 2 # 1. 6)-code over GF(3). We say that a subset D of { 1. It has a generator matrix G obtained from G in (13. This implies that the minimum weight and minimum distance are 5. From (13. By (13. the code is 2-error-correcting. 1) by adding a column (0 -1 -1 -1 -1 -1)T. Chapter 15). e. i=2 i. 3) we find the weight enumerator to be 12 (13. 4) 17 + 2646776 + 44097)3 + 2412 Remark. 2) we then have GGT = 0.. 85 . Then 12 1 Z is a 5-design with parameters (12. and since this weight cannot be 4 it must also be at least 5. 5) Theorem. Since the number of words is 36 and the volume of an S(x. Trivially a linear combination of more than 3 rows of G has weight ? 4 and hence ? 5. the proof would have been even more simple (cf.which support code words of weight 6 in C. e. 12 } supports a code word a in C if the non-zero coordinates occur in the positions of D. . (13. Clearly no code word can have weight 2 and in fact it is immediately clear that a linear com- bination of 2 rows of G produces a word of weight 5 or 6. 6. 2. i. .

x. (x. r) one of p. q. (iv) if in some code word (p. chapter 1). Therefore the weight of this code word is the sum of the weights of two nonzero words of C. The following observa- tions are obvious: (i) G24 is a linear code of dimension 12. We now construct a binary linear code G24 of length 24 by taking as code words linear combinations of vectors (a. By combining (i) to (iv) we see that G2 4 has minimum distance 8. G24 is a self-dual code. x). One easily checks that C n C' = 10. In the same way as we did above for the ternary Golay code we find that this code is perfect. Since S has 12 (5) = 792 5-subsets the proof is complete. q. 0. i. Consider the (8. 4). Proof. The result of Theorem (13. q. b E X E C'. i. 12)-code with minimum distance 7. a). This famous code is known as the binary Golay code (and G2 4 is the corresponding extended code). 5) is a special case of Theorem (14. b. r is 0 then (p. 1 J.(24. (iii) all basis vectors have weight divisible by 4 and therefore it follows from (ii) that this holds for all code words in G24. (ii) any two basis vectors are orthogonal. We mention it here because it motivated a search for 5-designs of this type. e. 13). 4) extended Hamming code C as defined in the example following Theorem (10. 86 . Our next topic is another remarkable perfect code which will give us a second construction of the Steiner system 5 . or one of p. If we shorten G24 by deleting one coordinate we obtain a binary (23. b). e. r is 1 and the weight is again 8. We have seen above that two code words a and b are supported by sets which intersect in at most 4 points unless a = ±b. 8. 1) (cf. chapter 1). where a E C. In chapter 15 we return to this question. Therefore (13. Let C' be the code obtained from C by reversing the order of the symbols. q. 4) implies that there are i X 264 X 6 = 792 distinct 5-subsets of S contained in sets of the collection 0. // Of course this is a 'well-known 5-design (cf. r) is a linear combination of basis vectors with x = 0 or 1. (0. 8.

(24. Since no two of these words have l's in the same five positions they cover 24 759(8) different five tuples. one of which is in the fixed position. Number of errors Number of parity checks that fail fixed position in error fixed position correct 1 253 77 2 176 112 3 141 125 4 1. Arguing in the same way for the other cases we find the following table. 1)) and both do not occur in 120 of the parity checks. We now compute the 253 parity checks. Since G2 4 is self-dual any word is also a parity check. The last topic in this chapter concerns self-dual codes associated with projective planes and related questions. If there are more than four errors in the received word we also find one of the numbers in the table and the precedure fails. 7. A first example of a code generated by a plane was given in chapter 11. Since the Golay code is perfect its weight enumerator can be determined (the reader should check this!) and one finds that G2 4 has 759 words of weight 8. The other two errors occur together in 21 of the parity checks (cf. 1). [53]) that the code G24 can be decoded by a variation of the method discussed in chapter 11. It was shown by Goethals (cf. 8.28 128 Therefore the number of parity checks which fail tells us the number of errors and also whether the fixed position is correct or not except in the case of four errors which can be detected but not corrected.(23. Let us now look at the words of weight 8 in G G. Assume that there are 3 errors. It follows that exactly 141 of the parity checks fail. We shall show that a projective plane generates a self-dual code. Hence these code words represent 5 the blocks of a 5 . 1). 87 . (1. Suppose that a received word has at most 4 errors. These 253 words represent on the remaining 23 positions the blocks of the derived design 4 . Consider a fixed position and take as parity checks all words of weight 8 in G2 4 with a 1 in this position.

Therefore dim C 1(n2 + n + 2). Ik P Define N 0 1 ). n). 6) Theorem. From (13. 6) we immediately have: (13.r = dim C1. 88 . 7) Theorem. n). 12). If n = 2 (mod 4) then the rows of the incidence matrix A of PG(2. 2+n) Then det ANT = det A = (n + 1)n2 (n Since all entries in the first k columns of ANT are even we find that 2k I det A. (i) 1 C () . interpreted in coding terms. We consider the sub- space C of (R(n2+n+l) over GF(2) which is generated by the rows of A. n) generate a code C. where n = 2 (mod 4). Since n is even the code C is self-orthogonal. From (i) and (ii) the theorem follows. 6) can also be proved using the invariant factors of A. The case of even n is more difficult. The method which has the advantage that it generalises to other characteristics will be illustrated in the proof of Theorem (13. a Proof. (ii) Let dim C = r and let k :=n 2 + n + 1 . It then follows that r> 2(n2 + n + 2). i. If n = 2 (mod 4) then the rows of the incidence matrix A of PG(2. These vectors generate the subspace of 6((n2+n+1) consisting of all words of even weight and this is C (because C obviously has no words of odd weight). for which C is self-dual. n) generate a code C with dimension (n2 + n + 2). If n is odd it is easy to see what the code C is. If we take the sum of the rows of A which have a I in a fixed position the result is a row with a 0 in that position and 1's elsewhere. // Remark.Let A be the incidence matrix of PG(2. e. Interpret A and N as rational matrices. We continue with a few other properties of the plane PG(2. Theorem (13. (13. Let H be a parity check matrix of rank k for C and assume the coordinate places have been permuted in such a way that H has the form (Ik P).

Proof. V E C. This would yield d > n + 3. (i) Let V E C and w(v) = n + 2. Each of the n lines * I through one of these 2a points meets v at least once more. n) through the pairs of points of v. ) Remark. then every line * l through this point meets v by (i). If there is a point of 1 not on v. Since n = 2 (mod 4) it follows that v is the sum of the lines of S. Let v * 0 be a code word with w(v) = d. d>n+1. Let S be the set of 2 (n + 1) (n + 2) distinct lines of PG(2. We leave it as an interesting exercise for the reader to check that Theorem (13.)d>n2+n+1. 6) has minimum weight n + 1 and every vector of minimum weight is a line in PG(2. i. a=0 or a=1.e. (13. each point of v is on n + 1 lines of S. // In chapter 1 we gave the definition of type II oval.(13. e. The code C of Theorem (13. Let 1 be a line and suppose v and l have 2a points in common. If w(v)=n+1 thenthereisa line 1 of PG(2.9) Theorem. 89 . // (Also see [4]. Therefore 2a+n:5n+2. Each point not in v is on a (n + 2) of these lines. i. Proof. 4) which was treated at the end of chapter 12. The vectors of weight n+2 in C are precisely the type II ovals of PG(2. In case (i) we find (n+1. n) which meets v in at least 3 points. Since every line has a 1 as overall parity check we see that: (i) if d is odd then v meets every line at least once and (ii) if d is even then every line through a fixed point of v meets v in a second point. 9) also holds for the dual of the code generated by PG(2. Fvery line meets v in an even number of points. n). (ii) Let V be a type II oval. In case (ii) we immediately have d > n + 1.e. 8) Theorem. n). i.

. that the ring of polynomials in and i which are invariant under this group is the free ring generated by 2 + 712 and 02)2 Let us now consider a possible plane of order 10. [88]) investigated the code words of weight 15 in C. Denniston concerning 6-arcs which are not contained in 7-arcs 90 . If q = 2. i. A15 and A16. Gleason (cf. 9) severely restricted the possibilities. This means that the polynomial is invariant under the linear transformation with matrix q. (13. where k = z n. Therefore we now know that if a plane of order 10 exists. 8) we know that the incidence matrix of this plane generates a code C for which C is a (112. Bruen and J. [89]). [52]. It was pointed out by A. MacWilliams.2 ( 1 ql1). G.1) ).. J. It was shown by A. C. By sub- stituting this in (13. In fact it was found that the plane had to contain a particular configuration of 1. A computer search showed that starting from such a configuration the plane could not be completed. =A10=0. N. 77) = q-kA(11 . 7) and (13.. all code words of C have even weight.. From (13. Fisher (cf. Sloane and J. A11 = 111. Thompson (cf.5 lines. 14) the following relation for the weight enumerator 71): (13. Therefore A13 = A14 = 0.k. 77 + (q . F. For a self-dual (n. 6). J. 10) or by using Gleason's results one sees that 17) is uniquely determined if we know A12. Recently F. It was assumed that the incidence matrix of the plane led to a code for which A15 * 0. [12]. [22]) that this same result also follows from a computer result of R. then (13. k) code C over GF(q) we find from Theorem (9. A.11)A15 = 0. Al A2=. The 2 transformations generate the dihedral group a) a. H. 56)-self-dual code and that the weight enumerator 77) of C has coefficients A0 = 1. M. Arguments of the same type as the ones we used in proving Theorems (13. i) is invariant under the trans- formation k . Since all the generators of C have weight 12 and any two words of C have an even number of 1's in common we see that all weights in C are divisible by 4. e. 8) and (13.

+ xNYN . Let A be the incidence matrix of a projective plane 7rof order n. y) := x1y1 + x2y2 + . such that PAQ = D. 11) has been proved again. i..... e. Now rkp(A) = rkp(D) is the number of invariant factors di which are not divisible by p. having determinants ±1). Remark. Further references: [100].(applied to the dual plane). i. invertible over Z (i. We have (det A)2 = det AAT = det(nI + J) = (n + 1)2nN-1.j) are orthogonal. We now generalise Theorem (13. We replace the usual inner +1). e. The dimension of C is equal to the p-rank rkp(A) of A (i. product on G where N = n2 + n + 1. Since p I n we can take the parity check -1.1 dN) with di c Z.. dim C > i (N + 1).1) of the invariant factors are divisible by p. its rank regarded as a matrix over GF(p)). .xN+IYN+l' (13. // This construction has been generalised by Lander [76]. The rows of (A . [8]. 1. so ±det A = det D = d d . so C S Cl. where D is a diagonal matrix diag(d1.. or Smith normal form (cf. C is self-dual relative to the Minkowski inner product.1) uniquely determines the weight enumerator of the code C if we assume that the plane of order 10 is extendable. that example (iv) is possible. by the Minkowski inner product (x. Then. and C the code generated by the rows of A over GF(p) with a parity check adjoined. if p II n. The result (13. satisfying diIdi+1 for 1 < i:5 N-1.12) Theorem. Cohn [34] p. There are integral matrices P and Q. 6). e. dN is exactly divisible by p2 (N-1) Thus I 2 at most z (N . Proof. In this way (1. We now appeal to the theory of invariant factors. e. Since we already know that C S C1 we must have C=C1. 279). who 91 . d2..

showed the following.

(13. 13) Theorem. Let n be a projective plane of order n with
incidence matrix A. Let N = n2 + n + 1, and let p be a prime such
that ps II n (s > 0). Then there is a sequence

{0} =C-1 sC0 c... ECS=(R (N+1)

of codes of length N + 1 over GF(p), with the following properties:
(i) each code C. is invariant under all automorphisms of 71;
(ii) Co is the extended code generated by the rows of A;
1
(iii) for -1 < i <_ s, C. Cs-1-i (relative to the Minkowski
inner product);
(iv) dim C.1 is equal to the number of invariant factors of A
not divisible by pi+1 for i < s;
(v) for 0 < i < s-1, Ci has minimum weight n + 2, and the
words of weight n + 2 in C0 correspond to lines or (possibly if p = 2)
ovals in 7T.

Remarks. (i) If s is odd, we see from (iii) that C, is a
Z (s-1)
self-dual code associated with 71.
(ii) Lander has also calculated the dimensions of the codes
C. in the case where n is Desarguesian, generalising the known result
that dim Co = 1 + (p2 1)s in this case (cf. [54]).
Theorem (13. 8) led to a different attempt to find a PG(2, 10),
involving ovals in two ways. Let A be the incidence matrix of any block
design 2 - (56, 11, 2), (four of which are known, cf. Chapter 5). Then
the rows of G := (I56 A) clearly generate a (112, 56) binary self-dual
code C. We discuss only the case where the 2 - (56, 11, 2) design
corresponds to the Gewirtz graph. In that case A is symmetric and
therefore (A I56) is also a generator matrix for C. It follows that in
order to show that C has minimum weight 12 it is sufficient to consider
the sum of i 6 rows of G. Let xj (1 < j < 56) be the number of ones
in the i chosen rows and the j-th column of A. Then we have Z xj = 11.i,
F (21) = 2(2). Therefore I xj(xj - 2) = 2i2 - 13i, i.e. at least 13i-2i 2
columns have exactly one 1, which shows that the weight of the sum of the

92

i rows is at least 2i(7 - i). Now this is 12 if and only if i = I (a row
of G) or i = 6. Furthermore, if the sum of 6 rows of G is 12 then
xj = 0 for 20 values of j, xj = 1. for 6 values of j, xj = 2 for the re-
maining 30 columns of A. So, the 6 rows of A have the property that
no point is in more than two of them, i. e. they correspond to the tangents
of an oval in 2 - (56, 11, 2). If we are lucky then the (112, 56)-code
corresponds to PG(2, 10). Then, we can reconstruct the plane, using
Theorem (13. 8), by considering the ovals of 2 - (56, 11, 2) through a
fixed point. The other designs with the same parameters can be treated
in a similar fashion. None of the known designs have produced the
desired result! (Also see [8], [9], [123]. )
The methods of this chapter were used recently to give a character-
isation of the biplane of order 2 (cf. [2]).

(13.14) Theorem. Let 5) be a symmetric 2 - (v, k, A)-design with
k = 0 (mod 4), A = 2 (mod 4), A Ik. Suppose 2) has ovals and that these
ovals also form a design 2 - (v, K, A). Then 5) is the unique
2 - (7, 4, 2).

Proof. First we observe that the design formed by the type II
ovals has K = (k + A)/A. Let A be the incidence matrix of 7) and let
C be the binary code generated by the rows of A. Clearly C c C1.
Since v is odd we have dim C < 2 (v - 1). We proceed as in the proof
of Theorem (13. 12). Let d1, d2, ... , dv be the invariant factors of A.
Then di Id 1+1 (1 ` i v-1), d1 d2 ... dv = det A = k(k - ) ' 2v - 1, and
dl d2 ... dv-1 is the g. c. d. of the minors of order v - 1 of A. If we
add all the rows of A we find a row with all entries k. It follows that
d1 ... d is is a divisor of (k - A)2 (v-1) Therefore at least 2 (v - 1)
of the invariant factors are odd and hence the rank of A over GF(2) is
at least 2 (v - 1). Since 1 E C but 1 % C (all code words of C have
even weight) we have proved that dim C = 2 (v - 1) and C1 is the direct
sum of C and 1.
Now consider a vector of weight d > 0 in C1. The same argument
as in Theorem (13. 9) shows that d >_ (k + A)/A, i. e. the ovals are vectors
of minimum weight in C1. Now, using the fact that the ovals also form

93

a design we can repeat the argument again and show that the minimum
weight of C is k (which is achieved by the blocks of D). Since the
minimum weight of C1 is less than k we see that each oval has the
form 1 + c, where c E C. It follows that the sum of two ovals is in C.
If these two ovals meet in a points we must have 2(k - - a) ? k.
Because we know that all words in C have weight divisible by 4 we find
a = 1 as only possibility (and hence A = 2). This means that the design
formed by the ovals of 0 is a projective plane of order z k on
v= 1+ i . k(k - 1) points, L e. k=4. //

94

. (ii) The monic polynomial of lowest degree in the ideal generated by c(x) is (c(x). Chapter 10). . Since xn . n-1 given by 7r. (iii) Vf (x) EV[c (x)f (x) = f (x) ].1). to denote the permutation o f 10.1) = (p1(x)g(x). We use the symbol IT . (14. We find c2 (x) + p1(x)p2 (x)g(x)h(x) = c(x). h(x)) = 1. 2) p1(x)g(x) + p2(x)h(x) = 1 (in GF(2)[x]). We multiply both sides of this equation by c(x). For every ideal V in (R (n) there is a unique poly- nomial c(x) E V. We consider binary cyclic codes of length n. 1. Proof.. (ii) c(x) generates V. n) = 1 then 7rj(c(x)) is the idempotent of 7rjV. c (x) is a unit for V. i. where n is odd. e.(k) := jk } (mod n) and also to denote the automorphism of (R(n) given by n. (iv) if (j. Since in (R (n) we have g(x)h(x) = 0. where c(x) := p1(x)g(x)..1 (in GF(2)[x]). called the idempotent of V. 1) Theorem. with the following proper- ties: (i) c(x) = c2(x).1 has no multiple zeros we have (g(x). we have proved (i). (i) Let g(x) be the generator of V and let g(x)h(x) = xn .14 Quadratic residue codes Before we can start with the actual topic of this chapter we must prove a theorem on cyclic codes (cf. Therefore there are polynomials p1(x) and p2 (x) such that (14. where we use the identification of vectors in 6t (n) and polynomials mod (xn .1). g(x)h(x)) = g(x).. xn . 95 .(xk) := xjk mod (xn .

(14. (ii) a is a primitive n-th root of unity in an extension field of 96 . We shall need example (14. For every polynomial q(x) the word q(x) {1 + f(x) } is in Hm. 4) Proposition. In the description of quadratic residue codes we shall use the following notation. Let f(x) = c(x)f1(x). if m = 4. 7f. and xh(x) = f(x) = (x + x2 + x4 + x8) + (x3 + xb + x9 + x12). i.(c(x)) is an idempotent and also a unit in 71JV. g. Then c(x)f(x) = c2(x)f1(x) = c(x)fl(x) = f(x). 2) now has the form p1(x)m1(x) + xsh(x) = 1.4) m1(x) generates the Hamming code which we now denote by H m Let xn . (i) n is an odd prime. 1) (ii) we have (14.1. (iii) By (ii) every f(x) E V is a multiple of c(x). It follows that there is an exponent s such that f(x) := xsh(x) is the idempotent of HI*11 and hence (14. By Theorem (14. 3) Example. 5) Notation. We leave it to the reader to check that f(x) has weight 2m-1. We know that h(x) generates the dual Hm of Hm (cf. n = 15 and m1(x) = x4 + x + 1 then (x8 + x2 + 1)m1(x) + x(x11 + x8 + x7 + x5 + x3 + x2 + x + 1) = 1. // (14. Then by Theorem (10. Let m1 (x) be the minimal polynomial of a primitive element a of GF(2m). Since the unit is unique we have proved (iv). Let n = 2m . We consider an example in detail. c(x) is a unit in V. This code H*m has dimension m and hence it consists of 0 and the n cyclic shifts of the word h(x). e.1 = m(x)h(x) in GF(2)[x]. E. (iv) Since is an automorphism of (R(n) the polynomial T. Clearly. the polynomial 1 + f(x) is the idempotent of Hm. 3) and proposition (14. 4) below in Chapter 16. Chapter 10).

1)g0(x) are both called quadratic residue codes (QR codes). remark following (14. Note that in the binary case the condition that q is a quadratic residue mod n is satisfied if n =±1 (mod 8). R1 denotes the set of nonzero elements of GF(n)\R°. 6) we obtain equivalent codes in the sense of Chapter 9 (cf.The cyclic codes of length n over GF(q) with generators g0(x) and (x . Clearly ae is a quadratic residue if e = 0 (mod 2) and a non- residue if e = 1 (mod 2).1 = (x .1)g0(x) consists of the words of even weight in the code with generator g0(x). The nonzero elements of GF(n) are powers of a primitive element a c GF(n). 10). Chapter 9). If j E R1 then the permutation 7T J maps R0 into R1 and vice versa. q is a quadratic residue mod n. Since q is a quadratic residue the set R0 is closed under multiplication by q (mod n).1)g0(x)gl(x)) q(n-1)/2=1 We assume that (mod n). e.ar). [79] Theorem 3.. 97 . ) In the binary case the code with generator (x .. (14.2.ar) (note that ° rcR 0 rER xn . If n=_-1 (mod 4) then -1 ER 1 and therefore the transformation x .x 1 maps a code word of the code with generator g0(x) into a code word of the code with generator g1(x). i.GF(q) (a suitably chosen later on). (Cf. then (1 1G 1) is a genera- tor matrix for the second code and the generator matrix for the extended r1 11 . (iv) g (x) := II (x .2). It is easily seen that if we replace R0 by R1 in (14. The extended quadratic residue code of length n + 1 is obtained by annexing an overall parity check to the code with generator 90M. 6) Definition. 11 code is 0 G (cf. (iii)R0 denotes the set of quadratic residues mod n. If G is a generator matrix for the first of these codes. g1 (x) := H (x .

ej= ek . 7) Theorem. 9(a1) = 1 if i ER 1 and finally 0(a°) = n21.4a for some a ? 0. 8) 9(x) := I xr rER 0 Clearly 9(x) is an idempotent in 63(n) (2 is a quadratic residue mod n). n-1 Proof. Hence. cixl is a code word of the QR code n-1 i=0 with generator go(x).1. if I ci # 0. We define (14. (ii) if n = -1 (mod 4) then d2 . If a in (14. By a suitable permutation r. This implies that c(x)c(x) = 1 + x + x 2 +. (ii) Now T. In the same way we have 9(a') = 9(a) if i E R0 and 9(a') + 9(a) = 1 if i E R. Then 9(a') = 0 if i E R0. (iii) Let c(x) _ x c(x) _ x 1. then i=1 i=1 e . If ei . This proves: (14.1)g0(x) if n = 1 (mod 8) and of the QR code with generator g0(x) if n = -1 (mod 8). We now choose a in such a way that 9(a) = 0. Define: 98 . d e. 9) Theorem. This proves (iii).d + 1 > n.ei = el . This proves (i). can be chosen such that (x) = c(x 1) and then c(x)c(x) has at most d2 . (iii) if n = -1 (mod 8) and q = 2 then d = 3 (mod 4). (i) Since F ci * 0 the polynomial c(x) is not divisible i= O by x . +xn-1 The number of nonzero coefficients of the left-hand side is at most d2. 5) (ii) is suitably chosen then the poly- nomial 9(x) of (14. n = d2 .1.d + 1 nonzero coefficients. 9(a) E GF(2). and if w(c) = d is the weight of this i=0 code word. then four at a time. again consider binary QR codes. if any terms cancel. 1.. then (i) d2?n. Hence { 9(a) } 2 = 9(a). 8) is the idempotent of the QR code with generator (x .d + 1 . d -e. e.ek. If c = c(x) _ I. // Now. i. we can transform c(x) into a polynomial c(x) which is divisible by g1 (x) and of course again not divisible by x .. e. Let C be the circulant with the code word 0 as its first row.e1. n-1 (14. i.

and 1 1- G := `cT CJ It now follows from (14. 7). n-1. They require that the coordinate in the extended position is multiplied by a factor in such a way that the resulting code is self-orthogonal (n = -1 (mod 4)) or orthogonal to the other QR code. To most readers this is probably a familiar property of G (compare with automorphism groups of certain Hadamard matrices). n) acting on the positions and show that the extended QR code remains invariant. Remark. 1. By the remark following (10. . Let us number the coordinate places of the code words in the ex- tended binary QR code using the coordinates of the projective line. 10) is also true for the q-ary codes. The group is generated by the transformations Sx := x + 1 and Tx := -x 1. [3]) slightly altered the definition of extended code for the non-binary case. e. n). 10) Theorem.. 0. It is a simple exercise to show that T maps the rows of G into linear com- binations of at most 3 rows of G (cf. 0 if n ° 1 (mod 8). The position of the overall parity check is We make the usual conventions about arithmetic operations: ±01 = 00 -1 f00 = 0. 10) this code is invariant under a doubly transitive group. 9) that the rows of G (which are not independent!) generate the extended binary QR code of length n + 1. The automorphism group of the extended binary QR code of length n + 1 contains PSL(2.8) we see that the binary QR code has odd minimum weight. 83). By (14. . Again. Thus we find: 99 . We have shown: (14. Since S leaves . if extended (n = 1 (mod 4)). Gleason and Prange (for a discussion cf.. They show that (14.+ a = . 1 if n=-1 (mod 8). [ 79] p. This allows us to apply Theorem (14. We now consider the permutations of PSL(2. i. .invariant and it is a cyclic shift on the other positions it leaves the code invariant.for a E GF(n). we consider an extended binary QR code.

9). In n = -1 (mod 4) then the extended binary QR code C of length n + 1 is self-dual. This follows from the remark preceeding (14. then d2 . e. defined after (14. e. By (14. chapter 13). 3) etc. form a 2-design. 12)Example. In this case the volume of S(x. 12)- code.d + 1 >_ n with equality if and only if the design is a projective plane of order d . We consider a very important example of QR codes. n21) cyclic code with minimum weight d and C1 S C. if equality holds in (14. A more combinatorial form of the square root bound (which is the name by which (14. U S(x. Remark.1. 10).(14.11) (ii).1) >_ 22. then there exists a projective plane of order d . The minimum weight of the binary QR code of length n with generator g0(x) is an odd number d for which (i) d2 > n if n = 1 (mod 8). 7) and (10. The reader familiar with the theory of difference sets and designs should have no difficulty checking that the proof of (14. a much larger group than the one guaranteed by Theorem (14. 11) is known) was recently found by Assmus. 11) Theorem. and Sachar (cf. Since d is odd we have d > 7. and if the subsets of the positions. i. i. The automorphism group of the extended binary Golay code is M2 4. 100 . 11) the minimum distance d of this code satisfies d(d .d + 1 >_ n if n = -1 (mod 8). Over GF(2) we have (x-1)(x11+x9+x7+x6+x5+1)(x11+x10+x6+x5+x4+x2+1) x23 -1= _ = (x-1)g0(x)gl W. The present representation shows that the code is cyclic. Mattson.1. 7) (ii) and (iii) implies that. It states that if C is an (n. using the form we found above for the generator matrix of C. (14. [100]). which support code words of minimum weight. The binary QR code C of length 23 with generator g0(x) is a (23. e) = 223 or in other words: the X EC code is perfect! Since it is known that the Golay code is unique we have found this code again (cf. 3) is 211. (ii) d2 . It can also be seen from the matrix G.

n) which support code words of weight d in A form a t-design. The proof is in 6 steps.t itself less than or equal to d . The statement for i = 0 implies that the complements of the blocks of D also form a t-design. Let vo be the largest integer v+q-2 satisfying o q_20 - [ q- < d. 2.k) dual code. and H. Let the minimum weights of these codes be d and e. An analogous statement holds for B. (14. 1) we have t i0j)ai = (fi)x (j = 0. 101 . where. Let t be an integer less than d. 1. 13)Theorem. Jr. . Assmus. Let A be an (n. If T is a fixed t-subset of S and C is any code then we denote by C' the code obtained by deleting the coordinates in T from the code words of C.t. . F. which shows that many codes can be used to yield t-designs..[ q. The theorem.t. Suppose the number of non-0 weights of Al which are less than or equal to n . the subsets of S := 11. is due to F. for each weight w with e < w < min { n . [3]). w o) . Denote by ai the number of blocks D of 0 with I D n T = i.1 ] < e. We now come to one of the most important theorems of this course. Furthermore. F. We denote by C0 the subcode of C which consists of the words of C which have zeros at all the positions of T.. t). for each weight v with d < v . k)-code over GF(q) and let Al be the (n. By solving these equations we find that ai does not depend on the choice of the set T. Then by (1. the subsets of S which support code words of weight w in Al forma t-design. Mattson. Let (S. (cf. (ii) We shall use the following property of t-designs. if q = 2. (i) By definition of v 0 two words of A with weight < vo which have the same support must be scalar multiples of each other (because there exists a linear combination of these two words with weight < d). and wo the largest integer satisfying wo . we take vo = wo = n. Then. For Al we write B. Proof. In the proof we shall use the following notation. Jr. ..v0. 0)) be a t-design and let T be a t-subset of S. n .

Let 5) be the collection of v-subsets of S which support words of weight v in A. yield a system of linearly independent equations which we can solve in principle. Since the minimum weight of A' is at least d .t we know d . Since A' = (B') 0 1 the same holds for A'. (iii) Consider any d .t in A' corres- ponding to words of weight v in A. 14).t in A' does not depend on T. By the induction hypothesis and (ii) the number of words of weight v . 14). The im- portant observation. (v) We first prove the second assertion of the theorem. however. Let 0 < v1 < v < .k . From the definition of d it follows that the resulting matrix still has rank k. (iv) Consider any t-subset T of S.t in the code B. Hence B'0 = (A') . Let 8 be the collection of w-subsets of S which support words of weight w in B.t. The dimension of B' is at least n . e. Let w min In .t (where 2 r d . By (iv) this number does not depend on T. . Consider the set 8' of complements of sets in 6.1 columns of the generator matrix of A and delete these. In the same way as in (v) we see that the number of sets in a) containing a given t-subset T of S is q-1 1 times the number of words of weight d . By (iv) this number does not depend on T. This is equal to the number of coefficients of the weight enumerator of Bo which we do not know yet.t. < v r< n .t in A'. (vi) To prove the first assertion of the theorem we start with v = d. Let d s v v0 and assume that the assertion of the theorem is true for all v' with d v' < v. The number of subsets of T containing a given t-subset T of S is q 1 1 times the number of words of weight v . For any t-subset T of S we find from (i) that the number of sets of containing T is q-1 1 times the number of words of weight w in B. Let a) be as before. again by Theorem (9. e.t) be the possible nonzero weights less than or equal to n . The MacWilliams relations. does not depend on the choice of T. the weight enu- merator of Bo. We now proceed by induction. By (iv) the total number of words of weight v .t coefficients of the weight enumerator of A'. i. k)- code. Then these are also the only possible nonzero weights for B B. i.t in A' corresponding to 102 . By (iii) A' is an (n-t. Theorem (9. Hence 6' is a t-design. By (ii) 8 is also a t-design. Clearly (B0)' C (A')1. . w0 }. is that the solution.

One can compute the weight enumerator of this code using Theorem (9. If we take v = 28. 4)-Hamming code. 20 or 24. Suppose d > 31. Let A be a (121. (14. Take n = 24. 13) is satisfied for t = 5.14)). 5). k = 4. Therefore the extended code C has minimum distance d = 12 and again it is self-dual. Then d = e = 6 (cf. The condition of (14. q = 2 and let A = Al be the extended binary Golay code G2 4 (cf. This yields 4 different 5-designs. 13). 17) Example. 28. Take t = 3. Since I S(x. 7) (iii) implies that the minimum distance is at least 11.16)Fxample. Hence O) is a t-design. (14. The only occurring weights are 0. There- fore the existence of the 5 . we find the result of Theorem (9. Since all weights in A are divisible by 3 the condition of the theorem is satisfied. 16. We now find the result of Theorem (13.18) Example. q = 2 and let A = Al be the extended (8. 8. (14. chapter 13. 14)Example. (14. 16. Hence the minimum distance is 11. Take n = 12. These examples explain the interest (from the point of view of 103 . (14. Let n = 47. 36 we find the complementary designs. and 24 are the only weights which occur. (13. Consider the binary QR code of length 47. Taking v = 4. 12. Taking v = d we find that the supports of the minimum weight words of A form a 5-design. Then (14. By (14. 6) I > 22 3 the code cannot be 6-error-correcting. q = 3 and let A = Al be the extended ternary Golay code. 13) is satisfied. Take t = 5. 4) the condition of (14. 32. 13) applies with t = 5. 12). 12)) we have already proved that 0. Then d = e = 4. // We shall look at several examples of this theorem.words of weight < v in A is also independent of T.(24. 16. We can take v = 12. k = 12. 20. By (13. k = 6. 1) Steiner system also follows from Theorem (14. 24. Take n = 8. 15)Example. 36 and 48. (14. 14). 61)-self-dual code over GF(3). 8. 11) the minimum distance d is at least 9. Hence (14. 12. 32.

Let G be the additive group of GF(q)..design theorists) in determining the minimum weight of quadratic residue codes. It follows that for each v the subsets which support code words of weight v form a 3-design. p-1). We shall only sketch the ideas and show a connection with certain well known designs. 13) is satisfied if t = 5. [118]). Hence if we take t = 3 the condition of (14. See also [86]. 30) extended QR code over GF(3) (cf. gIG which is to be considered as a formal expression. [31].21)-2-error- correcting primitive binary BCH-code C (cf.5)). This means that the condition of (14.. 1.1 . In Al only 3 weights occur. Several authors have generalized the idea of QR codes (cf. 221-2 ± 21-1.1. 20) c = cgxg. A survey for n < 59 can be found in [5]. (cg E F). As alphabet we may use any finite field F. The idea is to use an element of the automorphism group of a self-orthogonal code to map the code into a code of smaller length which is still self-orthogonal. In the generalization the length of the codes is q = pm (m > 1).. Consider a (221-1 . The minimum weight turned out to be 18. 104 . (10.13) which will turn up again later. This method was used successfully by Assmus and Mattson to determine the minimum weight of the (60. 13) is satisfied. [6]). q = 2 and let A be C. This is the idea of contractions of self-orthogonal codes. [37]. Very little is known in this area. We shall take m = 2. 14) one can show that in C1 the only weights which occur are 221-2.1 . Now take n = 221-1. k = 221-1 . where G= 10. 221-1 (14. Hence this QR code yields 5-designs on 30 points.21. . An elementary presentation of the theory is given in [83]. It would be beyond the scope of this course to go into a method of treating QR codes which looks very promising. . We give one more example of (14. We identify the positions of the code with elements of G and represent a code word as (14. Note that this corres- ponds to the representation of cyclic codes in the case that m = 1. 19)Fxample. Using Theorem (9.

Let . 22) tP 1(g) and for each h E G we define the character 4h by (14.iph establishes an iso- morphism between G and the group of characters. The generalized quadratic residue code (GQR code) of length q over F consists of all c = I cgxg such that Lu(c) = 0 for all u c U. 20) with the following rules for addition and multiplication. 21) a xg b xg a b )xg g g g1 +g2 =g g1 g2 where the summations are over all g in G. 24)Pf(I axg) g := F aiPf(g) g gEG gEG The reader who is familiar with character theory will see that we now have all the characters of G and that h . (D. From the independence of the characters one sees that A+ has 105 . We now define the character : G -+F by (14. *) is a ring. We denote this code by A+ and define B+ by replacing U by V. non-squares) in G. (14. Z agxg ® bgxg (ag + bg)xg. Since the proofs of all the facts to be stated below depend heavily on calculations with characters we leave out these proofs (which are straightforward and which can be found in [83]). For the definition of the generalized codes we denote by U (resp. 25) Definition. 23) i/h(g) V/1(gh) Finally the characters are extended to FG linearly by defining (14. Let a be a primitive element of GF(q). The group algebra FG consists of all expressions of the form (14. be a primitive p-th root of unity in some extension field F of F. V) the set of nonzero squares (resp. (14. Then (FG. The codes A and B are subcodes defined by the additional requirement gi0(c) = 0. Every element g E G can be represented as g = io + i1a with i0 and i1 in GF(p).

It then follows that Theorem (14. Subsequently. One can show that these codes are both invariant under PSL(2. The connection to QR codes did not appear in that approach. by adding an extra symbol to the code words such that the extended codes are dual codes.I under the action of PGL(2. B. a fairly difficult proof shows that the only words of minimum weight in A are the words in the orbit of (Z xg. [110]. 1)) can be generalized. 1) under gEK PSL(2.dimension i (q + 1). 1). and B. resp.(p2+1. Once again we find a design from the words of minimal weight in a code (and its dual). Further references for this chapter: [89]. 106 . 25) is the same as (14. 6)! In [83] it is shown that the idea of idempotent (cf. 7) also generalizes. It then follows that the union of the supports of code words of minimal weight in A. are the images of GF(p) U (. (14. q). and multiples of these words. p2) on the projective line of order p2 (represented by G u { -) ). So we have a complete generalization of the theory of QR codes. chapter 1). p+l. in this case one can show that if K = GF(p) then the word c= xg is in A .'. 10) one can extend A+ and B+ to Ao. The reader should now convince himself that if m = 1 Definition (14. q). (cf.. [37]) starting from the designs. In fact. In the same way as was mentioned in the remark following (14. Therefore we actually know that the minimum geK weight is equal to p. So these words form the Mobius plane 3 . The codes discussed above were first constructed (for F = GF(2)) by Delsarte (cf.

The results are due to V. they also lead to new 5- designs. b E GF(q)). g.15 . Let q be a power of an odd prime and let X be the quadratic character on GF(q). We prove a number of properties of these codes: 107 . If q is a power of an odd prime then we take for Cq 1 the Paley matrix defined above. The first construction is the well known Paley construction (cf. . dimension q + 1 as a (2q + 2. Then Cq+1 is a C-matrix of order q + I (cf. Pless ([94]. We define a symmetry code of (15. We note that Cq+7Cq+1 -I q+1 over GF(3) if q = -1 (mod 3) and that Cq+l is symmetric if q = 1 (mod 4). (15. [95 ]). i. [60]).a) (a. skew-symmetric if q = -1 (mod 4). Paley graphs in chapter 2). e. . Define Cq+1 by c := 0. ca b x(b . Let q = -1 (mod 6). 2) Definition. A C-matrix of order m is a matrix C with entries ±1 (outside the diagonal) and 0 (on the diagonal) such that CCT= (m-1)Im. e.a c a.Symmetry codes over GF(3) In this section we treat a sequence of codes which have a number of properties in common with QR codes. .o = X(-1).and the elements of GF(q). We number the rows and columns of a matrix of order q + 1 using the coordinates of the projective line of order q. q + 1)-code Sym2q+2 over GF(3) with generator G2q+2 :_ (Iq+1 Cq+l). where Cq+) is a C-matrix of order q + 1. 1) Definition. A lot is known about these matrices.

13. In describing code words in a symmetry code we shall denote by w1 (x). // (15. respectively the last q + 1 coordinates. Take q = 5. This follows from (15. We find r 0 1 1 1 1 1 1 0 1 -1 -1 1 1 1 0 1 -1 -1 G1 := I6 -1 1 -1 2 1 1 0 1 -1 -1 1 0 1 1 1 -1 -1 1 0 i. // (15. 4) Proposition. 1.Cq+1 = -I q+1 over GF(3). 7) Lemma. and G2*q+2 has rank q + 1. We have for every code word x in a symmetry code: (i) if wt (x) = 1 then w r (x) = q. 3) Proposition. 5) Proposition. (iii) if wl (x) = 3 then wr(x) ? [3(q . (15.3)/4]. 1)). e.)/2I leaves Sym2q+2 invariant. Sym12 is the extended ternary Golay code (cf. 2) (cf.(15. In the case of the binary Golay code it is also possible to find a generator matrix of the form of (15. // (15. wr(x) respectively. 4). Remark. q-1 Since Cq _ (-1) 2 Cq+1 we find G2q+2G*T Proof. 6) Example. If Cq+1 is a Paley matrix then (q+1)/2 G* G2q+2 :_ ((-1) Cq+llq+1) is also a generator matrix for Sym2q+2. 0 q+1 Proof. This follows from Cq+1. (ii) if wl (x) = 2 then w r (x) = (q + 3)/2. A Sym2q+2 is self-dual (and hence all weights are divisible by 3). Sym2q+2 is self-dual. 2q+2 = 0. the contribution to the weight of x due to the first q + 1. [79] p. 108 . Proof.01). . The linear transformation of (R (2q+2) with matrix r 0 Iq+l 1)(q+l.

. Proof. +. b.. 8) the existence of a code word with weight < 12 implies the existence of a code word x with w1 (x) = 4. It is easily checked by hand that no such x exists. 4). By (15. To prove (ii) we add the first two rows and find wr(x) = 2 + (1 + x23) + a + b = 2 (q + 3).. (ii). Then in a symmetry code: (i) There is a code word x with w1 (x) = wi.d=-x32.9) Example. 3) all weights are divisible by 3. Proof.--. 7) a code word x with wl (x) <_ 3 has weight >_ 12. (ii) For all code words x we have w(x) > 0.. First he should add one more line to the array in the proof of Lemma (15. Let q = 17. a b + c . By (15..+ ++..// (15. wr(x) = w2 iff there is a code word y with w1(y) = w2.+ ---j- +++1 0 1 0 0.. 8) Lemma. There are a few 109 .2. Let w1 and w2 be integers. 0 x31 x32 0 ++. 2 or 3 rows of the following array: a b c d 1000. Consider three rows of the generator matrix. Since C18 is a bordered circulant there are not very many different possibilities for such an x. Then 2 wr()?b+c+d=4{3(q-2)+xz3+x32+x21x31}'4(q-3). (ii) Cq+1 is non-singular. wr(y) = w1.. wr(x) = 5. Now add all three rows. and d denote the number of columns of the 4 differ- end types.c d=-x23. The reader should try this as an exercise. c. // (15. 0 0 + + ++. By (15.+ ++. Since multiplication of a column by -1 does not alter weights we may assume that x in (i). By the definition of C-matrix we have (over R): a+b+c+d=q. a+b . 7). a. (iii) is the sum of 1. a-b-c+d=-x21x31 Now (i) is obvious. We consider Sym3 6. 0 x 21 0 x 23 + + + + + + 0010. (i) This follows from (15.+++ At the right..

or 21. ea (a E GF(q)): e . 18. 10) Lemma. (15. oO bEGF(q) It follows that e. 15. These are new and so are the complementary designs with one possible exception. d=e=12.x + 1. q = 11. We consider Cq+1 and S as linear transformations of the space (R(q+l) over GF(3). q = 17. Of course q = 5 yields the well-known Steiner system 5 . 15.13). a EGF(q) eaCq+1 = X(-1)e + X(b . Let q be a power of an odd prime.) as shown in (13. In [95] the examples q = 5. vo=wo=23. We have for the standard basis vectors eu. 5). q = 23 and q = 29 are treated. Now we apply Theorem (14.Cq+1 = ea. The original work was done by computer but it is possible to do it by hand in half an hour and thus show that Sym3 6 has minimum weight 12. We find 5-designs on 36 points with blocks of size 12. Code words in Sym2q+2 have the form aT (Iq+1Cq+1)' aTA(Iq+lA-1Cq+1B) Now aT (Iq+1Cq+1)(0 B which is a code word. So we can take v=12. Proof.(12. It is not known whether the design with block soze 18 is self complementary. 1. Let G(q) denote the group of monomial transformations which leave Sym2q+2 invariant."SCq+1 and 110 . We now define monomial transformations in the same way as was done (in chapter 14) in the binary case for QR. We order these in the same way as for cyclic codes. codes.cases to be distinguished.. One should keep in mind that C-matrices are defined over R and that the remaining rows should be orthogonal to the ones we have.. 18or21. S is the permutation x . 6.a)e b . If t = 5 the condition of the theorem is satisfied. Here n = 36. q = 3. and let G(q) be the group of permutations induced by G(q). k = 18. The other designs are new. If A and B are monomial transformations of order q + 1 such that A-1Cq+1B = Cq+1 then l0 B) E G(q). To show this. We have defined Cq+1 by numbering rows and columns with coordinates of the projective line. one has to study the transformations which leave such codes invariant.

11). 13) we see that the permutations S. In [3 ] it is also shown that PSI. Remark. e0T* := e00. 23) is the automorphism group of this design. 15) Example. 9. (15. q). (15. 13) in [3] uses the (24. 12) extended quadratic residue code over GF(3).1)eb = eaSCq+l' bcGF(q) i.b2x (b * 0) can be shown to satisfy (15. One of the applications of our Theorem (14. and the monomial transformation T* generate a group R*. 12) P(b2)-1Cq+1P(b2) = Cq+l' The permutation Tx := x 1 which was used for QR codes is now replaced by a monomial transformation which behaves in the same way as T when considered as a permutation only. P(b2). 00). (15. 11) S-1Cq+1S = Cq+1' In the same way the permutation P(b2) defined by x '. 14) Theorem. 12). 6).a . In [95] it is shown that G(q) contains a subgroup iso- morphic to PGL(2. 13) _ Cq+1' From (15. We combine this knowledge with Lemma (15. It is easily seen that R*/ {I. It is known that G(5) is M12. (15. e. If we consider the prime q = 11 and construct the 111 . -11 is isomorphic to PSL(2. Define e00T* := (_l)(q-1)/ e0. Once again it is straightforward to check that T*-1Cq+1T* (15. q) x Z2. For other values of q it is not known what the group G(q) is. Since the minimum weight of this code is 9 one finds a 5-design with parameters (24. eaT* := X(a)e_l/a (a * 0. q). eaCq+lS = x(-1)e00 + I X(b . We denote it by T*. (2. + E X(b .a)eb+l b eGF (q) = X(-1)e. 10) to find: (15. G(q) contains a subgroup isomorphic to PSL(2.

In many cases Sym2q+2 can be con- tracted to the ternary Golay code. 14) these are not equivalent 5-designs since PSL(2. 11) is not contained in PSL(2. This also yields a 5-design. By Theorem (15. 23). 13) again. The two designs have the same parameters. 112 .corresponding symmetry code gym 24 we can apply Theorem (14. For more details and further information we refer the reader to [94] and [95]. The method of attack in [95] is another example of contraction which we briefly men- tioned at the end of Chapter 14.

This design can be extended to an (e + 2)- design. The theory of uniformly packed codes was developed by J. [106]). [114]). However. We now present some of the theory of binary nearly perfect codes due to J. (ii) Ta(V) := iX E T(v) 13 uEC[x E S(u. Snover (cf. we shall be concerned with the subclass of nearly perfect codes. Goethals and H. t). 1). Let C C (R (n) be a code with minimum distance d = 2t + 1. (iii) T(v) := T(v)\Ta(v). 113 . -M. A. n I which support code words of weight 2e + 1 in a binary perfect e-error-correcting code of word length n form an (e + 1)-design with parameters (n. A. Semakov. [80]. Goethals and S. Later in this chapter we shall describe some of the connections between uniformly packed codes. First. These two statements explain the interest of both group. [57]). For all v E C we define (16. 2e + 1. V. . van Tilborg. Zaitzef (cf. L. -M. It is easy to see that the collection of (2e + 1)-subsets of 11. 6)). 2. introduced by N. Zinovjev and G. Necessary conditions for the existence of such codes are very similar to those for perfect codes.. For a survey of most of what is known about these codes we refer to [115]. 1) (i) T(v) := S(v. e. i.. t)]). [79] (5. These codes also lead to t-designs. 2..and design-theorists in perfect codes. it has been shown that for e > 1 the Golay codes are the only nontrivial perfect codes if the size of the alphabet is a power of a prime (cf. C.16 Nearly perfect binary codes and uniformly packed codes The two 5-designs connected with the Golay codes have very inter- esting automorphism groups. t + 1)\S(v. We restrict ourselves to binary codes. These codes are a special case of the class of uniformly packed codes. V. a Steiner system (cf. V. so-called two-weight codes and strongly regular graphs. .

u) = 2t + 1. d(v. e. then d(v. u) < 2t + 1. For each v E C we have !s n-t (i) IT a(v) I (t) [t 11 (ii) IT 0(v) I (t + 1) . i. u) = 2t + 1 } . Therefore by (16.(t) [t + 1] Proof. (Note that in (ii) the code word u is uniquely determined by x. (16. To estimate IN 2t+1 (v) I observe that for two code t+1 words in N2t+1(v) there are at most t coordinate places where they both differ from v. V E C. (i ). If X E Ta(v) and u E C is such that x E S (u. t) I = . For each t-subset of the coordinate places this yields at most [t+ 1] elements of N2t±1(v). [721). t) Proof. Then IT aw I = 2t + 1)IN 2t+1 (v) 1. Hence n t 1) This proves (i). 2) Lemma. t). t) I = 2t+ 1) Let N2t+1(v) := {u E C I d(v. i=0 114 . 2) (ii) we have n/ICI n t t+l t+l)-(t)[tn- IvEC T(v)I> +l]}' t The proof is complete since I (R(n) I = 2n and I S(v. If C C (R(n) is a code with minimum distance d = 2t+1 then t t-Itn 2n n/1 ICI{i=0(i)+ (n By the definition of T0(v) we see that all the S(v. 3) Theorem. ) (16. a given x E R n) can be contained in at most [n/(t + 1)] distinct sets T0(v). But then ITa(v) n S(u. Then (ii) is obvious. where v E C and U T16 (v) are pairwise disjoint sets. Since code words V EC have distance at least 2t + 1. // N2t+1(v) I < [t + 1](t)/(2t + The following estimate is a specialised version of the Johnson bound (cf.

v+a)=t+1. 3) implies equality in (16. 5) Definition. x) > t] then d(v. // (16. This divisibility condition is a well-known neces- sary condition for the existence of binary perfect codes (cf. We shall show that nearly perfect codes are uniformly packed.(16. e. 3) are called nearly perfect. is a t-design with parameters (n. (16. Equality in (16. where a < (n . Binary codes for which equality holds in the inequality of (16.(n) and if VVEC[d(v.e)(q . 6) (ii) there are [(n ..1)/(e + 1). where u E Nd(v). U E C and d(u. Proof. // 115 . x) = t + 1 for exactly [n/(t + 1)] code words v. 4) Definition. If V E C then the collection Z1 of d-subsets of 11. x) = t + 1 for exactly [(n . n } which support the words u . 7) Theorem..v has coordinate 1. (16. . By Theorem (16.t)/(t + 1)]).v. x) = t then d(v. such that A is contained in the set of coordinate places. i. The last of these VEC implies (i). If (t + 1) 1 (n + 1) then a nearly perfect code with these parameters is perfect. An e-error-correcting code C of length n over GF(q) is called uniformly packed with parameters a and 6 iff for every word x E (R (n) (i) if x has distance e to C then x has distance e + 1 to exactly a code words. . Let A be a t-subset of the coordinate places and let a be the word with 1's in the positions of A. Remark. [79] (5.t)/(t + 1)] other code words v. (i) If X E 61. (ii) if x has distance > e to C then x has distance e + 1 to exactly 0 code words. 6) Theorem.t)/(t + 1)] code words u in Nd(v) such that d(u. In the following C denotes a nearly perfect binary code in 6l (n) with minimum distance d = 2t + 1. X = [(n . 8)). Proof. d. 2) (i) and (ii) and in the estimates for IN2t+1(v) I and U T0(v). where u . (ii) If x E 6i (n). 2. the first implies (ii). 2.

The (t + 1)-subsets of the blocks of D1 form a t- design. Hence Z2 contains all the (t + 1)- subsets of 11.. i. form a (t + 1)-design with the same as in Theorem (16.m. . A _ [(n-t)/(t+l)]. Consider C and delete the k-th coordinate of each word (k fixed). [96]). If V E C then the collection az of (t + 1)-subsets of { 7. . n. 2 . n } .2 and dimension 2m .. Let m ? 3. 2. but not perfect. Proof. 8) Theorem.t)/(t + 1)]. Let Hm be the cyclic Hamming code of word length n = 2m . Hence Ck is nearly perfect.. 7). n = 2m . is a t-design with X = (n .(16. .(t + 1)[(n . // (16. The design 11) 1 can be extended to a (t + 1)-design with parameters (n+l. . Apparently 22 is the t-design formed by all the other (t + 1)- subsets o f 11. minimum distance d and I Ck 1 = I C 1. We need quite a lot of prepara- tion... 3).t) . d+l. this code is nearly perfect.1 as described in (14. e. Proof. 3). v E C. Let Bm be the cyclic code with generator (x . 2. n I which are not contained in a subset belonging to the design D1. T(v) consists of the vectors u such that d(u.. We now introduce a sequence of codes which are of great combinatorial interest.1)/(x . the more pessimistic design-theorists begin to wonder whether there are any nearly perfect codes! (16. 2.v. By now.1)m1(x)m3(x).. The resulting code Ck again has word length n.v. Chapter 9) and delete any column. We obtain the parity check matrix of a linear code with d = 3. The Preparata Codes (cf.2 . where u E T(v). u E Nd+l(v). . We have equality in (16. w) > t for all w E Nd(v). n } which support the words u . 9) Theorem. The word 1 considered as a poly- nomial in GF(2) [x] (mod xn .. Consider the parity check matrix of a binary Hamming code (cf. This BCH-code is a subcode of Hm and by the remark following Theorem (10. I where u E C. We define the code Cm by 116 .. 10) Example..1).6) we see that this code has minimum distance at least 6. 7) that the collection of (d + 1)-subsets of 11. n+1 which support the words u . Since this is true for every k it follows from Theorem (16.1) is u(x) :_ (xn . .. v)=t+1 and d(u.

(16. 11) Cm := { (m(x), i, (x)+(m(1)+i)u(x)+s(x)) I m(x) EHm, iE {0,1 }, s(x) E m}'

This is obviously a linear code with word length 2n + 1 and dimension
equal to dim Hm + 1 + dim Bm = 2n - 3m.
We now prove:

(16. 12) Lemma. Cm has minimum distance > 5.

Proof. Since the code is linear we can consider weights.
(i) Let m(x) = 0, i = 0, s(x) * 0. Then w(s(x)) ? 6 because
s(x) E Bm.
(ii)Let m(x) = 0, i = 1. Then u(x) + s(x) * 0,
u(a) + s(a) = u(a3) + s(a3) = 0 and therefore w(u(x) + s(x)) ? 5 by
Theorem (10. 6).
(iii) m(x) Since m(x) + (m(1) + i)u(x) + s(x) E Hm we are
0.

finished unless m(x) +(m(1) + i)u(x) + s(x) = 0. In that case we find by
substituting x = a3 that m(a3) = 0, i. e. w(m(x)) > 5. //
Now consider S(0, 1) in polynomial form, i. e.

S(0, 1) = { 0, 1, x, x2, ... , xn-1 } C GF(2)[x] (mod xn - 1).
(2n+1)
In (R we consider the set of vectors

(16. 13) S := { (q(x), 0, q(x)f (x)) I q(x) E S(0, 1) 1,

where f(x) is the idempotent of Hm defined in (14. 3).

The Preparata code Km is the union of cosets of
(16. 14) Definition.
Cm given by Cm + S.
It is obvious that no 2 elements of S are in the same coset. Hence
IKmI = (n + 1)ICMi.

(16. 15) Theorem. For odd m the Preparata code Km is a nearly
perfect code with minimum distance 5.

Proof. (i) Since Km is nonlinear we must now consider any
two words of Km and determine the weight of their sum. By Lemma
(16. 12) it is sufficient to consider words from different cosets. Such a

117

sum has the form

(16. 16) (m(x), i, m(x)+(m(l)+i)u(x)+s(x))+(q1(x)+q2 (x), 0, (q1(x)+q2 (x))f(x)).

Since Hm is perfect there is a unique element xs a S(0, 1) such that

m'(x) := q1(x) + q2(x) + xs E Hnl.

Observe that m' (x) = 0 if q1(x) = 0 or q2 (x) = 0. We define

m"(x) := xs(1 + f(x)) + m'(x) + m'(l)u(x).

We remark that m"(x) E HM (cf. (14. 3)). Since all polynomials are
considered mod xn - 1 we have m'(x)f(x) = 0. We replace m(x)+m'(x)
by m(x) and rewrite (16. 16) as

(16. 17) (m(x), i, m(x) + (m(1)+i)u(x) + s(x)) + (xs, 0, xs) + (0, 0, m"(x)).

(ii) From (16. 17) we find that the weight W which we wish to
determine is

(16. 18) W = w(m(x)+xs) + i +w(m(x)+(m(1)+i)u(x) +S(X) +xs + m"(x)).

(ii) If m'(x) * 0, then m'(x) has the form xs + xj + xk, and
by definition m'(a) = 0. It follows that m'(a3) = a3s(ah + a2h), where
h = j - s * 0. If, on the other hand, m'(x) = 0 then m'(a3) can also be
written as a3s(ah + a2h), where h = 0.
(iv) Since m is odd the equation a3h = 1 has h = 0 as its only
solution (ah E GF(2m)). It follows that 1 + ah + a2h 0 for all h.
Therefore m"(a3) = a3s + m'(a3) = a3s(1 + ah + a2h) 0.
(v) Let O(x) be any word with 0(1) = 0, 0(a) = as,
0(a3) = a3s(ah + a2h) for some h. Then O(x) has weight at least 4.
To prove this, first observe that q(x) clearly is not 0 and that it has
even weight. If O(x) has weight 2 this would imply that there are two
elements x 1 and x 2 in GF(2m) such that x 1 + x2 = as and
a_sxl
xi3 + x2 = a3s(ah + a2h). From this we find that + ah and
a-sx2 + ah are the solutions of the equation 1 + + 2 = 0. We showed
above that this equation has no solutions in GF(2m) if m is odd.

118

(vi) Let s(x) E Bm. Define s'(x) := m"(x) + xs + s(x). Then
in (v) we can take O(x) = s'(x). We find

(16. 19)w(s'(x)) - 4.

In the remaining part of the proof we consider 5 different possible forms
of (16. 17).
(vii) Let m(1) = 0, i = 1. By (16. 18) we have

W ? 1 + w(s"(x)),

where s"(x) := u(x) + s(x) + m"(x). Since s"(1) = s"(a) = 0 and
s"(a3) * 0 by (iv), we find w(s"(x)) ? 4 from Theorem (10. 6).
(viii) Let m(1) = 0, i = 0, m(x) * 0. Then by Theorem (10. 6) we
have w(m(x)) j 4. Then (16. 18) yields

Wow(m(x))+w(m(x)+m"(x)+s(x))-224+3-2=5,
once again using Theorem (10.6) and the fact that m(1) + m"(1) + s(1) = 1
and m(a) + m"(a) + s(a) = 0.
(ix) Let m(x) = 0, i = 0. Then we find from (16. 18) and (16. 19)

W?1+w(s'(x))-5.
(x) Let m(1) = 1, i = 0. Let O(x) := m(x) + u(x) + s'(x). Then
O(x) # 0. From (16. 18) we find

W ? w(m(x) + xs ) + w(' (x))

By Theorem (t 0.6) w(m(x) + x s ) > 4 unless m(x) is a trinomial. In that
case O(x) satisfies the conditions of (v) and hence w(o(x)) ? 4. So in
both cases W - 5.
(xi) Let m(1) = 1, i = 1. From (16. 18) we find

W > w(m(x) + xs ) + 1 + w(m(x) + s'(x)).

In the same way as in (x) we are finished unless m(x) is a trinomial, say
m(x)=xs+XC+xd. In that case m(1)+s'(1)=1, and then W < 5 would imply
that s'(x)=xs+xc+xd+xe, i. e. m(x)+m" (x)+s(x)=xs+xeoHm, a contradiction.

119

We see that this implies that every word of weight 2 or 3 is at distance 2 or 3 from 0 and A code words of C (where A is the parameter of the 3-design). 21)Example. The blocks of this design can be considered as 112 words of weight 6 and length 16 with mutual distances at least 6. A. 19). C. The 112 blocks of the 3-design mentioned above are obtained by simultane- ous cyclic shifts on the positions 0 to 6 and 8 to 14 (we assume that the overall parity check is in position 15). Therefore C is a uni- 120 . 3). Because the definition is cyclic one only has to consider the coset of C m given by taking q(x) = 1 in (16. 3). It was shown by K. Let m = 2k . equality holds in (16. 4 3 4 ). In fact in [115] it is shown that for e ? 4 there does not exist a nontrivial uniformly packed code. 6. 20)Example. If m is odd we have IKmI { 1 + (2n1 2n++3 (2 2 1) (2 3 1) . 4) and example (14. (xii) We have already seen that IKmI = 2m22n-3m=22n-2m The word length is 2n + 1 and d = 2t + 1 = 5. van Tilborg [115] that the codes treated in (16. We first give an example of a uniformly packed code. In the extended code Km these words yield 16 words of weight 6 forming a 2-design corresponding to a Hadamard matrix of order 16. 10) and (16. This coset then has 6 words of weight 5 and 10 words of weight 6. 9) the words of k weight m 6 in this code form a 3-design with parameters (4k. Let C be the 2-error-correcting binary BCH-code of example (14.1. In this case Bm consists of 0 only and Cm has minimum distance 7. The steps (vii) to (xi) prove that Km has minimum distance 5. 13). It can be shown that such a code cannot have more than 112 words and 112 can be attained only by the blocks of a 3-design. // (16. The supports of the code words of weight 6 in the ex- tended code C form a 3-design. 15) are the only nearly perfect codes. Consider the extended Preparata code K which has word length 4k. Lindstrom [78] and H. As an exercise the reader should work out the example k = 2 using the example following (10. (16. For smaller values of e there are many interesting examples (such as G2 4).[2 1 ])} = 22n+1 1) + (2 2 1) + 1 3 i. e. By Theorem (16.

perfect) iff the number of nonzero weights in the dual code C1 is e + 1 (resp. If the columns of the generator matrix of a linear code C are pairwise linearly inde- pendent (i. c2. Define the graph r (C) by taking the code words as W 1. (16.22) this code is the dual of a uniformly packed code. W 2 vertices and joining x and y by an edge iff d(x.1)/(q . 23) Example. Let C be a two-weight projective code with weights (W 1 < W 2 ). The equation c x + c2 x2 + . Let H be the parity check matrix of the Hamming code of length (qm . (16.1. 121 . This code has dimension m and qm-1 or qm-1 .1) columns represent a subspace S = PG(k .1) zeros in the positions corresponding to S. y) = W1. + cmxm = 0 represents a hyperplane in PG(m-1. Let the columns be numbered in such a way that the first (qk .. The following two theorems. It follows that a linear combination of the rows of H has (qk . show how two- weight codes are related to certain strongly regular graphs. q). 22) then C1 has only two nonzero weights. Now we leave out the columns corresponding to S and generate a code with the rows of the remaining matrix. 24) Theorem.formly packed code with a = A. q). due to Delsarte [38]. Let C be an e. e..1)/(q .1) over GF(q). The columns of H repre- sent the points of PG(m-1. e).error..1) points.qk-1 the code words have weight By Theorem (16.1)/(q ..1)/(q . . Then C is uniformly packed (resp. (16.1) zeros or (qk-1 . If e = 1 in Theorem (16. Such a code is called a two-weight code. Let c = (c1. cm). The following theorem which follows from the general theory of uniformly packed codes makes it easier to decide whether a code is uniformly packed than using the definition. C1 has minimum weight > 3) the code C is called pro- jective. For a proof of the theorem we refer to [115]. Then r(C) is a strongly regular graph. q) I 1 which either contains S or meets S in (qk-1 .1)/(q . . /3 _ A + 1.correcting linear code. 22) Theorem.

2). 27) (Pl . Consider the matrix A = A1) in which the rows of A. Then there is a two-weight code C over GF(p) such that r = r(c). 2).1)N = -a . 2). In the proof it is shown that the eigenvalues. p1. 2 of C are related by (pk (16. For a proof we refer to [38]. Therefore b1 + b2 = qk . y I is an edge in r(C). p2. p prime. Take any column of A. with distance W. the word length N of C. If there are si words of weight w. 25) Theorem. Proof. Therefore we know the sum of the distances of the rows of A 1 to the first row of A1. The vectors with a zero in this column form a two-weight subcode of C and we can therefore calculate r1 and r2 in the same way as we calculated b1 and b2.p2 . (i = 1. and the weights W1. As an example we consider the partial geometry treated in chapter 12 which has parameters r = k = 6. with integral eigenvalues a.qk-1) 2 2 So we can calculate b1 and b2. we can determine s We have proved that if {x. Hence r1 and r2 are constants. Let it have r1 zeros in Al and r zeros in A2.1) (16. So. (i = 1. t = 2. This theorem allows us to provide connections between several of the topics treated so far.p2)(h . Let C be an (n. Vf2 and let there be bi words of weight W.qk-1 nonzero 2 entries.1 and s1w1 + s2w2 is the sum determined above. Let r be a strongly regular graph on n = pk vertices. (16. // The next theorem is the converse of Theorem (16. y) is not an edge is similar. V. The calculation in the case that {x.P2)Wi = (-p2 + (-1) . to the first row of A 1 then s1 + s2 = b1 . Then every column of A has qk . then there is a constant number of vertices z joined to x and to y. k)-code over GF(q) with weights W1. 1)pk-1. Every column of A 1 has r1 zeros. 24). (i= 1.1 and b1W1 + b W = n(qk . 26) (p1 . are all the code words of weight W. Let the elementary abelian p-group Gn be a regular group of automorphisms of r. This geometry was associated with a strongly regular graph r with 122 . Now consider A1.

4)-code over GF(3) in such a way that translating over a code word is obviously an automorphism. However. So r admits the elementary abelian group of order 81 as a group of automorphisms. 30.parameters 81.22) the dual of this code is uniformly packed. p2 = -6. Therefore Theorem (16. 9. The graph r was defined from a (5. By Theorem (16. 25) states that r is associated with a two- weight code of dimension 4 over GF(3) which has word length N = 25 and weights 15 and 18. we trust that it has been enough to convince the reader that this is a rewarding area for further study. 123 . 12 and pl = 3. A better understanding of this area is found in the theory of association schemes which we introduce in the next chapter. We have been rather sketchy in our treatment of uniformly packed codes and two-weight codes.

. though this is not significant here). that the number of triangles with given colouring on a given base depends only on the colouring and not on the base. where 1 <_ i s n. the number ni of q E X with { p. Delsarte. . the number aijk of r E X with {p. k). r } e Fj depends only on i. r } e ri. . and conversely. An association scheme consists of a set X together with a partition of the set of 2- element subsets of X into n classes F1. they generalise the square lattice and triangular graphs respectively. q) E rk. The first condition asserts that each graph ti is regular. and colour an edge of the complete graph on X with colour ci if it belongs to r . It is convenient to take a set of n 'colours' cl.. and k. A complementary pair of strongly regular graphs forms an association scheme with two classes. . the second. satisfying the con- ditions given p E X. q). In the Hamming scheme H(n. q E X with {p. and ri is the set of pairs of n-tuples which agree in n . For proofs. c.. i so ri is the c. Two important classes of association schemes are those which Delsarte calls the Hamming and Johnson schemes. we refer the reader to [39]. q } E I' (i) depends only on i. . X is the set of ordered n-tuples of elements from a set of car- dinality q (in practice often taken to be a finite field. X is the set of k-subsets of a v-set.. (ii) given p. In the Johnson scheme J(v.-coloured subgraph.i coordinate places. rn. and ri the set of 124 .17 Association schemes After a short account of the theory of association schemes. { q. in which many of the concepts of classical coding theory and design theory are generalised to classes of association schemes. with k <_ i v. this final chapter contains an outline of part of the thesis of P. Association schemes were introduced by Bose and Shimamoto [181 as a generalisation of strongly regular graphs. j.

we extend the range of definition of sub- scripts to include 0. which 'describes' the action of an arbitrary permutation group in the same way. graph. If G is a transitive permutation group on a set X with the property that any two points are interchanged by an element of G. the Hamming and Johnson schemes are easily seen to be metric.i)-subset. and define no = 1. (The main difference is that un- ordered pairs are replaced by ordered pairs. where 1 i k = n. G.pairs of k-subsets whose intersection is a (k . We shall see that these schemes have properties which are not typical of association schemes in general. Still weaker conditions suffice. (The condition on G can be weakened. 2) can be identified with the set of all subsets of a v-set. and ai 1 1+1 *0 for 1 < i < n . Higman [64] has defined and studied a more general combinatorial object. k). it is enough that the per- mutation character of G is 'multiplicity-free'.j I ski + j (this is just the 'triangle inequality'). for example 125 . the graph rk is also metrically regular. Note that H(v. for 1 <_ i <_ n. and so 'contains' all the Johnson schemes J(v. which he calls a 'coherent configuration'. To simplify formulae. but these are not easy to state. The Hamming schemes provide a setting for classical error-correcting coding theory. and the Johnson schemes may be re- garded as a setting for design theory.) A scheme is metric if and only i f aijk = 0 unless l i . but for n'> 2 metric schemes seem to be fairly rare among association schemes in general. ) An association scheme is called metric if ri is the set of pairs of points at distance i in the graph rl. aij0 = &ijni' aiOk = a0k = 6ik' (This can be interpreted to mean that r0 is the graph in which every vertex is joined to itself and nothing else!) Now the parameters satisfy a number of identities.) A strongly regular graph is metrically regular. (r1 is then called a metrically regular.) D. k). (In the scheme J(2k + 1. or perfectly regular. then the orbits of G on the set of 2-subsets of X form an association scheme on X. However.1.

if the scheme is metric.. j=0 n n tZ 0al itatjm kI 0al kmaijk (The last equation comes from counting in two ways. ) C. where ix. 1. As for strongly regular graphs.. 17. 1) have interpretations in terms of the structure of this algebra. = E aijkAk' pp k=0 so the commuting symmetric matrices A0. the first and fourth assert that A i Aj =AjA and (A1 AiAj = A1 (AiAj) respectively. paths with colour sequence cl. y } E rm. as indicated in Fig. cj joining x to y. Ci. 17. and Ao = I. An span an (n + 1)- dimensional real algebra called the Bose-Mesner algebra [17] or central- iser algebra [64] of the scheme. The relations (17. in particular.A.. . then Al generates the centraliser algebra.k = ni . 1 Let Ai be the adjacency matrix of ti. 1 x y c m Fig.1) a. then this algebra is just the set of matrices com- muting with all the permutation matrices representing the group. (The second name derives from the fact that if the scheme comes from a permutation group in the manner previously described. eigenvalues of the matrices Ai 126 .) Note that. Then n (17. 2) A. = a jk' nkaijk niakji n I ai.(17. .

) _ I µmPM(Ai)pm(A.)1 m I t= O a1 itatjm n k. and determine.0 al kma. and the map A. Conversely.. . one of these is simply the function associating with each matrix in the algebra its eigenvalue on a common eigenvector.).Mi induces an algebra iso- 1 morphism. thus there are 'rationality conditions' for arbitrary schemes. = I ai. So the irreducible representations can be found by simul- taneously diagonalising the Mi.. Then µo = 1 and n IXIsio = Trace(A) _ I µmPm(Ai M=O is a system of linear equations for the multiplicities µm .A.and their multiplicities are important. Suppose µm is the multiplicity of pm. suppose the pm and µm are known. ) Let po. (These matrices are usually much smaller than the Ai. For convenience we speak of (irreducible) representations or characters of the Bose-Mesner algebra.M. (Note that the matrices can be simultaneously diagonalised. M-0 127 . Define intersection matrices Mi. and depend only on the parameters aijk and not on the particular scheme considered.k n = kI 0 aijkMk)l m ' n so M. 0 < i <_ n. I. we can assume that p0(Ai) = ni for 0 <_ i <_ n (so po corresponds to the all-1 eigenvector). 1] = Trace(A. the parameters aijk.1S.kMk. The valencies ni are determined as p0(Ai). pn be the irreducible representations..) The irreducible representations and their multiplicities are determined by. We sketch the process. . by (Mi)l m = a 1 im' Then n (MiM. or alternatively are given by n X n..

we see that Q can also be defined as A-1PTA n .=1 and ai is the average valency of FiIY. Second. the IX I X (n + 1) matrix whose (x. so it is easier to diagonalise. m=0 If the scheme is metric..Ak) = . and its eigenvalues are distinct.. the process is simplified. we determine aijk: n X I nkai.Counting triangles with edges coloured ci.k = Trace(AiA. where n= (no. for any subset Y of the point set X of any scheme. Delsarte defines. .. . y) I {x. then the rows of B corresponding to points of Y are equal to a. with ao = 1 and 128 . Thus a. From the second-last paragraph. y } e Fi I for i> 0. diagonal entries the components of u. Du is the diagonal matrix with µ = (µo. . i) entry is Bx i = I Fi(x) n Y I . the theorem of Bannai and Ito [11] and Damerell [35] on Moore graphs is a good example. its inner distribution a= (a0. 11 µn). x EX. that is. x) Ix E X) and ri = {(x. A 1 generates the Bose-Mesner algebra. the 'character table' of the Bose-Mesner algebra.= Q x. or between a member of Y and a member of X.. let Q be the matrix defined by PQ = QP = IXII. and conversely. If Y is a regular subset. First. and for any vector u. so we only have to diagonalise the intersection matrix M1. an). by ai = IYI-1 IFi n Y2I where F. one is concerned with the distance (in the graph F between two members of Y. To formalise this. cj. . Y is said to be a regular subset if every graph Fi I Y is regular (with valency ai). k) entry pi(Ak). Let P be the (n + 1) X (n + 1) matrix with (i. this matrix is tridiagonal. An error-correcting code is simply a subset Y of the point set X of a Hamming scheme. ck. and the distribution matrix B of Y. Which (n + 1)-tuples can be inner distributions of subsets? Clearly the entries must be non-negative rational numbers. Almost all applications of the rationality conditions for association schemes have been to metric schemes. 0 :5i n. nn).

This fact is applied by Delsarte to find bounds on the sizes of sub- sets whose inner distributions satisfy various conditions.. we can find polynomials 0. the important Hamming and Johnson schemes are both P.. 3) Theorem. i= O (17.. Polynomial schemes appear to be fairly rare among association schemes in general. (D n. The actual polynomials are. 'b n. . zn. P = Q. Proof. . It follows that aQ has all its components non-negative. Delsarte proves a deeper necessary condition. and neither of the P-poly- nomial and Q-polynomial conditions implies the other. such that Pik 4) k(zi) for 0 s i. (This is clear for k = 0. Delsarte calls a scheme P-polynomial if there is a choice of z0. for the Johnson schemes. n y ai = I Y I an integer. zi = i. .. If a scheme is metric.. such that deg (D k = k and Ak = lb k(A1) for 0 <_ k < n... The P-polynomial condition has a simple combinatorial inter- pretation: (17. then there are polynomials o. .. and zi = i for the Q-polynomials. but. while the Q-polynomial condition depends on that of the representations p 1. with Q replacing P in the definition. 1. zi = i(v + 1 . Q-polynomial schemes are defined similarly. remarkably.) Note that the P-polynomial condition depends on the order- ing of the graphs ri. the Krawtchouk and Eberlein polynomials respectively for the two schemes. (For the Hamming schemes. . zn with z0 = 0 for which 4'k has degree k. and follows by induction from the relation A A k I = aklk-lAk-1 + aklkAk + aklk+lAk+1 129 . apart from trivial modi- fications. Clearly. . k n. for any choice of n + 1 distinct real numbers z0.. for 0 <_ k <_ n.i) for the P-polynomials. each of degree at most n. using the tech- niques of linear programming. .. An association scheme is P-polynomial if and only if it is metric.and Q-poly- nomial.2) BTB= (IYI/IXI) PTAaQP.

For any integer c. Apart from these. and Golay codes provide examples in the Hamming schemes.) If one of these bounds is attained. while those of radius r cover X. 4) Theorem. A pair of disjoint (2e + 1)-sets forms a perfect e-code in the Johnson scheme J(4e + 2). Now we have the following result: (17. Then e r E ni < 1XI/IYI E nl i=0 i=0 (In particular. e <_ r. The spheres of radius e centred on points of Y are i=0 pairwise disjoint. Hamming. though r may not be the smallest such integer. // A subset achieving the bound of (17.1)]. and the external distance r of Y to be the number of non-zero components of aQ excluding (aQ)0. Proof. It is also possible to give a purely algebraic proof. The result follows. Let Y be a subset of X with minimal distance d and external distance r. based on the P-poly- nomial concept and the linear programming method mentioned earlier. The converse can be proved by reversing the argument. The repetition. (2e + 1). The name 'external distance' is justified by the fact that any point of X lies at distance at most r from a point of Y. In the 130 . then so is the other.with the condition aklk+l * 0. Suppose we have a P-polynomial (= metric) scheme on X. put e = [2 (d . very few examples are known. For any subset Y of X. define the minimal distance of Y to be the index of the first non-zero component of its inner distribution a after a0.) Applying the character pi yields Pik = pi(Ak) = 4ik(pi(A1)). it is simply a subset Y of X with the property that any point of X is at distance at most e from a unique point of Y. the 'sphere of radius c' centred on a point x (the set of points at distance at most c from x) contains c I ni points. 4) is called a perfect a-code. Its minimal distance is 2e + 1 (rather than 2e + 2). The minimal distance is obviously the smallest distance between two distinct points of Y.

For other perfect 1-codes in metric schemes. k) with the alternative metric defined by the graph I' k. This should be compared with the 'sphere-packing condition'.(2k+1. so the first part of the theorem can be expressed in the form '*e(z) divides In(z) in R[z]'. .. By definition of P. which asserts that %Pe(0) (the size of a sphere of radius e) divides n(0) = IX I in Z. However. [30]. then the zeros of the sum polynomial "e (z) are contained in the set {z1. for which there is no simple equivalent of the 'metric' condition for P-polynomial schemes (17.. see [15].. it can be shown that a perfect 1-code is a (k-1) .. . and the inner distribution of Y is determined by e and the parameters of the scheme. as well as theorems about such codes. There is an analogy. zn }. i=0 whence XI-1 n E. . Given an association scheme. . 5) Theorem. we have n Ak = Z PikEi. or 'formal duality'. which is useful in suggesting results about Q- polynomial schemes.3). and conversely. 1) design. {z1. . Also. zn } is the set of roots of IF n(z). The theorem of Lloyd on classical perfect codes has been generalised to arbitrary metric schemes. If a perfect e-code Y exists. independently by Delsarte [39] and Biggs [14]. between certain pairs of concepts. can also be generalised to metric schemes. Notice that fe(z) depends only on e and the parameters of the scheme. Examples are known with k = 3 and k = 5. let E V .. k. En be the minimal idempotents of the Bose-Mesner algebra (t. (17.. concepts of design theory arise more naturally in Q-polynomial schemes. = I I QkiAk k=0 131 .Johnson scheme J(2k + 1. Thus. Other concepts from coding theory (such as uniformly packed codes). E i is the matrix in (t with eigenvalue 1 on the i-th eigenspace and 0 on the others.

. and so its eigenvalues bijk are bounded above and below by those of E1 0 Ej. then the subconstituents r r(p) and r r' (p) are both strongly 132 . The vanishing of Krein parameters also has combinatorial signifi- cance. = o F bijkEk k=0 for some real numbers bijk.The Hadamard product Ei G Fj is a principal submatrix of the Kronecher product F. Then (P1 + 1)(a + p1 + 2p1p2) < (a + P1)(P2 + 1)2.We have E 1 E . as noted at the end of chapter 5.MI 0 n mQmiQmjQmk' They are known as the Krein parameters. (17. the Krein bound takes the following form.. Suppose that r and its complement are connected. Remark. 7). 7) Proposition. is closed under Hadamard product. (5.E 1.Let r be a strongly regular graph whose adjacency matrix has eigenvalues a. and (P2 + 1)(a + p2+ 2p2p1) (a + P2)(P1 + 1)2. For example. we have n EEE. given by n zµkbijk _ l% . J it There is a second operation defined on 1. Since each Ai is a zero-one matrix. which is idempotent. (17. Since Q. p2. . p1. 6) Theorem. using a theorem of M. // For strongly regular graphs. 3) is a consequence of this result. that of Hadamard (or pointwise) product: (aij) o (bij) = (aijbij). 0 Ej. G. the defining properties ensure that Ai o Aj = 6ijAi. [29]. because the following result (the Krein bound) was proved. L.. see [28]. An are the minimal idempotents with respect to this multi- plication. if equality holds in either bound of (17. = b . Krein. by L.. whence A0. . Scott [101]. 0 <_ bijk <_ 1. Proof.

1. 4) does not apply here.. For example. (17.r. then the zeros of the sum polynomial *e(z) are contained in the set {z1. .4) and (17. < l y I 1 p. We proceed to the analogues of (17. 9) Theorem. then equality holds in (17. In a Q-polynomial scheme on X. A subset attaining these bounds is called a tight 2e-design. an association scheme is Q-polynomial if and only if bijk = 0 unless I i . define the maximum strength t of a subset Y of X to be one less than the index of the first non-zero component of aQ after (aQ)0. 8) Theorem. The meaning of the maximum strength is not obvious. In the same sense.regular.. i=0 1 i=0 1 (In particular. provided certain parameter sets are excluded. Let Y be a subset of X with maximal strength t and degree s. and the degree s of Y to be the number of non-zero components of a excluding a0. If a tight 2e-design Y exists. The Krein parameters bijk are formally dual to the intersection numbers aijk. Its maximal strength is 2e (and not (2e + 1). . where a is the inner dis- tribution of Y. for reasons which will become clear. Then e s I L. s . then so is the other. put e = 2 t. 7).) If one of these bounds is attained.j I k i + j while bi i+1 # 0 for 1 s i s n-1. but the algebraic proof can easily be dualised. The degree is the number of colours occurring in the restriction of the scheme to Y. Thus we have the duality t t-+ d . 5). if r I T (p) and r I I' (p) are strongly regular for some vertex p. . We now have the result (17. 133 . and must be dis- covered in each individual case. and the ) inner distribution of Y is determined by e and the parameters of the scheme. Conversely. The simple combinatorial proof of (17. the theories of metric and Q-polynomial schemes are formally dual. e < s. zn . for any vertex p.

This generalises the construction of strongly regular graphs from quasi-symmetric designs in chapter 3.' ps-1' and s-1 1Y1 . 10)' Theorem. 8)' Theorem. k). (i) A t-design has at least (e) blocks. i=0 The importance of results (17. k) is the largest integer t for which Y is a t-design. 16) of Ray-Chaudhuri and Wilson. which is Q-polynomial with multiplicities po. The degree of a design is the number of values taken by the size of the intersection of two distinct blocks.1) in that of J(v. For this. I pi. 10) Theorem. This can be proved directly.. we have (17.2)-design of degree s. (17. We mention another result. 11) Theorem. Let us then interpret the results in the context of t-designs.2. the first part of this result is the theorem (1. or by an argument based on the character theory of the symmetric group. Of course. With the hypotheses of (17. which dualises a lesser-known result for metric schemes. (ii) A 0-design of degree s has at most (S) blocks. (17. 8)-(17. . Thus. where e = [Zt]. The second part improves (3.. remembering that we are considering only designs with k < z v having no repeated blocks. by an argument involving embedding the Bose-Mesner algebra of J(v. we need the fact that the multiplicities in J(v. The maximum strength t of a subset Y of the Johnson scheme J(v. equality holds if and only if it is a 2s-design. suppose t >_ 2s . 6). k) are given by _ v v pi = (i) . the s relations on the blocks determined by cardinality of intersection form an association scheme. the multiplicities of the eigen- 134 .(i1)° 0 <_ i < k. k . In a (2s . Then the restriction of the given scheme to Y is an association scheme. 10) to us depends on the following fact: (17. equality holds if and only if its degree is e. 8).

given any t coordinate places. for which the P and Q matrices are the Q and P matrices respectively of the original schemes. 4.1 and t and the parameters r and s. An additive scheme can be regarded as a Schur-ring on A. Fxamples of 4-designs with degree 3 are the 5 . A final word on duality. 1) designs. To a subgroup Y of A corresponds a subgroup Y' of A' (the set of characters whose restriction to Y is trivial). 1) design. 10)' is 'best possible'. The inner distributions a and a' are related by iY la' = aQ. was given by Noda [92]. This is a 3 . where they are defined. JY' Ia = a'P. it is possible to define dual schemes. but the number of blocks disjoint from a given pair of disjoint blocks depends on whether the given blocks are parallel or not. who proved the first part of (17.(24.(11. The maximum strength of a subset is simply its maximum strength as an orthogonal array.(16. For consider the design of points and planes in AG(4. every t-tuple occurs equally often in those places in a member of Y. Maximum strength has a well-known interpretation in Hamming schemes as well.(23. as defined by Rao [97]. 5. 8. For certain schemes. (17. and this duality agrees with the concept defined by Tamaschke [111]. q) with the property that. 2). 14). Thus duality interchanges the parameters d . 17).values given by (17. (An orthogonal array of strength t is a subset Y of the point set of H(n. 4) (residual of the first) and 4 . 4 . 8. and in fact are isomorphic to their duals. 1). 135 . 10) agree with those found there. the relation I YIa' = aQ expresses the MacWilliams identities (9.) The analogue of (1. the deter- mination of tight 4-designs in Hamming schemes. 8) in this case. The point sets of the scheme and its dual are identified with A and its dual A' (the group of characters of A). notably the additive schemes (those admitting a transitive Abelian automorphism group A). The Hamming schemes are additive. and has degree 3 (since two blocks have at most two common points). for them.

307-24. Quart. [4] F. F. Algebraic theory of codes II. Mass. F. F. Aigner. [3] F. Jr. 485-92. On finite Moore graphs. Math. Research Lab. and H. F. Tokyo. Math. F. Bedford. New 5-designs. Aigner. Mattson. and J. On a combinatorial generaliza- tion of 27 lines associated with a cubic surface. Jr. F. F. Reidel. 138 (1974). D. J. Reidel. F. Jr. Ovals from the point of view of coding theory. Combinatorial Theory. (A) 27 (1979). [10] F.. D. 3 (1972). [2] F. F. van Lint. Assmus. A. Assmus. [6] E. J. Math. F. ). Bannai. Sci. Sachar. Mezzaroba and C. Fac. Jr. H. and H. Higher Combinatorics (M. Assmus. Z. Jr. H. Jr. Jr. 10 (1969). 28 (1977). 21-32. [9] F. Self- orthogonal Steiner systems and projective planes. Dordrecht (1977). . Mattson. Mattson. Higher Combinatorics (M. On weights in quadratic residue codes. 6 (1969). Mattson. Contractions of self- orthogonal codes. Jr. Jr. Combinatorial Theory. and H. Salwach. Assmus. Assmus. Szekeres. F. Math. 136 . ed. On tight designs. [5] E. Soc. and H. Ito. Report AFCRL-0013. Planes and biplanes. Jr. Bannai and T. (Oxford). [7] F. Jr. Univ. 89-96. Ovals in projective designs. of Sylvania Electronic Systems. ed. Discr. F. (1971). Discr. J. ). and Marcia Guza. Jr. 122-51. J. Assmus. Assmus. 20 (1973). Assmus. Ahrens and G. Math. Appl. Jr. J. Dordrecht (1977).References [1] R. 1-20. .. 205-12. F. J. [11] F. 213-16. W. 191-208. 433-48. F. J. and H. Mattson. [8] F. 3 (1972). Austral.

C. Cameron. Rome (1977). Combinatorial Theory (A). . 111 (1969). 215-20. C. 131 (1973). Bose. Bremner. Buekenhout. [24] F. Math. M. Shimamoto. C. [19] R. Extending symmetric designs. Austral. Accad. (1971). [27] P. Bruck. Combinatorial Theory (B) 15 (1973). Math. 151-84. Advances in Mathematics. 389-419. Classification and analysis of partially balanced incomplete block design with two associate classes. MacWilliams and N. [26] P. S. M. 289-96. R. C. and partially balanced designs. Teorie Combinatorie (Tomo I). . (1974). -Math. Math. Une caracterisation des espaces affines base sur la notion de droite. M. k-arcs and nets of order ten. IEEE Trans. J. Mesner. [25] P. S. M. Perfect codes in graphs. L. eds. A. Bose and D. 167 (1979). Strongly regular graphs. 21-38. Statist. Shrikhande and N. [20] A. Pacific J. Finite sets II: uniqueness and embedding. 66-79. (A). [14] N. Blocking sets. Cameron. Lincei. [21] R. Ann. Math. F. Sci. 14 (1973). 47 (1952). partial geometries.[12] F. Amer. N. L. Math. Fisher. Biggs. J. Z. Biggs. 409-14. [18] R. H. Inf. Edge-regular multigraphs and partial geometric designs. ). U. L. Mavron. Math. . 421-57. Lecture Notes 13. 353-6.. Z. P. J. [17] R. Finite groups of automorphisms. [23] F. 175 (1970). . Perfect codes and distance-transitive graphs. Bussemaker and J. Bruen and J. J.Y. 6. Lecture Notes. On doubly transitive permutation groups of degree prime squared plus one.. 30 (1959). Acad. 137 . Biggs. Seidel. Combinatorics (T. Symmetric Hadamard matrices of order 36. P. 13 (1963). J. 85-101. 1-8. Berlekamp. Sloane. S. McDonough and V. . Assoc. [15] N. J. [22] A. Bose and T. Biplanes. U. On linear associative algebras corresponding to association schemes of partially balanced designs. Pacific J. J. C. L. [13] N. Statist. S. P. J. 10 (1973). Osaka J. 367-71. C. Theory IT-18 (1972). Ann. C. [16] R. Bose. 317-20. Soc. C. Singhi. A diophantine equation arising from tight 4-designs. Naz. Gleason's theorem on self-dual codes. J. C. 13 (1963). L. . Cameron. J.

Vienna (1975). 138 . [40] P. J. Camb. Longo (ed. 196-206. 319-25. Seidel. On semipartial geometries. Polarities of generalized hexagons and perfect codes. [29] P. (A). Goethals and J. 26 (1978). Math. (New Ser. 25 (1978). G. Thas and S. .32. J. Geometriae Dedicata. London (1974). J. Orthogonal matrices with zero diagonal. Nederl. Canad. J. Algebra (vol. 55 (1978). J. Majority logic decodable codes derived from finite inversive planes. 242-50. Strongly regular graphs having strongly regular subconstituents. Global quadratic abelian codes.) 4 (1960). 23 (1971). Phil. A. Seidel. Record Peking Math. On Moore graphs. 81 (1978) (Indag. Record Peking Math. J. -M. Seidel. Thas. Debroey and J. n2 = 15. J. Damerell. in Information Theory (CISM Courses and Lectures no. 219). Cameron. Delsarte. (New Ser. [33] Chang Li-Chien. [28] P. 40 (1978)). J. Wetensch. [39] P. Algebra. spherical designs. Delsarte. A. 525-8. Wiley. 227-36.-M. . An algebraic approach to the association schemes of coding theory. Akad. Discrete Math. [35] R. 816. [30] P. The Krein condition. Philips Research Reports Supplements 1973. Camion. 74 (1973). 47-64. J. 18 (1971). [31] P. [37] P. . 5 (1976). ) 3 (1959). Proc.. Math. M. [38] P. The uniqueness and non-uniqueness of the trian- gular association schemes. M. J. 317-18. Kon. Springer-V erlag. ). J. J. [34] P. Cameron. Cameron. Inform. Association schemes of partially balanced designs with parameters v = 28. J. Soc. Cohn. 604-13. nl = 12. Norton algebras and permutation groups. Payne. Goethals and J. -M. no. Sci. 1). Delsarte. Goethals and J. 12-18. Proc. and 2 = 4. E. Control. [36] I. 257-80. Combinatorial Theory (A). [32] Chang Li-Chien. Sci. II. 10. Weights of linear codes and strongly regular normed spaces. 3 (1972). Delsarte.

363-88. M. Higman. H. F. [52] A. 915-34. Gewirtz. F.[41] P. Math. 157 (1964). 215-35. Bull. H. [46] M. [44] P. Dembowski. Ranks of incidence matrices of Steiner triple systems. Weight polynomials of self-dual codes and the MacWilliams identities. vol. 6 (1977). 11 (1971). Denniston. Some new 5-designs. eds). Math. 69 (1963).. Goethals. J. van Lint. 179-205. Math. Chapter 17 in A survey of combinatorial theory (J. Seidel. Soc. Driessen. Some combinatorial aspects of coding theory. Dembowski. Bull. and J. Amsterdam (1973). Canad. Discr. N. Goethals and J. Delsarte. Ars Combinatoria. [47] J. Vandensaval. Feit and G. Berlin- Heidelberg-New York (1968). Hubaut and M. H. 263-7. J. 3. [48] L. Amer. (to appear). Orthogonal matrices with zero diagonal. Math. Sos. Springer-Verlag. Srivastava et al. G. Seidel. Renyi and V. Une propriete extremale des plans projectifs finis dans une classe de codes equidistants. 1 (1976). 8 (1976). Math. 139 . 353-8. J. The nonexistence of certain generalized polygons. Combina- torial Theory (A). 211-5. M. Gleason. North-Holland Publishing Company. Algebra. J. [51] A. 6 (1973). Canad. Goethals. Mobiusebenen Gerader Ordnung. Deza. Studies Math. London Math. . . [49] P. 19 (1967). . [43] P. -M. . 434-46. 178-86. M. Geometriae Dedicata. Spherical codes and designs. Inversive planes of even order. [55] J. On the Golay perfect binary code. [50] W. Goethals and J. . J. 1 (1964). Math. Frdbs. Dembowski. [53] J. H. J. Hungar. Ann. 850-4. 33-42. M. . Linear codes supported by Steiner triple systems. [45] R. Z. [54] J. X. Soc. Actes congres Intern. -M. -M. Finite geometries. Frederix. 1001-10. Doyen. Graphs of maximal even girth. On a problem in graph theory. . [42] P. Math. A. 21 (1969). J. 1978. 1 (1966).

Designs derived from permutation groups. Higman. G. Goethals and S. Shult). 105 (1968). Math. J. Jr. Singleton. P. 492-7. J. R. F. Combinatorial considerations about_permutation groups. [67] D. Combinatorial theory.. Combinatorial Theory. [64] D. R. Develop. 352. Lane and D. Hall. [59] M. Hall. Higman and C. Hall. and F. R. 597-614. Blaisdell. Math. and W. 31 (1960). [58] J. [68] A. Hughes. Math. Canad. 3 (1973). 2 (1972). G. Connor. -M. Hall. [62] M. 87 (1965). S. J. Discr. Wales. J. T. Mathematical Institute. . 4 (1960). 22 (1970). 000. Janssen. and strongly regular graphs. Remark on Shult's graph extension theorem. . 8 (1970). L. Res. On t-designs and groups. Res. F. A. Hoffman and R. Hoffman. 35-41. W.[56] J. Strongly regular graphs derived from combinatorial designs. Hale. 761-78. Math. Goethals and J. IBM J. . generalized quadrangles. 153-226. [57] J. An embedding theorem for balanced incomplete block designs. Automorphisms of Steiner triple systems. Oxford (1971). Statist. Hall. 65-88. Nearly perfect binary codes. 497-504. On the uniqueness of the triangular association scheme. Amsterdam (1973). M. [65] D. Ann. (1967). 17 (1977). Higman. A. Discrete Math. C. A simple group of order 44. G. 12-22. [69] A. J. J. Jr. I. [63] N. Sims. M. [66] D. Develop. Equidistant codes with distance 12. 140 . 6 (1953). North-Holland Publishing Company. J. [70] D. Canad. Jr. 110-13. . Partial geometries. Atti del convegno geometria com- binatoria e sue applicazioni. . 71-83. 460-72. . -M. Z. Hiroshima Math. Seidel. Amer. Mass. . Snover. Jr. 4 (1960). J. Higman. Kolen and J. H. Math. IBM J. van Lint. Perugia (1971). G. On Moore graphs of diameters 2 and 3. [60] M. Math. Gagen. J. . J. Jr. Waltham. On the p-rank of the incidence matrix of a balanced or partially balanced incomplete block design and its applications to error-correcting codes. Hamada. Finite groups '72 (ed. . [61] M.

van Lint and A. Springer-Verlag. Amsterdam (1977). [79] J. J. [75] W. Berlin-Heidelberg-New York (1969). Mathematical Centre Tracts 55 (1974). AI. 28 (1977). IEEE Trans. 309-28. (Oxford). [74] W. J.[71] N. [84] J. H. 493-522. Moore geometries and rank 3 groups having µ=1. LMS Lecture Notes 13. [72] S. A new upper bound for error-correcting codes. Wetensch. Lecture Notes in Mathematics 84. Turku. Non-embeddable quasi-residual designs. [78] K. Nederl. M2 4 and the associated Steiner systems. 730-7. van Lint and F. Berlin (1971). P. H. Tight 4-designs. H. J. M2 3. H. Mavron. Lander. [73] W. Math. Constructions of strongly regular graphs. Springer-Verlag. Math. Equidistant point sets. Math. 24 (1978). Inform. M. M. Dimension and embedding theorems for geometric lattices. S. Transitive Erweiterungen endlicher Permutations- gruppen. van Lint. H. . 12 (1975). 201. van Lint. [80] J. Quart. [81] J. van Lint. Johnson. Univ. Lecture Notes in Math. Theory. Lemmens and J. MacWilliams. Schrijver. Generalized quadratic residue codes. McDonough and V. [85] H. J. [76] F. [82] J. Sloane. J. 169 (1975). to appear in Combinatorica. van Lint. North-Holland. To appear. (A). H. J. Kantor. two-weight codes and partial geometries by finite fields. On the Mathieu groups M2 2. Akad. Theory IT-8 (1962). Johnson. IEEE Trans. Algebra. Combinatorics (T. Proc. [83] J. W. C. [86] F. in Combinatorics part 1 (M. Coding theory. 169-76. 158-78. 81 (1978). U. Ito. H. The nonexistence of unknown nearly perfect binary codes. Combinatorial Theory (A). (1969). J. P. 141 . Ann. The theory of error- correcting codes. 269-75. 125 (1972). 494-512. eds). Kantor. 193-214. 24 (1973). Inform. Lindstrom. MacWilliams and N. . Recent results on perfect codes and related topics. 173-95. Osaka J. 17 (1974). [77] P. Hall and J. Equiangular lines. Seidel. C. 203-7. Z. Kon. Ser. Luneburg. van Lint editors). M. A.

Osaka J. 19 (1979). Math. 13 (1968). 45 (1977). [99] H. 14 (1973). Math. A. Combinatorial Theory. 803-19. A. 4 (1968). Symmetry codes over GF(3) and new 5-designs. Error-correcting codes associated with projective planes. 737-44. Sloane. Osaka J. T. Combinatorial Theory. Sachar. MacWilliams. The MacWilliams identities for nonlinear codes. K. Seidel. Wilson. J. J. 417-25. Statist. Some properties of character products. Algebra. Sloane and J. Thompson. 0) adjacency matrix having eigenvalue 3. Sloane and J. Information and control. Math. [97] C. Ito and R. Tutte) Acad. London Math. A. Combinatorial mathematics. Cambridge. Pless. [100] H. Enomoto. (1973). W. N. 281-98. Inform. 18 (1975). G. 259-65. [89] C. Bell System Tech. . Strongly regular graphs. Pa. Bethlehem. N. [101] L. 128-39. 378-400. J. Noda. Soc. 51 (1972). (2). Ph. Zametski. P. 1. Factorial experiments derivable from combinatorial arrangements of arrays. Ryser. 16 (1979). [95] V. Noda. Lehigh University. A class of optimum nonlinear double-error- correcting codes. I. Linear Alg. Combinatorial Theory ser A. 39-43. 66-78. [103] J. J. Ja. [92] R. Petrenjuk. 116-25. 119-42. [98] D. J. Scott. J. Massey. [93] A. On the existence of a projective plane of order 10. [94] V. Mass. Rao. [96] F. Seidel. Roy. J. J. On orthogonal arrays of strength 4 achieving Rao's bound. Recent progress in com- binatorics (ed. J. 1 (1968). Appl. . J. Wiley.D. J. L. M. Strongly regular graphs with (-1. J.[87] F. -M. Press (1969). On t-designs. 142 . Press. Goethals.. Mallows and N. Preparata. (1963). R. Symmetry codes and their invariant subcodes. Thesis. 22 (1973). MacWilliams. to appear. T. M. Pless. Tight 4-designs. Threshold decoding. 9 (1947). New York (1965). L. Ray-Chaudhuri and R. Soc. . Control. 185-97. N. L. [88] F. [102] J. J. 12 (1972). An upper bound for self- dual codes. [90] J. 12 (1975). 188-200. [91] H. J. J.

J. . Szeged (1978). Proc. Seidel. Z. W. The uniqueness of the L2 association scheme. 257-74. A representation of the Mathieu group M2 4 as a collineation group. Soc. Benjamin. Coll. [108] F. S. 781-98. Proc. Tietaviinen. Brauer and C. . translation in Problems of Information Transmission. Proc. Is there a (72. Boston (1916). Technological University Eindhoven (1976). di Mat. Sah). Dordrecht. [112] D. [111] O. On the nonexistence of perfect codes over finite fields. Tamaschke. Taylor. Appl. Shult. Thas. Theory of finite groups (ed. Lincei. ). H. Shrikhande. London Math. Math. 88-96. 350. On the isomorphism of two groups of orders 44. F. Pura ed Appl. Int. Seidel and D. SIAM J. Two-graphs. [116] J.[104] J. Reidel. van Tilborg. 30-9. Uniformly packed codes. Naz. Algebraic methods in graph theory. Zur Theorie der Permutationsgruppen mit Regularer Untergruppe I and II. [115] H. 251. 80 (1963). . Math. ed. A. A. Zinovjev. [117] O. 183-99. [1131 J. 143 . Statist. (1977). Combinatorics of partial geometries and generalized quadrangles. Theorie Combinatorie. . 36) d = 16 self-dual code? IEEE Trans. 7 (1971). Zaitzef. Semakov. C. . Roma (1973). New York (1969). on Information Theory IT-19. G. A. The graph extension theorem. F. V. 328-52 and 443-65. Uniformly packed codes. Probl. Projective geometry. 199-238. A. Higher Combinatorics (M. J. [107] S. 30 (1959). Ginn and Co. C. 000. R. Amer. 38-50 (Russian). Ann. Young. 35 (1977). Todd. A survey of two-graphs. Thesis. A. (1973). [109] C. V. . Math. 33 (1972). Aigner. 24 (1973). 278-84. 7 (1971). Sims. J. Peredaci Inform. D. (ser. Math. 4) 71 (1966). [105] J. Acc. [110] N. a second survey. F. Taylor. [114] A. Regular 2-graphs. Veblen and J. [106] N. Ann. V. Sloane. . Soc.

251-60. Quadratic residue codes and symplectic groups. [119] H. 265-75. J. Math. 256-64. [121] E. and Math. 29 (1974). [120] F. and C. 150-71. Algebra. J. Salwach and J. J. Math. Ward. Wielandt. Abh. 6. [123] C. Press. A. Internat. Internat. Salwach. Math. 2) designs. Seminar Hamburg 12 (1938). J. Acad. Assmus. New York- London (1964). F. N. . 2 (1979). and Math. Witt. Finite permutation groups. 2 (1979). The (16. J. The four known biplanes with k = 11. fiber Steinersche Systeme. .[118] H. Abh. Jr. Mezzaroba. Die 5-fach transitiven Gruppen von Mathieu. Witt. [122] F. Math. Seminar Hamburg 12 (1938). Sci. 144 . 261-81. Soc.

108 circuit 37 two-weight 113. q) 72 extended 64 plane 7 generalised quadratic residue 105 symplectic geometry 8 Hamming 69. 85. 115 cocycle 50 complement (of a graph) 16 code degree 133 BCH 71 Desarguesian 7 binary Golay 86. 100.Index adjacency matrix 17 code affine geometry 7 Euclidean geometry 81 permutation group. 100. 78 alphabet 61 linear 62 arc 11. 108 design cyclic 68 derived 8 dual 63 Hadamard 5 equidistant 81 pair 17 equivalent 63 145 . 115 Assmus-Mattson theorem 101 perfect 62. 86. 84 Bruck-Ryser-Chowla theorem 4 symmetry 107 centraliser algebra 22. 126 systematic 63 check polynomial 69 ternary Golay 85. AGL(1. 121 C-matrix 107 uniformly packed 113. 130 association scheme 124 Preparata 116 additive 135 primitive BCH 71 dual 135 projective 121 metric 125 projective geometry 81 polynomial 129 quadratic residue 97 Reed-Muller (RM) 78 block design 2 self-dual 84 Bose-Mesner algebra 126 self-orthogonal 81. 89 nearly perfect 113.

52 error 61 point 33 error-correcting code 62 pseudo-geometric 34 Fuclidean geometry 7 rank-3 20 extension regular 17 of a code 64 Shrikhande 34 of a design 8. 55 45. 17 block 26 inner distribution 128 Clebsch 19 inversive plane 10 complete 16 146 . 92 symmetric 3 Higman-Sims 30 t. L2(n) 19. 41 triangular. 39. 32. 33 external 130 Moore 37. 130 lattice. 51. 58 tight 13. 35 connected 37 quasi-symmetric 25 geometric 34 residual 8 Gewirtz 19. 37. 1 Hoffman-Singleton 38. 133 ladder 19 diameter 37 Latin square 36 distance 61. 59 group algebra 105 1-factor 56 finite geometry 76 Hadamard design 5 Fisher's inequality 3 matrix 5 Hamming distance 61 generalised quadrangle 35 scheme 124 generator matrix 63 hyperplane 6 polynomial 68 girth 37 idempotent 95. 33. 67 complete k-partite 19 quasi-residual 8. 34. 54 designed 71 line 25. 25. 121 of a symmetric design 9.design graph Paley 5. 131 graph 16 incidence matrix 2. of a permutation group 52. 51 strongly regular 17. 58 minimal 130 null 16 distribution matrix 128 Paley 19 Petersen 19. T(n) 19.

128 secant 11 sphere 61 square root bound 100 standard form 63 Steiner system 1. 89.Johnson bound 114 strength 133. 23. 82 path 3 7 passant 11 polarity 3. 85. 45 projective geometry 6 plane 6. 135 scheme 124 switching 24. 93 parity check 63 matrix 63 overall 64 partial geometry 33. 106 valency 17 majority decoding 74 weight 61 Mathieu groups 14. PGL(2. 131 147 . 50 Krein bound 132 tangent 11 threshold decoding 74 Latin square 16. 60 enumerator 65 Mobius plane 10. 36 two-graph 49 linear fractional group. 106 word 61 multiplicities 21. q) 10. 70. 88 rationality condition 21 regular 17. 127 length 61 net 36 null polarity 46 orthogonal check set 74 oval 11. 51.