Given υ objects 1, 2, · · ·, υ, a relation satisfying the following conditions is said to be an m-class partially balanced association scheme:

A. Any two objects are either 1st, 2nd, · · ·, or mth associates, the relation of association being symmetrical.

B. Each object α has n_i ith associates, the number n_i being independent of α.

C. If any two objects α and β are ith associates, then the number of objects that are jth associates of α and kth associates of β is p_jkⁱ and is independent of the pair of ith associates α and β.

The constants υ, n_i, p_jkⁱ are the parameters of the association scheme. A number of identities connecting these parameters were given by the Indian mathematicians Bose and K.R. Nair in 1939, but Bose and the U.S. mathematician D.M. Mesner in 1959 discovered new identities when m > 2.

A PBIB design is obtained by identifying the υ treatments with the υ objects of an association scheme and arranging them into b blocks satisfying the following conditions:

A. Each contains k treatments.

B. Each treatment occurs in r blocks.

C. If two treatments are ith associates, they occur together in λ_i blocks.

Two-class association schemes and the corresponding designs are especially important both from the mathematical point of view and because of statistical applications. For a two-class association scheme the constancy of υ, n_i, p₁₁¹, and p₁₁² ensures the constancy of the other parameters. Seven relations hold:

.

Sufficient conditions for the existence of association schemes with given parameters are not known, but for a two-class association scheme Connor and the U.S. mathematician Willard H. Clatworthy in 1954 obtained some necessary conditions:

Orthogonal Latin squares

A Latin square of order k is defined as a k × k square grid, the k² cells of which are occupied by k distinct symbols of a set X = 1, 2, . . ., k, such that each symbol occurs once in each row and each column. Two Latin squares are said to be orthogonal if, when superposed, any symbol of the first square occurs exactly once with each symbol of the second square. Two orthogonal Latin squares of order 4 are exhibited in Figure 2.

A set of mutually orthogonal Latin squares is a set of Latin squares any two of which are orthogonal. It is easily shown that there cannot exist more than k − 1 mutually orthogonal Latin squares of a given order k. When k − 1 mutually orthogonal Latin squares of order k exist, the set is complete. A complete set always exists if k is the power of a prime. An unsolved question is whether there can exist a complete set of mutually orthogonal Latin squares of order k if k is not a prime power.

Many types of experimental designs are based on Latin squares. Hence, the construction of mutually orthogonal Latin squares is an important combinatorial problem. Letting the prime power decomposition of an integer k be given, the arithmetic function n(k) is defined by taking the minimum of the factors in such a decomposition

.

Letting N(k) denote the maximum number of mutually orthogonal Latin squares of order k, the U.S. mathematician H.F. MacNeish in 1922 showed that there always exist n(k) mutually orthogonal Latin squares of order k and conjectured that this is the maximum number of such squares—that is, N(k) = n(k). There was also the long-standing conjecture of Euler, formulated in 1782, that there cannot exist mutually orthogonal Latin squares of order 4t + 2, for any integer t. MacNeish’s conjecture, if true, would imply the truth of Euler’s but not conversely. The U.S. mathematician E.T. Parker in 1958 disproved the conjecture of MacNeish. This left open the question of Euler’s conjecture. Bose and the Indian mathematician S.S. Shrikhande in 1959–60 obtained the first counterexample to Euler’s conjecture by obtaining two mutually orthogonal Latin squares of order 22 and then generalized their method to disprove Euler’s conjecture for an infinity of values of k = 2(mod 4). Parker in 1959 used the method of differences to show the falsity of Euler’s conjecture for all k = (3q + 1)/2, in which q is a prime power, q ≡ 3(mod 4). Finally these three mathematicians in 1960 showed that N(k) ≥ 2 whenever k > 6. It is pertinent to inquire about the behaviour of N(k) for large k. The best result in this direction is due to R.M. Wilson in 1971. He shows that N(k) ≥ k^1/17 − 2 for large k.

Orthogonal arrays and the packing problem

A k × N matrix A with entries from a set X of s ≥ 2 symbols is called an orthogonal array of strength t, size N, k constraints, and s levels if each t × N submatrix of A contains all possible t × 1 column vectors with the same frequency λ. The array may be denoted by (N, k, s, t). The number λ is called the index of the array, and N = λs^t. This concept is due to the Indian mathematician C.R. Rao and was obtained in 1947.

Orthogonal arrays are a generalization of orthogonal Latin squares. Indeed, the existence of an orthogonal array of k constraints, s levels, strength 2, and index unity is combinatorially equivalent to the existence of a set of k − 2 mutually orthogonal Latin squares of order s. For a given λ, s, and t it is an important combinatorial problem to obtain an orthogonal array (N, k, s, t), N = s^t, for which the number of constraints k is maximal.

Orthogonal arrays play an important part in the theory of factorial designs in which each treatment is a combination of factors at different levels. For an orthogonal array (λs^t, k, s, t), t ≥ 2, the number of constraints k satisfies an inequality

in which λs^t is greater than or equal to a linear expression in powers of (s − 1), with binomial coefficients giving the number of combinations of k − 1 or k things taken i at a time (i u).

Letting GF(q) be a finite field with q = p^h elements, an n × r matrix with elements from the field is said to have the property P_t if any t rows are independent. The problem is to construct for any given r a matrix H with the maximum number of rows possessing the property P_t. The maximal number of rows is denoted by n_t(r, q). This packing problem is of great importance in the theory of factorial designs and also in communication theory, because the existence of an n × r matrix with the property P_t leads to the construction of an orthogonal array (q^r, n, q, t) of index unity.

Again n × r matrices H with the property P_t may be used in the construction of error-correcting codes. A row vector c′ is taken as a code word if and only if c′H = 0. The code words then are of length n and differ in at least t + 1 places. If t = 2u, then u or fewer errors of transmission can be corrected if such a code is used. If t = 2u + 1, then an additional error can be detected.

A general solution of the packing problem is known only for the case t = 2, the corresponding codes being the one-error-correcting codes of the U.S. mathematician Richard W. Hamming. When t = 3 the solution is known for general r when q = 2 and for general q when r ≤ 4. Thus, n₂(r, 2) = (q^r − 1)/(q − 1), n₃(r, 2) = 2^r−1, n₃(3, q) = q + 1 or q + 2, according as q is odd or even. If q > 2, then n₃(4, q) = q² + 1. The case q = 2 is especially important because in practice most codes use only two symbols, 0 or 1. Only fairly large values of r are useful, say, r ≥ 25. The optimum value of n_t(r, 2) is not known. The BCH codes obtained by Bose and Ray-Chaudhuri and independently by the French mathematician Alexis Hocquenghem in 1959 and 1960 are based on a construction that yields an n × r matrix H with the property P_2u in which r ≤ mu, n = 2^m − 1, q = 2. They can correct up to u errors.