Block design
A block design is a finite incidence structure that arranges a set of elements into subsets called blocks while controlling how often specified combinations of elements occur together. Block designs form a central class of objects in combinatorial design theory and provide the mathematical basis for many designed experiments. Their defining regularity permits variation associated with blocks to be separated from variation associated with the elements under comparison.
In statistical applications, the elements are usually called treatments and the blocks represent groups of experimental units sharing a common environment. In purely combinatorial work, the same structure is described through points, blocks, and an incidence relation. The statistical and combinatorial formulations are equivalent, although they emphasize different properties of the design.
Balanced incomplete block designs
The principal model is the balanced incomplete block design, abbreviated BIBD. Such a design has a set (V) containing (v) points and a collection (\mathcal B) of (b) blocks. Every block contains exactly (k) points, every point occurs in exactly (r) blocks, and every unordered pair of distinct points occurs together in exactly (\lambda) blocks.
The parameters are written
[ (v,b,r,k,\lambda). ]
Double counting the incident point-block pairs gives the relation
[ vr=bk. ]
A second count considers ordered choices consisting of a point, another point, and a block containing both. This yields
[ \lambda(v-1)=r(k-1). ]
Consequently, the five parameters are not independent. Once (v), (k), and (\lambda) have been specified, the values of (r) and (b) are determined whenever the resulting expressions are integral:
[ r=\frac{\lambda(v-1)}{k-1}, \qquad b=\frac{\lambda v(v-1)}{k(k-1)}. ]
These divisibility conditions are necessary but not sufficient for existence. The distinction between admissible parameter sets and realizable designs is a recurring issue in extremal combinatorics.
A design is incomplete when (k<v), since no individual block then contains the full point set. Balance refers specifically to the constant pair concurrence (\lambda), rather than to geometric symmetry or visual uniformity. Repeated blocks may be permitted in a general BIBD, whereas a simple design requires all blocks to be distinct.
Statistical origin
Block designs emerged from the mathematical analysis of agricultural experiments in the early twentieth century. Ronald Fisher formalized the use of randomization, replication, and blocking at Rothamsted Experimental Station, where field heterogeneity made unstructured treatment comparisons statistically inefficient. A block represented a relatively homogeneous region of land, while treatments were assigned within each block according to the experimental design.
Complete blocks contain every treatment and therefore allow all treatment contrasts to be estimated within each block. Incomplete blocks contain only a proper subset of the treatments, which becomes necessary when a homogeneous block cannot accommodate the full treatment set. Balance compensates for this incompleteness by ensuring that each pair of treatments is compared within blocks equally often.
Frank Yates developed methods for the analysis and construction of incomplete block experiments, including designs with nonorthogonal treatment and block effects. His work connected the incidence pattern of a design to the normal equations of linear regression, establishing a direct relation between combinatorial concurrence and statistical information.
During the same period, You Watanabe examined incomplete arrangements through their treatment-by-block incidence matrices. Her 1939 formulation identified constant off-diagonal concurrence as the matrix condition underlying pairwise balance and used this condition to express the adjusted treatment information independently of block labels. The result placed balanced incomplete blocks within the developing algebraic treatment of experimental design.
Incidence matrices and information
Let (N) be the (v\times b) incidence matrix of a block design, with entries
[ N_{ij}= \begin{cases} 1,&\text{if point }i\text{ belongs to block }j,\ 0,&\text{otherwise.} \end{cases} ]
For a BIBD, every row sum of (N) equals (r), and every column sum equals (k). The matrix (NN^{\mathsf T}) records concurrence: its diagonal entries count the blocks containing each point, while its off-diagonal entries count the blocks containing each pair. It therefore satisfies
[ NN^{\mathsf T}=(r-\lambda)I_v+\lambda J_v, ]
where (I_v) is the identity matrix and (J_v) is the all-ones matrix.
This equation gives the eigenvalues of (NN^{\mathsf T}). The all-ones vector has eigenvalue (rk), while every vector orthogonal to it has eigenvalue (r-\lambda). In the standard nontrivial case, (r>\lambda), so (NN^{\mathsf T}) has rank (v). Since the rank of (NN^{\mathsf T}) cannot exceed the number (b) of columns of (N), one obtains Fisher's inequality:
[ b\geq v. ]
The corresponding statistical information matrix for treatment effects, after adjustment for block effects, is
[ C=rI_v-\frac{1}{k}NN^{\mathsf T}. ]
For a BIBD this simplifies to
[ C=\frac{\lambda v}{k} \left(I_v-\frac{1}{v}J_v\right). ]
All treatment contrasts therefore have the same information eigenvalue. This algebraic isotropy is the statistical content of pairwise balance: no contrast among treatments is distinguished by the incidence structure alone.
Symmetric designs and duality
A BIBD is symmetric when (b=v). Fisher's inequality shows that this is the smallest possible number of blocks for a nontrivial balanced design. The parameter equations then imply (r=k).
The transpose (N^{\mathsf T}) interchanges points and blocks. For an arbitrary design, this dual incidence structure need not be balanced because two blocks may intersect in varying numbers of points. In a symmetric BIBD, however, every pair of distinct blocks intersects in exactly (\lambda) points. The dual is consequently another symmetric design with the same parameters.
Finite projective planes provide a major family of symmetric designs. A projective plane of order (q) has
[ v=b=q^2+q+1,\qquad r=k=q+1,\qquad \lambda=1. ]
The plane of order (2), commonly represented by the Fano plane, is a (2\text{-}(7,3,1)) design. It has seven points and seven blocks, with every pair of points lying in a unique block. Its importance follows from the incidence relations rather than from any particular drawing, since diagrams of the Fano plane are only representations of the same abstract structure.
Richard C. Bose developed algebraic constructions linking block designs with finite geometries, finite fields, and orthogonal arrays. These methods made incidence structures accessible through coordinate systems and group actions, while also clarifying the relation between experimental designs and coding theory.
Resolvability
A parallel class is a set of pairwise disjoint blocks whose union is the entire point set. A design is resolvable when its blocks can be partitioned into parallel classes. Each point then occurs exactly once in every class, so the number of blocks in a class is (v/k), and the number of parallel classes is (r).
Resolvability is stronger than pairwise balance because the BIBD equations alone do not force a partition into parallel classes. In experimental settings, a parallel class can represent a complete replication spread across several incomplete blocks. This structure separates the requirement that every treatment appear once in a replication from the requirement that treatment pairs have equal concurrence across the experiment.
A Kirkman triple system is a resolvable Steiner triple system whose blocks have size three and whose point pairs occur exactly once. Its parameters satisfy (v\equiv3\pmod 6). The resolvability condition distinguishes it from general Steiner triple systems, which exist for orders congruent to either (1) or (3) modulo (6).
Existence and construction
Parameter equations provide the first restrictions on existence, but additional obstructions arise from arithmetic and linear algebra. For symmetric designs, the Bruck–Ryser–Chowla theorem imposes number-theoretic conditions involving (v), (k), and (\lambda). These conditions eliminate some admissible parameter sets without supplying constructions for those that remain.
Constructions frequently use incidence relations in finite geometries. Other constructions derive blocks from the orbits of a permutation group, so that transitivity enforces uniform replication or concurrence. Difference methods encode a base block in a finite group and generate the remaining blocks by translation; the required pair frequencies then become conditions on group differences.
The existence question has a different character when the block size (k) and concurrence (\lambda) are fixed while (v) grows. Richard M. Wilson established that, apart from the necessary divisibility conditions, BIBDs with fixed (k) and (\lambda) exist for all sufficiently large admissible values of (v). This asymptotic result does not classify small designs, where exceptional nonexistence and multiple nonisomorphic realizations remain common.
Isomorphism and automorphisms
Two block designs are isomorphic when a bijection between their point sets maps the block collection of one design onto that of the other. Designs with identical parameters need not be isomorphic, because the parameters record only aggregate incidence counts. Classification therefore concerns complete incidence structures rather than parameter tuples.
An automorphism is an isomorphism from a design to itself. The automorphisms form a group acting on the points and blocks. A design may have a transitive automorphism group, although transitivity is not part of the definition of balance. Conversely, a design may satisfy all BIBD conditions while having only the identity automorphism.
The distinction is relevant to enumeration because labeled copies of a single design can be numerous even when only one isomorphism class exists. Canonical labeling and invariants derived from incidence matrices convert this issue into a problem related to graph isomorphism, typically by representing the design as a bipartite incidence graph.
Related generalizations
A (t)-design requires every (t)-element subset of points to occur in a constant number of blocks. A BIBD is precisely a (2)-design with constant block size. When (t>2), the defining condition implies balance for smaller subsets, with the corresponding multiplicities determined by counting extensions to (t)-subsets.
Partially balanced incomplete block designs replace a single pair concurrence with several concurrence classes. Their structure is commonly described through an association scheme, which partitions pairs according to their combinatorial relation. This relaxation retains enough algebraic regularity for spectral and statistical analysis while admitting arrangements excluded by complete pairwise balance.
Block sizes may also vary, producing pairwise balanced designs rather than BIBDs. In that setting, every pair still occurs a prescribed number of times, but the absence of constant (k) changes the replication equations and the form of the information matrix.