In mathematics and abstract algebra, a relation algebra is a residuated Boolean algebra expanded with an involution called converse, a unary operation. The motivating example of a relation algebra is the algebra 2X² of all binary relations on a set X, that is, subsets of the cartesian square X2, with R•S interpreted as the usual composition of binary relations R and S, and with the converse of R as the converse relation.
Relation algebra emerged in the 19th-century work of Augustus De Morgan and Charles Peirce, which culminated in the algebraic logic of Ernst Schröder. The equational form of relation algebra treated here was developed by Alfred Tarski and his students, starting in the 1940s. Tarski and Givant (1987) applied relation algebra to a variable-free treatment of axiomatic set theory, with the implication that mathematics founded on set theory could itself be conducted without variables.
A relation algebra (L, ∧, ∨, −, 0, 1, •, I, ˘) is an algebraic structure equipped with the Boolean operations of conjunction x∧y, disjunction x∨y, and negation x−, the Boolean constants 0 and 1, the relational operations of composition x•y and converse x˘, and the relational constant I, such that these operations and constants satisfy certain equations constituting an axiomatization of a calculus of relations. Roughly, a relation algebra is to a system of binary relations on a set containing the empty (0), complete (1), and identity (I) relations and closed under these five operations as a group is to a system of permutations of a set containing the identity permutation and closed under composition and inverse. However, the first order theory of relation algebras is not complete for such systems of binary relations.
Following Jónsson and Tsinakis (1993) it is convenient to define additional operations x◁y = x•y˘, and, dually, x▷y = x˘•y . Jónsson and Tsinakis showed that I◁x = x▷I, and that both were equal to x˘. Hence a relation algebra can equally well be defined as an algebraic structure (L, ∧, ∨, −, 0, 1, •, I, ◁, ▷). The advantage of this signature over the usual one is that a relation algebra can then be defined in full simply as a residuated Boolean algebra for which I◁x is an involution, that is, I◁(I◁x) = x . The latter condition can be thought of as the relational counterpart of the equation 1/(1/x) = x for ordinary arithmetic reciprocal, and some authors use reciprocal as a synonym for converse.
Since residuated Boolean algebras are axiomatized with finitely many identities, so are relation algebras. Hence the latter form a variety, the variety RA of relation algebras. Expanding the above definition as equations yields the following finite axiomatization.
- B1: A ∨ B = B ∨ A
- B2: A ∨ (B ∨ C) = (A ∨ B) ∨ C
- B3: (A− ∨ B)− ∨ (A− ∨ B−)− = A
This axiomatization of Boolean algebra is due to Huntington (1933). Note that the meet of the implied Boolean algebra is not the • operator (even though it distributes over like a meet does), nor is the 1 of the Boolean algebra the I constant.
- B4: A•(B•C) = (A•B)•C
- B5: A•I = A
- B6: A˘˘ = A
- B7: (A•B)˘ = B˘•A˘
Converse and composition distribute over disjunction:
- B8: (A∨B)˘ = A˘∨B˘
- B9: (A∨B)•C = (A•C)∨(B•C)
B10 is Tarski's equational form of the fact, discovered by Augustus De Morgan, that A•B ≤ C− A˘•C ≤ B− C•B˘ ≤ A−.
- B10: (A˘•(A•B)−)∨B− = B−
These axioms are ZFC theorems; for the purely Boolean B1-B3, this fact is trivial. After each of the following axioms is shown the number of the corresponding theorem in Chapter 3 of Suppes (1960), an exposition of ZFC: B4 27, B5 45, B6 14, B7 26, B8 16, B9 23.
Expressing properties of binary relations in RA
The following table shows how many of the usual properties of binary relations can be expressed as succinct RA equalities or inequalities. Below, an inequality of the form A≤B is shorthand for the Boolean equation A∨B = B.
The most complete set of results of this nature is Chapter C of Carnap (1958), where the notation is rather distant from that of this entry. Chapter 3.2 of Suppes (1960) contains fewer results, presented as ZFC theorems and using a notation that more resembles that of this entry. Neither Carnap nor Suppes formulated their results using the RA of this entry, or in an equational manner.
|R is||If and only if:|
|Functional||R˘•R ≤ I|
|Left-total||I ≤ R•R˘ (R˘ is surjective)|
|Function||functional and left-total.|
||R•R˘ ≤ I (R˘ is functional)|
|Surjective||I ≤ R˘•R (R˘ is left-total)|
|Bijection||R˘•R = R•R˘ = I (Injective surjective function)|
|Transitive||R•R ≤ R|
|Reflexive||I ≤ R|
|Coreflexive||R ≤ I|
|Irreflexive||R ∧ I = 0|
|Symmetric||R˘ = R|
|Antisymmetric||R ∧ R˘ ≤ I|
|Asymmetric||R ∧ R˘ = 0|
|Total||R ∨ R˘ = 1|
|Connex||I ∨ R ∨ R˘ = 1|
|Idempotent||R•R = R|
|Preorder||R is transitive and reflexive.|
|Equivalence||R is a symmetric preorder.|
|Partial order||R is an antisymmetric preorder.|
|Total order||R is a total partial order.|
|Strict partial order||R is transitive and irreflexive.|
|Strict total order||R is a connex strict partial order.|
|Dense||R ∧ I− ≤ (R ∧ I−)•(R ∧ I−).|
The metamathematics of RA are discussed at length in Tarski and Givant (1987), and more briefly in Givant (2006).
RA consists entirely of equations manipulated using nothing more than uniform replacement and the substitution of equals for equals. Both rules are wholly familiar from school mathematics and from abstract algebra generally. Hence RA proofs are carried out in a manner familiar to all mathematicians, unlike the case in mathematical logic generally.
RA can express any (and up to logical equivalence, exactly the) first-order logic (FOL) formulas containing no more than three variables. (A given variable can be quantified multiple times and hence quantifiers can be nested arbitrarily deeply by "reusing" variables.) Surprisingly, this fragment of FOL suffices to express Peano arithmetic and almost all axiomatic set theories ever proposed. Hence RA is, in effect, a way of algebraizing nearly all mathematics, while dispensing with FOL and its connectives, quantifiers, turnstiles, and modus ponens. Because RA can express Peano arithmetic and set theory, Gödel's incompleteness theorems apply to it; RA is incomplete, incompletable, and undecidable. (N.B. The Boolean algebra fragment of RA is complete and decidable.)
The representable relation algebras, forming the class RRA, are those relation algebras isomorphic to some relation algebra consisting of binary relations on some set, and closed under the intended interpretation of the RA operations. It is easily shown, e.g. using the method of pseudoelementary classes, that RRA is a quasivariety, that is, axiomatizable by a universal Horn theory. In 1950, Roger Lyndon proved the existence of equations holding in RRA that did not hold in RA. Hence the variety generated by RRA is a proper subvariety of the variety RA. In 1955, Alfred Tarski showed that RRA is itself a variety. In 1964, Donald Monk showed that RRA has no finite axiomatization, unlike RA, which is finitely axiomatized by definition.
An RA is a Q-relation algebra (QRA) if, in addition to B1-B10, there exist some A and B such that (Tarski and Givant 1987: §8.4):
- Q0: A˘•A ≤ I
- Q1: B˘•B ≤ I
- Q2: A˘•B = 1
Essentially these axioms imply that the universe has a (non-surjective) pairing relation whose projections are A and B. It is a theorem that every QRA is a RRA (Proof by Maddux, see Tarski & Givant 1987: 8.4(iii) ).
Every QRA is representable (Tarski and Givant 1987). That not every relation algebra is representable is a fundamental way RA differs from QRA and Boolean algebras, which, by Stone's representation theorem for Boolean algebras, are always representable as sets of subsets of some set, closed under union, intersection, and complement.
1. Any Boolean algebra can be turned into a RA by interpreting conjunction as composition (the monoid multiplication •), i.e. x•y is defined as x∧y. This interpretation requires that converse interpret identity (ў = y), and that both residuals y\x and x/y interpret the conditional y→x (i.e., ¬y∨x).
2. The motivating example of a relation algebra depends on the definition of a binary relation R on a set X as any subset R ⊆ X², where X² is the Cartesian square of X. The power set 2X² consisting of all binary relations on X is a Boolean algebra. While 2X² can be made a relation algebra by taking R•S = R∧S, as per example (1) above, the standard interpretation of • is instead x(R•S)z = ∃y:xRy.ySz. That is, the ordered pair (x,z) belongs to the relation R•S just when there exists y ∈ X such that (x,y) ∈ R and (y,z) ∈ S. This interpretation uniquely determines R\S as consisting of all pairs (y,z) such that for all x ∈ X, if xRy then xSz. Dually, S/R consists of all pairs (x,y) such that for all z ∈ X, if yRz then xSz. The translation ў = ¬(y\¬I) then establishes the converse R˘ of R as consisting of all pairs (y,x) such that (x,y) ∈ R.
3. An important generalization of the previous example is the power set 2E where E ⊆ X² is any equivalence relation on the set X. This is a generalization because X² is itself an equivalence relation, namely the complete relation consisting of all pairs. While 2E is not a subalgebra of 2X² when E ≠ X² (since in that case it does not contain the relation X², the top element 1 being E instead of X²), it is nevertheless turned into a relation algebra using the same definitions of the operations. Its importance resides in the definition of a representable relation algebra as any relation algebra isomorphic to a subalgebra of the relation algebra 2E for some equivalence relation E on some set. The previous section says more about the relevant metamathematics.
4. Let be group. Then the power set is a relation algebra with the obvious boolean algebra operations, composition given by the product of group subsets, the converse by the inverse subset (), and the identity by the singleton subset . There is a relation algebra homomorphism embedding in which sends each subset to the relation . The image of this homomorphism is the set of all right-invariant relations on .
5. If group sum or product interprets composition, group inverse interprets converse, group identity interprets I, and if R is a one-to-one correspondence, so that R˘•R = R•R˘ = I, then L is a group as well as a monoid. B4-B7 become well-known theorems of group theory, so that RA becomes a proper extension of group theory as well as of Boolean algebra.
De Morgan founded RA in 1860, but C. S. Peirce took it much further and became fascinated with its philosophical power. The work of DeMorgan and Peirce came to be known mainly in the extended and definitive form Ernst Schröder gave it in Vol. 3 of his Vorlesungen (1890–1905). Principia Mathematica drew strongly on Schröder's RA, but acknowledged him only as the inventor of the notation. In 1912, Alwin Korselt proved that a particular formula in which the quantifiers were nested four deep had no RA equivalent. This fact led to a loss of interest in RA until Tarski (1941) began writing about it. His students have continued to develop RA down to the present day. Tarski returned to RA in the 1970s with the help of Steven Givant; this collaboration resulted in the monograph by Tarski and Givant (1987), the definitive reference for this subject. For more on the history of RA, see Maddux (1991, 2006).
- RelMICS / Relational Methods in Computer Science maintained by Wolfram Kahl
- Carsten Sinz: ARA / An Automatic Theorem Prover for Relation Algebras
- Alfred Tarski (1948) "Abstract: Representation Problems for Relation Algebras," Bulletin of the AMS 54: 80.
- Chris Brink; Wolfram Kahl; Gunther Schmidt (1997). Relational Methods in Computer Science. Springer. pp. 4 and 8. ISBN 978-3-211-82971-4.
- Tarski, A. (1941), p. 87.
- Korselt did not publish his finding. It was first published in Leopold Loewenheim (1915) "Über Möglichkeiten im Relativkalkül," Mathematische Annalen 76: 447–470. Translated as "On possibilities in the calculus of relatives" in Jean van Heijenoort, 1967. A Source Book in Mathematical Logic, 1879–1931. Harvard Univ. Press: 228–251.
- Carnap, Rudolf (1958). Introduction to Symbolic Logic and its Applications. Dover Publications.
- Givant, Steven (2006). "The calculus of relations as a foundation for mathematics". Journal of Automated Reasoning. 37 (4): 277–322. doi:10.1007/s10817-006-9062-x.
- Halmos, P. R. (1960). Naive Set Theory. Van Nostrand.
- Henkin, Leon; Tarski, Alfred; Monk, J. D. (1971). Cylindric Algebras, Part 1. North Holland.
- Henkin, Leon; Tarski, Alfred; Monk, J. D. (1985). Cylindric Algebras, Part 2. North Holland.
- Hirsch, R.; Hodkinson, I. (2002). Relation Algebra by Games. Studies in Logic and the Foundations of Mathematics. 147. Elsevier Science.
- Jónsson, Bjarni; Tsinakis, Constantine (1993). "Relation algebras as residuated Boolean algebras". Algebra Universalis. 30 (4): 469–78. doi:10.1007/BF01195378.
- Maddux, Roger (1991). "The Origin of Relation Algebras in the Development and Axiomatization of the Calculus of Relations" (PDF). Studia Logica. 50 (3–4): 421–455. doi:10.1007/BF00370681.
- Maddux, Roger (2006). Relation Algebras. Studies in Logic and the Foundations of Mathematics. 150. Elsevier Science. ISBN 9780444520135.
- Suppes, Patrick (1972) . "Chapter 3". Axiomatic Set Theory (Dover reprint ed.). Van Nostrand.
- Schmidt, Gunther (2010). Relational Mathematics. Cambridge University Press.
- Tarski, Alfred (1941). "On the calculus of relations". Journal of Symbolic Logic. 6 (3): 73–89. doi:10.2307/2268577. JSTOR 2268577.
- Tarski, Alfred; Givant, Steven (1987). A Formalization of Set Theory without Variables. Providence RI: American Mathematical Society.
- Yohji AKAMA, Yasuo Kawahara, and Hitoshi Furusawa, "Constructing Allegory from Relation Algebra and Representation Theorems."
- Richard Bird, Oege de Moor, Paul Hoogendijk, "Generic Programming with Relations and Functors."
- R.P. de Freitas and Viana, "A Completeness Result for Relation Algebra with Binders."
- Peter Jipsen:
- Vaughan Pratt:
- Priss, Uta:
- Kahl, Wolfram and Gunther Schmidt: Exploring (Finite) Relation Algebras Using Tools Written in Haskell. and Relation Algebra Tools with Haskell from McMaster University.