66 Robert Goldblattmodalities 〈 i 〉 for i ∈ I. The box modality [ i ] is defined to be ¬〈 i 〉¬. The relationp |= α, meaning “process p satisfies formula α”, is defined inductively, withp |= 〈 i 〉α iff for some i-experiment 〈p, p ′ 〉, p ′ |= α.Two processes are regarded as equivalent if there is no observable action that eithercan perform to distinguish them. Informally this means that to each observableaction that one can perform there is an action that the other can perform whichleads to an equivalent outcome, so each process can “simulate” the other. Spellingthis out,p is equivalent to q if, and only if,1. for every result p ′ of an experiment on p, there is an equivalent resultq ′ of an experiment on q; and2. for every result q ′ of an experiment on q, there is an equivalent resultp ′ of an experiment on p[Milner, 1980, p. 41]. As a definition of equivalence this appears to be circular,since the word “equivalence” occurs on both sides of the “if and only if”. Toformalise the idea, a sequence of equivalence relations ∼ n for n ≥ 0 is defined onP . For each relation S ⊆ P × P , define a relation E(S) by putting 〈p, q〉 ∈ E(S)if for every i ∈ I,1. 〈p, p ′ 〉 ∈ R i implies, for some q ′ , 〈q, q ′ 〉 ∈ R i and 〈p ′ , q ′ 〉 ∈ S; and2. 〈q, q ′ 〉 ∈ R i implies, for some p ′ , 〈p, p ′ 〉 ∈ R i and 〈p ′ , q ′ 〉 ∈ S.Put p ∼ 0 q for all p, q ∈ P , and inductively p ∼ n+1 q if 〈p, q〉 ∈ E(∼ n ). Thenp and q are defined to be observationally equivalent, written p ∼ q, if p ∼ n q forevery n.Now a relation R ⊆ P × P is image-finite if the set {p ′ : 〈p, p ′ 〉 ∈ R} is finite foreach p ∈ P . Hennessy and Milner gave a logical characterisation of observationalequivalence by showing that if each R i is image-finite, two processes are equivalentiff they satisfy the same formulas:p ∼ q iff for all formulas α, p |= α iff q |= α. (∗)Note that the operator E on relations is monotonic: R ⊆ S implies E(R) ⊆ E(S).This property implies, by induction, that ∼ n+1 ⊆ ∼ n , and so iteration of Egenerates a decreasing chain of relations∼ 0 ⊇ ∼ 1 ⊇ ∼ 2 ⊇ · · · ⊇ ∼ n ⊇ · · · · · ·Let ∼ ω = ⋂ {∼ n : n ≥ 0} be the intersection of the chain. Then in the image-finitecase, ∼ ω is the largest fixed point of the operator E, i.e. putting S =∼ ω givesthe largest solution to the equation S = E(S) (see [Hennessy and Milner, 1985,
Mathematical Modal Logic: A View of its Evolution 67Theorem 2.1]). In that case 〈p, q〉 ∈ S iff 〈p, q〉 ∈ E(S), legitimizing the circulardefinition of equivalence.The monotonicity of E alone is enough to guarantee that E has a largest fixedpoint (see section 7.4), but in the absence of image-finiteness this fixed point neednot be the relation ∼ ω . It may be a proper subrelation of ∼ ω that can only bereached by iterating E transfinitely often. Consequently this largest fixed pointhas become the general definition of the observational-equivalence relation ∼, andit is only in the image-finite case that ∼ is identified with ∼ ω .This analysis indicates that standard induction on natural numbers n (appliedto the relations ∼ n ) may not be effective as a method for proving equivalence ofprocesses. Instead, as was first realised by David Park, 60 a new kind of proof ruleis called for, based on the notion of a bisimulation. This is a relation S ⊆ P × Psatisfying S ⊆ E(S), i.e. 〈p, q〉 ∈ S implies (1) and (2) hold. The union of anycollection of bisimulations is a bisimulation, and so there is a largest bisimulation—the union of all of them–which turns out to be the same as the largest fixed point ofE. In other words, the observational relation ∼ is the largest bisimulation on anystructure (P, {R i : i ∈ I}). It is an equivalence relation in the mathematical sense(reflexive, symmetric and transitive) and is known as bisimulation equivalence orbisimilarity [Milner, 1989]. Itadmits an elegant proof technique; to show p ∼ q, it is necessary and sufficientto find some bisimulation containing the pair 〈p, q〉[Milner, 1983, p. 283]. In the general setting, when ∼ is not equal to ∼ ω , thesame modal-logical characterisation of bisimilarity as (∗) above can be obtainedby expanding the class of formulas to allow formation of the conjunction ∧ j∈J α jfor any set {α j : j ∈ J} (possibly infinite) of formulas.The term “bisimulation” was first used in [Park, 1981] for a relation of mutualsimulation between states of two automata, with motivation from an earliernotion of simulation of programs from [Milner, 1971]. Park showed that if twodeterministic automata are related by a bisimulation, then they accept the sameset of inputs. The concept and its use was systematically developed in [Milner,1983]. It is closely related to the notion of “p-relation” of van Benthem [1976a]mentioned in section 5.3. Segerberg’s p-morphisms are essentially bisimulations(between Kripke models) that are total and functional.Process algebra is now a substantial field, with many concepts and constructionsfor building processes, and many important variations on the notion of observationalequivalence or bisimilarity (see [Bergstra et al., 2001]). For any given familyof transition systems, i.e. systems of observation relations, we can seek to devisemodalities that generate formulas giving a logical characterisation of the bisimilarityrelations for those systems in the manner of (∗). This programme has beencarried out for many cases. Logics for more recently developed theories of “mobile”and “message-passing” processes are discussed in [Milner et al., 1993] and60 Information from Robin Milner, personal communication.