BRICS Basic Research in Computer Science

BRICSRS-03-1Sassone&Sobocinski:DerivingBisimulationCongruences:2-Categoriesvs.Pr

BRICS

Basic Research in Computer Science

Deriving Bisimulation Congruences:

2-Categories vs. Precategories

Vladimiro Sassone Paweł Sobocinski

BRICS Report Series RS-03-1

ISSN 0909-0878 January 2003

Copyright c2003, Vladimiro Sassone & Paweł Sobocinski.

BRICS, Department of Computer Science University of Aarhus. All rights reserved.

Reproduction of all or part of this work is permitted for educational or research use on condition that this copyright notice is included in any copy.

See back inner page for a list of recent BRICS Report Series publications.

Copies may be obtained by contacting:

BRICS

Department of Computer Science University of Aarhus

Ny Munkegade, building 540 DK–8000 Aarhus C

Denmark

Telephone: +45 8942 3360 Telefax: +45 8942 3255 Internet: BRICS@brics.dk

BRICS publications are in general accessible through the World Wide Web and anonymous FTP through these URLs:

http://www.brics.dk ftp://ftp.brics.dk

This document in subdirectoryRS/03/1/

Deriving Bisimulation Congruences:

2-categories vs precategories ^∗

Vladimiro Sassone Paweł Soboci´nski

University of Sussex University of Aarhus

Abstract

G-relative pushouts (GRPOs) have recently been proposed by the authors as a new foundation for Leifer and Milner’s approach to deriving labelled bisimulation congruences from reduction sys- tems. This paper develops the theory ofGRPOs further, arguing that they provide a simple and powerful basis towards a compre- hensive solution. As an example, we construct GRPOs in a cat- egory of ‘bunches and wirings.’ We then examine the approach based on Milner’s precategories and Leifer’s functorial reactive systems, and show that it can be recast in a much simpler way into the 2-categorical theory ofGRPOs.

Introduction

It is increasingly common for foundational calculi to be presented as re- duction systems. Starting from their common ancestor, theλ calculus, most recent calculi consist of a reduction system together with a contex- tual equivalence (built out of basic observations, viz. barbs). The strength of such an approach resides in its intuitiveness. In particular, we need not invent labels to describe the interactions between systems and their pos- sible environments, a procedure that has a degree of arbitrariness (cf.

∗Research supported by ‘DisCo: Semantic Foundations of Distributed Computa- tion’, EU IHP ‘Marie Curie’ contract HPMT-CT-2001-00290, and BRICS, Basic Re- search in Computer Science, funded by the Danish National Research Foundation.

early and late semantics of theπcalculus) and may prove quite complex (cf. [5, 4, 3, 1]).

By contrast, reduction semantics suffer at times by their lack of com- positionality, and have complex semantic theories because of their con- textual equivalences. Labelled bisimulation congruences based on la- belled transition systems (LTS) may in such cases provide fruitful proof techniques; in particular, bisimulations provide the power and manage- ability of coinduction, while the closure properties of congruences pro- vide for compositional reasoning.

To associate an LTS with a reduction system involves synthesising a compositional system of labels, so that silent moves (orτ-actions) re- flect the original reductions, labels describe potential external interac- tions, and all together they yield a LTS bisimulation which is a congru- ence included in the original contextual reduction equivalence. Proving bisimulation is then enough to prove reduction equivalence.

Sewell [19] and Leifer and Milner [13, 11] set out to develop a theory to perform such derivations using general criteria; a meta-theory of deriv- ing bisimulation congruences. The basic idea behind their construction is to use contexts as labels. To exemplify the idea, in a CCS-like calculus one would for instance derive a transition

a.P ^−|

¯ a.Q

IP|Q

because term a.P in context − |a¯.Q reacts to become P |Q; in other words, the context is a trigger for the reduction.

The first hot spot of the theory is the selection of the right triggers to use as labels. The intuition is to take only the “smallest” contexts which allow a given reaction to occur. As well as reducing the size of the LTS, this often makes the resulting bisimulation equivalence finer. Sewell’s method is based on dissection lemmas which provide a deep analysis of a term’s structure. A generalised, more scalable approach was later devel- oped in [13], where the notion of “smallest” is formalised in categorical terms as a relative-pushout (RPOs). Both theories, however, do not seem to scale up to calculi with non trivial structural congruences. Already in the case of the monoidal rules that govern parallel composition things become rather involved.

The fundamental difficulty brought about by a structural congruence

≡ is that working up to ≡ gives up too much information about terms

for the RPO approach to work as expected. RPOs do not usually exist in such cases, because the fundamental indication of exactly which oc- currences of a term constructor belong to the redex becomes blurred. A very simple, yet significant example of this is the category Bun of bunch contexts [13], and the same problems arise in structures such as action graphs [14] and bigraphs [15].

In [17] we therefore proposed a framework in which term structure is not explicitly quotiented, but the commutation of diagrams (i.e. equality of terms) is taken up to≡. Precisely, to give a commuting diagram r p≡ sq one exhibits a proofα of structural congruence, which we represent as a 2-cell (constructed from the rules generating≡and closed under all contexts).

k ^{p //}

q

l

r

m

ααα s //n

Since such proofs are naturally isomorphisms, we were led to consider G-categories, i.e., 2-categories where all 2-cells are iso, and initiated the study of G-relative pushouts (GRPOs), as a suitable generalisation of RPOs from categories toG-categories.

The purpose of this paper is to continue the development of the theory ofGRPOs. We aim to show that, while replacing RPOs at little further complication (cf. §2 and §3), GRPOs significantly advance the field by providing a convenient solution to simple, yet important problems (cf. §4 and §5). The theory ofGRPOs promises indeed to be a natural founda- tion for a meta-theory of ‘deriving bisimulation congruences.’

This paper presents two main technical results in support of our claims.

Firstly, we prove that the case of the already mentioned category Bun of bunch contexts, problematic for RPOs, can be treated in a natural way using GRPOs. Secondly, we show that the notions of precategory and functorial reactive system can be dispensed with in favour of a simpler GRPO-based approach.

The notion of precategory is proposed in [11, 12] to handle the ex- amples of Leifer in [11], Milner in [15] and, most recently, of Jensen and Milner in [7]. It consists of a category appropriately decorated by so-called “support sets” which identifies syntactic elements so as to keep track of them under arrow composition. Alas, such supported structures

are no longer categories – arrow composition is partial – which makes the theory laborious, and bring us away from the well-known world of categories and their theory. The intensional information recorded in pre- categories, however, allows one to generate a category “above” where RPOs exist, as opposed to the category of interest “below”, sayC, where they do not. The category “above” is related toC via a well-behaved functor, used to map RPOs diagrams from the category “above” to C, where constructing them would be impossible. These structures take the name of functorial reactive systems, and give rise to a theory to generate a labelled bisimulation congruences developed in [11].

The paper presents a technique for mapping precategories toG-categories so that the LTS generated usingGRPOs is the same as the LTS generated using the above mentioned approach. The translation derives from the precategory’s support information a notion of homomorphism, specific to the particular structure in hand, which constitutes the 2-cells of the de- rivedG-category. We claim that this yields an approach mathematically more elegant and considerably simpler than precategories; besides gen- eralising RPOs directly,GRPOs seem to also remove the need for further notions.

Structure of the paper. In §2 we review definitions and results pre- sented in [17]; §3 shows that, analogously to the 1-dimensional case, trace and failures equivalence are congruences provided that enoughGRPOs exist. In §4, we show that the category of bunch contexts is naturally a 2- category whereGRPOs exist; §5 shows how precategories are subsumed by our notion ofGRPOs. The exposition ends with a few concluding re- marks; §1 recalls basic notions of 2-categories, and can be safely skipped by those readers acquainted with the standard notations.

This paper is a version of [18] extended with proofs.

1 Preliminaries

Throughout the paper we assume a moderate knowledge of category the- ory and related terminology. In this section we fix notations and recall the basic elements and 2-categories. For a thorough introduction the reader is referred to [9]

We use Ord to denote the category of finite ordinals. We assume that Ord has chosen coproducts, namely the reader’s favourite definition of ordinal addition⊕. For any finite set x, let ord(x)be the finite ordinal of the same cardinality and t_x: x→ord(x)be a chosen isomorphism. There is an equivalence of categories F : Set_f →Ord. On objects it sends x to ord(x), on morphisms f : x→y to tyf t_x⁻¹: ord(x)→ord(y).

A 2-categoryCis a category whose homsets are categories and, cor- respondingly, whose composition maps are functors. Explicitly, a 2- categoryBconsists of the following.

• A class of objects X,Y,Z,....

• For any X,Y ∈C, a category C(X,Y). The objects C(X,Y) are called 1-cells, or simply arrows, and denoted by f : Y →X . Its morphisms are called 2-cells, are written α: f ⇒g : X →Y and drawn as

X

f ((

g

66α Y.

Composition inC(X,Y)is denoted by • and referred to as ‘verti- cal’ composition. Identity 2-cells are denoted by 1_f: f ⇒ f .

• For each X,Y,Z there is a functor.: C(X,Y)×C(Y,Z)→C(X,Z), the so-called ‘horizontal’ composition, which we often denote by mere juxtaposition. Horizontal composition is associative and ad- mits 1id_X as identities.

As a notation, we write αf and gα for, respectively, α1f and 1_gα.

We follow the convention that horizontal composition binds tighter than vertical composition.

In 2-categories, the order of composition of 2-cells is not important.

This is a consequence of the horizontal composition being functorial and can be axiomatised with the so called middle-four interchange law:

for f,f⁰,f⁰⁰: A→B and g,g⁰,g⁰⁰: B→C and α: f ⇒ f⁰, α⁰: f⁰⇒ f⁰⁰, β: g ⇒g⁰ and β⁰: g⁰ ⇒g⁰⁰ we have β⁰α0 •βα= (β0 •β)(α0 •α). As a consequence, given a diagram of 2-cells, there is at most one way to compose them and obtain a composite 2-cell. This primitive operation is sometimes referred to as pasting.

Two objects C, D of a 2-category C are equivalent when there are arrows f : C→D, g : D→C and 2-cellsα: id_C⇒g f ,β: f g⇒id_D. We refer to f and g as equivalences.

2 Reactive Systems and G RPOs

Lawvere theories [10] provide a canonical way to recast term algebras as categories. ForΣa signature, the (free) Lawvere theory onΣ, say C_Σ, has the natural numbers for objects and a morphism t : m→n, for t a n-tuple of m-holed terms. Composition is substitution of terms into holes.

Generalising from term rewriting systems on C_Σ, Leifer and Milner formulated a definition of reactive system [13], and defined a technique to extract labelled bisimulation congruences from them.

In order to accommodate calculi with non trivial structural congru- ences, as explained in the Introduction, we refine their approach as fol- lows.

Definition 2.1. AG-category is a 2-category where all 2-cells are iso- morphisms.

A G-category is a thus a category enriched over G, the category of groupoids.

Definition 2.2. AG-reactive system C consists of 1. aG-categoryC,

2. a subcategory D of reactive contexts; it is required to be closed under 2-cells to be and composition-reflecting,

3. a distinguished object I∈C,

4. a set of pairs

R

⊆^SC∈CC(I,C)×C(I,C)called the reaction rules.

The reactive contexts are those contexts inside which evaluation may occur. By composition-reflecting we mean that dd⁰∈D implies d and d⁰∈D, while the closure property means that given d∈Dandρ: d⇒d⁰ inCimplies d⁰∈D. The reaction relation Bis defined by taking

a Bdr if there exists hl,ri, d∈Dandα: dl⇒a.

As illustrated by the diagram below, this represents the fact that, up to structural congruence, a is the left-hand side l of a reduction rule in a reaction context d.

I

l NNNNN^aNNNN''

C ^α

d //C⁰

The notion ofGRPO formalises the idea of a context being the “small- est” that enables a reaction in aG-reactive system, and is a conservative 2-categorical extension of Leifer and Milner RPOs [13] (cf. [17] for a precise comparison).

For readers acquainted with 2-dimensional category theory, GRPOs are defined in Definition 2.3. This is followed by an elementary presen- tation in Proposition 2.4 taken from [17].

Definition 2.3 (GRPOs). Letρ: ca⇒db : W →Z be a 2-cell (see dia- gram below) in aG-categoryC. AG-relative pushout (GRPO) forρis a bipushout [8] of the pair or arrows(a,1): ca→c and(b,ρ): ca→d in the pseudo-slice categoryC/Z.

Z

X ^ρ

c}}}>>

} Y

`` d

AAAA

W ^b

>>

}} }} a

``AAAA (1)

Proposition 2.4. LetCbe aG-category. A candidateGRPO forρ: ca⇒ db as in diagram (1) is a tuplehR,e,f,g,β,γ,δisuch thatδb•gβ•γa=ρ – cf. diagram (i).

Z X

γ e //

β

c}}}}>>

}}

R

g δ

OO

Y

oo f

`` d

@@@@@@@

W

b

>>

~~

~~

~~ a

``AAAAAA

(i)

R⁰ X

ϕ e //

e{⁰{{{{==

{{

R

ψ h

OO

(ii)

f Y

oo

f⁰

aaCCCCCCC

Z R⁰

g⁰ τ

OO

R

aa g

CCCCCCC

oo h

(iii)

AGRPO forρis a candidate which satisfies a universal property. Namely, for any other candidate hR⁰,e⁰,f⁰,g⁰,β⁰,γ⁰,δ⁰i there exists a quadruple

hh,ϕ,ψ,τiwhere h : R→R⁰, ϕ: e⁰⇒he andψ: h f ⇒ f⁰ – cf. diagram (ii) – andτ: g⁰h⇒g – diagram (iii) – which makes the two candidates compatible in the obvious way, i.e.

τe•g⁰ϕ•γ⁰=γ δ0 •g⁰ψ•τ⁻¹f =δ ψb•hβ•ϕa=β⁰. Such a quadruple, which we shall refer to as mediating morphism, must be essentially unique. Namely, for any other mediating morphismhh⁰,ϕ⁰,ψ⁰,τ⁰i there must exist a unique two cellξ: h→h⁰which makes the two medi- ating morphisms compatible, i.e.:

ξe•ϕ=ϕ⁰ ψ•ξ⁻¹f =ψ⁰ τ0 •g⁰ξ=τ

Observe that whereas RPOs are defined up to isomorphism,GRPOs are defined up to equivalence (since they are bicolimits).

The definition below plays an important role in the following devel- opment.

Definition 2.5 (GIPO). Diagram (1) of Definition 2.3 is said to be a G- idem-pushout (GIPO) ifhZ,c,d,id_Z,ρ,1_c,1_diis itsGRPO.

The next two lemmas explain the relationships betweenGRPOs and GIPOs.

Lemma 2.6 (GIPOs fromGRPOs). IfhZ,c,d,u,α,η,µiis aGRPO for (i) below, as illustrated in (ii), then (iii) is aGIPO.

Z⁰ X α⁰

c~⁰~~~>>

Y

d⁰

__@@@@

W

a

``AAAA

b

>>

}} }}

(i)

Z⁰ X

c~⁰~~~>>

c^η//Z

uOO µ

Y

oo d

d⁰

__@@@@

W

a α

``AAAA

b

>>

}} }}

(ii)

Z

X ^α

c}}}>>

} Y

`` d

AAAA

W

a

``AAAA

b

>>

}} }}

(iii)

Lemma 2.7 (GRPOs fromGIPOs). If square (iii) above is aGIPO, (i) has aGRPO, andhZ,c,d,u,α,η,µiis a candidate for it as shown in (ii), thenhZ,c,d,u,α,η,µiis aGRPO for (i).

The following lemmas from [17] state the basic properties ofGRPOs.

Lemma 2.8. Suppose that diagram (ii) below has aGRPO.

U ^{a //}

b

V

d

e //W

g

X

α c //Y

β f //Z (i)

U ^{a //}

b

V

ge

X _{f c} ^//

βa•fα

Z (ii)

1. If both squares in (i) areGIPOs then the rectangle of (i) is aGIPO;

2. If the left square and the rectangle of (i) areGIPOs then so is the right square.

Lemma 2.9. Suppose that diagram (i) below is aGIPO.

Z

X ^α

c}}}>>

} Y

`` d

AAAA a W

``AAAA

b

>>

}} }}

(i)

Z

X ^α

c}}}>>

} Y

`` d

AAAA

W

`` a

AAAA a⁰

PP

ε b

>>

}} }}

(ii)

Z

X ^α

c}}}>>

} ^d Y

``AAAA^d0

oo ε

``AAAA

a W

``AAAA

b

>>

}} }}

(iii)

Then the regions obtained by pasting the 2-cells in (ii) and (iii) areGIPOs.

The previous lemma in particular implies that the following definition of labelled transition system derived from a G-reactive system is well defined.

Definition 2.10 (LTS). For C aG-reactive system whose underlying cat- egoryCis aG-category, defineGTS(C)as follows:

• the statesGTS(C)are iso-classes of arrows[a]: I→X in C;

• there is a transition[a] ^[f]I[a⁰]if there exists a 2-cell ρ, a rule hl,ri ∈

R

^{, and d}∈Dwith a⁰∼=dr and such that the diagram below is aGIPO.

Z X

f??

ρ Y

__ d

???

a I

__????

l

??





(2)

Henceforward we shall abuse notation and leave out the square brackets when writing transitions; ie. we shall write simply a ^f Ia⁰instead of [a] ^[f]I[a⁰].

Categories can be seen as a discreteG-categories (the only 2-cells are identities). Using this observation, eachG-concepts introduced above re- duces to the corresponding 1-categorical concept. For instance, aGRPO in a category is simply a RPO.

3 Congruence Results for G RPOs

The fundamental property that endows the LTS derived from a reduction system with a bisimulation which is a congruence is the following notion.

Definition 3.1 (RedexGRPOs). AG-reactive system C is said to have redexGRPOs if every square (2) in its underlyingG-category Cwith l the left-hand side of a reaction rulehl,ri ∈

R

^{, and d} ∈Dhas aGRPO.

In particular, the main theorem of [17] is as follows.

Theorem 3.2 (cf. [17]). Let C be a reactive system whose underlying G-categoryChas redexGRPOs. The largest bisimulation∼onGTS(C) is a congruence.

The next three subsections complement this result by proving the ex- pected corresponding theorems for trace and failure semantics, and by lifting them to the case of weak equivalences. Theorems and proofs in this section follow closely [11], as they are meant to show thatGRPOs are as viable a tool as RPOs are.

3.1 Traces Preorder

Trace semantics [16] is a simple notion of equivalence which equates processes if they can engage in the same sequences of actions. Even though it lacks the fine discriminating power of branching time equiv- alences, viz. bisimulations, it is nevertheless interesting because many safety properties can be expressed as conditions on sets of traces.

We say that a sequence f₁···f_nof labels ofGTS(C)is a trace of a if a ^f1 I··· ^fn Ia_n+1

for some a₁,...,a_n. The trace preorder.tr is then defined as a.tr b if all traces of a are also traces of b.

Theorem 3.3 (Trace Congruence). .tris a congruence.

Proof. Assume a.tr b. We prove that ca.tr cb for all contexts c∈C. Suppose that

ca=a¯₁ ^f1 Ia¯₂···a¯_n ^fn Ia¯_n+1. We first prove that there exist a sequence, for i=1,...,n,

·

αi

ai //

li

·

βi

gi

ci //·

f_i

· _d

i

//·

d_i⁰

//·

where a₁=a, c₁=c, c_i+1=d_i⁰, ¯a_i=c_ia_i, and each square is aGIPO.¹ The ith induction step proceeds as follows. Since ¯a_i ^fi Ia¯_i+1, there existsγi: ficiai⇒d¯ili, for somehli,rii ∈

R

^{and ¯d}i∈D, with ¯ai+1=d¯iri. Since C has redex GIPOs (cf. Definition 3.1), this can be split in two GIPOs: αi: g_ia_i ⇒d_il_i and βi: f_ic_i ⇒d⁰_ig_i (cf. diagram above). Take ai+1=diri, and the induction hypothesis is maintained. In particular, we obtained a trace

a=a₁ ^g1 Ia₂···a_n ^gn Ia_n+1

that, in force of the hypothesis a.tr b must be matched by a corre- sponding trace of b. This means that, for i=1,..,n, there existGIPOs α⁰_i: g_ib_i⇒e_il_i⁰, for somehl_i⁰,r_i⁰i ∈

R

^{and e}i∈D, once we take b_i+1to be e_ir⁰_i. We can then paste each of suchGIPOs together withβi: f_ic_i⇒d_i⁰g_i obtained above and, using Lemma 2.8, conclude that there existGIPOs

f_ic_ib_i⇒d_i⁰e_il_i⁰, as in the diagram below.

·

α⁰_i bi //

l⁰_i

·

βi

gi

ci //·

f_i

· _e1 ^//·

d_i⁰

//·

which means c_ib_i ^fi Id_i⁰e_ir_i⁰.

1Since the fact is not likely to cause confusion, we make no notational distinction between the arrows of C (e.g. inGRPOs diagrams) and the states and labels ofGTS(C), where the latter are iso-classes of the former.

As cb=c₁b₁, in order to construct a trace cb= ¯b₁ ^f1 I··· ^fn I ¯b_n+1 and complete the proof, we only need to verify that for i=1,...,n, we have that d_i⁰e_ir⁰_i=c_i+1b_i+1. This follows at once, as c_i+1=d_i⁰and b_i+1= e_ir⁰_i.

3.2 Failures Preorder

Failure semantics [6] enhances trace semantics with limited branch-inspecting power. More precisely, failure sets allow the testing of when processes renounce the capability of engaging in certain actions.

Formally, for a a state ofGTS(C), a failure of a is a pair(f₁···f_n,X), where f1···fnand X are respectively a nonempty sequence and a set of labels, such that:

• f₁···f_nis a trace of a, a ^f1 I··· ^fn Ia_n+1;

• a_n+1, the final state of the trace, is stable, i.e. a_n+16 B;

• a_n+1refuses X , i.e. a_n+16 ^x Ifor all x∈X .

The failure preorder.f is defined as a.f b if all failures of a are also failures of b.

Theorem 3.4 (Failures Congruence). .f is a congruence.

Proof. Assume a.fb to prove that ca.f cb for all contexts c∈C. The proof extends the previous one of Theorem 3.3.

Let (f₁···f_n,X), n>0, be a failure of ca. We proceed exactly as above to determine a matching trace cb= ¯b₁ ^f1 I··· ^fn I ¯b_n+1. In addition, we contextually need to prove that ¯b_n+1is stable and refuses X , exploiting the corresponding hypothesis on ¯a_n+1.

First, we claim that a_n+1 is stable. In fact, were it not, it would follow from c_n+1(=d_n⁰)∈Dthat also ¯a_n+1=c_n+1a_n+1 B, which is impossi- ble, since ¯an+1is stable. Secondly, an+1refuses both

Y ={g|there exists aGIPOδg: xcn+1⇒dg, for x∈X}and Z={g|there exists a 2-cellεg: dg⇒c_n+1, for d∈D},

which can be seen as follows. If a_n+1 ^g Ifor g∈Y , then there exists aGIPOα: ga_n+1⇒d⁰l, for some rulehl,ri, which could be pasted to- gether with δg to yield aGIPO xc_n+1a_n+1⇒dd⁰l, which is impossible

since it means that ¯a_n+1 ^x I, for x∈X . Similarly, if a_n+1 ^g I for g∈Z, pasting the correspondingGIPO with εg, we see that ¯a_n+1 B, contradicting the hypothesis that ¯a_n+1is stable.

If follows then from the hypothesis a.fb that b_n+1 is stable and refuses Y∪Z. It is then easy to complete the proof by transferring stability and X -refusal to ¯b_n+1. First, suppose that ¯b_n+1 B. This means that there exists a 2-cell dl⇒ ¯b_n+1. Since C has redex-GRPOs, we can factor c_n+1 out and obtain from this aGRPOsα: gb_n+1⇒d⁰l together with a 2-cell d⁰⁰g⇒c_n+1. But this would mean that b_n+1 ^g I, for g∈Z, which is a contradiction.

Suppose finally that ¯b_n+1 ^x I, for x∈X . Again, by definition of the transition relation, and exploiting the existence of redex-GRPOs, we find GRPOs xc_n+1⇒d⁰⁰g and gb_n+1⇒d⁰l, which mean that b_n+1 ^g I, for g∈Y .

3.3 Weak Equivalences

Theorems 3.2, 3.3, and 3.4 can be extended to weak equivalences, as outlined below.

For f a label of GTS(C) define a weak transition a ^fI b to be a mixed sequence of transitions and reductions a B^{∗ f} I B^∗ b.

Observe that this definition essentially identifies silent transitions in the LTS with reductions. As a consequence, care has to be taken to avoid interference with transitions^equiIsynthesised fromGRPOs and labelled by an equivalence. These transitions have essentially the same meaning as silent transitions (i.e. no context involved in the reduction), and must therefore be omitted in weak observations. This lead to consider the following definitions.

Definition 3.5 (Weak Traces and Failures). A sequence f₁···f_nof non- equivalence labels ofGTS(C)is a weak trace of a if

a ^f1Ia1···an−1 fn

Ian

for some a₁,...,a_n. The weak trace preorder is then defined accordingly.

A weak failure of a is a pair (f₁···f_n,X), where f₁···f_n and X are respectively a sequence and a set of non-equivalence labels, such that

f₁···f_n is a weak trace of a reaching a final state which is stable and refuses X . The weak trace preorder is defined accordingly.

Definition 3.6 (Weak Bisimulation). A symmetric relation

S

^onGTS(C) is a weak bisimulation if for all a

S

^b

a ^f Ia⁰ f not an equivalence, implies b ^fIb⁰with a⁰

S

^b0;

a Ba⁰ implies b B^∗b⁰with a⁰

S

^b0.

Using the definitions above Theorems 3.2, 3.3, and 3.4 can be lifted, respectively, to weak traces, failures and bisimulation.

It is worth remarking that the congruence results, however, only hold for contexts c∈D, as it is well known that non reactive contexts (i.e.

those c where ca B cb does not follow from a Bb, as e.g. the CSS context c=c₀+−) do not preserve weak equivalences. Alternative definitions of weak bisimulations are investigated in [11], and they are applicable mutatis mutandis toGRPOs.

4 Bunches and Wires

The category of “bunches and wires” was introduced in [13] as a skeletal algebra of shared wirings, abstracting over the notion of names in, e.g., theπcalculus. Although elementary, its structure is complex enough to lack RPOs.

A bunch context of type m₀→m₁ consists of an ordered set of m₁ trees of depth 1 containing exactly m₀holes. Leaves are labelled from an alphabet

K

^.

Definition 4.1. The category of bunch contexts Bun₀has

• objects the finite ordinals, denoted m₀,m₁,...

• arrows are bunch contexts c= (X,char,root): m₀→m₁, where X is a finite carrier, root : m₀+X→m₁is a surjective function linking leaves (X ) and holes (m₀) to their roots (m₁), and char : X→

K

^is

a leaf labelling function.

Composing c₀: m₀→m₁and c₁: m₁→m₂means filling the m₁holes of c₁with the m₁trees of c₀. Formally, c₁c₀is(X,root,char)where

X=X₀+X₁, root=root₁(root₀+id_X1), char= [char₀,char₁],

where + and [ , ] are, resp., coproduct and copairing. Identities are (/0,!,id): m₀→m₀.

A homomorphism of bunch contextsρ: c⇒c⁰: m₀→m₁is a func- tion ρ: X → X⁰ which respects root and char, i.e. root⁰ρ= root and char⁰ρ=char. An isomorphism is a bijective homomorphism. Isomor- phic bunch contexts are equated, making composition associative and Bun₀a category.

A bunch context c : m₀→m₁can be depicted as a string of m₁nonempty multisets on

K

+m₀, with the proviso that elements m₀ must appear ex- actly once in the string. In the examples, we represent elements of m0as numbered holes−i.

As we mentioned before, RPOs do not exist in Bun₀. Indeed, con- sider (i) below together with the two candidates (ii) and (iii). It is easy to show that these have no common “lower bound” candidate.

1

1

{K,−1} @@

1

{K,−1}

^^>>>

>>>>>>

0

{K}

@@

{K}

^^===

======

(i)

1

1

{K,−1} @@

{−1} //1

{K,−1}

OO

{−1} 1

oo

{K,−1}

^^>>>

>>>>>>

0

{K}

@@

{K}

^^===

======

(ii)

1

1

{K,−1} @@

{−1}{K}//2

{−1,−2}

OO

{K}{−oo 1} 1

{K,−1}

^^>>>

>>>>>>

0

{K}

@@

{K}

^^===

======

(iii) The point here is that by taking the arrows of Bun₀ up to isomorphism we lose information about how bunch contexts equal each other. Diagram (i), for instance, can be commutative in two different ways: the K in the bottom left part may corresponds either to the one in the bottom right or to the one in the top right, according to whether we read {K,−1}or {−1,K} for the top rightmost arrow. In order to track this information we endow Bun₀with its natural 2-categorical structure.

Definition 4.2. The 2-category of bunch contexts Bun has:

• objects the finite ordinals (cf. §1), denoted m0,m1,...

• arrows c= (x,char,root): m₀ →m₁ consist of a finite ordinal x, a surjective function root : m₀⊕x→m₁ and a labelling function char : x→

K

^.

• 2-cellsρare isomorphisms between bunches’ carriers.

Composition of arrows and 2-cells is defined in the obvious way. Notice that since⊕is associative, composition in Bun is associative. Therefore Bun is aG-category.

Replacing the carrier set X with a finite ordinal x allows us to avoid the unnecessary burden of working in a bicategory, which would arise because sum on sets is only associative up to isomorphism. Observe that this simplification is harmless since the set theoretical identity of the elements of the carrier is irrelevant. We remark, however, thatGRPOs are naturally a bicategorical notion and would pose no particular challenge in bicategories.

Theorem 4.3. Bun hasGRPOs.

Proof. In the following, we use only the fact that Bun is an extensive [2]

category with pushouts.

Suppose that we have an isomorphic 2-cellρ: ca⇒dl as illustrated below:

m₃ m1 ρ

cwww;;

w m2

cc d

GGGG

m₀. ^l

;;w

ww a w

ccGGGG

Usingρand the injections into the chosen coproduct in Ord (which are unlabelled below), we take four pullbacks obtaining the following diagram. Due to the extensivity of Ord, all the outside arrows are co- product injections.

x_c1

c1

l1 //x_l

x_a1

l2

oo

a1

x_c //x_a⊕x_c ^ρ//x_l⊕x_d x_c⊕x_a^ρoo x_aoo

x_c2

c2OO

d₁ //x_d

OO

x_a2

d₂

oo ^a2OO

Using the morphisms from the diagram above as building blocks, we can construct bijectionsγ: x_c→x_c1⊕x_c2,δ: x_a2⊕x_c2 →x_d andβ: xa⊕ xc₁ →x_l⊕xa₂ such that

x_l⊕δ.β⊕xc₂.xa⊕γ=ρ. (3)

Let root_c1 and root_a2 be the morphisms making (i) below m₀⊕x_a⊕x_c1

roota⊕x_c1

m0⊕β//m₀⊕x_l⊕x_a2^rootl⊕xa2//m₂⊕x_a2

root_a2

m₁⊕x_c1

root_c1 //m4

(i)

m₃ m1

γ c1 //

β czzzzz<<

zz

m4 c2 δ

OO

m2 a2

oo

bb d

DDDDDDD

m₀ ^l

<<

zz zz zz z a

bbDDDDDDD

(ii) into a pushout diagram. We can define charc₁, chara₂ and charc₂ in the obvious way.

Now consider the diagram below:

m₀⊕x_a⊕x_c1

roota⊕x_c1

mQ₀Q⊕QQxQaQ⊕QcQ₁QQQQQ((

m0⊕β//m₀⊕x_l⊕x_a2

(†) m0⊕xl⊕d2

((Q

QQ QQ QQ QQ QQ Q

root_l⊕x_a2

//m₂⊕x_a2

m₂⊕d₂

m₀⊕x_a⊕x_c

(‡) roota⊕xc

m0⊕ρ//m₀⊕x_l⊕x_d

root_l⊕xd

//m₂⊕x_d

root_d

m₁⊕x_c1

m₁⊕c₁ //m₁⊕x_c

rootc

//m₃.

Region (†) can be verified to be commutative using (3) while region (‡) commutes since ρ is a homomorphism. Using the pushout prop- erty, we get a unique function h : m₄ →m₃. Thus we define root_c2 = [h,root_ci]: m₄⊕x_c2→m₃. It is easy to verify that this function is surjec- tive, indeed, using rootc₁: m₁⊕x_c1→m₄we get a function[root_c1,x_c2]: m₁⊕ x_c→m₄⊕x_c2 such that

m₁⊕x_c

[root_c1,x_c2]

rootc

$$I

II II II II

m₄⊕x_c2

root_c2//m₃

is commutative. Thus surjectivity follows from the surjectivity of root_c. We shall verify thatβ: c₁a→a₂l, γ: c→c₂c₁ andδ: c₂a₂→d are homomorphisms. Because of the definition of char_c1, char_c2 and char_a2 it is clear that each of the functions preserves the character. Notice that