Event Model Decomposition

Introduction

One of the most important feature of the Event-B approach is the possibility to introduce new events during refinement steps, but a consequence is an increasing complexity of the refinement process when having to deal with many events and many state variables.

The main idea of the decomposition is to cut a model $M$ into sub-models $M_1, ..., M_n$, which can be refined separately and more comfortably than the whole.

The constraint that shall be satisfied by the decomposition is that these refined models might - the recomposition will never be performed in practice - be recomposed into a whole model $MR$ in a way that guarantees that $MR$ refines $M$. An event-based decomposition of a model is detailed in the Event Model Decomposition article: the events of a model are partitioned to form the events of the sub-models. In parallel, the variables on which these events act are distributed among the sub-models.

The purpose is here to precisely describe what is required at the Rodin platform level to integrate this event model decomposition, and to explain why. The details of how it could be implemented are out of scope.

Terminology

• Event model decomposition: The decomposition of a model, as defined in the modelling language, in sub-models. Other decomposition structures for Event-B are not considered here.

A model can contain contexts, machines, or both. The notion of model decomposition covers on the one hand the machine decomposition, and on the other hand the context decomposition, both being interdependent.

• Sub-machine: A machine built from a non-decomposed machine during the event model decomposition.
• Sub-context: A context built from a non-decomposed context during the event model decomposition.
• Shared variable: A variable of a given machine which is accessed by events distributed in distinct sub-machines (by opposition to private variable).
• Private variable: A variable of a given machine which is only accessed by events of the same sub-machine (by opposition to shared variable).
• External event: An event of a sub-machine which is built from an event of the non-decomposed machine, and which simulates the way the shared variables (between this sub-machine and another sub-machine) are handled in the non-decomposed machine (by opposition to internal event).
• Internal event: An event copied from the non-decomposed machine to a sub-machine, according to the end-user specified distribution (by opposition to external event).

Note that a variable is said to be accessed when it is read or written. More precisely, such an access may be performed by a predicate (invariant, guard, witness) or in an assignment (action).

High-level Specification

The high-level specification details how the event model decomposition shall be integrated into the Rodin platform as a new feature, by linking to the existing architecture.

Definition of the decomposition

It is necessary to first give a definition of the event model decomposition in the Rodin platform. Is it an Event-B project decomposition? Or, is it a decomposition performed from some well-identified machines and contexts of a given Event-B project?

The entry point for the decomposition is a machine $M$.

The machine $M$ to be taken as input for the decomposition shall be pointed by the end-user.

Configuration of the decomposition

The end-user shall be asked to parametrize the decomposition, and more precisely to:

• identify the machine to be taken as input for the decomposition.
• identify the sub-machines to be created.
• partition the events.

It is more suitable for the end-user to visualize the configuration, and as a consequence it shall preferably be performed through the Graphical User Interface of the Rodin platform.

The following dialog box, which fills these requirements, is given as an example. The left-hand side displays the non-decomposed model and the right-hand side the decomposed model. The second one is built by the user, by first adding machines and then copying events from left to right.

Execution of the decomposition

A Decompose action shall be added. It shall be enabled if and only if a machine is selected. It shall be available from the Project menu, from the toolbar, and in the contextual menu displayed when right-clicking on the selected project.

A new Event-B project shall be created for each sub-machine built during the configuration. The decomposition of the sub-machine shall first be completed, and then the non-decomposed context shall be partitioned.

As far as possible, the developments shall not be performed in the Event-B core; the dedicated extension points shall be used instead (eg. those provided for the static checker. See the plugin.xml file of the org.eventb.core package).

Generation of the proof obligations

A model to be decomposed is assumed to be proved, i.e. the proof obligations (PO) have all been handled successfully. The following conditions on PO shall be fulfilled during the decomposition:

• The decomposition shall not generate any new proof obligation.
• The proof obligations related to the non-decomposed model shall not be "propagated" to the decomposed models to be proved again. As a consequence, the Proof Obligation Generator (POG) shall be temporary disconnected until the decomposition is performed.

Low-level Specification

The low-level specification details through several steps how the event model decomposition shall be performed, and in which order. It establishes a distinction between the steps performed on the end-user's initiative, and the computed ones. It links when possible to the already implemented features of the Rodin platform which can be used at some steps.

The following sequence shall be observed to decompose a machine $M$ in sub-machines $M_1,...,M_n$:

1. The events of $M$ shall be partitioned in $M_1,...,M_n$, as indicated by the end-user (see Event partition).
2. The Rodin platform shall update the status of the events (see Event status).
3. The Rodin platform shall distribute the variables of $M$ in $M_1,...,M_n$, according to the event partition (see Variable distribution).
4. The Rodin platform shall distribute the invariants of $M$ in $M_1,...,M_n$, according to the variable distribution (see Invariant distribution).
5. The Rodin platform shall build external events in $M_1,...,M_n$, from the events of $M$, and according to the variable distribution (see External event construction).
6. The Rodin platform shall build the initialization events of $M_1,...,M_n$, according to the variable distribution (see Initialization event construction). I have a strong feeling that the INITIALISATION is just another external event.
7. The Rodin platform shall build the contexts seen by $M_1,...,M_n$, from the hierarchy of contexts associated to $M$ (see Context decomposition).

• Is there a step missing here? Copying the "internal" events accordingly to the M.

This order will be justified by itself subsequently, when going into the details of the decomposition.

Decomposition of a machine in sub-machines

The purpose of this paragraph is to precisely describe how to decompose $M$.

The refinement hierarchy for $M$ (see the $REFINES$ clauses) shall not be considered for the decomposition. A sub-machine $M_i$ may indeed be seen as a new abstract machine, which may be later refined if necessary. It is necessary to keep this assumption in mind when decomposing $M$.

About the variables

Some variables are needed by several sub-machines of the decomposition. As a consequence, these variables shall be replicated in the sub-machines. Beyond that, since it is not possible to ensure that such a variable will be refined in the same way in each sub-machine, they shall be given a special status (shared variable), with the limitation that they cannot be refined.

We will specify in this section how to introduce the notion of shared variable in the Rodin platform, and how to check the associated rules.

Defining a variable as shared

The following DTD excerpt describes the structure of a variable in the Rodin database:

<!ENTITY % NameAttDecl "name CDATA #REQUIRED">
<!ENTITY % CommentAttDecl "org.eventb.core.comment CDATA #IMPLIED">
<!ENTITY % IdentAttDecl "org.eventb.core.identifier CDATA #REQUIRED">

<!ELEMENT %variable; EMPTY>
<!ATTLIST %variable;
  %NameAttDecl;
  %CommentAttDecl;
  %IdentAttDecl;
  >

A first possibility to tag a variable as shared would be to add a shared specific attribute, which would be set to true if and only if the variable is shared:

<ENTITY % shared "org.eventb.core.shared CDATA #REQUIRED">

<!ELEMENT %variable; EMPTY>
<!ATTLIST %variable;
  ...
  %shared; (false|true) #REQUIRED
  >

Another possibility would be to define a more generic attribute, which could take different values, according to the nature of the variable:

<ENTITY % nature "org.eventb.core.nature CDATA #REQUIRED">

<!ATTLIST %variable;
  ...
  %nature; (0|1) #REQUIRED
  >

The second option, which has the main advantage to be more scalable, is retained here.

Ensuring that a shared variable is not data-refined

A shared variable shall always be present in the state space of any refinement of the component. The verification shall be added to those already performed by the static checker. The static checker shall have a way to determine if a given variable is shared or not.

Distributing the variables in the sub-machines of the decomposition

The first question raised by the distribution of the variables is whether it shall be the first stage of the decomposition, or not. Let's first suppose that the answer is "yes". The case where $e$ is an event that accesses a variable $v1$ associated to a sub-machine $M_1$ and a variable $v2$ associated to a sub-machine $M_2$ cannot be successfully handled: should $e$ be associated to $M_1$ or to $M_2$? Moreover, contrary to the events, the variables are not essentially bearers of meanings, and they cannot by themselves guide the decomposition.

As a consequence, it is pertinent to assume that the events have been first partitioned. The following cases have then to be taken into consideration when dealing with the variable distribution:

• If $v$ is a variable that is only accessed by events of a given sub-machine $M_i$, then $v$ is a private variable of $M_i$. It shall be moved to $M_i$.
• If $v$ is a variable that is accessed by events of distinct sub-machines $M_i, ..., M_j$, then $v$ is a shared variable. It shall be tagged as such and duplicated in all sub-machines.

If all the variables are shared at the conclusion of the distribution, the end user shall be notified (it certainly means that the decomposition was not judicious!).

Propagating the sharing status

A variable tagged as shared in the non-decomposed machine (when resulting from a previous decomposition) shall remain shared in the sub-machines.

About the events

It shall be possible to simulate the way the shared variables are handled in the non-decomposed machine. This is precisely the purpose of the so-called external events.

We will examine in this section how to define such events in the Rodin platform, how to construct them, and how to enforce the rules that apply (in particular, these events cannot be refined).

Partitioning the events in the sub-machines of the decomposition

The sub-machines $M_1,...,M_n$ shall be created and the events shall be partitioned, according to the end-user configuration.
At this step, the sub-machines shall only contain these internal events. In particular, the $SEES$ and $REFINES$ clauses of $M_1,...,M_n$ shall be empty. In the same manner, the $REFINES$ and $WITH$ (witnesses) clauses of the events shall be left out.
If an event has the extended status (i.e. it inherits the actions of the refined event), then it shall first be merged with the refined event before being copied in the sub-machine. Note that such a merge is performed by the pretty printer of the Rodin platform (compare the information displayed, on the one hand in the "Pretty Print" view, and on the other hand in the "Edit" view, for an extended event).

Propagating the event status

The convergence status of a given event shall be propagated in the sub-machines as described below:

• An event tagged as ordinary in the non-decomposed machine shall remain ordinary in the sub-machine.
• An event tagged as convergent in the non-decomposed machine shall become ordinary in the sub-machine.
• An event tagged as anticipated in the non-decomposed machine shall remain anticipated in the sub-machine.
• An external event shall always be declared as ordinary.

See the modelling language for precisions on the convergence status.

In the same manner, an event (internal or external) of a sub-machine shall always be declared as non-extended.

An event tagged as external in the non-decomposed machine (when resulting from a previous decomposition) shall remain external in the sub-machine.

Identifying an event as external

An attribute is already defined, which is introduced below, to precise the nature of an event. A first solution would be to add another masked value (eg. 4) to encode the external status.

<!ENTITY % convergence "org.eventb.core.convergence">

<!ATTLIST %event;
 ...
 %convergence; (0|1|2) #REQUIRED
 ...
 >

Another solution would be to add a distinct external attribute, which would be set to true if and only if the event is external:

<ENTITY % external "org.eventb.core.external CDATA #REQUIRED">

<!ATTLIST %event;
 ...
 %external; (false|true) #REQUIRED
 >

This solution is preferred because the notion of external event is totally orthogonal to the notion of convergence.

Ensuring that an external event cannot be refined

The verification shall be performed by the static checker. The static checker shall have a way to determine if a given event is external or not.

Constructing an external event

If $e$ is an event that modifies a shared variable $s$ (i.e. $s$ is listed among the free identifiers on the left-hand side of an assignment), then an external event that modifies $s$ shall be built from $e$ in each sub-machine where $s$ is accessed.

The construction of an external event depends on the source machine (i.e. the sub-machine containing the event $e$ from which the external event is to be built) and on the destination machine (i.e. the sub-machine where the external event is to be built).
Building an external event from a given event $e$ modifying a shared variable $s$ and duplicating it in each sub-machine where $s$ is accessed is indeed incorrect, as illustrated below: the sub-machine $M_3$ does not know the shared variable $s2$ and the sub-machine $M_1$ does not know the shared variable $s4$.

The construction of an external event highly relies on some rewriting rules. It is recommended to peruse them before reading further.

$e$ is an internal event of the source machine $M_s$, $v_i$ are private variables of $M_s$, $s_i$ are shared variables between $M_s$ and the destination machine $M_d$, $P, P_i, Q_i$ are before-after predicates, and $G$ is a predicate.

Generic construction
We first focus on the generic construction of an external event from an internal event whose action is expressed as follows:

 e
 WHERE 
   $G(v_1,...,v_n,s_1,...,s_m)~$      
 THEN                  
   $v_1,...,v_n,s_1,...,s_m \bcmsuch P(v1,...,v_n,s_1,...,s_m, v_1',...,v_n',s_1',...,s_m')$

1. The first step of the construction consists in replacing the private variables by parameters. Note that this step is purely fictive, because assigning an event parameter is not allowed!
Are the $v_1, .. v_n$ "private" variables. They are rather the variables that are not "accessed" by the distribution M, i.e. not "private", not "shared". e ANY $x_1,...,x_n~$ WHERE $G(x_1,...,x_n,s_1,...,s_m)~$ THEN $x_1,...,x_n,s_1,...,s_m \bcmsuch P(x1,...,x_n,s_1,...,s_m, x_1',...,x_n',s_1',...,s_m')$
2. The second step consists in adding guards to define the types of the parameters, if necessary. More precisely, a theorem shall be added for each parameter for which typing is required. The .bcm file associated to the non-decomposed machine shall be parsed in order to retrieve the typing information.
3. The third and last step of the construction consists in introducing an existential quantifier to resolve the invalid assignment. $external\_e$ is the newly built external event.
external_e ANY $x_1,...,x_n~$ WHERE $G(x_1,...,x_n,s_1,...,s_m)~$ THEN $s_1,...,s_m \bcmsuch \exists y_1,...y_n \qdot P(x1,...,x_n,s_1,...,s_m, y_1,...,y_n,s_1',...,s_m')$

Practically, not necessarily all the variables $v_1, .. v_n$ need to be moved to be parameters. Amongst those variables, only the one that appear in the guard $G$ and in the decomposed action, i.e. in $P$ need to be copied. Probably, this is a "derived construction".

Derived constructions
Then, it is possible to derivate the construction for other actions:

1. Each assignment of the action shall be handled separately (see transformation rules on Event-B actions).
2. The transformation rules shall be applied on each Event-B assignment.

Example

 e
 WHERE 
   $G(v_1,...,v_n,s_1,...,s_m)~$      
 THEN                  
   $\begin{array}{ll}v_1\!\!\! &\bcmsuch P_1(v_1,...,v_n,s_1,...,s_m,v_1')\\
    ...\\
    v_n\!\!\! &\bcmsuch P_n(v_1,...,v_n,s_1,...,s_m,v_n')\\
    s_1\!\!\! &\bcmsuch P_{n+1}(v_1,...,v_n,s_1,...,s_m,s_1')\\
    ...\\
    s_m\!\!\! &\bcmsuch P_{n+m}(v_1,...,v_n,s_1,...,s_m,s_m') \end{array}$

 external_e
 ANY $x_1,...,x_n~$
 WHERE 
   $G(x_1,...,x_n,s_1,...,s_m)~$      
 THEN                  
   $\begin{array}{ll}s_1\!\!\! &\bcmsuch P_{n+1}(x_1,...,x_n,s_1,...,s_m,s_1')\\
    ...\\
    s_m\!\!\! &\bcmsuch P_{n+m}(x_1,...,x_n,s_1,...s_m,s_m') \end{array}$

Decomposing the initialization event

An initialization event shall be built in each sub-machine from the initialization event of the non-decomposed machine, and according to the distribution of the variables among these sub-machines. The construction is detailed below. $initialization$ is the initial event and $e$ the built event, $x_i$ are variables (private or shared) of the sub-machine containing $e$, $y_i$ are variables of other sub-machines, and $P$ is a predicate.

 initialization    
 THEN                  
   $x_1,...,x_n,y_1,...,y_m \bcmsuch P(x_1',...,x_n',y_1',...,y_m')$

Only the variables of the considered sub-machine shall appear in the built initialization event; other variables shall become bound:

 e   
 THEN                  
   $x_1,...,x_n \bcmsuch \exists z_1,...,z_m \qdot P(x_1',...,x_n',z_1,...,z_m)$

The derived cases and simplification rules introduced during the construction of the external events apply here as well.

About the invariants

We will see in this section how to distribute the invariants among the sub-machines, once the variables have been distributed.

An invariant based on a $P(v_m,...,v_n)~$ predicate shall be copied in a sub-machine $M_i$ if and only if $M_i$ contains the $\{v_m,...,v_n\}~$ variables.

This distribution is illustrated in the figure below, where $M_0$ is an abstract machine, $M$ is a concrete machine extending $M_0$, $M_1$ and $M_2$ are the sub-machines resulting from the decomposition of $M$. $v4$ is a private variable of $M_2$, $v3$ is a variable shared between $M_1$ and $M_2$, and $P_i$ are predicates.

If an invariant is used for typing but it has not been copied, a theorem shall be added in the sub-machines for each variable for which typing is required (otherwise a problem will be detected by the static checker). Note that there is no contradiction with the requirements on proof obligations; no proof obligation (PO) is indeed generated for predicates $v \in T$, where $v$ is a variable and $T$ is a type.

In the same manner, if an invariant is used for well-definedness but is has not been copied, a theorem shall be added in the sub-machines. The associated proof obligation is the well-definedness (WD) obligation, and it is assumed to be proved; as a consequence, it does not have to be generated in the sub-machines.
For example, let $c$ be a constant. $P0$, $P1$ and $P2$ are defined as follows:
$\begin{array}{ll}P0\!\!\! &\defi v1 \in \nat \land v2 \in \nat\\ P1\!\!\! &\defi v3 = c \div v4\\ P2\!\!\! &\defi v4 = v1 + 1 \end{array}$
A theorem shall be added to indicate that $v4$ is not null.
The org.event.core.ast.Formula.getWDPredicate method shall be used to compute the WD predicate associated to a given predicate. If no WD predicate is required, this method returns true. The built WD predicate shall be inserted before the associated predicate in the list of invariants of the considered sub-machine.

Beyond that, a workaround exists if an invariant $inv$, based on a $P(v_m,...,v_n)~$ predicate, seems useful (for example an invariant between a concrete variable and some abstract variable) but it has disappeared in a sub-machine $M_i$ containing variables $\{v_p,...,v_q\}~$ (eg. $inv2$ and $inv3$ have both been excluded from the sub-machines by application of the stated rules). It is indeed possible to add in the non-decomposed machine a theorem based on $P~$, but where the variables $\{v_m,...,v_n\} \setminus \{v_p,...,v_q\}$ become bound, and then to perform again the decomposition. It will lead to a new proof obligation in the non-decomposed machine, which does not pose any difficulty.
For example, if a theorem $inv4: P4(v4)~$, with $P4 \defi \exists v1 \qdot P2(v1,v4)$, is added to the $M$ concrete machine, then it will be copied in the sub-machine $M_2$ during the decomposition. In order to prove the $P4~$ statement, the bound variable $v1$ shall obviously be instantiated with $v1$.

About the variants

As mentioned before, there is no convergent event in sub-machines. As a consequence, there is no need to take the variants into consideration when performing the decomposition.

Decomposition of a context in sub-contexts

The purpose of this paragraph is to specify how to decompose a context, according to the decomposition of a given machine $M$, and to establish how to link the sub-contexts to the sub-machines.

The hierarchy of contexts (see the $EXTENDS$ clauses of contexts and the $SEES$ clause of $M$) shall be first accumulated in a single context. More precisely, a new context $F$ shall be built (virtually or not), which contains all the carrier sets, constants and axioms of the hierarchy. This context is assumed to be the non-decomposed context from which the sub-contexts $C_1,...,C_n$ shall be built.
Note that it may be necessary to rename some axioms when flattening the hierarchy.

Then, an empty context $C_i$ shall be built for each sub-machine $M_i$, by respecting the following sequence: the constants shall be first included, then the carrier sets shall be added, and finally the axioms shall be considered.
$C_i$ shall be linked to $M_i$ through its $SEES$ clause. Note that, at the conclusion of the context decomposition, the sub-contexts $C_i$ that may be empty shall not be kept, and a $SEES$ clause shall not be added to the associated sub-machines $M_i$.

About the constants

A constant of a non-decomposed context $F$ shall be copied in a sub-context $C_i$ if and only if it appears in a predicate (invariant or guard) or in an assignment (action) of the associated sub-machine $M_i$.
Thus, if a constant is not accessed by a sub-machine (e.g. if such a constant was only accessed by an abstract machine extended by the non-decomposed machine), then it shall be left out during the decomposition. In parallel, if a constant is accessed by distinct sub-machines, then it shall be duplicated in the associated sub-contexts.

About the carrier sets

A carrier set shall be visible from any sub-machine $M_i$ which accesses it, explicitly or implicitly, through a predicate or an assignment. In other terms, a carrier set of a non-decomposed context $F$ shall be copied in a sub-context $C_i$ if and only if this set appears in a predicate or assignment of the associated sub-machine $M_i$, or types a constant previously copied to $C_i$.
As for the constants, a carrier set may be left out or duplicated.

About the axioms

We will see in this section how to distribute the axioms among the sub-contexts, once the constants and carrier sets have been copied.

As for the invariants, an axiom is copied in a sub-machine $M_i$ if and only if $M_i$ contains the referenced constants and sets. If an axiom is used for typing but it has not been copied, a theorem shall be added in the sub-contexts for each constant for which typing is required (otherwise a problem will be detected by the static checker).

Mathematical Approach

The purpose of this section is to mathematically formalize the Event-B decomposition previously specified, and by the way to remove the possible remaining ambiguity.

Let's define $\mathit{MACHINE}$ as the set of all machine handles, $\mathit{EVENT}$ the set of all events, and $\mathit{VAR}$ the set of all variables.

• The partition of the events of the non-decomposed machine among the different sub-machines (according to the end-user configuration) can be represented with a partial function:

$\mathit{partition}\in\mathit{EVENT}\pfun\mathit{MACHINE}$
For a given sub-machine $m$, $partition^{-1}[\{m\}]~$ is then the set of internal events of $m$.

• The access of a variable by a given event (according to the static-checker) can be expressed as:

$\mathit{access}\in\mathit{EVENT}\rel\mathit{VAR}$
For a given variable $v$, $(partition;access^{-1})[\{v\}]~$ is then the set of the sub-machines accessing $v$, and $v$ is a private variable of a sub-machine $m$ if and only if this set contains a single component (i.e. $card((partition;access^{-1})[\{v\}]) = 1~$); otherwise, and if and only if this set is not empty, $v$ is shared.
In parallel, for a given sub-machine $m$, $(access;partition^{-1})[\{m\}]~$ is the set of variables accessed by the events contained in $m$.

• The association of a variable with the events modifying this variable (according to the static-checker) can be specified as:

$\mathit{modify}\in\mathit{VAR}\rel\mathit{EVENT}$
For a given sub-machine $m$ and a variable $v \in (access;partition^{-1})[\{m\}]$, $modify[\{v\}]~$ is then the set of the events modifying $v$.

• The construction of the external events for a sub-machine can be represented with a relation:

$\mathit{extern}\in\mathit{MACHINE}\rel\mathit{EVENT}$
It is computed as follows: $extern = (modify;access;partition^{-1}) \setminus partition^{-1}$
Thus, the external events of a given sub-machine $m$ are events modifying the variables accessed by the internal events of $m$, but they are not internal events of $m$.

Example

The following example is taken from the Event Model Decomposition.

A non-decomposed machine has been decomposed in two sub-machines $M_1$ and $M_2$, as illustrated by the figure.
According to the terminology, $in\_a$ and $a\_2\_b$ are internal events of $M_1$, and $b\_2\_c$ and $out\_c$ are internal events of $M_2$. Concerning the variables, $a$ and $m$ are private variables of $M_1$, $c$ and $p$ are private variables of $M_2$, and $b$, $r$ and $s$ are shared variables.
The variables accessed by the internal events of $M_1$ are $a$, $m$, $b$, $r$ and $s$. The events modifying these variables are $in\_a$, $a\_2\_b$, which both are internal events of $M_1$, and $b\_2\_c$, which is an internal event of $M_2$. Thus, according to the definition given above, $b\_2\_c$ is an external event for $M_1$. In the same manner, $a\_2\_b$ is an external event for $M_2$.

N.B.: Note that the expression "is an external event for" is an extrapolation, and shall be literally interpreted as "should lead to the construction of an external event in".

Example

It seems to be interesting to illustrate the decomposition step by step through an example. The refinements leading to the non-decomposed model considered here are described in detail in the Event Model Decomposition. TO BE COMPLETED

The events

The variables

TO BE COMPLETED
...
...
...
...
...
...
...
...
...

The invariants

TO BE COMPLETED
...
...
...
...
...
...
...
...
...
...
...
...

Bibliography

• J.R. Abrial, The B-book: assigning programs to meanings, Cambridge University Press, 1996 (ISBN 0-521-49619-5).
• J.R. Abrial, Mathematical Models for Refinement and Decomposition, in The Event-B Book, to be published in 2009 (lien externe).
• J.R. Abrial, Event Model Decomposition, Version 1.3, April 2009.
• M. Butler, Decomposition Structures for Event-B, in Integrated Formal Methods iFM2009, Springer, LNCS 5423, 2009 (lien externe).