Chapter 27 Generalized rewriting
 27.1 Introduction to generalized rewriting
 27.2 Commands and tactics
 27.3 Extensions
 27.4 Strategies for rewriting
This chapter presents the extension of several equality related tactics to work over userdefined structures (called setoids) that are equipped with adhoc equivalence relations meant to behave as equalities. Actually, the tactics have also been generalized to relations weaker then equivalences (e.g. rewriting systems). The toolbox also extends the automatic rewriting capabilities of the system, allowing the specification of custom strategies for rewriting.
This documentation is adapted from the previous setoid documentation by Claudio Sacerdoti Coen (based on previous work by Clément Renard). The new implementation is a dropin replacement for the old one,^{1} hence most of the documentation still applies.
The work is a complete rewrite of the previous implementation, based on the type class infrastructure. It also improves on and generalizes the previous implementation in several ways:
 Userextensible algorithm. The algorithm is separated in two parts: generations of the rewriting constraints (done in ML) and solving of these constraints using type class resolution. As type class resolution is extensible using tactics, this allows users to define general ways to solve morphism constraints.
 Subrelations. An example extension to the base algorithm is the ability to define one relation as a subrelation of another so that morphism declarations on one relation can be used automatically for the other. This is done purely using tactics and type class search.
 Rewriting under binders. It is possible to rewrite under binders in the new implementation, if one provides the proper morphisms. Again, most of the work is handled in the tactics.
 Firstclass morphisms and signatures. Signatures and morphisms are ordinary Coq terms, hence they can be manipulated inside Coq, put inside structures and lemmas about them can be proved inside the system. Higherorder morphisms are also allowed.
 Performance. The implementation is based on a depthfirst search for the first solution to a set of constraints which can be as fast as linear in the size of the term, and the size of the proof term is linear in the size of the original term. Besides, the extensibility allows the user to customize the proof search if necessary.
27.1 Introduction to generalized rewriting
27.1.1 Relations and morphisms
A parametric relation R is any term of type forall (x_{1}:T_{1}) …(x_{n}:T_{n}), relation A. The expression A, which depends on x_{1} …x_{n}, is called the carrier of the relation and R is said to be a relation over A; the list x_{1},…,x_{n} is the (possibly empty) list of parameters of the relation.
An instance of a parametric relation R with n parameters is any term (R t_{1} …t_{n}).
Let R be a relation over A with n parameters. A term is a parametric proof of reflexivity for R if it has type forall (x_{1}:T_{1}) …(x_{n}:T_{n}), reflexive (R x_{1} …x_{n}). Similar definitions are given for parametric proofs of symmetry and transitivity.
A parametric unary function f of type forall (x_{1}:T_{1}) …(x_{n}:T_{n}), A_{1} > A_{2} covariantly respects two parametric relation instances R_{1} and R_{2} if, whenever x, y satisfy R_{1} x y, their images (f x) and (f y) satisfy R_{2} (f x) (f y) . An f that respects its input and output relations will be called a unary covariant morphism. We can also say that f is a monotone function with respect to R_{1} and R_{2}. The sequence x_{1},… x_{n} represents the parameters of the morphism.
Let R_{1} and R_{2} be two parametric relations. The signature of a parametric morphism of type forall (x_{1}:T_{1}) …(x_{n}:T_{n}), A_{1} > A_{2} that covariantly respects two instances I_{R1} and I_{R2} of R_{1} and R_{2} is written I_{R1} ++> I_{R2}. Notice that the special arrow ++>, which reminds the reader of covariance, is placed between the two relation instances, not between the two carriers. The signature relation instances and morphism will be typed in a context introducing variables for the parameters.
The previous definitions are extended straightforwardly to nary morphisms, that are required to be simultaneously monotone on every argument.
Morphisms can also be contravariant in one or more of their arguments. A morphism is contravariant on an argument associated to the relation instance R if it is covariant on the same argument when the inverse relation R^{−1} (inverse R in Coq) is considered. The special arrow > is used in signatures for contravariant morphisms.
Functions having arguments related by symmetric relations instances are both covariant and contravariant in those arguments. The special arrow ==> is used in signatures for morphisms that are both covariant and contravariant.
An instance of a parametric morphism f with n parameters is any term f t_{1} …t_{n}.
The signature of the function union A is set_eq A ==> set_eq A ==> set_eq A for all A.
Leibniz equality is a relation and every function is a morphism that respects Leibniz equality. Unfortunately, Leibniz equality is not always the intended equality for a given structure.
In the next section we will describe the commands to register terms as parametric relations and morphisms. Several tactics that deal with equality in Coq can also work with the registered relations. The exact list of tactic will be given in Sect. 27.2.2. For instance, the tactic reflexivity can be used to close a goal R n n whenever R is an instance of a registered reflexive relation. However, the tactics that replace in a context C[] one term with another one related by R must verify that C[] is a morphism that respects the intended relation. Currently the verification consists in checking whether C[] is a syntactic composition of morphism instances that respects some obvious compatibility constraints.
27.1.2 Adding new relations and morphisms
A parametric relation Aeq: forall (y_{1} : β_{!} …y_{m} : β_{m}), relation (A t_{1} …t_{n}) over (A : α_{i} > …α_{n} > Type) can be declared with the following command:
Add Parametric Relation (x_{1} : T_{1}) …(x_{n} : T_{k}) : (A t_{1} …t_{n}) (Aeq t′_{1} …t′_{m})
[reflexivity proved by refl]
[symmetry proved by sym]
[transitivity proved by trans]
as id.
after having required the Setoid module with the Require Setoid command.
The identifier id gives a unique name to the morphism and it is used by the command to generate fresh names for automatically provided lemmas used internally.
Notice that the carrier and relation parameters may refer to the context of variables introduced at the beginning of the declaration, but the instances need not be made only of variables. Also notice that A is not required to be a term having the same parameters as Aeq, although that is often the case in practice (this departs from the previous implementation).
In case the carrier and relations are not parametric, one can use the command Add Relation instead, whose syntax is the same except there is no local context.
The proofs of reflexivity, symmetry and transitivity can be omitted if the relation is not an equivalence relation. The proofs must be instances of the corresponding relation definitions: e.g. the proof of reflexivity must have a type convertible to reflexive (A t_{1} …t_{n}) (Aeq t′_{1} …t′_{n}). Each proof may refer to the introduced variables as well.
[reflexivity proved by @refl_equal A]
…
Some tactics (reflexivity, symmetry, transitivity) work only on relations that respect the expected properties. The remaining tactics (replace, rewrite and derived tactics such as autorewrite) do not require any properties over the relation. However, they are able to replace terms with related ones only in contexts that are syntactic compositions of parametric morphism instances declared with the following command.
Add Parametric Morphism (x_{1} : T_{1}) …(x_{k} : T_{k}) : (f t_{1} …t_{n})
with signature sig
as id.
Proof
…
Qed
The command declares f as a parametric morphism of signature sig. The identifier id gives a unique name to the morphism and it is used as the base name of the type class instance definition and as the name of the lemma that proves the welldefinedness of the morphism. The parameters of the morphism as well as the signature may refer to the context of variables. The command asks the user to prove interactively that f respects the relations identified from the signature.
Coq < Require Export Relation_Definitions.
Coq < Set Implicit Arguments.
Coq < Parameter set: Type > Type.
Coq < Parameter empty: forall A, set A.
Coq < Parameter eq_set: forall A, set A > set A > Prop.
Coq < Parameter union: forall A, set A > set A > set A.
Coq < Axiom eq_set_refl: forall A, reflexive _ (eq_set (A:=A)).
Coq < Axiom eq_set_sym: forall A, symmetric _ (eq_set (A:=A)).
Coq < Axiom eq_set_trans: forall A, transitive _ (eq_set (A:=A)).
Coq < Axiom empty_neutral: forall A (S: set A), eq_set (union S (empty A)) S.
Coq < Axiom union_compat:
forall (A : Type),
forall x x' : set A, eq_set x x' >
forall y y' : set A, eq_set y y' >
eq_set (union x y) (union x' y').
Coq < Add Parametric Relation A : (set A) (@eq_set A)
reflexivity proved by (eq_set_refl (A:=A))
symmetry proved by (eq_set_sym (A:=A))
transitivity proved by (eq_set_trans (A:=A))
as eq_set_rel.
Coq < Add Parametric Morphism A : (@union A) with
signature (@eq_set A) ==> (@eq_set A) ==> (@eq_set A) as union_mor.
Coq < Proof. exact (@union_compat A). Qed.
It is possible to reduce the burden of specifying parameters using (maximally inserted) implicit arguments. If A is always set as maximally implicit in the previous example, one can write:
reflexivity proved by eq_set_refl
symmetry proved by eq_set_sym
transitivity proved by eq_set_trans
as eq_set_rel.
Coq < Add Parametric Morphism A : (@union A) with
signature eq_set ==> eq_set ==> eq_set as union_mor.
Coq < Proof. exact (@union_compat A). Qed.
We proceed now by proving a simple lemma performing a rewrite step and then applying reflexivity, as we would do working with Leibniz equality. Both tactic applications are accepted since the required properties over eq_set and union can be established from the two declarations above.
eq_set (union (union S empty) S) (union S S).
Coq < Proof. intros. rewrite empty_neutral. reflexivity. Qed.
The tables of relations and morphisms are managed by the type class instance mechanism. The behavior on section close is to generalize the instances by the variables of the section (and possibly hypotheses used in the proofs of instance declarations) but not to export them in the rest of the development for proof search. One can use the Existing Instance command to do so outside the section, using the name of the declared morphism suffixed by _Morphism, or use the Global modifier for the corresponding class instance declaration (see §27.2.1) at definition time. When loading a compiled file or importing a module, all the declarations of this module will be loaded.
27.1.3 Rewriting and non reflexive relations
To replace only one argument of an nary morphism it is necessary to prove that all the other arguments are related to themselves by the respective relation instances.
When the relations associated to some arguments are not reflexive, the tactic cannot automatically prove the reflexivity goals, that are left to the user.
Setoids whose relation are partial equivalence relations (PER) are useful to deal with partial functions. Let R be a PER. We say that an element x is defined if R x x. A partial function whose domain comprises all the defined elements only is declared as a morphism that respects R. Every time a rewriting step is performed the user must prove that the argument of the morphism is defined.
27.1.4 Rewriting and non symmetric relations
When the user works up to relations that are not symmetric, it is no longer the case that any covariant morphism argument is also contravariant. As a result it is no longer possible to replace a term with a related one in every context, since the obtained goal implies the previous one if and only if the replacement has been performed in a contravariant position. In a similar way, replacement in an hypothesis can be performed only if the replaced term occurs in a covariant position.
Contrary to the previous implementation, no specific error message will be raised when trying to replace a term that occurs in the wrong position. It will only fail because the rewriting constraints are not satisfiable. However it is possible to use the at modifier to specify which occurrences should be rewritten.
As expected, composing morphisms together propagates the variance annotations by switching the variance every time a contravariant position is traversed.
27.1.5 Rewriting in ambiguous setoid contexts
One function can respect several different relations and thus it can be declared as a morphism having multiple signatures.
To declare multiple signatures for a morphism, repeat the Add Morphism command.
When morphisms have multiple signatures it can be the case that a rewrite request is ambiguous, since it is unclear what relations should be used to perform the rewriting. Contrary to the previous implementation, the tactic will always choose the first possible solution to the set of constraints generated by a rewrite and will not try to find all possible solutions to warn the user about.
27.2 Commands and tactics
27.2.1 First class setoids and morphisms
The implementation is based on a firstclass representation of properties of relations and morphisms as type classes. That is, the various combinations of properties on relations and morphisms are represented as records and instances of theses classes are put in a hint database. For example, the declaration:
Add Parametric Relation (x_{1} : T_{1}) …(x_{n} : T_{k}) : (A t_{1} …t_{n}) (Aeq t′_{1} …t′_{m})
[reflexivity proved by refl]
[symmetry proved by sym]
[transitivity proved by trans]
as id.
is equivalent to an instance declaration:
Instance (x_{1} : T_{1}) …(x_{n} : T_{k}) => id : @Equivalence (A t_{1} …t_{n}) (Aeq t′_{1} …t′_{m}) :=
[Equivalence_Reflexive := refl]
[Equivalence_Symmetric := sym]
[Equivalence_Transitive := trans].
The declaration itself amounts to the definition of an object of the record type Coq.Classes.RelationClasses.Equivalence and a hint added to the typeclass_instances hint database. Morphism declarations are also instances of a type class defined in Classes.Morphisms. See the documentation on type classes 20 and the theories files in Classes for further explanations.
One can inform the rewrite tactic about morphisms and relations just by using the typeclass mechanism to declare them using Instance and Context vernacular commands. Any object of type Proper (the type of morphism declarations) in the local context will also be automatically used by the rewriting tactic to solve constraints.
Other representations of first class setoids and morphisms can also be handled by encoding them as records. In the following example, the projections of the setoid relation and of the morphism function can be registered as parametric relations and morphisms.
Coq < Record Setoid: Type :=
{ car:Type;
eq:car>car>Prop;
refl: reflexive _ eq;
sym: symmetric _ eq;
trans: transitive _ eq
}.
Coq < Add Parametric Relation (s : Setoid) : (@car s) (@eq s)
reflexivity proved by (refl s)
symmetry proved by (sym s)
transitivity proved by (trans s) as eq_rel.
Coq < Record Morphism (S1 S2:Setoid): Type :=
{ f:car S1 >car S2;
compat: forall (x1 x2: car S1), eq S1 x1 x2 > eq S2 (f x1) (f x2) }.
Coq < Add Parametric Morphism (S1 S2 : Setoid) (M : Morphism S1 S2) :
(@f S1 S2 M) with signature (@eq S1 ==> @eq S2) as apply_mor.
Coq < Proof. apply (compat S1 S2 M). Qed.
Coq < Lemma test: forall (S1 S2:Setoid) (m: Morphism S1 S2)
(x y: car S1), eq S1 x y > eq S2 (f _ _ m x) (f _ _ m y).
Coq < Proof. intros. rewrite H. reflexivity. Qed.
27.2.2 Tactics enabled on user provided relations
The following tactics, all prefixed by setoid_,
deal with arbitrary
registered relations and morphisms. Moreover, all the corresponding unprefixed
tactics (i.e. reflexivity, symmetry, transitivity,
replace, rewrite)
have been extended to fall back to their prefixed counterparts when
the relation involved is not Leibniz equality. Notice, however, that using
the prefixed tactics it is possible to pass additional arguments such as
using relation.
setoid_rewrite [orientation] term [at occs] [in ident]
setoid_replace term with term
[in ident]
[using relation term]
[by tactic]
The using relation arguments cannot be passed to the unprefixed form. The latter argument tells the tactic what parametric relation should be used to replace the first tactic argument with the second one. If omitted, it defaults to the DefaultRelation instance on the type of the objects. By default, it means the most recent Equivalence instance in the environment, but it can be customized by declaring new DefaultRelation instances. As Leibniz equality is a declared equivalence, it will fall back to it if no other relation is declared on a given type.
Every derived tactic that is based on the unprefixed forms of the tactics considered above will also work up to user defined relations. For instance, it is possible to register hints for autorewrite that are not proof of Leibniz equalities. In particular it is possible to exploit autorewrite to simulate normalization in a term rewriting system up to user defined equalities.
27.2.3 Printing relations and morphisms
The Print Instances command can be used to show the list of currently registered Reflexive (using Print Instances Reflexive), Symmetric or Transitive relations, Equivalences, PreOrders, PERs, and Morphisms (implemented as Proper instances). When the rewriting tactics refuse to replace a term in a context because the latter is not a composition of morphisms, the Print Instances commands can be useful to understand what additional morphisms should be registered.
27.2.4 Deprecated syntax and backward incompatibilities
Due to backward compatibility reasons, the following syntax for the declaration of setoids and morphisms is also accepted.
Add Setoid A Aeq ST as ident
where Aeq is a congruence relation without parameters, A is its carrier and ST is an object of type (Setoid_Theory A Aeq) (i.e. a record packing together the reflexivity, symmetry and transitivity lemmas). Notice that the syntax is not completely backward compatible since the identifier was not required.
Add Morphism f:ident.
Proof.
…
Qed.
The latter command also is restricted to the declaration of morphisms without parameters. It is not fully backward compatible since the property the user is asked to prove is slightly different: for nary morphisms the hypotheses of the property are permuted; moreover, when the morphism returns a proposition, the property is now stated using a biimplication in place of a simple implication. In practice, porting an old development to the new semantics is usually quite simple.
Notice that several limitations of the old implementation have been lifted. In particular, it is now possible to declare several relations with the same carrier and several signatures for the same morphism. Moreover, it is now also possible to declare several morphisms having the same signature. Finally, the replace and rewrite tactics can be used to replace terms in contexts that were refused by the old implementation. As discussed in the next section, the semantics of the new setoid_rewrite command differs slightly from the old one and rewrite.
27.3 Extensions
27.3.1 Rewriting under binders
Warning: Due to compatibility issues, this feature is enabled only when calling the setoid_rewrite tactics directly and not rewrite.
To be able to rewrite under binding constructs, one must declare morphisms with respect to pointwise (setoid) equivalence of functions. Example of such morphisms are the standard all and ex combinators for universal and existential quantification respectively. They are declared as morphisms in the Classes.Morphisms_Prop module. For example, to declare that universal quantification is a morphism for logical equivalence:
Proper (pointwise_relation A iff ==> iff) (@all A).
Coq < Proof. simpl_relation.
1 subgoal
A : Type
============================
Proper (pointwise_relation A iff ==> iff) (all (A:=A))
1 subgoal
A : Type
x, y : A > Prop
H : pointwise_relation A iff x y
============================
all x <> all y
One then has to show that if two predicates are equivalent at every point, their universal quantifications are equivalent. Once we have declared such a morphism, it will be used by the setoid rewriting tactic each time we try to rewrite under an all application (products in Prop are implicitly translated to such applications).
Indeed, when rewriting under a lambda, binding variable x, say from P x to Q x using the relation iff, the tactic will generate a proof of pointwise_relation A iff (fun x => P x) (fun x => Q x) from the proof of iff (P x) (Q x) and a constraint of the form Proper (pointwise_relation A iff ==> ?) m will be generated for the surrounding morphism m.
Hence, one can add higherorder combinators as morphisms by providing signatures using pointwise extension for the relations on the functional arguments (or whatever subrelation of the pointwise extension). For example, one could declare the map combinator on lists as a morphism:
Proper ((eqA ==> eqB) ==> list_equiv eqA ==> list_equiv eqB) (@map A B).
where list_equiv implements an equivalence on lists parameterized by an equivalence on the elements.
Note that when one does rewriting with a lemma under a binder using setoid_rewrite, the application of the lemma may capture the bound variable, as the semantics are different from rewrite where the lemma is first matched on the whole term. With the new setoid_rewrite, matching is done on each subterm separately and in its local environment, and all matches are rewritten simultaneously by default. The semantics of the previous setoid_rewrite implementation can almost be recovered using the at 1 modifier.
27.3.2 Subrelations
Subrelations can be used to specify that one relation is included in another, so that morphisms signatures for one can be used for the other. If a signature mentions a relation R on the left of an arrow ==>, then the signature also applies for any relation S that is smaller than R, and the inverse applies on the right of an arrow. One can then declare only a few morphisms instances that generate the complete set of signatures for a particular constant. By default, the only declared subrelation is iff, which is a subrelation of impl and inverse impl (the dual of implication). That’s why we can declare only two morphisms for conjunction: Proper (impl ==> impl ==> impl) and and Proper (iff ==> iff ==> iff) and. This is sufficient to satisfy any rewriting constraints arising from a rewrite using iff, impl or inverse impl through and.
Subrelations are implemented in Classes.Morphisms and are a prime example of a mostly userspace extension of the algorithm.
27.3.3 Constant unfolding
The resolution tactic is based on type classes and hence regards userdefined constants as transparent by default. This may slow down the resolution due to a lot of unifications (all the declared Proper instances are tried at each node of the search tree). To speed it up, declare your constant as rigid for proof search using the command Typeclasses Opaque (see §20.6.7).
27.4 Strategies for rewriting
27.4.1 Definitions
The generalized rewriting tactic is based on a set of strategies that can be combined to obtain custom rewriting procedures. Its set of strategies is based on Elan’s rewriting strategies [102]. Rewriting strategies are applied using the tactic rewrite_strat s where s is a strategy expression. Strategies are defined inductively as described by the following grammar:

Actually a few of these are defined in term of the others using a primitive fixpoint operator:

The basic control strategy semantics are straightforward: strategies are applied to subterms of the term to rewrite, starting from the root of the term. The lemma strategies unify the lefthandside of the lemma with the current subterm and on success rewrite it to the righthandside. Composition can be used to continue rewriting on the current subterm. The fail strategy always fails while the identity strategy succeeds without making progress. The reflexivity strategy succeeds, making progress using a reflexivity proof of rewriting. Progress tests progress of the argument strategy and fails if no progress was made, while try always succeeds, catching failures. Choice is leftbiased: it will launch the first strategy and fall back on the second one in case of failure. One can iterate a strategy at least 1 time using repeat and at least 0 times using any.
The subterm and subterms strategies apply their argument strategy s to respectively one or all subterms of the current term under consideration, lefttoright. subterm stops at the first subterm for which s made progress. The composite strategies innermost and outermost perform a single innermost our outermost rewrite using their argument strategy. Their counterparts bottomup and topdown perform as many rewritings as possible, starting from the bottom or the top of the term.
Hint databases created for autorewrite can also be used by rewrite_strat using the hints strategy that applies any of the lemmas at the current subterm. The terms strategy takes the lemma names directly as arguments. The eval strategy expects a reduction expression (see §8.7) and succeeds if it reduces the subterm under consideration. The fold strategy takes a term c and tries to unify it to the current subterm, converting it to c on success, it is stronger than the tactic fold.
27.4.2 Usage
rewrite_strat s [in ident]:
Rewrite using the strategy s in hypothesis ident or the conclusion.
Error messages:
 Nothing to rewrite. If the strategy failed.
 No progress made. If the strategy succeeded but made no progress.
 Unable to satisfy the rewriting constraints. If the strategy succeeded and made progress but the corresponding rewriting constraints are not satisfied.
The setoid_rewrite c tactic is basically equivalent to rewrite_strat (outermost c).
 1
 Nicolas Tabareau helped with the gluing.