Apr 9, 2014 - Coq Basics

GADTs vs Inductive Datatypes

Consider the following definition of Nat-indexed Vector GADT in OCaml and Haskell: OCaml:

    (*  vec : * -> * -> * 
     *)
    type (_,_) vec =
    | Nil : ('a,zero) vec
    | Cons : 'a * ('a,'b) vec -> ('a,'b succ) vec

Haskell:

    data Vec :: * -> Nat -> * where
      VNil :: Vec a 'Zero
      VCons :: a -> Vec a n -> Vec a ('Succ n)

Observe that quantified type variables in the type of constructors are generalized by the type-constructor signature. Infact, this is not always true. For example, consider the GADT for higher-order abstract syntax:

  data Exp :: * -> * where
   -- other constructors elided
    App :: Exp (a -> b) -> Exp a -> Exp b

The type scheme of App contains a type variable (a) that is not generalized by the type-constructor signature.

  App :: a. b. Exp (a -> b) -> Exp a -> Exp b

This type is treated as existential type when an expression of type Exp b is destructed and matched agains App.

Therefore, the correct observation is this: All type variables generalized by the signature of type constructor are also generalized in type schemes of its value constructors.

Here is the nat-indexed vector inductive datatype in Coq:

Inductive vec : Type -> nat -> Type :=
  | vnil : forall A:Type, vec A O
  | vcons : forall A:Type, forall n:nat, A -> vec A n -> vec A (S n).

The above definition should be read as following: vec is a Type indexed by a Type and a nat. Here are the only ways to construct values of Vec X N, where X could be any Type and N could be any nat:

  • For any Type A, vnil A is a value of Vec A O
  • For any Type A and for any nat n, vcons A n accepts a value of type A, a value of type vec A n, and generates a value of type Vec A (S n)

Notice that it looks quite similar to Haskell definition, with type variables explicated. An empty list in explicitly typed haskell is VNil Int, whereas in Coq, it is vnil int. However, unlike Haskell, where there is term-type-kind hierarcy, Coq has full dependent types with infinite hierarchy of universes. One consequence of this is that we will frequently see arrows mapping members of one universe to members of a different universe. For example, if we Print vec., this is what we see:

    Inductive vec : Type (* Top.59 *) -> nat -> Type (* max(Set, (Top.61)+1, (Top.62)+1) *) :=
      | vnil : forall A : Type (* Top.61 *), vec A 0
      | vcons : forall (A : Type (* Top.62 *)) (n : nat),
                   A -> vec A n -> vec A (S n)

Evidently, vec maps a member of Type(59) universe and member of Set universe to a type universe at higher level. Compare this to the type of Vec type-constructor in haskell * -> * -> *, which simply says that when a is a type, and n is a type (i.e., a and n have kind *), then Vec a n is a type (i.e., it has kind *).

In the inductive definition of vec, observe that both constructors construct values of type vec A _. Therefore, we can as well generalize A at the type of type-constructor (here, this is a mere syntactic convenience):

    Inductive vec2 (A : Type) : nat -> Type :=
      | vnil2 : vec2 A O
      | vcons2 : forall n:nat, A -> vec2 A n -> vec2 A (S n). 

The above definition should be read as following: vec is a Type indexed by a Type A and a nat. Here are the only ways to construct values of Vec A N, where N could be any nat

Another consequence of having full dependent types in Coq is that the type of inductive type-constructors can be dependent as well. For eg, here is the inductive definition of sig datatype:

    Inductive sig (T : Type) (P : T -> Prop) : Type :=
        exist : forall x : T, P x -> sig T 

sig T P is the type of values of type T that satisfy proposition P. sig is a function on type, and propositions on that type. Observe that the type of second argument to type-constructor sig depends on the value of its first argument.

How did we arrive at that type definition? Let us say we want to define a refinement type for values of nat satisfying a predicate P. Effectively, we would like to index nat with a predicate P over natural numbers. Since P is a predicate, P : nat -> Prop. We define a new Type natST (read “nat, such that”) that is indexed by P : nat -> Prop:

    Inductive natST (P : nat -> Prop) : Type := ??

What should be its constructors? There has to be only one constructor that, given a nat (n), and a proof that it satisfies the predicate (i.e., proof of P n), returns evidence for natST P. In other words, we have an evidence for natST P if and only if we have a nat, and the evidence that it satisfies the predicate. Destructing a natST P value would produce a nat and a proof that it satisfies P.

Type natST helps us easily write type for division operation:

    Definition natPos : Type := natST (fun n:nat => n>0).
    Definition div : nat -> natPos -> nat * nat.
    Admitted. Defined.
    (* equivalent type of div: *)
    Definition div2 : forall n:nat, n>0 -> nat -> (nat*nat),
    Admitted. Defined.

Implicit Arguments

Following is to set the implicit arguments option in Coq.

    Set Implicit Arguments.

When the mode for automatic declaration of implicit arguments is on, the default is to automatically set implicit only the strict implicit arguments (eg: tyarg of Cons is strict implicit, whereas that of Nil is contextual implicit).

By default, Coq does not automatically set implicit the contextual implicit arguments. To tell Coq to infer also contextual implicit argument, use command

    Set Contextual Implicit.

Setting contextual implcit too may not infer ty annot for Nil. Why? How does coq know whether we want to partially applied Nil or fully apply it? Setting Maximal Implicit works, but that would prevent us from writing certain partial function applications. Therefore, we explciity set implicit args for Nil:

    Implicit Arguments Nil [X].

Notation, Associativity, and Precedence

The level is for precedence. Assume for instance that we want conjunction to bind more than disjunction. This is expressed by assigning a precedence level to each notation, knowing that a lower level binds more than a higher level. Hence the level for disjunction must be higher than the level for conjunction. 100 is the highest level. There is a special 200 level higher than 100.

Since connectives are the less tight articulation points of a text, it is reasonable to choose levels not so far from the higher level which is 100, for example 85 for disjunction and 80 for conjunction

The Omega Tactic

Omega is a decision procedure for linear arithmetic (also known as Presburger arithmetic). It can be used to discharge any silly goals on nats and integers like:

Omega can deal with nats. It converts nats to integers. Following operations on nat are recognized by omega:

  • Predicates, such as =, le, lt, gt, ge
  • Functions, such as +, minus, mult, pred, S, O

Omega needs to be imported:

    Require Import Omega.

apply and eapply tactics

apply H tries to unify the current goal with conclusion of H. If it succeeds, it generates a sub-goal for each assumption (argument) of H. For instance, in the following context:

     A : Type
    s1 : Ensemble A
    s2 : Ensemble A
    s3 : Ensemble A
    x : A
    H2 : In A s1 x
    ============================
     In A (Union A s1 (Union A s2 s3)) 

using apply Union_introl, where

    Union_introl
     : forall (U : Type) (B C : Ensemble U) (x : U),
       In U B x -> In U (Union U B C) 

unifies U with A, B with s1, C with (Union A s2 s3), and x with x to generate following sub-goal:

       ....
     ============================
        In A s1 

which is already an assumption.

The tactic eapply does the same, except that instead of failing due to inability to instantiate variables, it replaces them with existentials. When you have existential in you context, destruct it to skolemize. The ?6201 style variables in goal when you use eapply are actually existential variables.

The Auto Tactic

The auto tactic solves goals that are solvable by any combination of

  • intros, and
  • apply.

By default, auto only does apply on local hypotheses. To add to the global hint database of auto:

    Hint Resolve <Lemma-name>.
    Hint Resolve <Constructors of inductive prop>
    Hint Constructors A. <Hint Resolve for all constructors of A>
    Hint Unfold <A Definition>.

One can optionally declare a hint database using the command Create HintDb.

To see hints applicable in the current context, and list of all hints, respectively:

    Print Hint.
    Print Hint *.

Using auto tactic with database of hints named sets:

    auto with sets.

You can extened hint DB for one application of auto with using keyword. To extend with Union_introl:

    auto using Union_introl.

To extend with all constructors of Union:

    auto using Union.

More avatars of auto:

  • trivial tactic - non recursive auto. depth =0.
  • autounfold and autorewrite to use Hint Unfold and Hint Rewrite hints in the hint database.

Auto Examples

Completely manual proof for theorem union_is_associative:

    Theorem union_is_associative : forall A : Type, forall s1 s2 s3 : Ensemble A,
                         Union _ (Union _ s1 s2) s3 = Union _ s1 (Union _ s2 s3).
    Proof.
      intros; eapply Extensionality_Ensembles. unfold Same_set.
      unfold Included. split.
        intros; inversion H.
            inversion H0. Check Union_introl. apply Union_introl. assumption.
            apply Union_intror. apply Union_introl. assumption.
            apply Union_intror. apply Union_intror. assumption.
        intros. inversion H.
            apply Union_introl. apply Union_introl. assumption.
            inversion H0. 
              apply Union_introl. apply Union_intror; assumption. 
              apply Union_intror. assumption.
    Qed.

Proof with auto:

    Proof.
      intros. apply Extensionality_Ensembles. unfold Same_set.
      unfold Included. split; intros ? H; inversion H;
        [ inversion H0; auto with sets | auto with sets |
         auto with sets | inversion H0; auto with sets].
    Qed.

Even simpler auto proof:

    Proof.
      intros. apply Extensionality_Ensembles. unfold Same_set.
      unfold Included. split; intros ? H; inversion H; subst;
        try (inversion H0); auto with sets.
    Qed.

SF automation chapter (sfauto)says this on auto:

Note that proof search tactics never perform any rewriting step (tactics rewrite, subst), nor any case analysis on an arbitrary data structure or predicate (tactics destruct and inversion), nor any proof by induction (tactic induction). So, proof search is really intended to automate the final steps from the various branches of a proof. It is not able to discover the overall structure of a proof.

The Solve Tactical

Solve tactical takes a list of tactics, and tries them left to right, stopping when a tactic is successful. If none of the tactics are successful, the solve fails.

    solve [inversion H | inversion H1].

The difference between try solve [e1 | e2]' and try e1; try e2` is the short-cut behaviour of solve.

The Intuition Tactical

Intuition t applies propositional decision procedure to simplify the goal, and then applies tactic t on the resultant.

  • intuition fail succeeds if the goal is a propositional tautology. Fails otherwise. This is equivalent to tauto tactic.
  • Simply using intuition is equivalent to using intuition auto with \*.

Notes:

  • To instantiate a universally quantified hypothesis H on variable n, just do (H n). Eg: destruct (H n).
  • When there is ambiguity about which term rewrite rewrites when there are multiple rewritable terms in the goal, assert the required hypothesis as a separate assertion, and prove it (using apply).
  • It is futile to define strong lemmas with conjunction in consequent. Not only that applying them is difficult, but also that auto only tries such lemmas when current goal is a conjunction at top level. Instead, define two separate lemmas.

References:

Continue reading