I'm having some troubles in Coq when trying to perform case analysis on the result of a function (which returns an inductive type). When using the usual tactics, like elim, induction, destroy, etc, the information gets lost.
I'll put an example:
We first have a function like so:
Definition f(n:nat): bool := (* definition *)
Now, imagine we are at this step in the proof of a specific theorem:
n: nat
H: f n = other_stuff
------
P (f n )
When I apply a tactic, like let's say, induction (f n), this happens:
Subgoal 1
n:nat
H: true = other_stuff
------
P true
Subgoal 2
n:nat
H: false = other_stuff
------
P false
However, what I want is something like this instead:
Subgoal 1
n:nat
H: true = other_stuff
H1: f n = true
------
P true
Subgoal 2
n:nat
H: false = other_stuff
H1: f n = false
------
P false
In the way it actually works, I lose information, specially I lose any information about f n. In the problems I work with, I need to use the information that f n = true or f n = false, to use with other hypothesis, etc.
Is there a way to do the 2nd option?
I tried using stuff like cut(f n = false \/ f n = true) but it becomes very tiresome, specially when I have several of these "special" inductions in a row. I want to know if there is something that basically works exactly like the cut above, but with fewer tactics/proofs
The issue is that you perform induction on a constructed term, not a single variable. Keeping the information in your case has been proved to be a very difficult problem.
The usual work-around is to abstract your constructed term using the remember tactic. I don't have the exact syntax in mind right now but you should try something like
remember (f n) as Fn. (* this introduces an equality HeqFn : Fn = f n *)
revert f n HeqFn. (* this is useful in many cases, but not mandatory *)
induction Fn; intros; subst in *.
Hope it helps,
V.
Related
I understood destruct as it breaks an inductive definition into its constructors. I recently saw case_eq and I couldn't understand what it does differently?
1 subgoals
n : nat
k : nat
m : M.t nat
H : match M.find (elt:=nat) n m with
| Some _ => true
| None => false
end = true
______________________________________(1/1)
cc n (M.add k k m) = true
In the above context, if I do destruct M.find n m it breaks H into true and false whereas case_eq (M.find n m) leaves H intact and adds separate proposition M.find (elt:=nat) n m = Some v, which I can rewrite to get same effect as destruct.
Can someone please explain me the difference between the two tactics and when which one should be used?
The first basic tactic in the family of destruct and case_eq is called case. This tactic modifies only the conclusion. When you type case A and A has a type T which is inductive, the system replaces A in the goal's conclusion by instances of all the constructors of type T, adding universal quantifications for the arguments of these constructors, if needed. This creates as many goals as there are constructors in type T. The formula A disappears from the goal and if there is any information about A in an hypothesis, the link between this information and all the new constructors that replace it in the conclusion gets lost. In spite of this, case is an important primitive tactic.
Loosing the link between information in the hypotheses and instances of A in the conclusion is a big problem in practice, so developers came up with two solutions: case_eq and destruct.
Personnally, when writing the Coq'Art book, I proposed that we write a simple tactic on top of case that keeps a link between A and the various constructor instances in the form of an equality. This is the tactic now called case_eq. It does the same thing as case but adds an extra implication in the goal, where the premise of the implication is an equality of the form A = ... and where ... is an instance of each constructor.
At about the same time, the tactic destruct was proposed. Instead of limiting the effect of replacement in the goal's conclusion, destruct replaces all instances of A appearing in the hypotheses with instances of constructors of type T. In a sense, this is cleaner because it avoids relying on the extra concept of equality, but it is still incomplete because the expression A may be a compound expression f B, and if B appears in the hypothesis but not f B the link between A and B will still be lost.
Illustration
Definition my_pred (n : nat) := match n with 0 => 0 | S p => p end.
Lemma example n : n <= 1 -> my_pred n <= 0.
Proof.
case_eq (my_pred n).
Gives the two goals
------------------
n <= 1 -> my_pred n = 0 -> 0 <= 0
and
------------------
forall p, my_pred n = S p -> n <= 1 -> S p <= 0
the extra equality is very useful here.
In this question I suggested that the developer use case_eq (a == b) when (a == b) has type bool because this type is inductive and not very informative (constructors have no argument). But when (a == b) has type {a = b}+{a <> b} (which is the case for the string_dec function) the constructors have arguments that are proofs of interesting properties and the extra universal quantification for the arguments of the constructors are enough to give the relevant information, in this case a = b in a first goal and a <> b in a second goal.
I was curious about how the discriminate tactic works behind the curtain. Therefore I did some experiments.
First a simple Inductive definition:
Inductive AB:=A|B.
Then a simple lemma which can be proved by the discriminate tactic:
Lemma l1: A=B -> False.
intro.
discriminate.
Defined.
Let's see what the proof looks like:
Print l1.
l1 =
fun H : A = B =>
(fun H0 : False => False_ind False H0)
(eq_ind A
(fun e : AB => match e with
| A => True
| B => False
end) I B H)
: A = B -> False
This looks rather complicated and I do not understand what is happening here. Therefore I tried to prove the same lemma more explicitly:
Lemma l2: A=B -> False.
apply (fun e:(A=B) => match e with end).
Defined.
Let's again see what Coq has made with this:
Print l2.
l2 =
fun e : A = B =>
match
e as e0 in (_ = a)
return
(match a as x return (A = x -> Type) with
| A => fun _ : A = A => IDProp
| B => fun _ : A = B => False
end e0)
with
| eq_refl => idProp
end
: A = B -> False
Now I am totally confused. This is still more complicated.
Can anyone explain what is going on here?
Let's go over this l1 term and describe every part of it.
l1 : A = B -> False
l1 is an implication, hence by Curry-Howard correspondence it's an abstraction (function):
fun H : A = B =>
Now we need to construct the body of our abstraction, which must have type False. The discriminate tactic chooses to implement the body as an application f x, where f = fun H0 : False => False_ind False H0 and it's just a wrapper around the induction principle for False, which says that if you have a proof of False, you can get a proof of any proposition you want (False_ind : forall P : Prop, False -> P):
(fun H0 : False => False_ind False H0)
(eq_ind A
(fun e : AB => match e with
| A => True
| B => False
end) I B H)
If we perform one step of beta-reduction, we'll simplify the above into
False_ind False
(eq_ind A
(fun e : AB => match e with
| A => True
| B => False
end) I B H)
The first argument to False_ind is the type of the term we are building. If you were to prove A = B -> True, it would have been False_ind True (eq_ind A ...).
By the way, it's easy to see that we can simplify our body further - for False_ind to work it needs to be provided with a proof of False, but that's exactly what we are trying to construct here! Thus, we can get rid of False_ind completely, getting the following:
eq_ind A
(fun e : AB => match e with
| A => True
| B => False
end) I B H
eq_ind is the induction principle for equality, saying that equals can be substituted for equals:
eq_ind : forall (A : Type) (x : A) (P : A -> Prop),
P x -> forall y : A, x = y -> P y
In other words, if one has a proof of P x, then for all y equal to x, P y holds.
Now, let's create step-by-step a proof of False using eq_ind (in the end we should obtain the eq_ind A (fun e : AB ...) term).
We start, of course, with eq_ind, then we apply it to some x - let's use A for that purpose. Next, we need the predicate P. One important thing to keep in mind while writing P down is that we must be able to prove P x. This goal is easy to achieve - we are going to use the True proposition, which has a trivial proof. Another thing to remember is the proposition we are trying to prove (False) - we should be returning it if the input parameter is not A.
With all the above the predicate almost writes itself:
fun x : AB => match x with
| A => True
| B => False
end
We have the first two arguments for eq_ind and we need three more: the proof for the branch where x is A, which is the proof of True, i.e. I. Some y, which will lead us to the proposition we want to get proof of, i.e. B, and a proof that A = B, which is called H at the very beginning of this answer. Stacking these upon each other we get
eq_ind A
(fun x : AB => match x with
| A => True
| B => False
end)
I
B
H
And this is exactly what discriminate gave us (modulo some wrapping).
Another answer focuses on the discriminate part, I will focus on the manual proof. You tried:
Lemma l2: A=B -> False.
apply (fun e:(A=B) => match e with end).
Defined.
What should be noted and makes me often uncomfortable using Coq is that Coq accepts ill-defined definitions that it internally rewrites into well-typed terms. This allows to be less verbose, since Coq adds itself some parts. But on the other hand, Coq manipulates a different term than the one we entered.
This is the case for your proof. Naturally, the pattern-matching on e should involve the constructor eq_refl which is the single constructor of the eq type. Here, Coq detects that the equality is not inhabited and thus understands how to modify your code, but what you entered is not a proper pattern-matching.
Two ingredients can help understand what is going on here:
the definition of eq
the full pattern-matching syntax, with as, in and return terms
First, we can look at the definition of eq.
Inductive eq {A : Type} (x : A) : A -> Prop := eq_refl : x = x.
Note that this definition is different from the one that seems more natural (in any case, more symmetric).
Inductive eq {A : Type} : A -> A -> Prop := eq_refl : forall (x:A), x = x.
This is really important that eq is defined with the first definition and not the second. In particular, for our problem, what is important is that, in x = y, x is a parameter while y is an index. That is to say, x is constant across all the constructors while y can be different in each constructor. You have the same difference with the type Vector.t. The type of the elements of a vector will not change if you add an element, that's why it is implemented as a parameter. Its size, however, can change, that's why it is implemented as an index.
Now, let us look at the extended pattern-matching syntax. I give here a very brief explanation of what I have understood. Do not hesitate to look at the reference manual for safer information. The return clause can help specify a return type that will be different for each branch. That clause can use the variables defined in the as and in clauses of the pattern-matching, which binds respectively the matched term and the type indices. The return clause will both be interpreted in the context of each branch, substituting the variables of as and in using this context, to type-check the branches one by one, and be used to type the match from an external point of view.
Here is a contrived example with an as clause:
Definition test n :=
match n as n0 return (match n0 with | 0 => nat | S _ => bool end) with
| 0 => 17
| _ => true
end.
Depending on the value of n, we are not returning the same type. The type of test is forall n : nat, match n with | 0 => nat | S _ => bool end. But when Coq can decide in which case of the match we are, it can simplify the type. For example:
Definition test2 n : bool := test (S n).
Here, Coq knows that, whatever is n, S n given to test will result as something of type bool.
For equality, we can do something similar, this time using the in clause.
Definition test3 (e:A=B) : False :=
match e in (_ = c) return (match c with | B => False | _ => True end) with
| eq_refl => I
end.
What's going on here ? Essentially, Coq type-checks separately the branches of the match and the match itself. In the only branch eq_refl, c is equal to A (because of the definition of eq_refl which instantiates the index with the same value as the parameter), therefore we claimed we returned some value of type True, here I. But when seen from an external point of view, c is equal to B (because e is of type A=B), and this time the return clause claims that the match returns some value of type False. We use here the capability of Coq to simplify pattern-matching in types that we have just seen with test2. Note that we used True in the other cases than B, but we don't need True in particular. We only need some inhabited type, such that we can return something in the eq_refl branch.
Going back to the strange term produced by Coq, the method used by Coq does something similar, but on this example, certainly more complicated. In particular, Coq often uses types IDProp inhabited by idProp when it needs useless types and terms. They correspond to True and I used just above.
Finally, I give the link of a discussion on coq-club that really helped me understand how extended pattern-matching is typed in Coq.
In mathematics, we often proceed as follows: "Now let us consider two cases, the number k can be even or odd. For the even case, we can say exists k', 2k' = k..."
Which expands to the general idea of reasoning about an entire set of objects by disassembling it into several disjunct subsets that can be used to reconstruct the original set.
How is this reasoning principle captured in coq considering we do not always have an assumption that is one of the subsets we want to deconstruct into?
Consider the follow example for demonstration:
forall n, Nat.Even n => P n.
Here we can naturally do inversion on Nat.Even n to get n = 2*x (and an automatically-false eliminated assumption that n = 2*x + 1). However, suppose we have the following:
forall n, P n
How can I state: "let us consider even ns and odd ns". Do I need to first show that we have decidable forall n : nat, even n \/ odd n? That is, introduce a new (local or global) lemma listing all the required subsets? What are the best practices?
Indeed, to reason about a splitting of a class of objects in Coq you need to show an algorithm splitting them, unless you want to reason classically (there is nothing wrong with that).
IMO, a key point is getting such decidability hypotheses "for free". For instance, you could implement odd : nat -> bool as a boolean function, as it is done in some libraries, then you get the splitting for free.
[edit]
You can use some slightly more convenient techniques for pattern matching, by enconding the pertinent cases as inductives:
Require Import PeanoNat Nat Bool.
CoInductive parity_spec (n : nat) : Type :=
| parity_spec_odd : odd n = true -> parity_spec n
| parity_spec_even: even n = true -> parity_spec n
.
Lemma parityP n : parity_spec n.
Proof.
case (even n) eqn:H; [now right|left].
now rewrite <- Nat.negb_even, H.
Qed.
Lemma test n : even n = true \/ odd n = true.
Proof. now case (parityP n); auto. Qed.
In using the Coq apply ... with tactic, the examples I have seen all involve explicitly giving the names of variables to instantiate. For example, given a theorem about the transitivity of equality.
Theorem trans_eq : forall (X:Type) (n m o : X),
n = m -> m = o -> n = o.
To apply it:
Example test: forall n m: nat,
n = 1 -> 1 = m -> n = m.
Proof.
intros n m.
apply trans_eq with (m := 1). Qed.
Note that in the last line apply trans_eq with (m := 1)., I have to remember that the name of the variable to instantiate is m, rather than o or n or some other names y.
To me, whether m n o or x y z are used in the original statement of the theorem shouldn't matter, because they are like dummy variables or formal parameters of a function. And sometimes I can't remember the specific names I used or somebody else put down in a different file when defining the theorem.
Is there a way by which I can refer to the variables e.g. by their position and use something like:
apply trans_eq with (#1 := 1)
in the above example?
By the way, I tried: apply trans_eq with (1 := 1). and got Error: No such binder.
Thanks.
You can specialize the lemma with the right arguments. The _ is used for all arguments that we don't want to specialize (because they can be inferred). The # is required to specialize implicit arguments.
Example test: forall n m: nat,
n = 1 -> 1 = m -> n = m.
Proof.
intros n m.
apply (#trans_eq _ _ 1).
Qed.
You can omit the binder names after with, so in your case do apply trans_eq with 1.
Example test: forall n m: nat,
n = 1 -> 1 = m -> n = m.
Proof.
intros n m.
apply trans_eq with 1; auto.
Qed.
I've changed your original example a little to conclude the proof.
Why this works
To understand why this works, check the manual under Bindings:
Tactics that take a term as an argument may also accept bindings to
instantiate some parameters of the term by name or position. The
general form of a term with bindings is termtac with bindings where
bindings can take two different forms:
bindings::= (ident | natural := term)+
| one_term+
What is shown in this example is the form one_term, which is described as follows:
in the case of apply, or of constructor and its variants, only instances for the dependent products that are not bound in the conclusion of termtac are required.
Which is why only one term needs to be supplied.
I am learning Coq. I am stuck on a quite silly problem (which has no motivation, it is really silly). I want to build a function from ]2,+oo] to the set of integers mapping x to x-3. That should be simple... In any language I know, it is simple. But not in Coq. First, I write (I explain with a lot of details so that someone can explain what I don't understand in the behaviour of Coq)
Definition f : forall n : nat, n > 2 -> nat.
I get a subgoal
============================
forall n : nat, n > 2 -> nat
which means that Coq wants a map from a proof of n>2 to the set of integers. Fine. So I want to tell it that n = 3 + p for some integer p, and then return the integer p. I write :
intros n H.
And I get the context/subgoal
n : nat
H : n > 2
============================
nat
Then i suppose that I have proved n = 3 + p for some integer p by
cut(exists p, 3 + p = n).
I get the context/subgoal
n : nat
H : n > 2
============================
(exists p : nat, 3 + p = n) -> nat
subgoal 2 (ID 6) is:
exists p : nat, 3 + p = n
I move the hypothesis in the context by
intro K.
I obtain:
n : nat
H : n > 2
K : exists p : nat, 3 + p = n
============================
nat
subgoal 2 (ID 6) is:
exists p : nat, 3 + p = n
I will prove the existence of p later. Now I want to finish the proof by exact p. So i need first to do a
destruct K as (p,K).
and I obtain the error message
Error: Case analysis on sort Set is not allowed for inductive
definition ex.
And I am stuck.
You are absolutely right! Writing this function should be easy in any reasonable programming language, and, fortunately, Coq is no exception.
In your case, it is much easier to define your function by simply ignoring the proof argument you are supplying:
Definition f (n : nat) : nat := n - 3.
You may then wonder "but wait a second, the natural numbers aren't closed under subtraction, so how can this make sense?". Well, in Coq, subtraction on the natural numbers isn't really subtraction: it is actually truncated. If you try to subtract, say, 3 from 2, you get 0 as an answer:
Goal 2 - 3 = 0. reflexivity. Qed.
What this means in practice is that you are always allowed to "subtract" two natural numbers and get a natural number back, but in order for this subtraction make sense, the first argument needs to be greater than the second. We then get lemmas such as the following (available in the standard library):
le_plus_minus_r : forall n m, n <= m -> n + (m - n) = m
In most cases, working with a function that is partially correct, such as this definition of subtraction, is good enough. If you want, however, you can restrict the domain of f to make its properties more pleasant. I've taken the liberty of doing the following script with the ssreflect library, which makes writing this kind of function easier:
Require Import Ssreflect.ssreflect Ssreflect.ssrfun Ssreflect.ssrbool.
Require Import Ssreflect.ssrnat Ssreflect.eqtype.
Definition f (n : {n | 2 < n}) : nat :=
val n - 3.
Definition finv (m : nat) : {n | 2 < n} :=
Sub (3 + m) erefl.
Lemma fK : cancel f finv.
Proof.
move=> [n Pn] /=; apply/val_inj=> /=.
by rewrite /f /= addnC subnK.
Qed.
Lemma finvK : cancel finv f.
Proof.
by move=> n; rewrite /finv /f /= addnC addnK.
Qed.
Now, f takes as an argument a natural number n that is greater than 2 (the {x : T | P x} form is syntax sugar for the sig type from the standard library, which is used for forming types that work like subsets). By restricting the argument type, we can write an inverse function finv that takes an arbitrary nat and returns another number that is greater than 2. Then, we can prove lemmas fK and finvK, which assert that fK and finvK are inverses of each other.
On the definition of f, we use val, which is ssreflect's idiom for extracting the element out of a member of a type such as {n | 2 < n}. The Sub function on finv does the opposite, packaging a natural number n with a proof that 2 < n and returning an element of {n | 2 < n}. Here, we rely crucially on the fact that the < is expressed in ssreflect as a boolean computation, so that Coq can use its computation rules to check that erefl, a proof of true = true, is also a valid proof of 2 < 3 + m.
To conclude, the mysterious error message you got in the end has to do with Coq's rules governing computational types, with live in Type, and propositional types, which live in Prop. Coq's rules forbid you from using proofs of propositions to build elements that have computational content (such as natural numbers), except in very particular cases. If you wanted, you could still finish your definition by using {p | 3 + p = n} instead of exists p, 3 + p = n -- both mean the same thing, except the former lives in Type while the latter lives in Prop.