Theorem Proving

Equality

Idris allows propositional equalities to be declared, allowing theorems about programs to be stated and proved. Equality is built in, but conceptually has the following definition:

data (=) : a -> b -> Type where
   Refl : x = x

Equalities can be proposed between any values of any types, but the only way to construct a proof of equality is if values actually are equal. For example:

fiveIsFive : 5 = 5
fiveIsFive = Refl

twoPlusTwo : 2 + 2 = 4
twoPlusTwo = Refl

The Empty Type

There is an empty type, \bot, which has no constructors. It is therefore impossible to construct an element of the empty type, at least without using a partially defined or general recursive function (see Section Totality Checking for more details). We can therefore use the empty type to prove that something is impossible, for example zero is never equal to a successor:

disjoint : (n : Nat) -> Z = S n -> Void
disjoint n p = replace {P = disjointTy} p ()
  where
    disjointTy : Nat -> Type
    disjointTy Z = ()
    disjointTy (S k) = Void

There is no need to worry too much about how this function works — essentially, it applies the library function replace, which uses an equality proof to transform a predicate. Here we use it to transform a value of a type which can exist, the empty tuple, to a value of a type which can’t, by using a proof of something which can’t exist.

Once we have an element of the empty type, we can prove anything. void is defined in the library, to assist with proofs by contradiction.

void : Void -> a

Simple Theorems

When type checking dependent types, the type itself gets normalised. So imagine we want to prove the following theorem about the reduction behaviour of plus:

plusReduces : (n:Nat) -> plus Z n = n

We’ve written down the statement of the theorem as a type, in just the same way as we would write the type of a program. In fact there is no real distinction between proofs and programs. A proof, as far as we are concerned here, is merely a program with a precise enough type to guarantee a particular property of interest.

We won’t go into details here, but the Curry-Howard correspondence [1] explains this relationship. The proof itself is trivial, because plus Z n normalises to n by the definition of plus:

plusReduces n = Refl

It is slightly harder if we try the arguments the other way, because plus is defined by recursion on its first argument. The proof also works by recursion on the first argument to plus, namely n.

plusReducesZ : (n:Nat) -> n = plus n Z
plusReducesZ Z = Refl
plusReducesZ (S k) = cong (plusReducesZ k)

cong is a function defined in the library which states that equality respects function application:

cong : {f : t -> u} -> a = b -> f a = f b

We can do the same for the reduction behaviour of plus on successors:

plusReducesS : (n:Nat) -> (m:Nat) -> S (plus n m) = plus n (S m)
plusReducesS Z m = Refl
plusReducesS (S k) m = cong (plusReducesS k m)

Even for trivial theorems like these, the proofs are a little tricky to construct in one go. When things get even slightly more complicated, it becomes too much to think about to construct proofs in this ‘batch mode’.

Idris provides interactive editing capabilities, which can help with building proofs. For more details on building proofs interactively in an editor, see Theorem Proving.

Totality Checking

If we really want to trust our proofs, it is important that they are defined by total functions — that is, a function which is defined for all possible inputs and is guaranteed to terminate. Otherwise we could construct an element of the empty type, from which we could prove anything:

-- making use of 'hd' being partially defined
empty1 : Void
empty1 = hd [] where
    hd : List a -> a
    hd (x :: xs) = x

-- not terminating
empty2 : Void
empty2 = empty2

Internally, Idris checks every definition for totality, and we can check at the prompt with the :total command. We see that neither of the above definitions is total:

*theorems> :total empty1
possibly not total due to: empty1#hd
    not total as there are missing cases
*theorems> :total empty2
possibly not total due to recursive path empty2

Note the use of the word “possibly” — a totality check can, of course, never be certain due to the undecidability of the halting problem. The check is, therefore, conservative. It is also possible (and indeed advisable, in the case of proofs) to mark functions as total so that it will be a compile time error for the totality check to fail:

total empty2 : Void
empty2 = empty2
Type checking ./theorems.idr
theorems.idr:25:empty2 is possibly not total due to recursive path empty2

Reassuringly, our proof in Section The Empty Type that the zero and successor constructors are disjoint is total:

*theorems> :total disjoint
Total

The totality check is, necessarily, conservative. To be recorded as total, a function f must:

  • Cover all possible inputs
  • Be well-founded — i.e. by the time a sequence of (possibly mutually) recursive calls reaches f again, it must be possible to show that one of its arguments has decreased.
  • Not use any data types which are not strictly positive
  • Not call any non-total functions

Directives and Compiler Flags for Totality

By default, Idris allows all well-typed definitions, whether total or not. However, it is desirable for functions to be total as far as possible, as this provides a guarantee that they provide a result for all possible inputs, in finite time. It is possible to make total functions a requirement, either:

  • By using the --total compiler flag.
  • By adding a %default total directive to a source file. All definitions after this will be required to be total, unless explicitly flagged as partial.

All functions after a %default total declaration are required to be total. Correspondingly, after a %default partial declaration, the requirement is relaxed.

Finally, the compiler flag --warnpartial causes to print a warning for any undeclared partial function.

Totality checking issues

Please note that the totality checker is not perfect! Firstly, it is necessarily conservative due to the undecidability of the halting problem, so many programs which are total will not be detected as such. Secondly, the current implementation has had limited effort put into it so far, so there may still be cases where it believes a function is total which is not. Do not rely on it for your proofs yet!

Hints for totality

In cases where you believe a program is total, but Idris does not agree, it is possible to give hints to the checker to give more detail for a termination argument. The checker works by ensuring that all chains of recursive calls eventually lead to one of the arguments decreasing towards a base case, but sometimes this is hard to spot. For example, the following definition cannot be checked as total because the checker cannot decide that filter (< x) xs will always be smaller than (x :: xs):

qsort : Ord a => List a -> List a
qsort [] = []
qsort (x :: xs)
   = qsort (filter (< x) xs) ++
      (x :: qsort (filter (>= x) xs))

The function assert_smaller, defined in the Prelude, is intended to address this problem:

assert_smaller : a -> a -> a
assert_smaller x y = y

It simply evaluates to its second argument, but also asserts to the totality checker that y is structurally smaller than x. This can be used to explain the reasoning for totality if the checker cannot work it out itself. The above example can now be written as:

total
qsort : Ord a => List a -> List a
qsort [] = []
qsort (x :: xs)
   = qsort (assert_smaller (x :: xs) (filter (< x) xs)) ++
      (x :: qsort (assert_smaller (x :: xs) (filter (>= x) xs)))

The expression assert_smaller (x :: xs) (filter (<= x) xs) asserts that the result of the filter will always be smaller than the pattern (x :: xs).

In more extreme cases, the function assert_total marks a subexpression as always being total:

assert_total : a -> a
assert_total x = x

In general, this function should be avoided, but it can be very useful when reasoning about primitives or externally defined functions (for example from a C library) where totality can be shown by an external argument.

[1]Timothy G. Griffin. 1989. A formulae-as-type notion of control. In Proceedings of the 17th ACM SIGPLAN-SIGACT symposium on Principles of programming languages (POPL ‘90). ACM, New York, NY, USA, 47-58. DOI=10.1145/96709.96714 http://doi.acm.org/10.1145/96709.96714