Consider these algorithms and data structures:

• sort a sequence of numbers;
• finite maps from numbers to (arbitrary-type) data
• finite maps from any ordered type to (arbitrary-type) data
• priority queues: finding/deleting the highest number in a set

To prove the correctness of such programs, we need to reason about less-than comparisons (for example, on integers) and about "these two sets/sequences have the same contents". In this chapter, we introduce some techniques for reasoning about:

• less-than comparisons on natural numbers
• permutations (rearrangements of lists)

Then, in later chapters, we'll apply these proof techniques to reasoning about algorithms and data structures.

The Less-Than Order on the Natural Numbers

These Check and Locate commands remind us about Propositional and the Boolean less-than operators in the Coq standard library.

We write x < y for the Proposition that x is less than y, and we write x <? y for the computable test that returns true or false depending on whether x<y. The theorem that lt is related in this way to ltb is this one:

For some reason, the Coq library has <? and <=? notations, but is missing these two:

Relating Prop to bool

The reflect relation connects a Proposition to a Boolean.

That is, reflect P b means that P↔True if and only if b=true. The way to use reflect is to make, for each of your operators, a lemma like these next three:

Here's an example of how you could use these lemmas. Suppose you have this simple program, (if a <? 5 then a else 2), and you want to prove that it evaluates to a number smaller than 6. You can use ltb_reflect "by hand":

For the ReflectT constructor, the guard a <? 5 must be equal to true. The if expression in the goal has already been simplified to take advantage of that fact. Also, for ReflectT to have been used, there must be evidence H that a < 5 holds. From there, all that remains is to show a < 5 entails a < 6. For the ReflectF constructor, the guard a <? 5 must be equal to false. So the if expression simplifies to 2 < 6, which is immediately provable by omega. The omega tactic, which is capable of automatically proving some theorems about inequalities, succeeds in finishing both subgoals. We'll explain omega in more detail later in this chapter. A shorter version of the above proof wouldn't need to do the assert and remember: we can directly skip to destruct.

But there's another way to use ltb_reflect, etc: read on.

A Tactic for Boolean destruction

We're now going to build a tactic for that you'll want to use, but you won't need to understand the details of how to build it yourself. Let's put several of these reflect lemmas into a Hint database, called bdestruct because we'll use it in our boolean-destruction tactic:

Here is the tactic, the body of which you do not need to understand. Invoking bdestruct on Boolean expression b does the same kind of reasoning we did above: reflection and destruction. It also attempts to simplify negations involving inequalities in hypotheses.

Here's a brief example of how to use bdestruct. There are more examples later.

Linear Integer Inequalities

In our proofs about searching and sorting algorithms, we sometimes have to reason about the consequences of less-than and greater-than. Here's a contrived example.

Now, there's a hard way to prove this, and an easy way. Here's the hard way.

With bogus subtraction, this omega_example1 theorem even True? Yes it is; let's try again, the hard way, to find the proof.

And here's the easy way.

Here we have used the omega tactic, made available by importing Coq.omega.Omega as we have done above. Omega is an algorithm for integer linear programming, invented in 1991 by William Pugh. Because ILP is NP-complete, we might expect that this algorithm is exponential-time in the worst case, and indeed that's true: if you have N equations, it could take 2^N time. But in the typical cases that result from reasoning about programs, omega is much faster than that. Coq's omega tactic is an implementation of this algorithm that generates a machine-checkable Coq proof. It "understands" the types Z and nat, and these operators: < = > ≤ ≥ + - ¬, as well as multiplication by small integer literals (such as 0,1,2,3...) and some uses of ∨ and ∧. Omega does not understand other operators. It treats things like a×b and f x y as if they were variables. That is, it can prove f x y > a×b → f x y + 3 ≥ a×b, in the same way it would prove u > v → u+3 ≥ v. Now let's consider a silly little program: swap the first two elements of a list, if they are out of order.

In this program, we wrote a>?b instead of a>b. Why is that?

We cannot compute with elements of Prop: we need some kind of constructible (and pattern-matchable) value. For that we use bool.

The name ltb stands for "less-than boolean."

Instead of defining an operator Nat.geb, we just defined (above) the notation for greater-or-equal-boolean as a less-or-equal-boolean with the arguments swapped.

Here's a theorem: maybe_swap is idempotent -- that is, applying it twice gives the same result as applying it once.

What do we do here? We must proceed by case analysis on whether a>b.

Now what? Look at the hypotheses H: b<a and H₀: a<b above the line. They can't both be true. In fact, omega "knows" how to prove that kind of thing. Let's try it:

omega didn't work, because it operates on comparisons in Prop, such as a>b; not upon comparisons yielding bool, such as a>?b. We need to convert these comparisons to Prop, so that we can use omega. Actually, we don't "need" to. Instead, we could reason directly about these operations in bool. But that would be even more tedious than the omega_example1 proof. Therefore: let's set up some machinery so that we can use omega on boolean tests.

Let's try again, a new way:

This is where we left off before. Now, watch:

The omega tactic noticed that above the line we have an arithmetic contradiction. Perhaps it seems wasteful to bring out the "big gun" to shoot this flea, but really, it's easier than remembering the names of all those lemmas about arithmetic!

Moral of this story: When proving things about a program that uses boolean comparisons (a <? b), use bdestruct. Then use omega. Let's review that proof without all the comments.

Permutations

Another useful fact about maybe_swap is that it doesn't add or remove elements from the list: it only reorders them. We can say that the output list is a permutation of the input. The Coq Permutation library has an inductive definition of permutations, along with some lemmas about them.

We say "list al is a permutation of list bl", written Permutation al bl, if the elements of al can be reordered (without insertions or deletions) to get the list bl.

You might wonder, "is that really the right definition?" And indeed, it's important that we get a right definition, because Permutation is going to be used in the specification of correctness of our searching and sorting algorithms. If we have the wrong specification, then all our proofs of "correctness" will be useless. It's not obvious that this is indeed the right specification of permutations. (It happens to be true, but it's not obvious!) In order to gain confidence that we have the right specification, we should use this specification to prove some properties that we think permutations ought to have.

Exercise: 2 stars, standard (Permutation_properties)

Think of some properties of the Permutation relation and write them down informally in English, or a mix of Coq and English. Here are four to get you started:

• 1. If Permutation al bl, then length al = length bl.
• 2. If Permutation al bl, then Permutation bl al.
• 3. [1;1] is NOT a permutation of [1;2].
• 4. [1;2;3;4] IS a permutation of [3;4;2;1].

YOUR ASSIGNMENT: Add three more properties. Write them here: Now, let's examine all the theorems in the Coq library about permutations:

Which of the properties that you wrote down above have already been proved as theorems by the Coq library developers? Answer here:

Let's use the permutation rules in the library to prove the following theorem.

That example illustrates a general method for proving permutations involving cons :: and append ++. You identify some portion appearing in both sides; you bring that portion to the front on each side using lemmas such as Permutation_app_comm and perm_swap, with generous use of perm_trans. Then, you use perm_skip to cancel a single element, or Permutation_app_head to cancel an append-chunk.

Exercise: 3 stars, standard (permut_example)

Use the permutation rules in the library (see the Search, above) to prove the following theorem. These Check commands are a hint about useful lemmas; you won't necessarily need all of them.

Exercise: 1 star, standard (not_a_permutation)

Prove that [1;1] is not a permutation of [1;2]. Hints are given as Check commands.

Back to maybe_swap. We prove that it doesn't lose or gain any elements, only reorders them.

Now let us specify functional correctness of maybe_swap: it rearranges the elements in such a way that the first is less than or equal to the second.

Summary: Comparisons and Permutations

To prove correctness of algorithms for sorting and searching, we'll reason about comparisons and permutations using the tools developed in this chapter. The maybe_swap program is a tiny little example of a sorting program. The proof style in maybe_swap_correct will be applied (at a larger scale) in the next few chapters.

Exercise: 2 stars, standard (Forall_perm)

To close, a useful utility tactic and lemma. First, the tactic. Coq's inversion H tactic is so good at extracting information from the hypothesis H that H becomes completely redundant, and one might as well clear it from the goal. Then, since the inversion typically creates some equality facts, why not then subst ? This motivates the following useful tactic, inv.

Second, the lemma. You will find inv useful in proving it. Forall is the Coq library's version of the All proposition defined in Logic, but defined as an inductive proposition rather than a fixpoint. Prove this lemma by induction; but is it induction on al, or on bl, or on Permutation al bl, or on Forall f al ?