Istari

Definitions

Basic definitions

To define a constant, one may use the define command:

define /double/
/ fn x . x + x /
/ nat -> nat /;

The first argument is the constant being defined, the second gives the definition, and the third gives its type. Istari will then begin a proof wherein you can prove that double has the type nat -> nat.

You can also define a constant that takes arguments. For example, the same definition could be written:

define /double x/
/ x + x /
/ nat -> nat /;

Note that the type argument is unchanged from before; it is the type of double, not the type of double x.

Implicit arguments

A definition can also take implicit arguments, such as:

define /double_list {a} l/
/ ` append a l l /
/ intersect (i : level) . forall (a : U i) . list a -> list a /;

Here a is an implicit argument. Whenever double_list is used, the parser will insert evars for each implicit argument. One can suppress the insertion of implicit arguments using the tick symbol (`).

In the example, tick is used to suppress the implicit argument of append, allowing us to specify it explicitly.

Raw definitions

One can also define a constant without a type, using the command defineRaw. Then, if desired, one can attach a type to the constant after the fact using by proving a typing lemma and attaching it using recordTyping:

lemma "double_type"
/ double : nat -> nat /;

... do the proof ...

qed ();
recordTyping "double_type";

There is a hazard to using raw definitions: Constant definitions are not permitted to contain evars. With an ordinary definition, any evars (such as those arising from calling functions with implicit arguments) would typically be instantiated in the course of establishing the constant’s type. But in a raw definition one must be sure to avoid creating evars.

Recursive definitions

One can define a recursive function using a regular definition and fix, but Istari offers a more convenient syntax:

definerec /length {a} l/
/ list_case l 0 (fn h t . succ (length a t)) /
/ intersect (i : level) . forall (a : U i) . list a -> nat /;

Note that length’s implicit argument a is still explicit during the definition.

A recursive definition not only creates the constant, it also creates an unrolling rewrite. If one uses the unfold rewrite on (say) ` length a l, one gets the underlying fixpoint term:

fix (fn length1 a l . list_case l 0 (fn h t . succ (length1 a t))) a l

but using unroll results in a much cleaner term:

list_case l 0 (fn h t . succ (` length a t))

One can also make recursive definitions without a type using the command definerecRaw, subject to the same free evar hazard discussed above.

Mutually recursive definitions

One can define mutually recursive definitions using a recursive definition of a tuple, but again Istari offers a more convenient syntax:

definemutrecRaw /x/
/
  snork1 y =
    if Nat.eqb y 0 then
      x
    else
      snork2 y

  and
  snork2 y = snork1 (y - 1) * 2
/;

Here, x is a pervasive argument that is available to all of the mutually defined functions. Thus snork1 and snork2 have the type nat -> nat -> nat (although this has to be proven, of course).

The definition also creates an appropriate unrolling rewrite. If one uses unfold /snork2/ on the goal:

snork2 4 2 = 16 : nat

one obtains:

snork1 4 (2 - 1) * 2 = 16 : nat

For mutually recursive definitions, there is only the raw version. To give mutually recursive definitions a type, one must prove a typing lemma and then install it using recordTyping. Again, this is subject to the free evar hazard.

Datatype definitions

One defines datatypes with the typedef command. For example:

typedef
/
  datatype
    intersect (i : level) .
    forall (a : U i) .
    U i
  of
    tree : nat -> type =
    | Empty : tree 0
    | Node : forall (n : nat) . a -> forest n -> tree (succ n)
    and
    forest : nat -> type =
    | Nil : forest 0
    | Cons : forall (m n : nat) . tree m -> forest n -> forest (m + n)
/;

A datatype definition begins with “invisible” pervasive arguments preceded by the intersect keyword. They are intersect-bound in the types of the datatypes and constructors. (This is typically used for level variables.) Next are the visible pervasive arguments, preceded by the forall keyword. They are arguments to all the datatypes and constructors. Next is the universe to which the datatypes belong. In the example, the datatypes belong to the same universe (U i) to which the type argument a belongs.

Next are the datatype and their constructors. Each datatype can take additional “index” arguments. In the example, both datatypes take nat. Thus, they each have the type:

intersect (i : level) . forall (a : U i) . nat -> U i

The constructors can take “external” and “internal” arguments. The external argument’s types are unrestricted, except they cannot mention the datatypes being defined. The internal argument’s types must be the datatypes being defined. External arguments can be dependent (as in m and n in the example), but internal arguments must be non-dependent.

In the example, the constructors have type:

Empty : intersect (i : level) . forall (a : U i) . tree a 0
Node  : intersect (i : level) . forall (a : U i) (n : nat) . a -> forest a n -> tree a (succ n)
Nil   : intersect (i : level) . forall (a : U i) . forest a 0
Cons  : intersect (i : level) . forall (a : U i) (m n : nat) . tree a m -> forest a n -> forest a (m + n)

Defining a datatype incurs several typing obligations. With typedef, those obligations are discharged automatically by the typechecker if possible. With typedefRaw they are all handed over to the user.

For more discussion of how datatypes are used, see the datatypes page.

Reductions

To define new reductions to be used automatically by the normalization engine, one may use the reductions command:

reductions
/
  length _ (nil _) --> 0 ;
  length a (cons _ h t) --> succ (length a t) ;
  unrolling length
/;

To establish that the specified reductions are correct, the prover places the left- and right-hand-side into normal form and compares them. The unrolling clause indicates that, while normalizing, the prover should unroll length. One can also employ an unfolding clause to indicate that the prover should unfold certain identifiers. An unrolling clause will be applied only once (to avoid looping), and only on the left-hand-side. In contrast, an unfolding clause will be applied at every opportunity.

The left-hand-sides of reductions take a special syntax: they must consist of a constant followed by a list of arguments. Amidst the arguments is exactly one principal argument (nil _ and cons _ h t in the example). The principal argument must be parenthesized, even if it takes no arguments.

The right-hand-side uses the ordinary term syntax, except that implicit arguments are not inserted. (An evar on the right-hand-side would never be correct.)

Note that the process of verifying reductions is purely syntactic, so it can be sensitive to subtle changes. For example, this variation on the second reduction, though deceptively similar, is incorrect:

  length _ (cons a h t) --> succ (length a t) ;  (* incorrect *)

The right-hand-side of a reduction should not contain an instance of the left-hand-side (even beneath a lambda). Such a reduction (if valid) will be accepted, but it will result in normalization looping. Since unification tries to avoid normalizing, the non-terminating behavior might not arise for some time, making its cause not obvious when it finally does.

Inductive function definitions

To define an inductive function, one can make an ordinary definition using a datatype iterator. However, Istari provides a command to streamline the process:

defineInd /a/
/
  treesize : tree [a] __ -> nat of
  | Empty . 0
  | Node _ x f . succ (forestsize f)
  | One _ t . treesize t
  
  and
  forestsize : forest __ -> nat of
  | Nil . 0
  | Cons _ _ t f . treesize t + forestsize f
/
/
  intersect i . forall (a : U i) n . tree a n -> nat
  and
  intersect i . forall (a : U i) n . forest a n -> nat
/;

This defines multiple inductive functions simultaneously, installs the appropriate reductions, and attempts to prove the typing lemmas. The syntax of the inductive functions is discussed here. (Note that the keyword fnind is dropped.)

The first argument to defInd (in the example, a) gives a list of arguments that each inductive function should take. Those arguments are available to the pervasive arguments ([a]), and to the arms of the inductive functions (not used in this example). The second argument gives the inductive functions, and the third argument gives their types, as a list separated by and.

In addition to installing the appropriate reductions, it writes them to the registry. For example:

treesize _ _ (Empty _) --> 0

is written using the name treesize_Empty.