# PML 2 : Integrated Program Verification in ML

We present the PML 2 language, which provides a uniform environment for programming, and for proving properties of programs in an ML-like setting. The language is Curry-style and call-by-value, it provides a control operator (interpreted in terms of classical logic), it supports general recursion and a very general form of (implicit, non-coercive) subtyping. In the system, equational properties of programs are expressed using two new type formers, and they are proved by constructing terminating programs. Although proofs rely heavily on equational reasoning, equalities are exclusively managed by the type-checker. This means that the user only has to choose which equality to use, and not where to use it, as is usually done in mathematical proofs. In the system, writing proofs mostly amounts to applying lemmas (possibly recursive function calls), and to perform case analyses (pattern matchings).

Comments

There are no comments yet.

## Authors

• 3 publications
• ### Check Your (Students') Proofs-With Holes

Cyp (Check Your Proofs) (Durner and Noschinski 2013; Traytel 2019) verif...
09/02/2020 ∙ by Dennis Renz, et al. ∙ 0

read it

• ### Programming Metamorphic Algorithms: An Experiment in Type-Driven Algorithm Design

In dependently typed programming, proofs of basic, structural properties...
10/30/2020 ∙ by Hsiang-Shang Ko, et al. ∙ 0

read it

• ### Analysis of MiniJava Programs via Translation to ML

MiniJava is a subset of the object-oriented programming language Java. S...
12/30/2020 ∙ by Martin Mariusz Lester, et al. ∙ 0

read it

• ### Zero-Cost Coercions for Program and Proof Reuse

We introduce the notion of identity coercions between non-indexed and in...
02/02/2018 ∙ by Larry Diehl, et al. ∙ 0

read it

• ### The Marriage of Univalence and Parametricity

Reasoning modulo equivalences is natural for everyone, including mathema...
09/11/2019 ∙ by Nicolas Tabareau, et al. ∙ 0

read it

• ### Classical Proofs as Parallel Programs

We introduce a first proofs-as-parallel-programs correspondence for clas...
09/10/2018 ∙ by Federico Aschieri, et al. ∙ 0

read it

• ### To Memory Safety through Proofs

We present a type system capable of guaranteeing the memory safety of pr...
10/29/2018 ∙ by Hongwei Xi, et al. ∙ 0

read it

##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction: joining programming and proving

In the last thirty years, significant progress has been made in the application of type theory to computer languages. The Curry-Howard correspondence, which links the type systems of functional programming languages to mathematical logic, has been explored in two main directions. On the one hand, proof assistants such as Agda [agda] or Coq [coq] are based on very expressive logics [mltt, coc]. To establish their consistency, the underlying programming languages need to be restricted to provably terminating programs. As a result, they forbid the most general forms of recursion. On the other hand, functional programming languages such as Haskell, SML or OCaml are well-suited for programming, as they impose no restriction on recursion. However, their type systems are inconsistent when considered as logics, which means that they cannot be used for proving mathematical formulas.111This particular point will be explained in more detail in Section 5.

The aim of is to provide a uniform environment in which programs can be designed, specified and proved. The idea is to combine a full-fledged ML-like programming language, with an enriched type system allowing the specification of computational behaviours.222On might argue that is not a full-fledged ML-like language as it does not have mutable references. It is nonetheless effectful as it provides a control operator similar to Scheme’s call/cc. The obtained system can thus be used as ML for type-safe general programming, and as a proof assistant for proving properties of ML programs. The uniformity of the framework implies that programs can be incrementally refined to obtain more and more guarantees. In particular, there is no syntactic distinction between programs and proofs. The only difference is that the latter must be typed-checked against the consistent core of the system, which only accepts programs that can be proved terminating. In the current implementation, programs must also be proved terminating to be directly accepted by the type-checker. It would however be possible to accept programs that do not pass the termination check, and we would then need to make sure that such programs are not used to write proofs.333For example, we could have two different function types: one used for functions whose termination has been established, and another one that that may be used for any function (and that would be a supertype of the former). Note however that the system can already be used to reason about arbitrary programs, including untyped ones and those whose termination cannot be established (an example will be given in Section 5.3).

### 1.1 Program verification principles

In , program properties may be specified with types containing equations of the form t ≡ u, where t and u are terms of the language itself. By quantifying over the free variables of these terms, we can express properties such as the following.

// "add" is commutative.
nnat, mnat, add n m  add m n
// "reverse" is involutive.
a, llista⟩, reverse (reverse l)  l
// "sort" produces sorted lists.
llistnat⟩, sorted (sort l)  true
// All natural numbers are equal to "Zero".
nnat, n  Zero.
// Sorted lists are not affected by "sort".
llistnat⟩, (sorted l  true)  (sort l  l)

Of course, such a specification may be inaccurate, in which case it will not be provable. Note that it is possible to observe complex behaviours using predicates such as sorted, which correspond to boolean-valued functions.

The language relies on two main ingredients for proving that programs meet their specifications. First, the type system of the language can be considered as a (classical) logic through the Curry-Howard correspondence, although it is only consistent for terminating programs. This (usual) part of the type system provides basic reasoning principles, which are used to structure the proofs. The second ingredient is an automatic decision procedure for the equational theory of the language. It is used to manage a context of equational assumptions, and eventually prove equations in this context. The decision procedure is driven by the type-checker, without any direct interaction with the user. As a consequence, the user only has to care about the structure of the proof, and not about the details of where equations should be applied. In fact, equality types of the form t ≡ u are computationally irrelevant in the system. More precisely, t ≡ u is equivalent to the unit type when the denoted equality holds (and can be proved), and it is empty otherwise. As a consequence, a proof will generally consist of a (possibly recursive) program that calls other programs and performs pattern matching but eventually just returns a completely uninteresting result. Nonetheless, writing proofs in this way is a very similar experience to writing functional programs. We can hence hope that our approach to program verification will feel particularly intuitive to functional programmers.

### 1.2 Previous work on the language

is based on many ideas introduced by Christophe Raffalli in the PML language [pml]. Although this first version of the system was very encouraging on the practical side, it did not stand on solid theoretical grounds. On the contrary, is based on a call-by-value, classical realizability model designed by the author [lepigre2016]. This framework provides a satisfactory way of combining call-by-value evaluation and effects with dependent function types.444Due to the soundness issues explained in previous work [lepigre2016], the application of dependent functions is usually restricted to value arguments. This is commonly called value restriction in the context of ML. The proposed solution relies on a relaxed form of value restriction (called semantical value restriction), which takes advantage of our notion of program equivalence and its decision procedure.555Intuitively, two terms are (observationally) equivalent if they have the same computational behaviour (i.e., they both converge or they both diverge) in every possible evaluation context [lepigre2016, lepigrePhD]. In particular, it allows the application of a dependent function to a term which is not a value, under the condition that it can be proved equivalent to some value. This is especially important because dependent functions are an essential component of . Indeed, they enable a form of typed quantification without which many program properties could not be expressed (see Section 3).

Another important specificity of ’s type system is that it relies on the notion of local subtyping, which was introduced in joint work with Christophe Raffalli [subml]. This framework can be used to give a syntax-directed formulation of the typing and subtyping rules of the system666This means that exactly one typing rule applies for every term constructor, and only one subtyping rule applies for every pair of type constructors (up to commutation and circular proof construction)., despite its Curry-style nature. In particular, it provides a very general notion of infinite (circular) proof, that is used for handling inductive and coinductive types in subtyping derivations, and recursion and termination checking in typing derivations. Of course, infinite proofs are only valid if they are well-founded (this is ensured using the size-change principle [scp]). The combination of local subtyping [subml] and the realizability model of the system has been addressed in the author’s PhD thesis [lepigrePhD, Chapter 6].

Last but not least, the implementation of [implem] was initiated as part of the author’s thesis [lepigrePhD], and continues with the collaboration of Christophe Raffalli. The implemented system is intended to remain very close to the theoretical type system, and every part of the implementation is justified by the formal semantics of [lepigrePhD]. Note that all the examples given in this paper are accepted by the version 2.0.2_types2017 of , which can be downloaded at the following URL.

### 1.3 Disclaimer: the aim of this paper

This document is intended to be an introductory paper for the system. Its aim is not to give the details of the realizability semantics, nor to prove new theoretical results, but rather to list the principles and ideas on which is based. In particular, Section 6 contains an extensive description of several ideas that we would like to investigate in the near future, and that will be necessary for achieving the goals of completely. For technical details, the reader should refer to the author’s thesis [lepigrePhD], and related papers [lepigre2016, subml].

## 2 Functional programming in PML2

Our first goal in designing was to obtain a practical, functional programming language. Out of the many possible technical choices, we decided to consider a call-by-value language similar to OCaml or SML, as they have proved to be highly practical and efficient. Our language provides polymorphic variants [polyvar] and SML-style records, which are convenient for encoding data types. As an example, the type of lists can be defined as follows,777Note that  list  is used without its type parameter in the type of the  Cns  constructor. This is due to the fact that the “ type rec ” syntax is desugared to an inductive type (or least fixed point), and  list⟨a⟩  is actually defined as “ μ list, [Nil; Cns of {hd : a; tl : list}] ”. In particular, the current version of does not support polymorphically recursive types. together with the corresponding iter and append functions.

type rec lista = [Nil ; Cns of {hd : a ; tl : list}]
val rec iter : a, (a  {})  lista  {} =
fun f l {
case l {
Nil     {}
Cns[c]  f c.hd; iter f c.tl
}
}
val rec append : a, lista  lista  lista =
fun l1 l2 {
case l1 {
Nil     l2
Cns[c]  Cns[{hd = c.hd ; tl = append c.tl l2}]
}
}

Note that the iter and append functions are polymorphic, which means, for instance, that they can be applied to lists with elements of an arbitrary type. In our syntax, this is explicitly materialised using universally quantified type variables. Note also that the iter function relies on the type {}, which contains records with no fields. It plays the same role as OCaml’s unit type, and its unique inhabitant is denoted {} as well. As in System F [girard, reynolds], polymorphism can be used anywhere in types, and it is not limited to let-polymorphism (or prenex polymorphism) as in most ML-like languages.

### 2.1 Control operator and classical logic

The programming languages of the ML family generally include effectful operations such as references (i.e., mutable variables). Our system is no exception since it provides a control operator similar to call/cc.888This instruction can be used to capture the current continuation (or evaluation context), so that it can be restored later. It was first introduced in the Scheme language. On the programming side, it may be used to encode a form of exception mechanism. For instance, we can define the following exists function, which tests whether there is an element satisfying a given predicate in a given list, and stops as soon as possible if such an element is found.

val exists : a, (a  bool)  lista  bool =
fun pred l {
save k {
iter (fun e { if pred e { restore k true } }) l;
false
}
}

Here, the continuation is saved in a variable k before calling the iter function, and it is restored with the value true if an element satisfying the predicate is found. In this case, the evaluation of iter is simply aborted. To obtain a similar behaviour without using a continuation would require the user to write an independent recursive function (i.e., one that does not rely on iter). A more interesting example that cannot be written without a control operator will be given in Section 4.

As is now well-known, control operators such as ours can be used to give a computational content to classical theorems, thus extending the Curry-Howard correspondence to classical logic [griffin]. It is hence possible to define programs with a type corresponding to Peirce’s law, or to the law of the excluded middle.

val peirce : a b, ((a  b)  a)  a =
fun x {
save k {
x (fun y { restore k y })
}
}
// Disjoint sum (logical disjunction) and (logical) negation
type eithera,b = [InL of a ; InR of b]
type nega = a  x,x
val excl_mid : a, {}  eithera, nega⟩⟩ =
fun _ {
save k {
InR[fun x { restore k InL[x] }]
}
}

Note that the definition of excl_mid requires a dummy function constructor due to the call-by-value evaluation strategy. Indeed, excl_mid would not be a value if it did not start with an abstraction, and it would thus save its continuation right away, unlike peirce which must first be given an argument to trigger the computation. This is related to value restriction [wright1, wright2], which is required in presence of control operators [harper].999Value restriction is a sufficient (but not necessary) condition for correctness.

From a computational point of view, manipulating continuations using control operators can be understood as “cheating”. For example excl_mid (or rather, excl_mid {}) saves the continuation and immediately returns a (possibly false) proof of neg⟨a⟩. Now, if this proof is ever applied to a proof of a (which would result in absurdity), the program backtracks and returns the given proof of a. This interpretation has been well-known for a long time, and an account is given in the work of Wadler [wadler, Section 4], for example.

### 2.2 Non-coercive subtyping

In , subtyping plays a very important role as it allows us to give a mostly syntax-directed presentation of the system [subml, lepigrePhD]. Although it is less widespread than polymorphism in mainstream languages, subtyping can be exploited to improve code modularity. Note that we here consider a non-coercive form of subtyping, which means that if a is a subtype of b, then any value of type a is also a value of type b (i.e., no coercion is required).

In the system, there are many forms of subtyping that may interact. In particular, subtyping is used to handle all the connectives that do not have algorithmic contents (i.e., no counterpart in the syntax of terms). Such connectives include quantifiers as well as inductive types, but also the equality types of . Subtyping also plays an important role with variants and records. For instance, it implies that a record can always have more fields than required. Moreover, subtyping enables many commutations of connectives.

As a first example, we can show that the type corresponding to the (classical) double negation elimination principle can in fact be seen as an instance of Peirce’s law. Indeed, it can be defined as follows in .

val dneg_elim : a, negnega⟩⟩  a =
peirce

It is relatively easy to see that the type of peirce is indeed a subtype of that of dneg_elim. The corresponding subtyping derivation is sketched below.101010The proof does not contain all the necessary information to ensure its validity. The reader should refer to the author’s thesis [lepigrePhD, Figure 6.5] to fill in the missing details. a0 ⇒ ∀x,x a0 ⇒ ∀x,x a0 a0 ∀x,x a0 (a0 ⇒ ∀x,x) ⇒ ∀x,x (a0 ⇒ ∀x,x) ⇒ a0 a0 a0 ((a0 ⇒ ∀x,x) ⇒ a0) ⇒ a0 ((a0 ⇒ ∀x,x) ⇒ ∀x,x) ⇒ a0 ∀b, ((a0 ⇒ b) ⇒ a0) ⇒ a0 ((a0 ⇒ ∀x,x) ⇒ ∀x,x) ⇒ a0 ∀a, ∀b, ((a ⇒ b) ⇒ a) ⇒ a ((a0 ⇒ ∀x,x) ⇒ ∀x,x) ⇒ a0 ∀a, ∀b, ((a ⇒ b) ⇒ a) ⇒ a ∀a, ((a ⇒ ∀x,x) ⇒ ∀x,x) ⇒ a

Intuitively, universal quantification on the right of the inclusion can be eliminated by introducing a fresh constant. On the left, variables that are quantified over can be replaced by anything. As usual, the subtyping rule for handling the arrow type reverses the inclusion between the domains due to the contra-variance of the arrow type.

We will now consider the extension of the type of lists with an additional constructor allowing constant time concatenation. In , the corresponding type of “append lists” can be defined in such a way that it admits the type of regular lists as a subtype.

type rec alista =
[Nil ; Cns of {hd : a; tl : alist} ; App of alist × alist]
// Constant time "append" function.
val alist_append : a, alista  alista  alista =
fun l1 l2 { App[(l1,l2)] }

Although regular lists are a special case of “append lists”, the converse is not true. To transform an “append list” into a list, it is necessary to define the following recursive, flattening function.

val rec alist_to_list : a, alista  lista =
fun l {
case l {
Nil     Nil
Cns[c]  Cns[{hd = c.hd ; tl = alist_to_list c.tl}]
App[c]  append (alist_to_list c.1) (alist_to_list c.2)
}
}

Another example of extension for a pre-existing type can be obtained by defining the type of red-black trees as a subtype of binary trees. More precisely, a red-black tree can be represented as a tree whose nodes have an extra color field. Of course, the presence of additional information in the form of a new record field does not prevent the use of tree functions such as binary search.

### 2.3 Toward the encoding of a module system

Despite its many different features, remains a fairly small system, which can be implemented rather concisely. Its design is based on the principle that every feature should be orthogonal. For instance, there is only one notion of product type in : records. This is not the case in OCaml, for instance, which provides tuples, records, objects, modules, which all have common product type characteristics.

In , modules can be easily encoded using a combination of records for storing the values, functions for building functors, and existentials for type abstraction. However, the implementation does not yet provide a specific syntax for modules. For instance, there is still no way of “opening” a module so that its values are accessible in the scope. It is nonetheless possible to work with the target of the encoding directly. For example, we can define a type corresponding to the signature (or interface) of a simple module providing an abstract representation for the stack data structure with the corresponding operations.111111Here, corresponds to the sort of types with one type parameter. In particular, is the sort of types (or propositions), and we will later encounter the sort of program values .

type stack_sig = stack: ο  ο,
{ empty : a, stacka⟩;
push  : a, a  stacka  stacka⟩;
pop   : a, stacka  [None ; Some of a × stacka⟩] }

An implementation of this interface can then be defined by giving a corresponding record value. For example, we can implement stacks with lists as follows.

val stack_impl : stack_sig =
{ empty = (Nil : a, lista⟩);
push  = fun e s { Cns[{hd = e; tl = s}] };
pop   = fun s {
case s {
Nil     None
Cns[c]  Some[(c.hd, c.tl)]
}
} }

Note that we need to give at least some type annotation for the system to know what to instantiate the existential with. This could be done in a more systematic way with a syntax requiring the user to give the intended definition for the stack type.

It is possible to define a dot-projection operation in order to access abstract types, so that it is possible to write stack_impl.stack to refer to the type of stacks. More details are given in previous work [subml], and in the corresponding implementation.

## 3 Verification of ML programs

is not only a programming language, but also a proof assistant focusing on program verification. Its proof mechanism relies on equality types of the form t ≡ u, where t and u are arbitrary (possibly untyped) terms of the language itself. Such an equality type is inhabited by the term {}121212Recall that it denotes a record with no fields, or the unique inhabitant of a one-element type. if the denoted equivalence is true, and it is empty otherwise. Equivalences are managed using a partial decision procedure that is driven by the construction of programs. An equational context is maintained by the type checker to keep track of the equational assumptions during the construction of proofs. This context is extended when new equations are learnt (e.g., when a lemma is applied), and an equation is proved by deriving a contradiction (e.g., two different variants that are equated) from its negation.

Terms not only appear in (equality) types, but also play the role of objects in the underlying logic. In particular, they can be quantified over in types, and thus form one particular domain of discourse. In fact, our system is based on a higher-order logic with several atomic sorts (including types and terms), which means that many different kinds of objects can be quantified over (universally and existentially) in our types. We can for example quantify over types with one type parameter (of sort ), as in the signature used for the stack module given in the previous section.

### 3.1 (Un)typed quantification and unary natural numbers

To illustrate the proof mechanism, we will consider very simple examples of proofs on unary natural numbers. Their type is given below, together with the corresponding addition function defined using recursion on its first argument.

type rec nat = [Zero ; S of nat]
val rec add : nat  nat  nat =
fun n m {
case n {
Zero  m
S[k]  S[add k m]
}
}

As a first example, we will show that for all n we have add Zero n ≡ n. This property is expressed using the type ∀n:, add Zero n ≡ n, and it is proved as follows.131313Here, the domain of the quantification is the set of values of the language, whose sort is . It is not limited to natural numbers, and also encompasses booleans and functions for example.

val add_Zero_n : n:ι, add Zero n  n =
{} // immediate

The proof is immediate (i.e., only {}) as we have add Zero n ≡ n by definition of add. Note that this equivalence holds for every value n, whether it corresponds to an element of the type nat or not. For instance, it can be used to show add Zero true ≡ true since the term add Zero true evaluates to true. Here, it is crucial that n ranges only over values of the language, as otherwise the definition of add could not be unfolded. Indeed, since we are in call-by-value, it is only possible to effectively apply a function when its arguments are all values.

Let us now show that for every n we have add n Zero ≡ n. Although this property looks similar to add_Zero_n, the following proof is invalid.

// val add_n_Zero : ∀n:ι, add n Zero  n =
//   {} // invalid

Indeed, the equivalence add n Zero ≡ n does not hold when n is not a unary natural number. In this case, the computation of add n Zero produces a runtime error while that of n does not. As a consequence, we need to rely on a form of quantification that only ranges over unary natural numbers. This can be achieved with the type ∀n∈nat, add n Zero ≡ n, which corresponds to a (dependent) function taking as input a natural number n and returning a proof of add n Zero ≡ n. This property can then be proved using induction (i.e., using a recursive function) and case analysis (i.e., pattern matching) with the following program.

val rec add_n_Zero : nnat, add n Zero  n =
fun n {
case n {
Zero  {}
S[k]  add_n_Zero k
}
}

If n is Zero, then we need to show add Zero Zero ≡ Zero, which is immediate by definition of add. In the case where n is S[k] we need to show add S[k] Zero ≡ S[k]. By definition of add, this reduces to S[add k Zero] ≡ S[k]. We can then use the induction hypothesis add_n_Zero k to learn add k Zero ≡ k and conclude the proof. The dependent product type (or typed quantification) constructor is not primitive in . It is encoded using a membership type of the form t∈a which contains all the elements of type a that are equivalent to the term t (it can be seen as a form of singleton type). The dependent function type ∀x∈a, b is then encoded as ∀x:, x∈a ⇒ b, which corresponds to the relativised quantification scheme (see previous work [lepigre2016, lepigrePhD]).

It is important to note that, in our system, a program that is considered as a proof needs to go through a termination checker. Indeed, a looping program could be used to prove anything otherwise.141414More details will be given in Section 5. For example, the following proof is rejected.

// val rec add_n_Zero_loop : ∀n∈nat, add n Zero  n =
//   fun n {
//     add_n_Zero_loop n
//   }

It is however easy to see that add_Zero_n and add_n_Zero are terminating, and hence valid. In the following, we will only consider programs that can be automatically proved terminating by the system.

### 3.2 Building up an equational context

There are two main ways of learning new equations in the system. On the one hand, when a term t is matched in a case analysis, a given branch can only be reached when the corresponding pattern C[x] matches. In this case we can extend the equational context with t ≡ C[x]. On the other hand, it is possible to invoke a lemma by calling the corresponding function. In particular, this must be done to use the induction hypothesis in proofs by induction like in add_Zero_n or the following lemma.

val rec add_n_S_m : n mnat, add n S[m]  S[add n m] =
fun n m {
case n {
Zero  {}
S[k]  add_n_S_m k m
}
}

In this case, the equation corresponding to the conclusion of the used lemma is directly added to the context. Of course, more complex results can be obtained by combining more lemmas. For example, the following proves the commutativity of addition using a proof by induction with add_n_Zero and add_n_S_m.

val rec add_comm : n mnat, add n m  add m n =
fun n m {
case n {
Zero  add_n_Zero m
S[k]  add_comm k m; add_n_S_m m k
}
}

Note that terms can be put in sequence with a semicolon. In the above proof, the recursive call add_comm k m is performed first, before calling add_n_S_m m k. They are also type-checked in that order, and the corresponding equations are added to the context one after the other as a side-effect to type-checking. Here, the order in which equations are added is not significant (the resulting equational context is the same either way), but that is not always the case (lemmas may require some equations to hold to be applied).

### 3.3 Detailed proofs using type annotations

Although the above proof of commutativity is perfectly valid, it might not be easy enough to read by a human. This problem arises in most proof assistants. For instance, it is almost impossible to understand a Coq [coq] proof without replaying it step by step in a compatible editor. In , it is possible to annotate proofs to highlight the corresponding thought process. For example, we can reformulate add_comm as follows.

val rec add_comm : n mnat, add n m  add m n =
fun n m {
case n {
Zero  show add Zero m  add m Zero using add_n_Zero m; qed
S[k]  show add k m  add m k using add_comm k m;
deduce add S[k] m  S[add m k];
show add S[k] m  add m S[k] using add_n_S_m m k; qed
}
}

Note that no addition to the system is required for such annotations to be supported, it is only syntactic sugar. For instance, qed is a synonym of {}, and show u1 ≡ u2 using p is translated to p : u1 ≡ u2, which amounts to a type coercion.

Many examples of proofs and programs are provided with the implementation of the system. Each of the examples given here has been automatically checked upon the generation of the document, they are hence correct with respect to the implementation.

### 3.4 Mixing proofs and programs

We will now see that the programming and the proving features of can be mixed when constructing proofs or programs. In fact, there is no obvious distinction between the world of the usual programs, and the world of proofs (remember that proofs are programs in ). For instance, it is possible to combine proofs with programs for them to transport properties (e.g., addition carrying its own commutativity). This can be achieved using restriction types, which are in fact used to encode equality types. In , the type a | t ≡ u is equivalent to a if t ≡ u is true, and to the empty type otherwise. The type t ≡ u is thus encoded as {} | t ≡ u, where {} is the unit type. Intuitively, the restriction type can be seen as a form of conjunction with no algorithmic contents.

When combined with existential quantification and the membership type, restriction can be used to encode a set type syntax similar to that of NuPrl [nuprl]. Indeed, we can define {x ∈ a | t ≡ u}, which contains all the elements of type a such that t ≡ u holds, as ∃x:, x∈(a | t ≡ u). This provides a very useful scheme for defining the set of terms of a

that satisfy some property. For example, we can encode the type of vectors (i.e., lists of a given length) by taking every list

l that has size s. The type of vectors will hence have two parameters: the type of the contained elements and a term giving the size of vectors.

val rec length : a:ο, lista  nat =
fun l {
case l {
Nil     Zero
Cns[c]  S[length c.tl]
}
}
type veca:ο, s:τ⟩ = {l  lista | length l  s}

In the definition of vec, the second parameter must have sort (the sort of terms) and not (the sort of values). Indeed, it is often required to work with vectors whose sizes are of the form add n m (see the definition of the app function below). There is no constraint on the type of s in the definition of vec. This means that it is possible to consider the type of vectors of size true for example, but it will be empty since the length function only returns natural numbers. One of the main advantages of this approach is that it is compatible with subtyping.

Let us stress that vectors can always be used as lists, independently of their size. The type of vectors is a subtype of the type of lists, as shown by the following function.

val vec_to_list : a:ο, s:τ, veca,s  lista =
fun x { x }

Note that we will never need to use the function vec_to_list to turn a vector into a list. A vector can be seen as a list directly, without relying on any form of coercion.

We will now define a concatenation function app on vectors. It produces a vector whose length is the sum of the lengths of its two arguments. Note that we are first required to define the length_total function for a technical reason that will be explained in Section 5.151515We have good hopes of simplifying this particular point in future work, for example by automatically obtaining  length_total  from the definition of  length  as they have a similar structure.

val rec length_total : a:ο, llista⟩, v:ι, v  length l =
fun l {
case l {
Nil     {}
Cns[c]  length_total c.tl
}
}
val rec app : a:ο, m n:ι, veca, m  veca, n  veca, add m n =
fun l1 l2 {
case l1 {
Nil     l2
Cns[c]  length_total c.tl; Cns[{hd = c.hd; tl = app c.tl l2}]
}
}

Thanks to the Curry-style nature of our system, the sizes of the argument vectors do not need to be provided as arguments. This may be surprising for readers that are used to manipulating equivalent types in Agda or Coq, for example.

In , the proof mechanism can also be used to eliminate unreachable code. Indeed, if an equational contradiction is triggered only by learning equations along the way, then the code in that branch cannot be accessed during evaluation. In this case, a special value ✂ (to be pronounced “scissors”) can be used. Note that reachability information would be particularly useful to efficiently compile programs down to assembly code.

## 4 Programs extracted from classical proofs

We will now consider an example of a program that can only be written in a classical setting (i.e., with control operators). We are going to define a function on streams of natural numbers called extract

, that extracts a substream of odd numbers or a substream of even numbers from its input. This will prove that such a substream exists for all steams of natural numbers.

161616Intuitively, we will have shown that every stream of natural numbers contains either infinitely many odd numbers or infinitely many even numbers (and possibly both). First, we need to define odd and even numbers using our set type syntax.

val rec is_odd : nat  bool =
fun n {
case n {
Zero  false
S[m]
case m {
Zero  true
S[p]  is_odd p
}
}
}
type odd  = {vnat | is_odd v  true }
type even = {vnat | is_odd v  false}

As for the length function of the previous section, we will need to show that the is_odd function is total for a technical reason (see Section 5 for more details). Intuitively, this will allow us to reason by cases on the oddness (or evenness) of a given number of the input stream. Indeed, the totality of is_odd implies that this function always produces a result value, and hence that we can pattern match on its result.

val rec odd_total : nnat, v:ι, is_odd n  v =
fun n {
case n {
Zero  {}
S[m]
case m {
Zero  {}
S[p]  odd_total p
}
}
}

We also need to define the type of streams, together with a related type corresponding to streams with an explicit size annotation (or ordinal) s. Intuitively, this size annotation indicates the number of elements that are available in the stream (see Section 5 for more details on sized-types).

type corec streama = {}  {hd : a; tl : stream}
type sized_streams,a = ν_s stream, {}  {hd : a; tl : stream}

We can now define the extract_aux function, that will be used to define extract on the next page. Note that it relies on abort, which logically amounts to the ex falso quodlibet principle. Size annotations are also required on the type of extract_aux, for our type-checking algorithm to prove its termination.

val abort : y, (∀x,x)  y = fun x { x }
val rec extract_aux : a b,
negsized_streama,even⟩⟩
negsized_streamb,odd ⟩⟩  negstreamnat⟩⟩ =
fun fe fo s {
let {hd ; tl} = s {};
use odd_total hd;
if is_odd hd {
fo (fun _ {
{hd = hd; tl = save oc {
abort (extract_aux fe (fun x { restore oc x }) tl)}}
})
} else {
fe (fun _ {
{hd = hd; tl = save ec {
abort (extract_aux (fun x { restore ec x }) fo tl)}}
})
}
}

Intuitively, the extract_aux function looks at the head of its third argument (a stream of natural numbers), and depending on whether this number is odd or even, the function calls one of its first two arguments. They can be understood as partially constructed stream of even or odd numbers, in the form of continuations.171717Logical negation is intuitively used to type continuations represented in the form of a function. A continuation of type  neg⟨a⟩  can thus be called with a value of type  a  in any context since it yields a logical contradiction (or an element of type  ∀x,x ). The read number is then added to this stream, and a recursive call is made to continue the construction. It may seem surprising that our prototype implementation is able to establish the termination of extract_aux as an element is added to one of two streams at each call. Moreover, this example does not satisfy the usually required semi-continuity condition [abel]. It is here accepted because our termination test depends more finely on the structure of programs than previous approaches [subml].

The extract function can then be defined as follows, to complete the construction. The function starts by saving two continuations, corresponding to the constructors InL and InR of the return type, and then calls extract_aux on the input stream.

val extract : streamnat  eitherstreameven⟩, streamodd⟩⟩ =
fun s {
save a {
InL[save ec { restore a InR[save oc {
abort (extract_aux (fun x { restore ec x})
(fun x { restore oc x }) s)
} ] } ]
}
}

The very fact that we can write extract proves that it is possible to extract a stream of odd numbers or a stream of even numbers from any stream of natural numbers.

Of course, it is only possible to observe a finite prefix of a stream using a terminating program. As a consequence, we may want to consider a finite version of extract, whose result is a vector of a given size n instead of a stream.

val rec prefix : a, nnat, streama  veca,n =
fun n s {
case n {
Zero  Nil
S[k]  let {hd ; tl} = s {};
Cns[{hd ; tl = prefix k tl}]
}
}
val finite_extract : nnat,
streamnat  eitherveceven,n⟩, vecodd,n⟩⟩ =
fun n s {
case extract s {
InL[s]  InL[prefix n s]
InR[s]  InR[prefix n s]
}
}

It is possible to give an equivalent definition of finite_extract in an intuitionistic setting (i.e., without using a control operator). Indeed, at most 2 × n elements of the input stream need to be considered to construct the result.

Of course, the type of extract or finite_extract does not directly imply that the result of these functions is a substream of their input. It is nonetheless easy to convince oneself that this is indeed the case, and we could certainly prove it in with some effort.

To conclude this section, we will consider the result returned by extract (or rather finite_extract) on two particular streams. The former will be the stream of all natural numbers, which can be defined as follows, and is called naturals.

val rec naturals_from : nat  streamnat =
fun n _ {
{hd = n; tl = naturals_from S[n]}
}
val naturals : streamnat = naturals_from Zero

The latter will be a stream of ones, prefixed by three zeroes. It can be defined as follows, and is called three_zeroes_then_ones.

val rec ones : streamnat =
fun _ { {hd = S[Zero]; tl = ones} }
val three_zeroes_then_ones : streamnat =
fun _ { {hd = Zero; tl =
fun _ { {hd = Zero; tl =
fun _ { {hd = Zero; tl = ones} }} }} }

The results of finite_extract on naturals and three_zeroes_then_ones may be displayed using the following printing functions.

val rec print_nat : nat  {} =
fun n {
case n {
Zero  print "0"
S[k]  print "S"; print_nat k
}
}
val rec print_list : a, (a  {})  lista  {} =
fun pelt l {
case l {
Nil           print "\n"
Cns[{hd;tl}]  pelt hd; print " "; print_list pelt tl
}
}
val print_res : eitherlistnat⟩, listnat⟩⟩  {} =
fun e {
case e {
InL[l]  print "  InL "; print_list print_nat l
InR[l]  print "  InR "; print_list print_nat l
}
}

Although the above is boiler-plate code, it is provided so that the examples in this document are completely self-contained, and can be type-checked and evaluated by without any modification. We have now all the components that are required to run some tests, and to display prefixes of the streams produced by the extract function. We will display, for each example, prefixes of increasing size. We will thus rely on the following test function, taking as input a stream of natural numbers, and showing the result of applying the finite_extract function on this stream with various prefix lengths (from zero to four).

val test : streamnat  {} =
fun s {
print_res (finite_extract Zero s);
print_res (finite_extract S[Zero] s);
print_res (finite_extract S[S[Zero]] s);
print_res (finite_extract S[S[S[Zero]]] s);
print_res (finite_extract S[S[S[S[Zero]]]] s)
}

Let us first consider the output that is produced by applying the above test function to three_zeroes_then_ones, which contains three zeroes followed by infinitely many ones.

InL
InL 0
InL 0 0
InR S0 S0 S0
InR S0 S0 S0 S0

As one should expect, the computation of the smallest prefixes yields a list of even numbers. However, if more elements of the input stream are read, the extract function eventually backtracks and produces a list of odd numbers instead. Indeed, the input stream only contains three even numbers.

Note that one could expect the fourth line of the output to show a list of even numbers with three zeroes. The produced result is due to the definition of extract, which looks ahead one element further than strictly necessary in the input stream. It would be possible to avoid doing so, but the function would be even more complex than it already is. Note that this also has consequences on the result obtained by running test on naturals.

InL
InL 0
InL 0 SS0
InL 0 SS0 SSSS0
InL 0 SS0 SSSS0 SSSSSS0

Here, one would expect the prefixes to alternate between lists of even numbers, and lists of odd numbers. Indeed, the stream of all the natural numbers both contain an infinite sub-stream of even numbers, and an infinite sub-stream of odd numbers.

## 5 Termination and internal totality proofs

We will now look more deeply into the relation between proofs and termination checking in . Technically, the termination of programs (and thus of proofs) is established using circular proof techniques introduced in joint work with Christophe Raffalli [subml] and adapted in the author’s thesis [lepigrePhD]. The idea is to type recursive programs, or more precisely the fixed-point combinator used by , using a simple unfolding rule. In other words, instances of the fixed-point construction of the language are typed assuming that they can be typed, thus leading to a circular structure. Of course, proofs constructed in this way may be invalid (i.e., not well-founded). To rule out such invalid proofs, a test based on the size-change principle [scp] is used. When it is able to show that the structure of a proof is indeed well-founded, termination then follows from a standard semantic proof by realizability.181818In this case, adequacy can still be proved by well-founded induction on the structure of the typing proof.

### 5.1 Termination and consistency

As mentioned in the introduction, practical functional programming languages like OCaml or Haskell cannot be used to prove mathematical formulas, since their type system is not consistent when seen as a logic. More precisely, the “empty type” is inhabited by a simple looping program in these systems. Any formula can thus be proved through the ex falso quodlibet principle, as demonstrated by the following piece of Haskell code.191919Note that the  Rank2Types  (or  RankNTypes ) extension is required for the definition of  Empty .

type Empty = forall a. a
bad :: Empty
bad = bad
ex_falso :: Empty  a
ex_falso e = e

A similar example can also be given in OCaml, but a slightly more complex definition is required for bad.202020Note that  empty  is encoded using a polymorphic record field. This is due to the call-by-value evaluation strategy of the language, which restricts the use of the let rec construct to the definition of functions.

type empty = { any : a.’a }
let bad : empty =
let rec bad_aux : unit  empty = fun ()  bad_aux () in
bad_aux ()
let ex_falso : type a. empty  a =
fun e  e.any

Of course, a similar example can be written in . This is why the current implementation requires every program (not only proofs) to pass the termination check.

As can be used to prove program equivalences, the inconsistency that would be introduced by possible non-termination would allow proving any program equivalence by inhabiting the corresponding type. Moreover, a non-terminating program would allow invalid program equivalences to be added to the equational context. The following invalid program (first given in Section 3) gives an example of such a scenario.

// val rec add_n_Zero_loop : ∀n∈nat, add n Zero  n =
//   fun n {
//     add_n_Zero_loop n
//   }

Here, the recursive call add_n_Zero_loop n brings into the equational context the equivalence add n Zero ≡ n, which exactly corresponds to the goal of the proof.

The example add_n_Zero_loop must be rejected because the underlying program (and hence proof structure) is non terminating (and hence not well-founded). The management of equivalences being correct by construction, incorrect equations can only be proved in a contradictory equational context. In the example, a faulty equation is learned when the non-well-founded recursive call is made.

### 5.2 Sized types

The termination checking technology used by is based on a notion of size that is attached to typing judgments. In fact, inductive and coinductive types are annotated using an ordinal size indicating the number of times their definition can be unfolded (this is usual in the context of sized types [abel, hugues, sacchini13, subml]). Inductive or coinductive types, such as lists or streams, can then be seen as sized types annotated by a large enough (limit) ordinal.212121In practice, is sufficient for most of the usual data types, but this is not true in general. Nonetheless, there exists an ordinal that is large enough for all the inductive and coinductive types to converge [subml].

In practice, the implementation of

introduces sizes automatically when typing recursive functions. This means that it replaces (some) inductive and coinductive types, which carry a limit ordinal, with universal quantification over all possible ordinals. Note that it is only possible to do so when the obtained type is more general that the one that was given by the user. To enforce this invariant, we only introduce quantification on inductive types in negative position, and on coinductive types in positive position. In practice, this heuristic works well on simple functions, but the user is sometimes required to annotate the functions with explicit quantifications for termination to be established. Moreover, manual annotation may lead to more precise types, leading to more examples passing the termination check. For example, the following

map function is accepted by the implementation.

val rec map : a b, (a  b)  lista  listb =
fun fn l {
case l {
Nil     Nil
Cns[c]  Cns[{hd = fn c.hd; tl = map fn c.tl}]
}
}

However, if the user writes a complex recursive function containing a recursive call through the map function, it will not be possible to establish its termination (despite the fact that map does not change the size of the list it is applied to). To solve this problem, the user may rather use a more precise sized type, which is a subtype of the former type.

type slists:κ, a:ο⟩ = μ_s slist, [Nil ; Cns of {hd : a; tl : slist}]
val rec map : s, a b, (a  b)  slists,a  slists,b =
fun fn l {
case l {
Nil     Nil
Cns[c]  Cns[{hd = fn c.hd; tl = map fn c.tl}]
}
}

The type slist⟨s,a⟩ should not be confused with the type of vectors vec⟨a,s⟩ defined in Section 3.4. Although they are both subtypes of regular lists, the former carries an ordinal (of sort ) that can be used by the termination checker to establish size relations, while the latter contains a term (of sort ) corresponding to the size of the list (as computed by the length function) which cannot be used by the termination checker. Note however that these two types could be easily combined. The above type of map only enforces that the output list is at most as long as the input list. For instance, we could give the same type to a function taking the same two arguments and always returning an empty list. A similar scheme can be applied to insertion sort for example, but not for quick sort, as discussed in previous work [subml]. Indeed, a richer language of ordinals would be required to express the fact that the partition function preserves the number of elements of its input.222222It is nonetheless possible to show that quick sort is size-preserving using a proof.

### 5.3 Proof by equivalence to a terminating function

Although the current implementation of checks the termination of all programs (not only proofs), it is possible to use the specific features of the system to write termination proofs. Indeed, the equivalence relation on which the system relies can be used to substitute one term with another, provided that they are equivalent. This means that if we want to establish the termination of a program that does not pass the termination check directly, then we can instead establish the termination of any equivalent program. We will here consider the example of the well-known McCarthy 91 function, whose termination cannot be established by most of the existing termination criteria (if not all).

include lib.nat
include lib.nat_proofs
def mccarthy91_hard =
fix fun mccarthy91 n {
if gt n u100 {
minus n u10
} else {
mccarthy91 (mccarthy91 (add n u11))
}
}
// val mccarthy91 : nat  nat =
//   mccarthy91_hard

Note that here, the value mccarthy91_hard is not defined as a usual, type checked and termination checked value, but as a value object (using the def keyword). This means that this function can be manipulated as an object of the logic, but not evaluated directly. Note also that we rely on some functions (and constants) defined in the standard library of . The minus function computes the difference, and the gt function tests whether its first argument is strictly greater than its second argument.

Although is not able to prove the termination of the commented version of mccarthy91, we can give the following alternative (but equivalent) definition.

val mccarthy91_easy : nat  nat =
fun n {
if gt n u100 {
minus n u10
} else {
u91
}
}

This second definition passes our termination check (it is not even recursive), but it does not really correspond to the traditional definition of the McCarthy 91 function, which is a shame. We can nonetheless write a proof showing that these definitions are (pointwise) equivalent, which will then allow us to replace one with the other. To do so, we first need to show that mccarthy91_hard n has value u91 for all numbers that are not greater than u100.

val hard_aux: nnat, gt n u100  false  mccarthy91_hard n  u91 =
fun n eq {
{- ... -} // Can be done by enumerating the domain.
}

We do not give the full proof for lack of space, but it can be easily completed since the domain of quantification is finite. One simply needs to explore the domain by pattern matching on n, obtaining a trivial proof for all numbers less or equal to u100. In the case of numbers greater that u100, the presence of an additional successor produces a contradiction with the hypothesis gt n u100 ≡ false which allows the enumeration to remain finite. This brute-force approach, although it could be easily automated, yields a proof that it rather inefficient. A better solution would be to write a proof by “induction”, which is what one would do on paper.

Using the hard_aux lemma, we can then show that the two implementations of the McCarthy 91 function produce the same result on every natural number as follows.

val hard_is_easy : nnat, mccarthy91_easy n  mccarthy91_hard n =
fun n {
use gt_total n u100;
if gt n u100 {
deduce mccarthy91_easy n  minus n u10;
deduce mccarthy91_hard n  minus n u10;
qed
} else {
deduce mccarthy91_easy n  u91;
show mccarthy91_hard n  u91 using hard_aux n {};
qed
}
}

The proof is straightforward232323None of the  deduce  annotations are necessary, they are only provided for clarity. The  gt_total  lemma is defined in the standard library, more detail about its purpose will be given in the next section. since the two implementations have the same structure, and they share the same “then” branch. In the case of the “else” branch, hard_aux can be used to conclude. We can then type check and prove the termination of the original version of the McCarthy 91 function as follows.

val mccarthy91 : nat  nat =
fun n {
check mccarthy91_easy n  // Term used for type-checking.
for mccarthy91_hard n  // Actual term used in the definition.
because hard_is_easy n // Proof that they are equal.
// The above really is "mccarthy91_hard n" (up to erasure).
}

The annotation used in the definition of mccarthy91 instructs the type-checker to substitute mccarthy_hard n with mccarthy_easy n in the construction of the typing proof. This is only possible because these two terms are equivalent (when n has type nat), as witnessed by hard_is_easy n. However, the term used for the computation will indeed be mccarthy_hard n after the annotations are erased.

As all the types of are closed under equivalence, it is always possible to replace a term by another equivalent term. This technique can not only be used for proving termination of functions such as mccarthy91, but also for typing terms that would not be typable otherwise (but that are, for example, more efficient). Note that we did not prove mccarthy_hard ≡ mccarthy_easy, which may not even be true. Indeed, equivalence considers these two terms as untyped, and it is very well possible that they can be distinguished by a certain evaluation context.242424In , the equivalence  t ≡ u  being provable implies that  t  and  u  are observationally equivalent, which means that they have the same “observable behaviour” in every possible evaluation context. Note that we only observe termination, versus divergence or runtime error [lepigrePhD, lepigre2016]. A simpler example arises when comparing different implementations of the identity function on natural numbers: we have (fun n { case n { Zero ⇒ n | S[_] ⇒ n } }) k ≡ (fun n { n }) k for all k in nat, but fun n { case n { Zero ⇒ n | S[_] ⇒ n } } ≡ fun n { n } is false. These two functions can be distinguished using the argument false, which yields a pattern matching failure on the former, while the latter successfully returns false.

### 5.4 Internal totality proofs

In this last section, we will give more explanations about the so-called “totality proofs” that are currently required in . A function is said to be total if it computes some value, when applied to any value of its domain. In , the totality of functions can be expressed inside the system using an existential quantification. We can thus write internal totality proofs such as the following.

val rec add_total : n mnat, v:ι, add n m  v =
fun n m {
case n {
Zero  qed
S[k]  use add_total k m; qed
}
}

Note that the value that is obtained by applying the function is not relevant here, nor is its type. We could however modify the definition of add_total to make sure that an element of type nat is returned. In this case, we could even use add_total as add.

The reason why totality proofs are required in the system is strongly related to the call-by-value evaluation strategy of the language. Indeed, in call-by-value, a function can only be applied when all of its arguments are values. More precisely, it only makes sense to reduce a -redex if the term in argument position is a syntactic value. To understand where the notion of totality is really required, let us consider the following proof example showing the associativity of addition.

val rec add_assoc : m n pnat, add m (add n p)  add (add m n) p =
fun m n p {
use add_total n p;
case m {
Zero  qed
S[k]  use add_assoc k n p; use add_total k n; qed
}
}

Ignoring the first call to add_total, the proof starts by a case analysis on variable m. Let us consider the Zero case, which already illustrates very well the necessity for the totality proof. In this branch, the automatic decision procedure learns the equation m ≡ Zero. As a consequence, the goal simplifies to add Zero (add n p) ≡ add (add Zero n) p, and even as add Zero (add n p) ≡ add n p since both Zero and n are values (the function can thus be applied). However, the left-hand side of the equation cannot reduce further because add n p is not a value. We can then only proceed using the totality proof produced by add_total n p, which gives us a value v such that add n p ≡ v. As a consequence, this allows us to obtain add Zero (add n p) ≡ add n p as follows.

add Zero (add n p) ≡ add Zero v ≡ v ≡ add n p

It is clear that the totality proof corresponding to a given function has a similar structure as the definition of the function itself. We may thus hope that totality proofs can be generated and called automatically, at least in most cases.

## 6 Future work

The current implementation of already allows for several convincing examples, some of which cannot be expressed in other systems. They include, for example, the extract function of Section 4 or the mccarthy91 function of Section 5.3. However, theoretical work and implementation work remain to be done for the language to become fully practical, both as a programming language and as a proof assistant.

### 6.1 Mixing termination and non-termination

As mentioned earlier, termination checking is only necessary for programs that are considered as proofs. In the theory, proving that a program terminates amounts to showing that its typing derivation has a well-founded circular structure. In this case, a standard semantic proof can be used to prove normalisation [subml], the essential point being that the adequacy of the type system can be established by induction on the circular structure of proofs252525Circularity is introduced by the typing rule for the fixed-point combinator., provided that they are well-founded.

Alternatively, it is possible to type programs with standard (non-circular) typing proofs,262626This feature is not available in the current implementation, which only accepts terminating programs. to the expense of losing normalisation since our termination criterion works by analysing the circular structure of proofs. Note that lack of termination checking implies the loss of soundness, but type-safety is nonetheless preserved. As it is hard to automatically prove the termination of programs, it is clear that a user will not want to be restricted to programs that can be proved terminating. For this purpose, it is important to allow arbitrary (type-safe) programs to be written, if only to prove them terminating later (examples of such programs can be found in previous work [pml]).

As programs that can be proven terminating can be typed in both ways, it is natural to consider a way of mixing the two approaches in the theory. This has actually already been implemented in a particular branch of our implementation (called totality) [implem]. The corresponding extension of the theory has also been checked informally.

### 6.2 Other forms of effects, mutation

One of the distinguishing features of is the possibility for programs to manipulate their own continuation (or evaluation context). This is achieved using a construct similar to Scheme’s call/cc, or rather Michel Parigot’s -abstraction [parigot], which triggers a form of effect. As shown in Section 2.1, it can be used to realize theorems which only hold classically, by extracting a program from their proofs [raffalli_dickson].

Although is the first proof system based on a programming language with effects and a classical realizability model [lepigrePhD], one may argue that control structures only have a limited interest for writing practical ML programs.272727They can however be used to encode a form of exception mechanism. Other forms of effects however, for example input/output directives or mutable cells, are essential to ML programmers. Although it should be relatively easy to extend the system with the former, the latter poses a real technical challenge. Indeed, it is not yet known how to account for mutation in a classical realizability model.

### 6.3 Subject reduction and strong safety

The theory of is based on a realizability model, which has the major advantage of being flexible. More precisely, the adequacy lemma, which is the keystone of the development, only needs to be modified locally to encompass a new typing or subtyping rule. However, we have not yet proved any subject reduction result for the system, and thus we only have a weak form of type safety.

### 6.4 Extensible variants and records (better inference)

The current type-system of requires a relatively small amount of type annotations (at least for programs). Nonetheless, the system relies on unification in several places, and it may happen that the system guesses the wrong types. This situation arises most often with variant and record types, for which some fields or constructors might be left out. This problem can be solved using extensible variant types and record types, but we will need to make sure that this does not pose any problem in the theory.

### 6.5 Support for mutually recursive function

In the current implementation, lacks the possibility of defining mutually recursive functions. Although it is always possible to encode mutual recursion using additional parameters, this method does not perform very well when combined with our termination checking technology. We thus need to consider a different fixed-point instruction for our abstract machine, which does not seem to pose any theoretical problem. The idea is to replace the current fixed-point instruction with a term constructor , binding the term variable into the value , and with the reduction rule .282828Term variables should not be confused with value variables (or -variables). In particular, the former can be substituted with any term, while the latter can only be substituted with values. Mutual recursion can then be encoded using a value that is a record containing several -abstractions, which can still be typed using a simple unfolding rule (as in previous work [subml, lepigrePhD]).

### 6.6 Certificates using proof traces for equivalences

For now, it is not possible to formally check the proofs produced by in another system. Although the system already records the proof trees that are produced during type-checking, the decision procedure for program equivalence yet lacks the ability of producing a proof trace. However, there is no theoretical evidence that it would not be possible for the decision procedure to record enough information for an external prover (for example Coq [coq] or Dedukti [dedukti]) to check the proofs produced by .

## 7 Similar systems

To conclude this paper, we will compare to other proof systems and languages that can be used to formalise and prove program properties, or that rely on similar principles.

### 7.1 Dependent types in ML

To our knowledge, the combination of call-by-value evaluation, side-effects and dependent products has never been achieved before. At least not for a dependent product fully compatible with effects and call-by-value. For example, the Aura language [aura] forbids dependency on terms that are not values in dependent applications. Similarly, the language [fstar] relies on (partial) let-normal forms to enforce values in argument position. Daniel Licata and Robert Harper have defined a notion of positively dependent types [licata] which only allow dependency over strictly positive types. Finally, in languages like ATS and DML [ats, dml], dependencies are limited to a specific index language.

### 7.2 Tools based on intuitionistic type theory

The most actively developed proof assistants following the Curry-Howard correspondence are Agda and Coq [agda, coq]. The former is based on Martin-Löf’s dependent type theory and the latter on Coquand and Huet’s calculus of constructions [coc, mltt]. These two constructive theories provide dependent types, which allow the definition of very expressive specifications. Contrary to , Coq and Agda do not directly give a computational interpretation to classical logic. Classical reasoning can only be done through a negative translation or with the definition of axioms such as the law of the excluded middle. In particular, these two languages are not effectful. However, they are logically consistent, which means that they only accept terminating programs. As termination checking is a difficult (and undecidable) problem, many terminating programs are rejected. Although this is not a problem for formalizing mathematics, this makes programming tedious. In , only proofs really need to be shown terminating, and it is in any case possible to reason about non-terminating and even untyped programs as they can be manipulated as objects in types.

### 7.3 NuPrl and refinement types

The NuPrl system [nuprl] has many similarities with on the theoretical side, although it is inconsistent with classical logic. NuPrl accommodates an observational equivalence relation similar to ours (Howe’s squiggle relation [howe]), which is partially reflected in the syntax of the system. Being based on a Kleene-style realizability model, NuPrl can also be used to reason about untyped terms. Another major difference between and NuPrl is that the latter is based on refinement types, which means that it does not have an automatic way of building typing derivations for programs. Indeed, typing derivations are built interactively using a specific interface, and the user must say what typing rule should be applied first.

### 7.4 Partially consistent languages

The TRELLYS project [trellys] aims at providing a language in which a consistent core interacts with type-safe dependently typed programming with general recursion. Although the language is call-by-value and effectful, it suffers from value restriction like Aura [aura]. The value restriction does not appear explicitly but is encoded into a well-formedness judgement appearing as the premise of the typing rule for application. Apart from value restriction, the main difference between the language of the TRELLYS project and ours resides in the calculus itself. Their calculus is Church-style (or explicitly typed) while ours is Curry-style (or implicitly typed). In particular, their terms and types are defined simultaneously, while our type system is constructed on top of an untyped calculus.

### 7.5 Systems aimed at program verification in ML

Several systems have been proposed for verifying ML programs. ProPre [propre] relies on a notion of algorithms, corresponding to equational specifications of programs. It is used in conjunction with a type system based on intuitionistic logic. Although it is possible to use classical logic to prove that a program meets its specification, the underlying programming language is not effectful. Similarly, the PAF! system [baro] implements a logic supporting proofs of programs, but it is restricted to a purely functional subset of ML. Another approach for reasoning about purely functional ML programs is given in the work of Yann Regis-Gianas [regisgianas], where Hoare logic is used to specify program properties. Finally, it is also possible to reason about ML programs (including effectful ones) by compiling them down to higher-order formulas [chargueraud10, chargueraud11], which can then be manipulated using an external prover such as Coq [coq]. In this case, the user is required to master at least two languages, contrary to our system in which programming and proving take place in a uniform framework.