Associated Type Synonyms

Manuel M. T. Chakravarty

Gabriele Keller

University of New South Wales

Programming Languages and Systems
{chak,keller}@cse.unsw.edu.au

Abstract
Haskell programmers often use a multi-parameter type class in
which one or more type parameters are functionally dependent on
the first. Although such functional dependencies have proved quite
popular in practice, they express the programmers intent somewhat
indirectly. Developing earlier work on associated data types, we
propose to add functionally-dependent types as type synonyms to
type-class bodies. These associated type synonyms constitute an
interesting new alternative to explicit functional dependencies.
Categories and Subject Descriptors D.3.3 [Programming Languages]: Language Constructs and Features
General Terms Design, Languages, Theory
Keywords Type classes; Type functions; Associated types; Type
inference; Generic programming

1. Introduction
Suppose you want to define a family of containers, where the representation type of the container defines (or constrains) the type of its
elements. For example, suppose we want containers supporting at
least insertion, union, and a membership test. Then a list can contain elements of any type supporting equality; a balanced tree can
only contain elements that have an ordering; and a bit-set might
represent a collection of characters. Here is a rather natural type for
the insertion function over such collections:
insert :: Collects c Elem c c c
The type class Collects says that insert is overloaded: it will work
on a variety of collection types c, namely those types for which
the programmer writes an instance declaration for Collects. But
what is Elem? The intent is obviously that Elem c is the element
type for collection type c; you can think of Elem as a typelevel function that transforms the collection type to the element
type. However, just as insert is non-parametric (its implementation
varies depending on c), so is Elem. For example, Elem [e] is e, but
Elem BitSet is Char .
The core idea of this paper is to extend traditional Haskell type
classes with the ability to define associated type synonyms. In our
example, we might define Collects like this:

Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. To copy otherwise, to republish, to post on servers or to redistribute
to lists, requires prior specific permission and/or a fee.
ICFP05 September 2628, 2005, Tallinn, Estonia.
c 2005 ACM 1-59593-064-7/05/0009. . . $5.00.
Copyright

Simon Peyton Jones

Microsoft Research Ltd.
Cambridge, UK
[email protected]

class Collects c where

type Elem c
-- Associated type synonym
empty :: c
insert :: Elem c c c
toList :: c [Elem c]
The type definition says that c has an associated type Elem c,
without saying what that type is. This associated type may then
be used freely in the types of the class methods. An instance
declaration gives an implementation for Elem, just as it gives an
implementation for each method. For example:
instance Eq e Collects [e] where
{type Elem [e] = e; . . .}
instance Collects BitSet where
{type Elem BitSet = Char ; . . .}
instance (Collects c, Hashable (Elem c))
Collects (Array Int c) where
{type Elem (Array Int c) = Elem c; . . .}
Haskell aficionados will recognise that associated type synonyms
attack exactly the same problem as functional dependencies, introduced to Haskell by Mark Jones five years ago [15], and widely
used since then in surprisingly varied ways, many involving typelevel computation. We discuss the relative strengths of the two approaches in detail in Section 6. It is too early to say which is better; our goal here is only to describe and characterise a new point
in the design space of type classes.
Specifically, our contributions are these:
We explore the utility and semantics of type synonym declara-

tions in type classes (Section 2).

We discuss the syntactic constraints necessary to keep type in-

ference in the presence of associated type synonyms decidable

(Section 3).
We give a type system that supports associated type synonyms
and allows an evidence translation to an explicitly typed core
language in the style of System F (Section 4).
We present a type inference algorithm that can handle the nonsyntactic equalities arising from associated type synonyms; the
algorithm conservatively extends Jones algorithm for qualified
types [12] (Section 5).
This paper is a natural development of, and is complementary to,
our earlier work on associated data types [1], in which we allow
a class declaration to define new algebraic data types. We discuss
other related type systemsin particular, functional dependencies,
HM(X), and ML modulesin detail in Sections 6 and 7. We developed a prototype implementation of the type checker, which we
make available online.1
1 http://www.cse.unsw.edu.au/chak/papers/CKP05.html

2. Applications of Associated Type Synonyms

We begin informally, by giving several examples that motivate
associated type synonyms, and show what can be done with them.
2.1

Formatting: type functions compute function types

The implementation of a string formatting function whose type

depends on a format specifier seems a natural application for dependent types and meta programming [26]. Although Danvy [4]
demonstrated that Standard MLs type system is powerful enough
to solve this problem, type functions enable a more direct solution [10], using an inductive definition instead of explicit continuation passing style. The following implementation with associated
synonyms is based on [22]. Format specifiers are realised as singleton types:2
data I f = I f
-- Integer value
data C f = C f
-- Character value
data S f = S String f -- Literal string
formatSpec :: S (I (S (C String)))
formatSpec = S "Int: " $ I $ S ", Char: " $ C $ "."
-- Example format: "Int: %d, Char: %c."
The singleton type declarations reflect the structure of a format
specifier value in their type. Consequently, we can use a specifiers type to calculate an appropriate type for a sprintf function
applied to that specifier. We implement this type level calculation
by defining an associated synonym Sprintf in a class Format in
the following way:
class Format fmt where
type Sprintf fmt
sprintf :: String fmt Sprintf fmt
instance Format String where
type Sprintf String = String
sprintf prefix str = prefix ++ str
instance Format a Format (I a) where
type Sprintf (I a) = Int Sprintf a
sprintf prefix (I a) = i. sprintf (prefix ++ show i) a
instance Format a Format (C a) where
type Sprintf (C a) = Char Sprintf a
sprintf prefix (C a) = c. sprintf (prefix ++ [c]) a
instance Format a Sprintf (S a) where
type Sprintf (S a)
= Sprintf a
sprintf prefix (S str a) = sprintf (prefix ++ str ) a
sprintf :: Format fmt fmt Sprintf fmt
sprintf = sprintf ""
New format-specifier types (such as I and S above) can be introduced by the programmer at any time, simply by defining the type,
and giving a matching instance declaration; that is, the definition of
sprintf is open, or extensible.
Notice how important it is that the associated type is a synonym:
it is essential that Sprintf fmt is a function type, not a data type.
2.2

Generic data structures

The collections abstraction Collects from Section 1 is an example of a generic data structureothers include sequences, graphs,
and so on. Several very successful C++ libraries, such as the Standard Template Library [29] and the Boost Graph Library [28], provide highly-parameterised interfaces to these generic data structures, along with a wide range of implementations of these interfaces with different performance characteristics. Recently, Garcia
et al. [8] published a qualitative comparison of six programming
2 The infix operator f x in Haskell is function application f x at a lesser
$
precedence.

languages when used for this style of programming. In their comparison Haskell, including multi-parameter type classes and functional dependencies, was rated very favourably, except for its lack
of support for associated types.
Here is part of the interface to a graph library, inspired by their
paper; although, we have simplified it considerably:
type Edge g = (Node g, Node g)
-- We simplify by fixing the edge representation
class Graph g where
type Node g
outEdges :: Node g g [Edge g]
class Graph g BiGraph g where
inEdges :: Node g g [Edge g]
Using an associated type synonym, we can make the type of nodes,
Node g, a function of the graph type g. Basic graphs only support
traversals along outgoing edges, whereas bi-graphs also support going backwards by following incoming edges. A graph representation based on adjacency lists would only implement the basic interface, whereas one based on an adjacency matrix can easily support
the bi-graph interface, as the following instances illustrate:
data AdjList v = AdjList [[v ]]
instance Enum v Graph (AdjList v ) where
type Node (AdjList v ) = v
outEdges v g = [(v , w ) | w g!!fromEnum v ]
type AdjMat = Array.Array (Int, Int) Bool
instance Graph AdjMat where
type Node AdjMat = Int
outEdges v g = let ((from, ), (to, )) = bounds g
in [w | w [from..to], g!(v , w )]
instance BiGraph AdjMat where
inEdges v g = let ((from, ), (to, )) = bounds g
in [w | w [from..to], g!(w , v )]
By making Edge, as well as Node, an associated type synonym
of Graph and by parameterising over traversals and the data structures used to maintain state during traversals, the above class can
be made even more flexible, much as the Boost Graph Library, or
the skeleton used as a running example by Garcia et al. [8].

3. The programmers-eye view

In this section, we give a programmers-eye view of the proposed
language extension. Formal details follow later, in Section 4.
We propose that a type class may declare, in addition to a set of
methods, a set of associated type synonyms. The declaration head
alone is sufficient, but optionally a default definitionmuch like
those for methodsmay be provided. If no default definition is
given, an optional kind signature may be used; otherwise, the result
kind of a synonym application is assumed to be . An associated
type synonym must be parametrised over all the type variables of
the class, and these type variables must come first, and be in the
same order as the class type variables.
Each associated type synonym introduces a new top-level type
constructor. The kind of the type constructor is inferred as usual in
Haskell; we also allow explicit kind signatures on type parameters:
class C a where
type S a (k :: ) ::
Instance declarations must give a definition for each associated type
synonym of the class, unless the synonym has been given a default
definition in the class declaration. The definition in an instance
declaration looks like this:
instance C [a] where
type S [a] k = (a, k a)

The part to the left of the = is called the definition head. The
head must repeat the type parameters of the instance declaration
exactly (here [a]); and any additional parameters of the synonym
must be simply type variables (k , in our example). The overall
number of parameters, called the synonyms arity, must be the
same as in the class declaration. All applications of associated
type synonyms must be saturated; i.e., supplied with as many type
arguments as prescribed by their arity.
We omit here the discussion of toplevel data type declarations
involving associated types, as we covered these in detail previously [1]. In all syntactic restrictions in this section, we assume
that any toplevel type synonyms have already been replaced by their
right-hand sides.
3.1

Equality constraints

Suppose we want to write a function sumColl that adds up the

elements of a collection with integer elements. It cannot have type
sumColl :: (Collects c) c Int
-- Wrong!
sumColl c = sum (toList c)
because not all collections have Int elements. We need to constrain
c to range only over Int-element collections. The way to achieve
this is to use an equality constraint:
sumColl :: (Collects c, Elem c = Int) c Int
sumColl c = sum (toList c)
As another example, suppose we wanted to merge two collections,
perhaps with different representations, but with the same element
type. Then again, we need an equality constraint:
merge :: (Collects c1, Collects c2, Elem c1 = Elem c2)
c1 c2 c2
merge c1 c2 = foldr insert c2 (toList c1)
Without loss of generality, we define an equality constraint to
have the form (S = ), where S is an associated type synonym, are as many type variables as the associated class has
parameters, and the and are arbitrary monotypes. There is
no need for greater generality than this; for example, the constraint ([S a] = [Int]) is equivalent to (S a = Int); the constraint ([S a] = Bool ) is unsatisfiable; and (a = Int) can be
eliminated by replacing a by Int. These restrictions are stronger
than they would have to be. However, they allow us later on to
characterise well-formed programs on a purely syntactical level.
3.2

Constraints for associated type synonyms

Does this type signature make sense?

funnyFst :: (Elem c, c) Elem c
Recall that Elem is a partial function at the type level, whose
domain is determined by the set of instances of Collects. So it
only makes sense to apply funnyFst at a type that is an instance of
Collects. Hence, we reject the signature, requiring you to write
funnyFst :: Collects c (Elem c, c) Elem c
to constrain the types at which funnyFst can be called. More precisely, each use of an associated type synonym in a programmerwritten type signature gives rise to a class constraint for its associated class; and that constraint must be satisfied by the context of
the type signature, or by an instance declaration, or a combination
of the two. This validity check for programmer-supplied type annotations is conveniently performed as part of the kind checking
of these annotations, as we will see in Section 4. Kind checking is
only required for programmer-supplied type annotations, because
inferred types will be well-kinded by construction.
3.3

Instance declarations

Given that associated type synonyms amount to functions on types,

we need to restrict their definitions so that type checking remains

tractable. In particular, they must be confluent; i.e., if a type expression can be reduced in two different ways, there must be further
reduction steps that join the two different reducts again. Moreover,
type functions must be terminating; i.e., applications must reach
an irreducible normal form after a finite number of reduction steps.
The first condition, confluence, is already standard on the level of
values, but the second, termination, is a consequence of the desire
to keep type checking decidable.
Similar requirements arise already for vanilla type classes as
part of a process known as context reduction. In a declaration
instance (1 , . . . , n ) C 1 m
we call C 1 m the instance head and (1 , . . . , n ) the instance context, where each i is itself a class constraint. Such an
instance declaration implies a context reduction rule that replaces
the instance head by the instance context. The critical point is that
the constraints i can directly or indirectly trigger other context
reduction rules that produce constraints involving C again. Hence,
we have recursive reduction rules and the same issues of confluence
and termination as for associated type synonyms arise. Haskell 98
carefully restricts the formation rules for instance declarations such
that the implied context reduction rules are confluent and terminating. It turns out, that we can use the same restrictions to ensure
these properties for associated type synonyms. In the following, we
discuss these restrictions, but go beyond Haskell 98 by allowing
multi-parameter type classes. We will also see how the standard
formation rules for instances affect the type functions induced by
associated synonym definitions.
Restrictions on instance heads. Haskell 98 imposes the following three restrictions. Restriction (1): Heads must be constructorbased; i.e., the type patterns in the head may only contain variables
and data type constructors, synonyms are not permitted. Restriction (2): Heads must be specific; i.e., at least one type parameter
must be a non-variable term. Restriction (3): Heads must be nonoverlapping; i.e., there may be no two declarations whose heads
are unifiable.
Given that the heads of associated synonyms must repeat the
type parameters of the instance head exactly, the above three restrictions directly translate to associated synonyms. Restriction (1)
is familiar from the value level, and we will discuss Restriction (2)
a little later. The value level avoids Restriction (3) by defining that
the selection of equations proceeds in textual order (i.e., if two
equations overlap, the textually earlier takes precedence). However,
there is no clear notion of textual order for instance declarations,
which may be spread over multiple modules.
Restrictions on instance contexts. Haskell 98 imposes one more
restriction. Restriction (4): Instance contexts must be decreasing.
More specifically, Haskell 98 requires that the parameters of the
constraints i occurring in an instance context are variables. If
we have multi-parameter type classes, we need to further require
that these variable parameters of a single constraint are distinct.
Restriction (4) and (2) work together to guarantee that each context
reduction rule simplifies at least one type parameter. As type terms
are finite, this guarantees termination of context reduction.
In the presence of associated types, we generalise Restriction (4) slightly. Assuming 1 , . . . , n are each either a type variable or an associated type applied to type variables, a context constraint i can either be a class constraint of the form D 1 n or
be an equality constraint of the form S 1 m = .
The right-hand sides of the associated type synonyms of an instance are indirectly constrained by Restriction (4), as they may
only contain applications of synonyms whose associated class appears in the instance context. So, if we have
instance (1 , . . . , n ) C where
type SC = [SD ]

and SD is associated with class D, then one of the i must be

D . In other words, as a consequence of the instance context restriction, associated synonym applications must have parameters
that are either distinct variables or other synonyms applied to variables. Hence, the reduction of associated synonym applications terminates for the same reason that context reduction terminates.
This might seem a little restrictive, but is in fact sufficient for
most applications. Strictly speaking we, and Haskell 98, could be
a bit more permissive and allow that if there are n occurrences
of data type constructors in the type parameters of an instance
head, each constraint in the instance context may have up to n 1
occurrences of data type constructors in its arguments. Moreover,
we may permit repeated variable occurrences if the type checker
terminates once it sees the same constraint twice in one reduction
chain. Work on term rewriting system (TRS) [19] has identified
many possible characterisations of systems that are guaranteed to
be confluent and terminating, but the restrictions stated above seem
to be a particularly good match for a functional language.
3.4

Ambiguity

This celebrated function has an ambiguous type:

echo :: (Read a, Show a) String String
echo s = show (read s)
The trouble is that neither argument nor result type mention a, so
any call to echo will give rise to the constraints (Read a, Show a),
with no way to resolve a. Since the meaning of the program depends on this resolution, Haskell 98 requires that the definition is
rejected as ambiguous.
The situation is much fuzzier when functional dependencies [15] are involved. Consider
class C a b | a b where
...
poss :: (C a b, Eq b) a a
Is the type of poss ambiguous? It looks like it, because b is not
mentioned in the type after the . However, because of the
functional dependency, fixing a will fix b, so all is well. But
the dependency may not be so obvious. Suppose class D has no
functional dependency, but it has an instance declaration like this:
class D p q where
...
instance C a b D [a] b where
...
poss2 :: (D a b, Eq b) a a
Now, suppose poss2 is applied to a list of integers. The call will
give rise to a constraint (D [Int] t), which can be simplified by the
instance declaration to a constraint (C Int t). Now the functional
dependency for C will fix t, and no ambiguity arises. In short, some
calls to poss2 may be ambiguous, but some may not.
It does no harm to delay reporting ambiguity. No unsoundness
arises from allowing even echo to be defined; but, in the case of
echo, every single call will result in an ambiguity error, so it is
better to reject echo at its definition. When the situation is less
clear-cut, it does no harm to accept the type signature, and report
errors at the call site. Indeed, the only reason to check types for
ambiguity at all is to emit error messages for unconditionallyambiguous functions at their definition rather than at their use.
Associated type synonyms are easier, because functionallydependent types are not named with a separate type variable. Here
is how class C and poss would be written using an associated
synonym S instead of a functionally-dependent type parameter:
class C a where
type S a
poss :: (C a, Eq (S a)) a a

So, just as in Haskell 98, a type is unconditionally ambiguous if

one of its constraints mentions no type variable that is free in the
value part of the type.
There is a wrinkle, however. Consider this function:
poss3 :: (C a) S a S a
It looks unambiguous, since a is mentioned in the value part of
the type, but it is actually unconditionally ambiguous. Suppose we
apply poss3 to an argument of type Int. Can we deduce what a
is? By no means! There might be many instance declarations for C
that all implement S as Int:
instance ... C where
type S = Int
(In fact, this situation is not new. Even in Haskell 98 we can have
degenerate type synonyms such as
type S a = Bool
which would render poss3 ambiguous.)
The conclusion is this: When computing unconditional ambiguityto emit earlier and more-informative error messageswe
should ignore type variables that occur under an associated type
synonym. For poss3, this means that we ignore the a in S a, and
hence, there is no occurrence of a left to the right of the double
arrow, which renders the signature unconditionally ambiguous. An
important special case is that of class method signatures: Each
method must mention the class variable somewhere that is not
under an associated synonym. For example, this declaration defines
an unconditionally-ambiguous method op, and is rejected:
class C a where
type S a
op :: S a Int

4. The Type System

In this section, we formalise a type system for a lambda calculus
including type classes with associated type synonyms. This type
system is based on Jones Overloaded ML (OML) [13, 14] and
is related to our earlier system for associated data types [1]. Like
Haskell 98 [9, 7], our typing rules can be extended to give a typedirected translation of source programs into an explicitly-typed
lambda calculus akin to the predicative fragment of System F. We
omit these extended rules here, as the extension closely follows our
earlier work on associated data types [1].
The key difference between type checking in the presence of
associated data types compared to associated type synonyms is
the treatment of type equality. In the conventional Damas-Milner
system as well as in its extension by type classes with associated
data types, type equality is purely syntactici.e, types are equal
iff they are represented by the same term. When we add associated
type synonyms, type equality becomes more subtle. More precisely,
the equations defining associated type synonyms in class instances
refine type equality by introducing non-free functions over types.
The treatment of this richer notion of equality in a Damas-Milner
system with type classes during type checking and type inference
constitutes the main technical contribution of this paper.
4.1 Syntax
The syntax of the source language is given in Figure 1. We use
overbar notation extensively. The notation n means the sequence
1 n ; the n may be omitted when it is unimportant. Moreover, we use comma to mean sequence extension as follows:
an , an+1 , an+1 . Although we give the syntax of qualified and
quantified types and constraints in an uncurried way, we also some-

Symbol Classes
, , htype variablei
T
htype constructori
Sk
hassociated type synonym, arity ki
D
htype classi
x, f, d hterm variablei
Source declarations
(whole program)
pgm
cls; inst; val
cls
class .D D where (class decl)
tsig; vsig
inst
instance . where (instance declaration)
atype; val
val
x = e
(value binding)
(assoc. type signature)
tsig
type S k k
vsig
x ::
(method signature)
(k1)
= (assoc. type synonym)
atype type S k
Source terms
e x | e1 e2 | x.e | let x = e1 in e2 | e ::
Source types
, T | | 1 2 |

Sk k

.
Constraints
c D
= =

c | =

. | . =

(monotype)
(associated type)
(qualified type)
(type scheme)
(class constraint)
(equality constraint)
(simple constraint)
(qualified constraint)
(constraint scheme)

Environments
x : (type environment)

(instance environment)
U =
(set of equality constraints)
Figure 1: Syntax of expressions and types
times use equivalent curried notation, thus:
n
n
n .

1 n
1 n
1 n .

We accommodate function types 1 2 by regarding them

as the curried application of the function type constructor to two
arguments, thus () 1 2 . We use (Fv x ) to denote the free
variables of a structure x, which maybe an expression, type term,
or environment.
The unusual features of the source language all concern associated type synonyms. A class declaration may contain type
declarations as well as method signatures, and correspondingly an
instance declaration may contain type definitions as well as
method implementations. These type synonyms are the associated
type synonyms of the class, and are syntactically distinguished: S is
an associated type synonym constructor, while T is a regular type
constructor (such as lists or pairs). In the declaration of an associated type synonym, the type indexes come first. The arity of a type
synonym is the number of arguments given in its defining tsig. The
arity is given by a superscript to the constructor name, but we drop
it when it is clear from the context. The syntax of types includes

, the saturated application of an associated type. Note that such a

saturated application can be of higher kind, if the result kind in
the defining tsig is not .
In the syntax of Figure 1, and in the following typing rules, we
make two simplifying assumptions to reduce the notational burden:
1. Each class has exactly one type parameter, one method, and one
associated type synonym.
2. There are no default definitions, neither for methods nor synonyms. A program with default definitions can be rewritten into
one without, by duplicating the defaults at instances not providing their own versions.
3. We elide all mention of kinds, as we exactly follow Jones
system of constructor classes [14].
Lifting the first restriction is largely a matter of adding (a great
many) overbars to the typing rules.
4.2 Type checking
Figures 2 and 3 present the typing rules for our type system. Our
formalisation is similar to [9] in that we maintain the context reduction rules as part of the instance environment. The main judgement
has the conventional form | e : meaning in type environment , and instance environment , the term e has type .
Declarations are typed by Figure 3, where all the rules are standard
for Haskell, except for Rule (inst).
The instance environment is a set of constraint schemes
that hold in e. A constraint scheme takes one of two forms
(Figure 1): it is either a class constraint scheme ., or an equality
scheme . = . The instance environment is populated firstly by
class and instance declarations, which generate constraint schemes
using the rules of Figure 3; and secondly by moving underneath a
qualified type (rule (I ) of Figure 2). The latter adds only a simple
constraint , which can be a class constraint c or an equality
constraint = ; these simple constraints are special cases of the two
forms described earlier3 .
The typing rules are almost as for vanilla Haskell 98, with two
major differences. The first is in the side conditions that
check the well-formedness of types, in rules (I ) and (E ), for
reasons we discussed in Section 3.2. The rules for this judgement
are also in Figure 2. The judgement needs the instance environment so that it can check the well-formedness of applications of
associated-type applications (Rule (wfsyn )), using the entailment
judgement described below. In the interest of brevity, the
presented rules elide all mention of kinds, leaving only the wellformedness check that is distinctive to a system including associated types. More details concerning the implications of the judgement are in our previous work [1]. It is worthwhile to note
that Rule (sig) does not contain well-formedness judgement, although it mentions a user-supplied type. This type is also produced
by the typing judgement in the hypothesis, and the judgement always produces well-formed types. So, the rule will never apply to
a malformed type.
The second major difference is Rule (conv ), which permits type
conversion between types 1 and 2 if we have 1 = 2 .
The auxiliary judgement defines constraint entailment,
which in Haskell 98 only handles type classes, but which we extend
here with additional rules for type equality constraints. These rules
are also given in Figure 2 and include the four standard equality
axioms (eqrefl ), (eqsymm ), (eqtrans ), and (eqsubst ). The last of
these allows equal types to be wrapped in an arbitrary context: for
example, if 2 = 3 , then Tree (List 2 ) = Tree (List 3 ).
3 Rule (I ), and the syntax of types, does not allow one to quantify over
constraint schemes, an orthogonal and interesting possible extension[11].

|e:
(x : )
(var )
|x:

1 | e1 : 1
2 | [x : 1 ] e2 : 2
(let)
1 , 2 | let x = e1 in e2 : 2

| e1 : 2 1
| e2 : 2
(E)
| e1 e2 : 1

| [x : 1 ] e2 : 2
1
(I)
| x.e2 : 1 2

|e:

(E)
|e:

, | e :
(I)
|e:
|e:
6 Fv Fv
(I)
| e : .

| e : 1
1 = 2
(conv )
| e : 2

| e : .

(E)
| e : [ /]

|e:
Fv =
(sig)
| (e :: ) :

(mono)

.
(spec)
[ /]

(mp)

2 = 1
(eqsymm )
1 = 2

(eqrefl )

[1 /]
1 = 2
(eqsubst )
[2 /]

1 = 2
2 = 3
(eqtrans )
1 = 3

D 0

S is an associated type of D

S k 0 (k1)

(wfsyn )

1
2
(wfapp )
1 2

6 Fv
(wfspec )
.

(wfvar )

,
(wfmp )

(wfcon )

Figure 2: Typing rules for expressions

| cls : ,
, D
(cls)
n
class .D D where
n
| type S ::
: [.D D ], [f : .D ]
f ::

c , i | inst :
for each D where (.D D ) c , we have c , i . D
c ,

(f : .D )

c , i , | e : [ /]

instance . D where
: [. D , .S = ]
c , i | type S =
f =e

| val :
|e:
(val )
| (x = e) : [x : ]

(inst)

pgm
= c , i
| cls : c , c

= c , v

| inst : i
cls; inst; val

Figure 3: Typing rules for declarations

| val : v

Much of this is standard for extensions of the Hindley-Milner

system with non-syntactic type equality [24, 22, 21, 23, 32, 2].
Novel in our system is the integration of entailment of type class
predicates with type equalities. In particular, our system allows
equality schemes, such as
a. Sprintf (I a) = Int Sprintf a
in the constraint context of the typing rules. Equality schemes are
introduced by Rule (inst) of the declaration typing rules from Figure 3. This rule turns the definitions of associated type synonyms
into equality schemes, such as the Sprintf scheme above, which
are used in the inference rules of judgement from Figure 2.
An important property of the system is that the well-formedness
of types is invariant under the expansion of associated type synonyms. Specifically, Rule (inst) ensures that each equality scheme
.S = is paired with a class constraint scheme .
D , where D is the class S is associated with. Hence, all the
premises for the validity of are fulfilled whenever the equality
is applicable. Rule (inst) ensures this with the premise c , .
Let us consider an example: The Format class of Section 2.1 has
an instance
instance Format a Format (I a) where
type Sprintf (I a) = Int Sprintf a
Hence, Rule (inst) adds the two constraint schemes
C = a.Format a Format (I a)
= = a. Sprintf (I a) = Int Sprinft a
to the context . Now consider an expression e of type
Format (I String) Char Sprintf (I String)
Note that the context Format (I String) is required for the type
to be well-formed according to judgement . In order to
remove the constraint Format (I String) from the type of e by
Rule (E ), we need to have F ormat (I String), which
according to C and Rule (mp) requires F ormat String.
The result of applying = to Char Sprintf (I String),
i.e., Char Int Sprintf String, is well-formed only if
F ormat String holds. That it holds is enforced by C .
There are two more noteworthy points about Rule (inst).
Firstly, the rule checks that the superclass constraints of the class
D are fulfilled in the first premise. Secondly, when checking the
validity of the right-hand side of the associated synonym, namely ,
we assume only c (the superclass environment). If we would add
i , we would violate Restriction (4) of Section 3.3 and potentially
allow non-terminating synonym definitions.

5. Type Inference
The addition of Rule (conv ) to type expressions and of equality
axioms to constraint entailment has a significant impact on type
inference. Firstly, we need to normalise type terms involving associated types according to the equations defining associated types.
Secondly, type terms involving type variables often cannot be completely normalised until some type variables are further instantiated. Consequently, we need to extend the standard definition of
unification in the Hindley-Milner system to return partially-solved
equality constraints in addition to a substitution.
To illustrate this, reconsider the definition of the class Collects
with associated type Elem from Section 1. The unification problem
(Int, a) = (Elem c, Bool ) implies the substitution [Bool/a],
but also the additional constraint Int = Elem c. We cannot decide whether the latter constraint is valid without knowing more
about c, so we call such constraints pending equality constraints. In
the presence of associated types, the type inference algorithm has
to maintain pending equality constraints together with class predicates.

In this section, after fixing some constraints on source programs,

we will first discuss a variant of Hindley-Milner type inference with
predicates. The inference algorithm depends on a unification procedure that includes type normalisation (Section 5.3). Afterwards,
we will use the unification procedure to test the subsumption relation of type schemes (Section 5.4). We conclude the section with
some formal properties of our inference system for associated type
synonyms.
5.1 Well-formed programs
To achieve decidable type inference computing principal types,
we impose the following restrictions on the source programs as
produced by Figure 1 (cf., Restrictions (1) to (4) in Section 3):
Instance heads must be constructor-based, specific, and non-

overlapping.
Instance contexts must be decreasing.

To guarantee a coherent (i.e., unambiguous) translation of welltyped source programs to an explicitly-typed language in the style
of System F, which implements type classes by dictionaries, we
require two further constraints:
Equality constraints (in programmer-supplied type annotations)

must be of the form S = (i.e., the first argument must be

a type variable).
If (. ) is a method signature in a class declaration
for D , we require that 6 Fv .
(These restrictions are not necessary if the semantics of the program is given by translation to an untyped intermediate language,
or perhaps one with a richer type system than System F; see for example [30].) Finally, to stay with Haskells tradition of rejecting
unconditionally-ambiguous type signatures, in the sense of Section 3.4, we require two more constraints:
If (. ) is a method signature in a class declaration

for D , we require that Fixv .

Similarly, we require of all signatures e :: . that Fv

Fixv .
Here Fixv , the set of fixed variables of a signature , is defined
as follows:
Fixv T
= {}
Fixv
= {}
Fixv (1 2 )
= Fixv 1 Fixv 2
Fixv
= {}
Fixv (( = ) ) = Fixv Fixv
Fixv (D )
= Fixv
Fixv (.)
= Fixv \ {}
Intuitively, the fixed variables of a type (or signature) are those
free variables that will be constrained when we unify the type (or
type component of the signature) with a ground type; provided it
matches.
A program that meets all of the listed constraints is well-formed.
The declaration typing rules of Figure 3 determine, for a given program, the validity of a program context P and typing environment
. Both are inputs to type inference for expressions and, if they are
without superfluous elements, we call them well-formed. As in the
typing rules, we implicitly assume all types to be well-kinded. It
is straight-forward to add the corresponding judgements to the presented rules.
We call an expression well-formed if any signature annotation
of the form (e :: ) obeys the listed constraints. For the rest of
the paper, we confine ourselves to well-formed programs, program
contexts, typing environments, and expressions. This in particular

means that the rewrite system implied by the equality schemes in a

program context P is confluent and terminating.
5.2

Type inference

We begin with type inference for type classes with associated

types. Figure 4 displays the rules for the inference judgement
W

, U | T e : . Given a type environment and an expression

e as inputs, it computes the following outputs (1) a set of class
constraints , (2) a set of pending equality constraints U , (3) a
substitution T , and (4) a monotype . The judgement implicitly
also depends on the program context P , initially populated by
the instance declarations and remains constant thereafter. Because
class and instance declarations are, in effect, explicitly typed,
their type inference rules are extremely similar to those in Figure 3;
so, we do not give them separately here.
Our inference rules generalise the standard rules for Haskell
in two ways: (1) the inference rules maintain a set of equality
constraints U and (2) unification produces a set of pending equality
constraints in addition to a substitution. Let us look at both aspects
in some more detail.
Type inference for Haskell needs to maintain a set of of constraints, called the constraint context. In Haskell 98, the constraint
context is a set of type class predicates. We extend the context to
also include equality constraints, just as in the typing rules. However, as we need to treat these two kinds of constraints differently
during the inference process, we partition the overall context into
the two subsets and U denoting the class constraints and equality
constraints, respectively. Hence, the phrase , U to the left of the
vertical bar in the inference judgement captures the whole context.
In particular, Rules (varW ) and (sigW ) result in contexts [/],
which are implicitly partitioned into the two components and U
in the hypotheses of other inference rules. During generalisation
by the function Gen(; ) in Rule (letW ) and (sigW ), both class
and equality constraints are moved into the signature context of the
produced signature.
In Rule (EW ), the two sets of equality constraints U1 and
U2 are both extracted out of the constraint context, to be fed to the
unification process in conjunction with the new equality constraint
T2 1 = 2 . Unification simplifies the equalities as far as
possible, producing, in addition to a substitution R, a set of pending
equality constraints U . This set is used in the conclusion of the
inference rule.
Rules (EW ) and (sigW ) make use of two auxiliary judgements for unification and subsumption, respectively. Both of these
judgements depend on the program context P and are the subject
of the following two subsections.
5.3

Unification

Associated type synonyms extend Haskell 98s purely syntactic

type equality; e.g, Elem BitSet = Char holds under the definitions for the class Collects discussed before. To handle nonsyntactic equalities during unification, we exploit the properties of
well-formed programs. In particular, the type functions in wellformed programs are confluent and terminating; i.e., type terms
have unique normal forms. Hence, we can determine whether two
type terms are equal by syntactically comparing their normal forms.
The judgement
, whose inference rules are given by
Figure 5, defines a one-step reduction relation on type terms under
a constraint environment . The reduction relation on types is used
by the one-step unification judgement U 1 = 2
U ; R.
This judgement holds for an equality 1 = 2 iff the corresponding
unification problem can be reduced to a simpler unification problem
in the form of a set of equality constraints U and a substitution
R. The symbol in the inference rules represents an empty set of
equality constraints and the identity substitution, respectively. The

repeated application of one-step unification, as performed by the

reflexive and transitive closure U U U ; R, reduces a set
of equality constraints U to a set of pending equality constraints
U and a substitution R. If U cannot be reduced any further, we
also write U U ! U ; R, hence turning the transitive closure
of one-step unification into a deterministic function. Note that, due
to the syntactic restrictions on , Rule (app U ) and (red U ) are not
overlapping.
Unification is performed under a constraint environment only
because this environment is required by associate type synonym
expansion. It is easy to see that, where two types 1 and 2 are free
!
of associated synonyms, the relation U {1 = 2 }
; R
coincides with standard syntactic unification as employed in type
inference for vanilla Haskell 98.
Just like Jones [14], we only need first-order unification despite
the presence of higher-kinded variables, as we require all applications of associated type synonyms to be saturated.
Soundness of unification. The judgements of Figure 5 enjoy the
following properties.
L EMMA 1 (Soundness of type reduction). Given a well-formed
constraint environment and a type with , we have that

implies = .
L EMMA 2 (Soundness of one-step unification). Given a well-formed
constraint environment , the judgement U 1 = 2
U; R
implies that R(1 = 2 ) iff U .
T HEOREM 1 (Soundness of unification). Given a well-formed constraint environment and a set of equalities U , then U U
U ; R implies that RU iff U .
The proofs proceed by rule induction. Moreover, Theorem 1 requires that R1 iff R2 follows from 1 iff
2 for any substitution R, which is a basic property of constraint entailment.
5.4 Subsumption
To handle type annotations, such as in expressions of the form
e :: , during type inference, we need to have an algorithm deciding
type subsumption. We say a type scheme 1 subsumes 2 , written
1 2 iff any expression that can be typed as 1 can also be
typed as 2 .
The subsumption check can be formulated as a constraint entailment problem in the presence of equality constraints. If we want to
check
(1 .1 1 ) (2 .2 2 ),
in the context of a program context P , we need to demonstrate
that, given a set of new type constructors T of the appropriate kind,
there exists a substitution R = [/1 ], such that
P , [T /2 ]2 R1 , R1 = [T /2 ]2
where we define the entailment of a set of predicates as the
conjunction of all i .
Given that we implement the entailment of equalities by unification during type inference, it is not surprising that we can implement subsumption via a combination of unification and the usual
backchaining approach to entailment; a process called context reduction when used to simplify type class contexts. In Figure 6, we
define the procedure computing entailment using the same structure as we used for the unification procedure: Firstly, we define
a one-step entailment judgement E
P ; R that reduces
a constraint (which may be a class constraint or equality constraint) to a simpler set of constraints P and a substitution R.

, U | T e :
(x : . )

fresh

(var W )

[/] | x : [/]
W

fresh
P U U1 , U2 , (T2 1 = 2 )

1 , U1 | T1 e1 : 1

2 , U2 | T2 (T1 ) e2 : 2

U; R

(EW )

R(T2 1 , 2 ), U | (RT2 T1 ) e1 e2 : R
W

, U | T [x : ] e :

fresh

(IW )

, U | T x.e : T
W

1 , U1 | T1 e1 : 1

2 , U2 | T2 (T1 [x : ]) e2 : 2

= Gen(T1 ; (1 , U1 ) 1 )

(let W )

2 , U2 | (T2 T1 ) (let x = e1 in e2 ) : 2
W

, U | T e : 1

P (. 2 )

= Gen(T ; (, U ) 1 )

fresh
Fv ( 2 )

[/] | T (e :: . 2 ) : [/]2

Gen(; ) = .

(sig W )

, where = Fv \ Fv

Figure 4: Type inference rules

1
1 2
i

1ik

Sk

i
k

2
1 2

1
(lapp R )
1 2

j = j , j 6= i
Sk

U =

U 2 = 1
U 1 = 2

(refl U )

U 1 2 = 1 2

{1 = 1 , 2 = 2 };

U;

U; R

U; R
(symm U )
U; R
(app U )

U U
U U

(. = )
(syn R )
[/]
[/]

(arg R )

U 1 = 2

2
(rapp R )
1 2

6 Fv
(var U )
U =
; [ /]

2
(red U )
U = 1
{2 = 1 };

U ; R

U 1 = 2
U ; R
U R U, U
U U, 1 = 2 U ; R R

Figure 5: Unification in the presence of type functions

U ; R

 E

E D

(. D )
(bchain E )
E [/]D
[/];

E P
U
U P

P;

P;R

(red E )
{D };

U 1 = 2
E 1 = 2

U; R
(eq E )
U; R

P ; R

P ; R
U R P, P
U P, P ; R R

P ; R

Figure 6: Constraint entailment with equality constraints

Rule (bchainE ) formalises standard backchaining, Rule (redE ) enables the reduction of associated type synonyms in class parameters, and Rule (eqE ) invokes one-step unification for equality constraints. Secondly, we define the reflexive and transitive closure

E P
P ; R. Finally, we turn the closure into a deter!
ministic function by requiring that for P E P
P ; R, the

resulting set of constraints P cannot be reduced further.

Now, we can define subsumption as follows:
P , [T /2 ]2 E {1 , 1 = [T /2 ]2 } ! ; R
T new type constructors
()
P (1 .1 1 ) (2 .2 2 )
In other words, subsumption succeeds if context reduction can
resolve the constraint set entirely.
Soundness of subsumption. The subsumption rule enjoys the
following property.
L EMMA 3 (Soundness of subsumption). Let P be a well-formed
constraint environment, and 1 = 1 .1 1 and 2 =
2 .2 2 signatures with P 1 and P 2 , where 2 is
closed. Then, P 1 2 implies that, given a set of new type
constructors T of the appropriate kind, there exists a substitution
R = [/1 ], such that P , [T /2 ]2 R1 , R1 = [T /2 ]2 .

6. Functional Dependencies
Associated type synonyms cover much the same applications as
functional dependencies. It is too early to say which of the two
is better; hence, we will simply contrast the properties of both.
The comparison is complicated by the existence of at least three
variants of functional dependencies: (1) The system described by
Mark Jones [15], (2) the generalised system implemented in the
Glasgow Haskell Compiler (GHC) [31], and (3) an even more
permissive system formalised by Stuckey & Sulzmann [30]. Like
our associated type synonyms, the first two of these systems permit
an implementation by a dictionary translation to an explicitly-typed
language akin to System F. Stuckey and Sulzmanns system, based
on an encoding into constraint handling rules, gains additional
generality by giving up on a type-preserving translation and on
separate compilation.
As pointed out by Duck et al. [6], GHCs system is not decidable, as it permits recursive instance declarations that, for some programs that should be rejected, leads to non-termination of type inference. Jones original system is decidable. The Stuckey-Sulzmann
system, and the associated type synonyms we describe here, both
ensure decidability by a suitable set of restrictions on the admissible constraint handling rules and associated type synonyms, respectively; both systems can handle more general rule sets if decidable
inference is not required.

The lemma is an immediate consequence of Theorem 1.

5.5

Formal properties of type inference

T HEOREM 2 (Soundness of type inference). Assume a well-formed

program context P , typing environment , and expression e. If
W
, U | T e : , then (P , , U ) | T e : .
In proving this theorem, we followed Jones [12] who introduces a
syntax-directed system as a bridge between the declarative typing
rules and the algorithmic inference rules, as first suggested by
Clement at al. [3]. The tricky bit is to define the inference rule
for function elimination in the syntax-directed system such that
a correspondence to the inference system can be established. We
define the rule as follows:
| e1 : 1
| e2 : 2
P , 1 = 2 1
(ES )
| e1 e2 : 1
In addition, we need to use Theorem 1 and Lemma 3.
We believe that our inference algorithm is complete and that we
can obtain principal types by performing a further generalisation
step on its outputs. However, we leave the proof to further work
which we plan to do by again following Jones scheme for qualified
types.

6.1 Type substitution property

We usually require from a type system that, if we can type an
expression e with a type scheme ., then we can also type it
with an instance [ /]. Interestingly, this property does not hold
for Jones and GHCs version of functional dependencies. Here is
an example:
class C a b | a b where
foo :: a b
instance C Bool Int where
foo False = 0
foo True = 1
Under these definitions, consider the following three signatures of
the function bar :
bar :: C a
b a
b
-- (1). . . most general type
bar :: C Bool b Bool b
-- (2). . . subst. instance
bar ::
Bool Int -- (3). . . apply fun-dep
bar = foo
With signature (1) and (3) the program type checks fine, but with
(2) it is rejected by both GHC and Hugs. This is despite (2) clearly
being a substitution instance of (1), and (2) and (3) being equivalent
(i.e., each one subsumes the other). The Stuckey-Sulzmann system
accepts all three signatures.

The corresponding declarations with associated types would be

class C a where
type B a
foo :: a B a
instance C Bool where
type B Bool = Int
foo False
= 0
foo True
= 1
and we get the following signatures:
bar :: C a a
Ba
-- (1). . . most general type
bar ::
Bool B Bool -- (2). . . substitution inst.
bar ::
Bool Int
-- (3). . . reduce assoc syn.
bar = foo
Here we immediately get a ground constraint (which we omit as
usual) by the substitution and all three signatures are accepted. The
crucial difference is that with associated type synonyms, we do not
need a type variable to represent the functionally dependent type.
6.2

Readability

As observed by Garcia et al. [8], if functional dependencies are

used for the type of generic programming outlined in Section 2.2
on a full scale graph library and not the cut down version we
used for illustration purposestype classes quickly accumulate
a significant number of parameters. This affects readability for
reasons similar to why records with named fields are preferred over
tuples of large arity. Associated types, just like records with named
fields, refer to parameters by name and not by position. The same
observation led to the addition of traits classes [20] in C++, which
use a form of associated type synonyms in C++ classes.
As discussed in Section 3.4, if a class with functional dependencies is used in the context of an instance declaration of a class that
does not have functional dependencies, the arising dependencies
can be quite subtle and depend indirectly on the types of arguments
a function is applied to.
6.3

Expressiveness

An immediate question is whether associated types can generally

be encoded as functional dependencies and vise versa. This question does not have a simple answer. Duck et al. [6, Example 4] use
the following class head:
class Zip a b c | a c b, b c a
It is not immediately clear how to translate this to associated types.
However, as Oleg Kiselyov [17] demonstrates, the simpler class
header
class Zip a b c | c a b
is sufficient for this and similar applications. This second declaration can also be readily captured as an associated type synonym. In
fact, it appears from inspecting publicly available code as if functional dependencies are normally used in a way that can be directly
translated into associated types. In particular, even sophisticated
uses of functional dependencies, such as HList [18] and a type safe
evaluator, can be mirrored with associated types.
Conversely, a direct encoding of associated types with functional dependencies is not possible with Jones and GHCs system
due to the limitation discussed in Section 6.1. A more sophisticated
encoding similar to the translation to an explicitly type passing language in the style of System F, as discussed in [1], does not require functional dependencies; plain multi-parameter type classes
are sufficient. The Stuckey-Sulzmann system is, however, more
powerful and seems to be able to directly handle all signatures involving associated types.

7. Other related work

Associated data types. In our previous work on associated data
types [1], we allowed a class declaration to define a new data
type (rather than a new type synonym). Is it necessary to have
associated type synonyms as well as associated data types? At first,
it might seem as if notbut both mechanisms are subtly different
and cannot completely emulate each other.
Lets first try realising associated type synonyms as associated
data types. Recall, for example, the list instance for Collects:
instance Eq e Collects [e] where
type Elem [e] = e
insert x xs = x : xs
The elements of [e] are of type e. It would be extremely inconvenient to have to wrap the elements of the list with a constructor, so
that they were a fresh type:
instance Eq e Collects [e] where
data Elem [e] = ListElem e
insert (ListElem x ) xs = x : xs
In fact, it is not only inconvenient, as now signatures, such as
merge :: (Collects c1, Collects c2, Elem c1 = Elem c2)
c1 c2 c2
do not make any sense anymore. If Elem is an associated data
type, the equality constraint Elem c1 = Elem c2 is unsatisfiable.
Moreover, the Format class crucially rests on Sprintf being a
synonym; otherwise, the right-hand sides would not be function
types anymore.
Now, how about the conversecan associated synonyms emulate associated data types? They almost can. Declarations of the
form
class ArrayElem e where
data Array e
index :: Array e Int e
instance ArrayElem Int where
data Array Int = ArrInt UnboxedIntArray
index = indexUnboxedIntArray
can be rewritten as
class ArrayElem e where
type Array e
index :: Array e Int e
data ArrayInt = ArrInt UnboxedIntArray
instance ArrayElem Int where
type Array Int = ArrayInt
index = indexUnboxedIntArray
The translation becomes more involved for more complicated
instances, but the principle remains the same. The difference
between the two versions is the following. If we pass to the
index :: Array e Int e function a value of type Array Int,
then for an associated data type, we know that e = Int. However,
for an associated synonym, we cannot draw that conclusion. The
equation Array e = Array Int may have more than one solution,
depending on the available ArrayElem instances.
More generally, the introduction of a new type by an associated
data type implies a bijection between collection type and element
type, much like a bi-directional functional dependency. This is too
strong a requirement for, for example, the class Collects. It needs
a uni-directional dependence, and that is just what an associated
synonym does. In particular, two collections represented differently
may still contain elements of the same type.
Open versus closed type functions. Each associated type synonym constitutes a function over types, which is defined by the set
of type synonym equations contained in the instance declarations
of the class the synonym is associated with. Since Haskell permits

to add new instances to an existing class at any point, associated

type synonyms are open type functionsin the same way as class
methods are open functions on the value level.
Using associated synonyms, we can define computations on the
type level, and such computations, expressed using functional dependencies, have become quite popular with hard-core Haskell programmers. The standard example is addition on Peano numerals:
data Zero
-- empty type
data Succ a
-- empty type
class Nat n where
type Add n m ::
instance Nat Zero where
type Add Zero m = m
instance Nat n Nat (Succ n) where
type Add (Succ n) m = Succ (Add n m)
This is, for example, useful to define bounded data types, such as
fixed length vectors:
class Nat n VecBound n where
dataVec n a
appVec :: (VecBound m, VecBound (Add n m))
Vec n a Vec m a Vec (Add n m) a
instance VecBound Zero where
data Vec Zero a = Nil
appVec Nil ys = ys
instance VecBound n VecBound (Succ n) where
data Vec (Succ n) a = Cons a (Vec n a)
appVec (Cons x xs) ys = Cons x (appVec xs ys)
The signature of appVec guarantees that the length of the result
vector is the sum of the lengths of the argument vectors. The type
Vec is an associated data type.
This definition of Add is unsatisfactory in two ways: (1) In its
use of empty data types (i.e. ones with no constructors) and (2) the
use of type classes to define Add . After all, we would not use type
classes to define summation on Peano numerals on the value level!
Instead, we would use an algebraic data type and a closed function.
On the type level, this corresponds to an algebraic kind definition
and a closed type function, which we might denote as follows:
kind Nat
= Zero | Succ Nat
type Add Zero
m = m
type Add (Succ n) m = Succ (Add n m)
This is not only more concise, it also captures our intentions much
better: it ensures that Add is only applied to types of kind Nat,
and it asserts that we will not extend the definition of Nat and
Add in the future. Closed kind and type function definitions in a
Haskell-like language have been proposed by Sheard for the mega
system [25].
Closed type functions are definitely usefulbut so are open
type functions! Indeed, open type functions are an extremely natural complement to the value-level open functions expressed by type
classes. Although we have not addressed closed type functions in
this paper, we would like to note that the type system and type inference algorithm presented in Sections 4 and 5 can handle closed
type functions as well as open type functions. All we have to require is that the equations defining closed type functions are total
and obey the same restrictions as those for open type functions,
which we outlined in Section 3.3, so that they are confluent and
terminating.
Functional logic overloading, HM(X), and constraint handling
rules. Neubauer et al. [22] proposed a variant of type classes that
define functions, instead of predicates, for use in type contexts. In
their system, constraint environments do not contain any predicates
anymore, but only equalities. Moreover, one type class defines exactly one function, whereas a type class can have multiple associ-

ated type synonyms in our system. Multiple associated synonyms

are, for example, useful for a more fully fledged version of our
Graph class. Neubauer et al. consider both open and closed functions as well as overlapping definitions, whereas we only consider
open functions without overlapping definitions. An evidence translation for Neubauer et al.s system is future work, whereas the translation of associated synonyms follows our earlier work on associated data types.
Neubauer et al. base their formal system on work by Odersky et al. [23], which introduces Hindley-Milner type inference
parametrised with a wide range of constraint systems, called
HM(X). The HM(X) system is based on type subsumption, of
which equality is a special case.
Stuckey and Sulzmann [30] have developed the HM(X) framework further by fixing the constraint system to be based on constraint handling rules (CHRs). The resulting system is no longer
restricted to regular equational theories and has been successfully
used to encode a general version of functional dependencies. The
price for this generality is the loss of the ability to translate to an
explicit type-passing language, as is standard in dictionary translations of type classes. Moreover, separate compilation is problematic as whole-program information is required for the translation.
Nevertheless, we expect to be able to implement type classes with
associated types via HM(CHR) and indeed this seems to be an interesting item for future work.
Guarded recursive data types and generalised abstract data types.
Generalised abstract data types (GADTs), also known as guarded
recursive data types and first-class phantom types, constrain the
construction of data types by equality constraints in data type
definitions [32, 2, 16, 25]. These constraints are available in the
corresponding branches of case expressions, which enables typing
expressions that would otherwise have been too specific. There
appears to be a connection between these equalities and equality
constraints, but they are handled differently during type inference.
Moreover, GADTs are closed and not associated with type classes.
ML modules. There is a significant overlap in functionality between Haskell type classes and Standard ML modules [5] . Associated types increase this overlap even more, and our equality
constraints appear to bear a relation to sharing constraints. Nevertheless, there are also interesting differences, and both constructs
have a large design space. We regard a more precise comparison of
the two as interesting future work.
C++ traits classes. Typedefs in C++ class templates are of a similar nature as associated data types [20], and there are many C++
libraries [29, 28, 8] that demonstrate the usefulness of this mechanism. Type checking in C++ is, however, of an entirely different
nature. In particular, template definitions are not separately type
checked, but only after expansion at usage occurrences. Moreover,
C++ does not have type inference. In exchange, C++ does not constrain the resulting type functions to be terminating or confluent,
but permits arbitrary computations at compile time.
Siek and Lumsdaine [27] recently proposed an interesting combination of Haskell-style type classes with the C++ approach to
generic programming. They propose a language FG that adds a notion of concepts to System F and includes support for associated
types of the same flavour as those introduced in the present paper.
The main difference between Siek and Lumsdaines work and ours
is that Siek and Lumsdaine trade type inference for a more expressive type language.

8. Conclusions and further work

In this paper and its predecessor, we have explored the consequences of enabling a Haskell type class to contain data and type

definitions, as well as method definitions. Doing so directly addresses Garcia et als primary concern about large-scale programming in Haskell. It also fills out Haskells existing ability to define
open functions at the value level using type classes, with a complementary type-level facility.
There is clearly a big overlap between functional dependencies
and associated type synonyms, and no language would want both.
We do not yet have enough experience to know what difficulties (either of expressiveness or convenience), if any, programmers would
encounter if functional dependencies were replaced by associated
type synonymsbut we regard that as an attractive possibility.
Acknowledgements. We particularly thank Martin Sulzmann for
his detailed and thoughtful feedback, which helped to identify
and characterise several potential pitfalls. We are also grateful to
Roman Leshchinskiy and Stefan Wehr who suggested significant
improvements to the presentation and the formal system; moreover,
Stefan Wehr throughly improved and extended the prototype type
checker. We also thank the anonymous referees for their helpful
comments. The first two authors have been partly funded by the
Australian Research Council under grant number DP0211203.

References
[1] Manuel M. T. Chakravarty, Gabriele Keller, Simon Peyton Jones,
and Simon Marlow. Associated types with class. In Martin Abadi,
editor, Conference Record of POPL 2005: The 32nd ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, pages
113. ACM Press, 2005.
[2] James Cheney and Ralf Hinze. First-class phantom types. Technical
Report TR2003-1901, Cornell University, 2003.
[3] Dominique Clement, Thierry Despeyroux, Gilles Kahn, and Joelle
Despeyroux. A simple applicative language: mini-ML. In LFP 86:
Proceedings of the 1986 ACM Conference on LISP and Functional
Programming, pages 1327, New York, NY, USA, 1986. ACM Press.
[4] Olivier Danvy. Functional unparsing. J. Funct. Program., 8(6):621
625, 1998.
[5] Derek Dreyer, Karl Crary, and Robert Harper. A type system for
higher-order modules. In Proceedings of the 30th ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, pages
236249, 2003.
[6] Gregory J. Duck, Simon Peyton Jones, Peter J. Stuckey, and Martin
Sulzmann. Sound and decidable type inference for functional
dependencies. In ESOP;04, LNCS. Springer-Verlag, 2004.
[7] Karl-Filip Faxen. A static semantics for Haskell. Journal of
Functional Programming, 12(4+5), 2002.
[8] Ronald Garcia, Jaakko Jarvi, Andrew Lumsdaine, Jeremy Siek, and
Jeremiah Willcock. A comparative study of language support for
generic programming. In Proceedings of the 18th ACM SIGPLAN
Conference on Object-Oriented Programing, Systems, Languages,
and Applications, pages 115134. ACM Press, 2003.
[9] Cordelia V. Hall, Kevin Hammond, Simon L. Peyton Jones, and
Philip L. Wadler. Type classes in Haskell. ACM Trans. Program.
Lang. Syst., 18(2):109138, 1996.
[10] Ralf Hinze. Formatting: A class act. Journal of Functional
Programming, 13:935944, 2003.
[11] Ralf Hinze and Simon Peyton Jones. Derivable type classes. In
Graham Hutton, editor, Proceedings of the 2000 ACM SIGPLAN
Haskell Workshop, volume 41.1 of Electronic Notes in Theoretical
Computer Science. Elsevier Science, 2001.
[12] Mark P. Jones. A theory of qualified types. In ESOP92: Symposium
proceedings on 4th European symposium on programming, pages
287306, London, UK, 1992. Springer-Verlag.
[13] Mark P. Jones. Simplifying and improving qualified types. In
FPCA 95: Conference on Functional Programming Languages and
Computer Architecture. ACM Press, 1995.
[14] Mark P. Jones. A system of constructor classes: Overloading
and implicit higher-order polymorphism. Journal of Functional

Programming, 5(1), 1995.

[15] Mark P. Jones. Type classes with functional dependencies. In
Proceedings of the 9th European Symposium on Programming (ESOP
2000), number 1782 in Lecture Notes in Computer Science. SpringerVerlag, 2000.
[16] Simon Peyton Jones, Geoffrey Washburn, and Stephanie Weirich.
Wobbly types: type inference for generalised algebraic data types.
http://research.microsoft.com/Users/simonpj/papers/
gadt/index.htm, 2004.
[17] Oleg Kiselyov. Functions with the variable number (of variously typed) arguments. http://okmij.org/ftp/Haskell/
vararg-fn.lhs, 2004.
[18] Oleg Kiselyov, Ralf Lammel, and Keean Schupke. Strongly typed
heterogeneous collections. In Haskell 04: Proceedings of the ACM
SIGPLAN Workshop on Haskell, pages 96107, New York, NY, USA,
2004. ACM Press.
[19] J. W. Klop. Term rewriting systems. In S. Abramsky D. M. Gabbay
and T. S. E. Maibaum, editors, Handbook of Logic in Computer
Science, volume 2, pages 1116. Oxford University Press, 1992.
[20] Nathan C. Myers. Traits: a new and useful template technique. C++
Report, June 1995.
[21] Matthias Neubauer and Peter Thiemann. Type classes with more
higher-order polymorphism. In ICFP 02: Proceedings of the
seventh ACM SIGPLAN international conference on Functional
programming, pages 179190, New York, NY, USA, 2002. ACM
Press.
[22] Matthias Neubauer, Peter Thiemann, Martin Gasbichler, and Michael
Sperber. Functional logic overloading. In POPL 02: Proceedings
of the 29th ACM SIGPLAN-SIGACT symposium on Principles of
programming languages, pages 233244. ACM Press, 2002.
[23] Martin Odersky, Martin Sulzmann, and Martin Wehr. Type inference
with constrained types. Theory and Practice of Object Systems, 5(1),
1999.
[24] Didier Remy. Extending ML type system with a sorted equational
theory. Research Report 1766, INRIA Rocquencourt, 1992.
[25] Tim Sheard. Languages of the future. In OOPSLA 04: Companion
to the 19th Annual ACM SIGPLAN Conference on Object-Oriented
Programming Systems, Languages, and Applications, pages 116119,
New York, NY, USA, 2004. ACM Press.
[26] Tim Sheard and Simon Peyton Jones. Template meta-programming
for Haskell. ACM SIGPLAN Notices: PLI Workshops, 37(12):6075,
2002.
[27] Jeremy Siek and Andrew Lumsdaine. Essential language support for
generic programming. SIGPLAN Not., 40(6):7384, 2005.
[28] Jeremy G. Siek, Lie-Quan Lee, and Andrew Lumsdaine. The Boost
Graph Library User Guide and Reference Manual. Addison-Wesley,
2001.
[29] A. A. Stepanov and M. Lee. The standard template library. Technical
Report X3J16/94-0095, WG21/N0482, ISO Programming Language
C++ Project, 1994.
[30] Peter J. Stuckey and Martin Sulzmann. A theory of overloading.
ACM Transactions on Programming Languages and Systems, 2004.
To appear.
[31] The GHC Team. The Glasgow Haskell Compiler. http://haskell.
org/ghc/documentation.html.
[32] Hongwei Xi, Chiyan Chen, and Gang Chen. Guarded recursive
datatype constructors. In Proceedings of the 30th ACM SIGPLANSIGACT Symposium on Principles of Programming Languages, pages
224235. ACM Press, 2003.