Lariat
Version 0.3
Lariat is a project to define a total abstract data type for
proper lambda terms, consisting of four basic operations:
app
, abs
, var
, and destruct
.
This repository presents the definition of these operations. It also contains implementations of this abstract data type in various programming languages, currently including:
The version of the Lariat defined by this document is 0.3. This version number will be promoted to 1.0 once vetted sufficiently.
Table of Contents
Background
There are several approaches to representing lambda terms in software.
The naive approach is to represent them just as they are written on paper. In this approach, whether a variable, such as x, is free or bound depends on whether it is inside a lambda abstraction λ x or not. If you need to manipulate it (or the abstraction it's bound to), you might need to rename it so that it doesn't conflict with another variable also called x that is perhaps free or perhaps bound to a different lambda abstraction.
This is tiresome and error-prone. So other approaches were developed.
One such alternate approach is De Bruijn indexes (Wikipedia), where variables are represented not by names, but by numbers. The number indicates which lambda abstraction the variable is bound to, if any; a 1 indicates the immediately enclosing lambda abstraction, a 2 indicates the lambda abstraction just above that, and so on. If the number exceeds the number of enclosing lambda abstractions, then it is a free variable.
But this, too, has some drawbacks, so people have devised a number of other approaches:
- "maps" (Viewing Terms through Maps (PDF), Sato et al., 1980) (see also these slides (PDF) from 2012)
- "nominal techniques" (A New Approach to Syntax (PDF), Gabbay and Pitts, 1999)
- "locally nameless" (I am not a number (PDF), McBride and McKinna, 2004)
- "bound" (bound: Making de Bruijn Succ Less, Kmett, 2013)
among others.
But the point I would like to make in this article is this: At some level of abstraction it does not matter which approach is chosen as long as the approach satisfies the essential properties that we require of lambda terms.
To this end, this article presents an abstract data type (ADT) for lambda terms, which we call Lariat, consisting of four operations. The actual, concrete data structure in which they are stored, and the actual, concrete mechanism by which names become bound to terms, are of no consequence (and may well be hidden from the programmer) so long as the implementation of the operations conforms to the stated specification.
This ADT is designed for simplicity and elegance rather than performance. It is a minimal formulation that does not necessarily make any of commonly-used manipulations of lambda terms efficient.
This ADT has two properties, intended to contribute to its elegance. Firstly, it can represent only proper lambda terms; that is, it is not possible for a lambda term constructed by the Lariat operations to contain an invalid bound variable.
Secondly, it is total in the sense that all operations are defined for all inputs that conform to their type signatures. There are no conditions (such as trying to pop from an empty stack in a stack ADT) where the result is undefined, nor any defined to return an error condition. This totality does, however, come at the cost of the operations being higher-order and with polymorphic types.
For more background information, see the Discussion section below.
Names
Lambda terms are essentially about name binding, and in any explication of name binding, we must deal with names. As of 0.3, Lariat requires only two properties of names.
Firstly, it must be possible to compare two names for equality. This is required for operations that replace free variables that have a given name with a value -- there must be some way for them to check that the free variable has the name that they are seeking.
Secondly, given a set of names, it must be possible to generate a new name that
is not equal to any of the names in the set (a so-called "fresh" name). This is
required to properly implement the destruct
operation. If names are modelled
as character strings, obtaining a fresh name could be as simple as finding the
longest string of a set of strings, and prepending "a"
to it.
Note that, although neither of these properties is exposed as an operation, it would be reasonable for a practical implementation of Lariat to expose them so. It would also be reasonable to provide other operations on names, such as constructing a new name from a textual representation, rendering a given name to a canonical textual representation, and so forth. From the perspective of Lariat itself these are ancillary operations, and as such will not be defined in this document.
Terms
We now list the four operations available for manipulating terms.
var(n: name): term
Given a name n, return a free variable with the name n.
Note: A free variable is a term; it can be passed to any operation that expects a term.
app(t: term, u: term): term
Given a term t and a term u, return an application term which contains t as its first subterm and u as its second subterm.
Note: An application term is a term that behaves just like an ordered pair of terms.
abs(n: name, t: term): term
Given a name n and a term t, return an abstraction term containing t', where t' is a version of t where all free variables named n inside t' have been replaced with bound variables. These bound variables are bound to the returned abstraction term.
Note: we may consider a bound variable to be a term, but the user of the abstract data type cannot work with bound variables directly, so it is unlike all other kinds of terms in that respect. A bound variable is always bound to a particular abstraction term. In the case of
abs
, the abstraction term to which variables are bound is always the term returned by theabs
operation.Note: an abstraction term contains one subterm. This subterm cannot be extracted directly, as it may contain bound variables, which the user cannot work with directly.
destruct(t: term, f1: fun, f2: fun, f3: fun): X
Given a term t and three functions f1, f2, and f3 (each with a different signature, described below), choose one of the three functions based on the structure of t, and evaluate it, returning what it returns.
If t is a free variable, evaluate f1(n) where n is the name of the free variable t.
If t is an application term, evaluate f2(u, v) where u is the first subterm of t and v is the second subterm of t.
If t is an abstraction term, evaluate f3(u, n) where u is a version of t where all bound variables in u that were bound to u itself have been replaced by n, where n is a fresh name (i.e. a name that does not occur in any free variable in any subterm of u).
Note: as stated above, a bound variable is always bound to an abstraction term. The bound variables that are replaced by the
destruct
of an abstraction term are always and only those that are bound to the abstraction term beingdestruct
ed.Note: the
destruct
operation's signature shown above was abbreviated to make it less intimidating. The full signature would bedestruct(t: term, f1: fun(n: name): X, f2: fun(u: term, v: term): X, f3: fun(u: term, n: name): X): X
Note: see the section on "Names" above for the basic requirements for obtaining a fresh name.
Some Examples
We will now give some concrete examples of how these operations can be used. But first, we would like to emphasize that Lariat is an ADT for lambda terms, not the lambda calculus. Naturally, one ought to be able to write a lambda calculus normalizer using these operations (and this will be one of our goals in the next section), but one is not restricted to that activity. The terms constructed using the Lariat operations may be used for any purpose for which terms-with-name-binding might be useful.
Example 1
A common task is to obtain the set of free variables present
in a lambda term. This is not difficult; we only need to
remember that every time we destruct
an abstraction term, we
introduce a fresh free variable of our own, to keep track of
these, and make sure not to include any of them when we
report the free variables we found.
Note: In the following pseudocode,
+
is the set union operator.
let freevars = fun(t, ours) ->
destruct(t,
fun(n) -> if n in ours then {} else {n},
fun(u, v) -> freevars(u, ours) + freevars(v, ours),
fun(u, n) -> freevars(u, ours + {n})
)
Example 2
Given an abstraction term and a value, return a version of
the body of the abstraction term where every instance of the
variable bound to the abstraction term is replaced by the given
value. We can call this operation resolve
.
let resolve = fun(t, x) ->
destruct(t,
fun(n) -> t,
fun(u, v) -> t,
fun(u, n) -> replace_all(u, n, x)
)
where replace_all = fun(t, m, x) ->
destruct(t,
fun(n) -> if n == m then x else var(n),
fun(u, v) -> app(replace_all(u, m, x), replace_all(v, m, x))
fun(u, n) -> abs(n, replace_all(u, m, x))
)
Note that this operation was specifically not called subst
,
because the name subst
is often given to a process that
replaces free variables, while this operation replaces
bound ones. It was also specifically not named beta
because it does not require that t and x come from the same
application term.
Example 3
The next task is to write a beta-reducer. We destruct
the term twice, once to ensure it is an application term,
and again to ensure the application term's first subterm
is an abstraction term. Then we use resolve
, above, to
plug the application term's second subterm into the
abstraction term.
let beta = fun(r) ->
destruct(r,
fun(n) -> var(n),
fun(u, v) ->
destruct(u,
fun(_) -> app(u, v),
fun(_, _) -> app(u, v),
fun(u) -> resolve(u, v)
),
fun(t) -> t
)
In fact, we could merge this implementation with the
implementation of resolve
and this would save a call
to destruct
; but this would be merely an optimisation.
It is left as an exercise to any reader who may be so
motivated to undertake it.
Example 4
The next task would be to search through a lambda term,
looking for a candidate application term to reduce, and
reducing it. The pseudocode below returns a pair
[bool, term]
where the boolean value indicates whether
the term has been rewritten by the call or not. It
implements a leftmost-outermost reduction strategy.
let reduce = fun(t) ->
if is_beta_reducible(t) then
[true, beta(t)]
else
destruct(t,
fun(n) -> [false, var(n)],
fun(u, v) ->
let
[has_rewritten, new_u] = reduce(u)
in
if has_rewritten then
[true, app(new_u, v)]
else
let
[has_rewritten, new_v] = reduce(v)
in
if has_rewritten then
[true, app(u, new_v)]
else
[false, app(new_u, new_v)],
fun(t) -> [false, t]
)
From there it's just a hop, a skip, and a jump to a proper lambda term normalizer:
let normalize(t) ->
let
[has_rewritten, new_t] = reduce(t)
in
if has_rewritten then
normalize(new_t)
else
t
Discussion
Prior work: Paulson's exercise
The idea of formulating an ADT for lambda terms is not a new one. In Chapter 9 of "ML for the Working Programmer", 1st ed. (1991), Lawrence Paulson develops an implementation of lambda terms in ML and notes that:
Signature LAMBDA_NAMELESS is concrete, revealing all the internal details. [...] An abstract signature for the λ-calculus would provide operations upon λ-terms themselves, hiding their representation.
So the idea is an established one; but if so, why does one see so few instances of it out in the wild? I think it's this: most lambda term manipulation code sees actual use only academic contexts, most usually in such things as theorem provers. These are contexts that don't greatly benefit from the software engineering principle of being able to swap out one implementation of an interface with an alternative implementation. Indeed, in a theorem proving context, an extra level of abstraction may just be another burden that the mechanical reasoning methods need to deal with, with no other benefit. So concrete data types are used, because concrete data types are sufficient.
In the context of Lariat, however, the ADT is the object of study in its own right.
Now, here's the part I elided from the above quoted paragraph:
Many values of type term are improper: they do not correspond to real λ-terms because they contain unmatched bound variable indices. [...] abstract returns improper terms and subst expects them.
And at the end of the section, he poses Exercise 9.16:
Define a signature for the λ-calculus that hides its internal representation. It should specify predicates to test whether a λ-term is a variable, an abstraction, or an application, and specify functions for abstraction and substitution.
However, as he mentioned earlier, these operations produce and expect improper terms; so he appears to be asking for an abstract representation of lambda terms that includes improper lambda terms. [Footnote 1] I would argue that such an ADT has a lot less value as an abstraction than an ADT in which only proper lambda terms can be represented. [Footnote 2]
Although it was not in direct response to this exercise (which I hadn't seen for years until I came across it again), it was consideration of this point -- how does one formulate an ADT that represents only proper lambda terms? -- that led me to formulate Lariat.
The role of destruct
var
, app
, and abs
construct terms, while destruct
takes them apart.
Constructing terms is the easy part; it's taking them apart properly that's
hard.
destruct
is a "destructorizer" in the sense described in
this article on Destructorizers.
In fact, this use case of "taking apart" lambda terms was one
of the major motivations for formulating the destructorizer
concept.
Although it was not specifically intended, destruct
is also what permits
the ADT to be "total" in the sense that there are no operations that are
undefined.
Equality modulo renaming of bound variables
When working with lambda terms, one is often concerned with
comparing two lambda terms for equality, modulo renaming of bound
variables. We haven't introduced such an operation because it should
be possible to build such an operation using destruct
; basically,
render the two terms as text (or some other concrete representation),
then compare the texts for equality. (This does however require that
thete is an operation for rendering a name to its textual representation,
and also that the procedure for obtaining a fresh name is deterministic,
so that the fresh names generated when destruct
ing two equal abstractions,
match up in both of the terms.)
Of course, such an operation could be provided as a native
operation for performance or convenience. (This is one of the nice
things about ADTs -- they can be sub-ADTs of a larger ADT.) Similarly,
although we have shown that we can implement freevars
using the operations
of the ADT, the definition of the destruct
operation essentially requires
that something equivalent to it already exists, and it could be exposed to
the user as well.
The possibility of an algebraic formulation
The ADT that has been described in this document has been described quite precisely (I hope) but not formally. A direction that this work could be taken in would be to produce a definition of Lariat that is actually formal, i.e. in the form of an equational theory, or equivalently, an algebra.
There are reasons to believe this is not impossible. In The Lambda Calculus is Algebraic (PDF) (Selinger, 1996) an algebra equivalent to the lambda calculus is formed by treating free variables as "indeterminates" (although I must admit I'm not entirely certain what is meant by that). Additionally, section 1.3 of Language Prototyping: An Algebraic Specification Approach (1996; van Deursen, Heering, Klint eds., borrowable online at archive.org) gives a definition of the lambda calculus in the algebraic definition language ASF+SDF, which comes fairly close to conventional equational logic (although it does contain extras such as conditional equations).
However, in Lariat, destruct
is a "higher-order" operation,
in the sense that it takes functions as parameters, and this may
well complicate the task of defining an equational theory based
on Lariat, or it may complicate the resulting equational theory.
We'll talk about that in the next section.
Variation: Partial Lariat
To support the effort of formulating an algebra based on Lariat,
or for any other purposes which it may suite, it's worth looking at the possibility
of replacing destruct
with a set of "first-order" operations.
When I first started working out Lariat, I thought that using a
destructorizer would be essential to the problem of being able to
destruct an abstraction term and have the result be a proper lambda term.
It's not as essential as I thought. What destruct
does when given
an abstraction term is, basically, to form a free variable with a
fresh name (one that does not occur in the abstraction term) and
resolve
(as defined in the examples above) the abstraction term with it.
If the ADT were to have discrete operations for picking a fresh name
given a lambda term, and for resolve
, these could be applied
"manually", and in this manner the user could destruct abstraction
terms just the same as destruct
does.
There are subtle differences: resolve
would need to be an intrinsic operation
which is exposed by the ADT, rather that derived from the basic operations
of Lariat. It also gives the user the freedom to apply resolve
with whatever
they wish, while in Lariat, destruct
can only apply this action with a fresh
variable that it itself has chosen, which is significantly more restrictive.
A version of Lariat without destruct
would also need operations
for testing if a term is an abstraction term, vs. an application
term, vs. a free variable. The operation of extracting the first
and second values from an application term (basically, the theory
of ordered pairs) would not be sensibly defined for abstraction
terms or free variables, and so this version of the ADT would be
partial rather than total, thus the name "Partial Lariat".
Footnotes
Footnote 1
Either that or, based on his remark about an "abstract signature for the λ-calculus", he intended the operations in this exercise to be on the level of the lambda calculus, i.e. beta-reduction and normalization? But that's not what he wrote, and lacking a copy of the 2nd edition to see if this has been corrected, I shall take him at his word.
Footnote 2
For more information on this philosophy, see "Parse, don't Validate"; in particular, LCF-style Natural Deduction illustrates how it applies to theorem objects in an LCF-style theorem prover; and it applies here too.
Commit History
@master
git clone https://git.catseye.tc/Lariat/
- Arrange licensing info in repo according to REUSE 3.2 convention. Chris Pressey 3 months ago
- Arrange licensing info in repo following REUSE 3.0 convention. Chris Pressey 10 months ago
- Place under a 3-clause BSD license. Chris Pressey 1 year, 26 days ago
- Update links. Chris Pressey 1 year, 1 month ago
- Merge pull request #2 from catseye/develop-0.3 Chris Pressey (commit: GitHub) 2 years ago
- Many edits, especially to discussion. Add footnotes section. Chris Pressey 2 years ago
- Now `equal` is not an exposed operation either. Chris Pressey 2 years ago
- Add note about where deterministic fresh would be required. Chris Pressey 2 years ago
- Checkpoint various editing. Chris Pressey 2 years ago
- Obtaining a fresh name is no longer exposed as an operation. Chris Pressey 2 years ago