Lambda Calculus

They taught us Turing machines in my computer science classes. Despite being purely theoretical, Turing machines are important:

  • A state machine reading and writing symbols on an infinite tape is a useful abstraction of a CPU reading from and writing to RAM.

  • Many high-level programming languages adhere to the same model: code writes data to memory, then later reads it to decide a future course of action.

  • We immediately see how to measure algorithmic complexity.

  • Encoding a Turing machine on a tape is straightforward, and gently guides us to the equivalence of code and data, with all its deep implications.

All the same, I wish they had given equal weight to an alternative model of computation known as lambda calculus:

Why Lambda Calculus?

Lambda calculus is historically significant. Alonzo Church was Alan Turing’s doctoral advisor, and his lambda calculus predates Turing machines.

But more importantly, working through the theory from its original viewpoint exposes us to different ways of thinking. Aside from a healthy mental workout, we find lambda calculus is sometimes superior:

  • simple: Here’s how to multiply two numbers in lambda calculus: \(\lambda m.\lambda n.\lambda f.m(n f)\). Spare a thought for students struggling to make Turing machines do simple tasks.

  • practical: With a little syntax sugar, lambda calculus becomes a practical programming language. Already, our factorial example above is shorter than equivalent code in many high-level languages! In contrast, sweetened Turing machines would probably still be unpalatable.

  • one-line universal program: Here’s a lambda calculus self-interpreter: \( (\lambda f.(\lambda x.f(xx))(\lambda x.f(xx)))(\lambda em.m(\lambda x.x)(\lambda mn.em(en))(\lambda mv.e(mv))) \). In contrast, universal Turing machines are so tedious that classes often skip the details and just explain why they exist.

  • representing data with functions leads to rich algebras where a little notation goes a long way. For example, we can draw intricate diagrams with a few lines.

  • solves the halting problem: By adding types, we can ensure lambda calculus programs always halt. It’s unclear how we can similarly tame Turing machines.

  • provably correct: More generally, typed lambda calculus turns out to be deeply connected to the foundations of mathematics. Sufficiently advanced types make bugs impossible to express, that is, every syntactically correct program is also semantically correct. This connection is harder to see from a Turing-machine viewpoint.

As the importance of software grows in our world, so does the importance of the advantages of lambda calculus, and in particular, its connections with the foundations of mathematics. Computer science without lambda calculus is like engineering without physics.

A better analogy would be mathematics without proofs, but proofs have been part of mathematics for so long that it may be difficult to imagine one without the other.

Why Lambda?

See The impact of lambda calculus in logic and computer science by Henk Barendregt, and History of Lambda-calculus and Combinatory Logic by Felice Cardone and J. Roger Hindley. It seems its true name should be "hat calculus". We’ll find that lambdas are redundant, but I suppose we need a symbol of some sort to avoid calling it just "calculus".

A talk by Dana Scott contains a conflicting anecdote: Church’s son-in-law wrote a postcard inquiring why lambda was chosen, and the written response was "eeny meeny miney moe", that is, Church chose this Greek letter for no reason in particular!

In the same talk, Scott suggests Kleene, despite being a lambda calculus expert, was perhaps responsible for the dominance of Turing machines. Kleene worried that audiences were mystified by lambda calculus in his lectures, so he switched to Turing machines.

Beta reduction

Unlike Turing machines, everyone already knows the basics of lambda calculus. In school, we’re accustomed to evaluating functions. In fact, one might argue they focus too much on making students memorize and apply formulas such as \( \sqrt{a^2 + b^2} \) for \(a = 3\) and \(b = 4\).

In lambda calculus, this is called beta reduction, and we’d write this example as:

\[(\lambda a b . \sqrt{a^2 + b^2}) 3 \enspace 4\]

This is almost all there is to lambda calculus! Only, instead of numbers, we plug in other formulas. The details will become clear as we build our interpreter.

I was surprised this substitution process learned in childhood is all we need for computing anything. A Turing machine has states, a tape of cells, and a movable head that reads and writes; how can putting formulas into formulas be equivalent?

We use code to help answer the question, which requires a bit of boilerplate:

{-# LANGUAGE CPP #-}
#ifdef __HASTE__
import Haste.DOM
import Haste.Events
#else
import System.Console.Haskeline
#endif
import Data.Char
import Data.List
import Text.Parsec

Terms

Lambda calculus terms can be viewed as a kind of binary tree. A lambda calculus term consists of:

  • Variables, which we can think of as leaf nodes holding strings.

  • Applications, which we can think of as internal nodes.

  • Lambda abstractions, which we can think of as a special kind of internal node whose left child must be a variable. (Or as a internal node labeled with a variable with exactly one child.)

data Term = Var String | App Term Term | Lam String Term

When printing terms, we’ll use Unicode to show a lowercase lambda (λ). Conventionally:

  • Function application has higher precedence, associates to the left, and their child nodes are juxtaposed.

  • Lambda abstractions associate to the right, are prefixed with a lowercase lambda, and their child nodes are separated by periods. The lambda prefix is superfluous but improves clarity.

  • With consecutive bindings (e.g. "λx.λy.λz."), we omit all lambdas but the first, and omit all periods but the last (e.g. "λx y z.").

For clarity, we enclose lambdas in parentheses if they are right child of an application.

instance Show Term where
  show (Lam s t)  = "\955" ++ s ++ showB t where
    showB (Lam x y) = " " ++ x ++ showB y
    showB expr      = "." ++ show expr
  show (Var s)    = s
  show (App x y)  = showL x ++ showR y where
    showL (Lam _ _) = "(" ++ show x ++ ")"
    showL _         = show x
    showR (Var s)   = ' ':s
    showR _         = "(" ++ show y ++ ")"

As for input, since typing Greek letters can be nontrivial, we follow Haskell and interpret the backslash as lambda. We may as well follow Haskell a little further and accept -> in lieu of periods, and support line comments.

Any alphanumeric string is a valid variable name.

Typing a long term is laborious, so we support a sort of let statement. The line

true = \x y -> x

means that for all following terms, the variable true is shorthand for the term on the right side, namely \x y -> x. There is an exception: if the left child of a lambda abstraction is the variable true, then this variable shadows the original let definition in its right child. It is good style to avoid such name collisions.

Our parser accepts empty lines, let definitions, or terms that should be evaluated immediately.

data LambdaLine = Blank | Let String Term | Run Term

line :: Parsec String () LambdaLine
line = between ws eof $ option Blank $
    try (Let <$> v <*> (str "=" >> term)) <|> (Run <$> term) where
  term = lam <|> app
  lam = flip (foldr Lam) <$> between lam0 lam1 (many1 v) <*> term
  lam0 = str "\\" <|> str "\955"
  lam1 = str "->" <|> str "."
  app = foldl1' App <$> many1 ((Var <$> v) <|> between (str "(") (str ")") term)
  v   = many1 alphaNum <* ws
  str = (>> ws) . string
  ws = spaces >> optional (try $ string "--" >> many anyChar)

Evaluation

If the root node is a free variable or a lambda, then there is nothing to do. Otherwise, the root node is an App node, and we recursively evaluate the left child.

If the left child evaluates to anything but a lambda, then we stop, as a free variable got in the way somewhere.

Otherwise, the root node is a reducible expression or redex, namely, it is an App node whose left child is some lambda term \( \lambda v . M \). We perform beta reduction as follows. We traverse \(M\) and replace every occurrence of \(v\) with the right child of the root node.

While doing so, we must handle a potential complication. A reduction such as (\y -> \x -> y)x to \x -> x is incorrect. To prevent this, we rename the first x and find (\y -> \x1 -> y)x reduces to \x1 -> x.

More precisely, we say a variable v is bound if it appears in the right subtree of a lambda abstraction node whose left child is v. Otherwise v is free. If a substitution would cause a free variable x to become bound, then we rename bound occurrences of x to before proceeding.

We store the let definitions in an associative list named env, and perform lookups on demand to see if a given string is a variable or shorthand for another term.

These on-demand lookups and the way we update env means recursive let definitions are possible. Thus our interpreter actually runs more than plain lambda calculus; a true lambda calculus term is unable to refer to itself. (Haskell similarly permits recursion via let expressions.) This is harmless, as below we’ll reveal how to achieve recursion with pure lambda calculus.

The quote business is a special feature that will be explained later.

eval env (App (Var "quote") t)              = quote env t
eval env (App m a) | Lam v f <- eval env m  =
  eval env $ beta env (v, a) f
eval env (Var v)   | Just x <- lookup v env = eval env x
eval _   term                               = term

beta env (v, a) t = case t of
  Var s   | s == v       -> a
          | otherwise    -> Var s
  Lam s m | s == v       -> Lam s m
          | s `elem` fvs -> Lam s1 $ rec $ rename s s1 m
          | otherwise    -> Lam s (rec m)
          where s1 = newName s $ v : fvs `union` fv env [] m
  App m n                -> App (rec m) (rec n)
  where rec = beta env (v, a)
        fvs = fv env [] a

fv env vs (Var s) | s `elem` vs            = []
                  -- Handle free variables in let definitions.
                  -- Avoid repeatedly following recursive lets.
                  | Just x <- lookup s env = fv env (s:vs) x
                  | otherwise              = [s]
fv env vs (App x y)                        = fv env vs x `union` fv env vs y
fv env vs (Lam s f)                        = fv env (s:vs) f

To pick a new name, we append "_1" if the name contains no underscore, and otherwise increment the number after the underscore until we’ve avoided names of other free variables.

Our parser rejects underscores in variable names, so the new name never clashes with a name chosen by the user.

newName :: String -> [String] -> String
newName x ys = head $ filter (`notElem` ys) $ (s ++) . ('_':) . show <$> [1..] where
  s = takeWhile (/= '_') x

Renaming a free variable is a tree traversal that skips lambda abstractions that bind them:

rename :: String -> String -> Term -> Term
rename x x1 term = case term of
  Var s   | s == x    -> Var x1
          | otherwise -> term
  Lam s b | s == x    -> term
          | otherwise -> Lam s (rec b)
  App a b             -> App (rec a) (rec b)
  where rec = rename x x1

Our eval function terminates once no more top-level function applications (beta reductions) are possible, reaching weak head normal form. We recursively call eval on child nodes to reduce other function applications throughout the tree, resulting in the normal form of the lambda term. The normal form is unique up to variable renaming (which is called alpha-conversion).

norm :: [(String, Term)] -> Term -> Term
norm env term = case eval env term of
  Var v   -> Var v
  Lam v m -> Lam v (rec m)
  App m n -> App (rec m) (rec n)
  where rec = norm env

A term with no free variables is called a closed lambda expression or combinator. When given such a term, our function’s output contains no App nodes.

That is, if it ever outputs something. There’s no guarantee that our recursion terminates. For example, it is impossible to reduce all the App nodes of:

omega = (\x -> x x)(\x -> x x)

In such cases, we say the lambda term has no normal form. We could limit the number of reductions to prevent our code looping forever; we leave this as an exercise for the reader.

In an application App m n, the function eval tries to reduce m first. This is called a normal-order evaluation strategy. What if we reduced n first, a strategy known as applicative order? More generally, instead of starting at the top level, what if we picked some sub-expression to reduce first? Does it matter?

Yes and no. On the one hand, the Church-Rosser theorem implies the order of evaluation is unimportant in the sense that if terms \(b\) and \(c\) are both derived from term \(a\), then there exists a term \(d\) to which both \(b\) and \(c\) can be reduced. In particular, if we reach a term where no further reductions are possible, then it must be the normal form we defined above.

On the other hand, some strategies may loop forever instead of normalizing a term that does in fact possess a normal form. It turns out this never happens with normal-order evaluation: the normalization theorem guarantees that if a normal form exists, then we eventually reach it after repeatedly reducing the leftmost redex.

This is intuitively evident, as at each step we’re doing the bare minimum. Reducing m before n means we ignore arguments to a function until they are needed, which explains other names for this strategy: call by need, or lazy evaluation. Technically, we should also memoize to avoid repeating computations, otherwise it’s just call by name, but this only matters when considering efficiency.

User interface

We’ve saved the worst for last:

#ifdef __HASTE__
main = withElems ["input", "output", "evalB"] $ \[iEl, oEl, evalB] -> do
  let
    prep s = do
      Just button <- elemById $ s ++ "B"
      button `onEvent` Click $ const $ setInput s
    setInput s = do
      Just para   <- elemById $ s ++ "P"
      getProp para "value" >>= setProp iEl "value" >> setProp oEl "value" ""
  setInput "exp"
  mapM_ prep $ words "exp fact quote sur sha"
  evalB `onEvent` Click $ const $ do
    let
      run (out, env) (Left err) =
        (out ++ "parse error: " ++ show err ++ "\n", env)
      run (out, env) (Right m) = case m of
        Blank      -> (out, env)
        Run term   -> (out ++ show (norm env term) ++ "\n", env)
        Let s term -> (out, (s, term):env)
    es <- map (parse line "") . lines <$> getProp iEl "value"
    setProp oEl "value" $ fst $ foldl' run ("", []) es
#else
repl env = do
  ms <- getInputLine "> "
  case ms of
    Nothing -> outputStrLn ""
    Just s  -> case parse line "" s of
      Left err  -> do
        outputStrLn $ "parse error: " ++ show err
        repl env
      Right Blank -> repl env
      Right (Run term) -> do
        outputStrLn $ show $ norm env term
        repl env
      Right (Let v term) -> repl ((v, term):env)

main = runInputT defaultSettings $ repl []
#endif

A Lesson Learned

Until I wrote an interpreter, my understanding of renaming was flawed. I knew that we compute with closed lambda expressions, that is, terms with no free variables, so I had thought this meant I could ignore renaming. No free variables can become bound because they’re all bound to begin with, right?

In an early version of this interpreter, I tried to normalize:

(\f x -> f x)(\f x -> f x)

My old program mistakenly returned:

\x x -> x x

It’s probably obvious to others, but it was only at this point I realized that the recursive nature of beta reductions implies that in the right subtree of a lambda abstraction, a variable may be free, even though it is bound when the entire tree is considered. With renaming, my program gave the correct answer:

\x x1 -> x x1

Booleans, Numbers, Pairs

When starting out with lambda calculus, we soon miss the symbols of Turing machines. We endlessly substitute functions in other functions. They never bottom out. Apart from punctuation, we only see a soup of variable names and lambdas. No numbers nor arithmetic operations. Even computing 1 + 1 seems impossible!

The trick is to use functions to represent data. This is less intuitive than encoding Turing machines on a tape, but well worth learning. The original and most famous scheme is known as Church encoding. We’ll only summarize briefly. See:

Booleans look cute in the Church encoding:

true = \x y -> x
false = \x y -> y
and = \p q -> p q p
or = \p q -> p p q
if = \p x y -> p x y
ifAlt = \p -> p  -- So "if" can be omitted in programs!
not = \p -> p false true
notAlt = \p x y -> p y x

Integers can be encoded in a unary manner:

0 = \f x -> x
1 = \f x -> f x
2 = \f x -> f (f x)
-- ...and so on.

We can perform arithmetic on them with the following:

succ = \n f x -> f(n f x)
pred = \n f x -> n(\g h -> h (g f)) (\u -> x) (\u ->u)
add = \m n f x -> m f(n f x)
sub = \m n -> (n pred) m
mul = \m n f -> m(n f)
exp = \m n -> n m
is0 = \n -> n (\x -> false) true
le = \m n -> is0 (sub m n)
eq = \m n -> and (le m n) (le n m)

The predecessor function is far slower than the successor function, as it constructs the answer by starting from 0 and repeatedly computing the successor. There is no quick way to strip off one layer of a function application.

We can pair up any two terms as follows:

pair = \x y z -> z x y
fst = \p -> p true
snd = \p -> p false

From such tuples, we can construct lists, trees, and so on.

Admittedly, the predecessor function is complicated, probably more so than the a typical Turing machine implementation. However, this is an artifact of the Church encoding. With the Scott encoding, we have a fast and simple predecessor function:

0 = \f x -> x
succ = \n f x -> f n
pred = \n -> n (\x -> x) 0
is0 = \n -> n (\x -> false) true

Instead of unary, we could encode numbers in binary by using lists of booleans. Though more efficient, we lose the elegant spartan equations for arithmetic that remind us of the Peano axioms.

Recursion

Because our interpreter cheats and only looks up a let definition at the last minute, we can recursively compute factorials with:

factrec = \n -> if (is0 n) 1 (mul n (factrec (pred n)))

But we stress this is not a lambda calculus term. If we tried to expand the let definitions, we’d be forever replacing factrec with an expression containing a factrec. We’d never eliminate all the function names and reach a valid lambda calculus term.

Instead, we need something like the Y combinator. The inner workings are described in many other places (I’ve attempted explaining it myself; see also my ZuriHac 2023 talk) , so we’ll content ourselves with definitions, and observing they are indeed lambda calculus terms.

Y = \f -> (\x -> f(x x))(\x -> f(x x))
fact = Y(\f n -> if (is0 n) 1 (mul n (f (pred n))))

Thus we can simulate any Turing machine with a lambda calculus term: we could concoct a data structure to represent a tape, which we’d feed into a recursive function that carries out the state transitions.

Lambda calculus with lambda calculus

Mogensen describes a delightful encoding of lambda terms with lambda terms. If we denote the encoding of a term \(T\) by \(\lceil T\rceil\), then we can recursively encode any term with the following three rules for variables, applications, and lambda abstractions, respectively:

\[\begin{align} \lceil x \rceil &= \lambda a b c . a x \\ \lceil M N \rceil &= \lambda a b c . b \lceil M\rceil \lceil N\rceil \\ \lceil \lambda{x}.M \rceil &= \lambda a b c . c (\lambda x . \lceil M\rceil) \end{align}\]

where \(a, b, c\) may be renamed to avoid clashing with any free variables in the term being encoded. In our code, this translates to:

quote env term = case term of
  Var x   | Just t <- lookup x env -> rec t
          | otherwise              -> f 0 (\v -> App v $ Var x)
  App m n                          -> f 1 (\v -> App (App v (rec m)) (rec n))
  Lam x m                          -> f 2 (\v -> App v $ Lam x $ rec m)
  where
    rec = quote env
    fvs = fv env [] term
    f n g = Lam a (Lam b (Lam c (g $ Var $ abc!!n)))
    abc@[a, b, c] = renameIfNeeded <$> ["a", "b", "c"]
    renameIfNeeded s | s `elem` fvs = newName s fvs
                     | otherwise    = s

With this encoding the following lambda term is a self-interpreter, that is, \(E \lceil M \rceil\) evaluates to the normal form of \(M\) if it exists:

E = Y(\e m.m (\x.x) (\m n.(e m)(e n)) (\m v.e (m v)))

Also, the following lambda term R is a self-reducer, which means \(R \lceil M \rceil\) evaluates to the encoding of the normal form of \(M\) if it exists:

P = Y(\p m.(\x.x(\v.p(\a b c.b m(v (\a b.b))))m))
RR = Y(\r m.m (\x.x) (\m n.(r m) (\a b.a) (r n)) (\m.(\g x.x g(\a b c.c(\w.g(P (\a b c.a w))(\a b.b)))) (\v.r(m v))))
R = \m.RR m (\a b.b)

Unlike the self-interpreter, the self-reducer requires the input to be the encoding of a closed term. See Mogensen’s paper for details.

Scary quotes

Why did we step outside lambda calculus and hard-code the implementation of quote? Maybe we can define it within lambda calculus. Let’s suppose so. Then consider the expression:

quote ((\y.y) x)

If we evaluate the outermost function application, we get:

λa b c.b(λa b c.c(λy a b c.a y))(λa b c.a x)

On the other hand, if we first evaluate the sub-expression ((\y.y) x), then it reduces to quote x, which is:

λa b c.a x

This violates the Church-Rosser theorem. In short, id x = x but quote(id x) /= quote x. Thus quote is not a function, and should be seen as a sort of macro; a labour-saving abbreviation lying outside of lambda calculus.

We named quote after a similar primitive in the Lisp language, which suffers from the same affliction. The Right Way to reify is to sacrifice brevity:

Var=\m.\a b c.a m
App=\m n.\a b c.b m n
Lam=\f.\a b c.c f

Then quote ((\y.y) x) can be expressed in pure lambda calculus as:

App (Lam (\y.Var y)) (Var x)

This is less convenient, but it’s more comprehensible than the raw encoding. Most importantly, we’re back on firm theoretical ground.

A shallow encoding

In Linear-Time Self-Interpretation of the Pure Lambda Calculus, Mogensen describes a shorter self-interpreter: E=\q.q(\x.x)(\x.x).

To encode, we pick two unused variables, say a and b, and prepend \a b. to the term. Then we replace each application m n with a m n and each lambda \x.m with b(\x.m).

In Breaking Through the Normalization Barrier: A Self-Interpreter for F-omega, Matt Brown and Jens Palsberg describe perhaps the shortest possible self-interpreter: E=\q.q(\x.x).

To encode a term for this self-interpreter, we pick a new variable, say i, and prepend \i. to the term. Then we replace each application m n with i m n.

For example, the term succ 0, that is, (\n f x.f(n f x))(\f x.x), becomes:

\i.i(\n f x.i f(i(i n f)x))(\f x.x)

We can avoid threading the i variable throughout the encoding by inserting another layer of abstraction, so we can reuse our notation above:

Var=\x.\i.x
App=\m n.\i.i(i m i)(i n i)
Lam=\f.\i x.f x i
E=\q.q(\x.x)  -- A self-interpreter.

This encoding is shallow in the sense that only a self-interpreter can do anything useful with it.

These days, for clarity, it may be better to use the term self-recognizer as the definition of "self-interpreter" varies.


Ben Lynn blynn@cs.stanford.edu 💡