Lambda Calculus

This is a note and simple tutorial of lambda-calculus. I studied lambda-calculus for my final project in the Department of Computer Science.

The Untyped $\lambda$-calculus

$\lambda$-calculus is equivalent to Turing machines in computational power.

Conversion rules

The $\lambda$-calculus provides several conversion rules for transforming one $\lambda$-calculus into an equivalent one. The conversion rules are defined as follows:

  • $\beta$-conversion:
    $$(\lambda x.e)e’ \Leftrightarrow e[e’/x]$$
    Just subsitute $x$ with $e’$ in $e$ like normal mathematic functions.

  • $\alpha$-conversion:
    $$\lambda x.e \Leftrightarrow \lambda y.e[y/x] \quad y \notin
    FV(e)$$

  • $\eta$-conversion:
    $$\lambda x.(e \space x) \Leftrightarrow e \quad x \notin FV(e)$$
    Just eliminate a redundant $\lambda$-abstraction

  • $\delta$-conversion:
    This rule define conversion of built-in constrants and functions, for example,
    $$(times \space 3 \space 4) \Leftrightarrow 12$$

We divide the set of $\lambda$-terms into $\alpha$-equivalence classes; this means that any two $\lambda$-terms that can be transfromed into one another via $\alpha$-conversion are in the same equivalence class.

Reduction rules

The conversion rules express that two terms are equivalent, reduction rules are used to evaluate a term. There two reduction rules.

  • $\beta$-reduction:
    $$(\lambda x.e)e’ \Rightarrow e[e’/x]$$

  • $\delta$-reduction:
    $$(times \space 3 \space 4) \Rightarrow 12$$

Evaluation strategies

Serveral different evaluation strategies for the $\lambda$-calculus which have been studied over the years by programming language designers and theorists.

For example, consider the term:
$$(\lambda x.x) \space ((\lambda x.x) \space (\lambda z . (\lambda x.x)z))$$

And we have a definition of identity function, which just do nothing but return its argument:
$$id = \lambda x.x$$

Thus our target can be rewrite to a more readable form:
$$id \space (id \space (\lambda z.id \space z))$$

  • Full $\beta$-reduction
    Under this strategy, we might choose, for example, bo begin with the innermost redex(reducible expression), then do the one in the middle, then the outermost:
    $$\begin{equation}
    \quad id \space (id \space (\lambda z.\underline{id \space z})) \
    \rightarrow id \space \underline{(id \space (\lambda z.z)}) \
    \rightarrow \underline{(id \space (\lambda z.z)} \
    \rightarrow \lambda z.z
    \end{equation}$$

  • Normal order strategy
    Under this strategy, the leftmost, outermost redex is always reduced first:
    $$\begin{equation}
    \quad \underline{id \space (id \space (\lambda z.id \space z))} \
    \rightarrow \underline{id \space (\lambda z.id \space z)} \
    \rightarrow \lambda z.\underline{id \space z} \
    \rightarrow \lambda z.z
    \end{equation}$$

  • Call-by-name
    The call-by-name strategy is yet more restrictive, allowing no reductions inside abstractions.
    $$\begin{equation}
    \quad \underline{id \space (id \space (\lambda z.id \space z))} \
    \rightarrow \underline{id \space (\lambda z.id \space z)} \
    \rightarrow \lambda z.id \space z
    \end{equation}$$

  • Call-by-value
    This strategy is like to evaluate the parameters first.
    $$\begin{equation}
    \quad id \space \underline{(id \space (\lambda z.id \space z))} \
    \rightarrow \underline{id \space (\lambda z.id \space z)} \
    \rightarrow \lambda z.id \space z
    \end{equation}$$

Programming in the $\lambda$-calculus

  • Church Booleans
    Boolean values is a language feature tahat can easily be encoded in the $\lambda$-calculus. Define the terms tru and fls as follows:
    $$\begin{split}
    tru &= \lambda t. \lambda f. t; \
    fls &= \lambda t. \lambda f. f;
    \end{split}$$

    According to the definition, the function tru accepts two arguments and returns the first one while fls returns the second one.
    Now, we can define some boolean operatos like logical conjunction as functions:
    $$\begin{split}
    and &= \lambda b. \lambda c. b \space c \space fls; \
    or &= … \
    not &= …
    \end{split}$$

  • Church Numerals
    We define the Church numerals $c_0, c_1, c_2, etc.$ as follows:
    $$\begin{split}
    c_0 &= \lambda s. \lambda z.z; \
    c_1 &= \lambda s. \lambda z. s \space z; \
    c_2 &= \lambda s. \lambda z. s \space (s \space z);
    \end{split}$$

    Each number $n$ is represented by a combinator $c_n$ that takes two argument, $s$ and $z$(for “successor” and “zero”). Then, we can define the successor function on Church numerals as follows:
    $$succ = \lambda n. \lambda s. \lambda z. s \space (n \space s \space z);$$
    So,
    $$\begin{split}
    succ \space c_1 &= \lambda n. \lambda s. \lambda z. s \space (n \space s \space z) \space c_1 \
    &= \lambda s. \lambda z. s \space ((\lambda s. \lambda z. s \space z) \space s \space z) \
    &= \lambda s. \lambda z. s \space (s \space z) \
    &= c_2
    \end{split}$$

  • Enrich the calculus

  • Resursion

The Simply Typed $\lambda$-Calculus

Typed $\lambda$-calculi are like untyped one, except that every bound identifier is given a type, which means each variables and functions are typed. The typed $\lambda$-calculus introduce the notion of type correctness of a term, which one would like to check before trying to evaluate the term.

Introduction

We are going to consider the successor function as an example. A successor is a function that you put n than you got n + 1 as result, here is a sample practice in c language:

1
2
3
int succ(int n) {
return n + 1;
}

In Typed $\lambda$-calculus we define the successor function as:
$$succ = \lambda n:int.n+1$$

We can see that the variable n is given a type $int$.

Now, we assume the addition operator + is typed as:
$$+: int \times int \rightarrow int$$

With the two definitions above, we could obtain the type:
$$succ: int \rightarrow int$$

Where $\rightarrow$ is the function type constructor and $\times$ the tuple type constructor used for multiple function arguments. We could then define:
$$twice = \lambda f: int \rightarrow int.\lambda x: int.f(f \space x)$$

and the term:
$$(twice \space succ) \space 4$$

would be type-correct and result in 6. On the other hand,
$$twice \space 7$$

would not be type-correct, since the type of 7 is $int$ rather than $$int \rightarrow int$$