7

Let $h \in A\rightarrow (B\rightarrow C)$
I'm trying to understand the following reduction:

$$\lambda x\in A. \lambda y \in B.(h(x))(y) \\= \lambda x\in A.h(x) \\= h$$

Apparently, this is done by using "Eta-Reduction". Can you help me understand the use of this rule here?

Sriotchilism O'Zaic
  • 2,228
AnnieOK
  • 2,920
  • I really don't know what I'm talking about, but using this link, to the first step take $\lambda y.My$ as $\lambda y\in B.(h(x))(y)$ and perform $\eta$-reduction to on this formula to get $$\lambda x\in A. \lambda y \in B.(h(x))(y) \underset{\eta}{\longrightarrow} \lambda x\in A.h(x)$$ (because $\lambda y\in B.(h(x))(y)\underset{\eta}{\longrightarrow} h(x)$). Similarly $\lambda x\in A.h(x) \underset{\eta}{\longrightarrow}h$. – Git Gud Jun 16 '14 at 13:10
  • I repeat, I don't know what I'm talking about, but to me $$\lambda x\in A. \lambda y \in B.(h(x))(y) \underset{\eta}{\longrightarrow} \lambda y\in B.h(y)\underset{\eta}{\longrightarrow} h$$ feels better somehow. – Git Gud Jun 16 '14 at 13:11
  • @GitGud Your first comment seems to be the right answer, but I can't understand your second comment. You seem to be $\eta$-converting a function with domain $A$ to a function with domain $B$ and then back to a function with domain $A$. Surely, $\eta$-conversion (or anything that deserves to be called "conversion") preserves types. – Andreas Blass Jun 16 '14 at 13:15
  • @AndreasBlass Thank you for the input. I wasn't sure $h$ is a function, my motivation for the second comment was too look at $(h(x))(y)$ as a predicate - (I have no idea if predicate is a $\lambda$-calculus term or not) - first as a predicate depending on $x$, so I eliminate $x$ first. Then a predicate on $y$ remains and I eliminate $y$. – Git Gud Jun 16 '14 at 13:21
  • It took me a few seconds, but then it hit me. Thanks! – AnnieOK Jun 16 '14 at 13:23

2 Answers2

13

Consider some function $F$. This function takes an argument $x$ and yields $F x$.

Now consider the following function $G$: $$\lambda y. F y$$ This function also takes an argument $x$ and yields $F x$.

$G$ and $F$ might be completely different lambda-terms, but it is the case that $Fx=Gx$ for every $x$, so the functions represented by $F$ and $G$ are equal.

For example, consider $$\begin{align} F & = \lambda x.x \\ G & = \lambda x.(\lambda y.y)x \\ \end{align}$$

$F$ is the identity function. $G$ is the function that takes an argument $x$ and applies the identity function to it. $F$ is not the same term as $G$: One is four symbols long, the other ten symbols. But the behavior of $F$ and $G$ is the same, in the sense that for every term $x$, $F x $ has a normal form if and only if $ G x$ does, and if so these normal forms are the same.

When this happens, we say that $F$ and $G$ are $\eta$-equivalent, or that $F$ is the $\eta$-reduction of $G$, or that $G$ is the $\eta$-expansion of $F$, or that $F$ and $G$ are $\eta$-conversions of one another.

MJD
  • 65,394
  • 39
  • 298
  • 580
  • This reminds me of the warning I get in Haskell when I define a function f which takes n parameters and just calls a function g with the same n parameters. – sliiime Feb 22 '24 at 15:07
7

For Eta reduction, see Hendrik Pieter Barendregt, The Lambda Calculus Its Syntax and Semantics (rev ed - 1984), page 63 :

3.3.1. Definition : (i) $\eta:\lambda x.Mx \rightarrow M$ provided $x \notin FV(M)$.

The point of $\beta \eta$-reduction is that it axiomatizes provable equality in the extensional $\lambda$-calculus.

This amount to saying that, whenever the function $M$ is to be used, any argument will simply be passed to $M$. Hence, this function is, in essence, equal to $M$. One caveat of $\eta$-abstraction is that a name conflict must be avoided. What this means is if $x$ is a free variable (i.e., not scoped in a function abstraction) in $M$ itself, then this $\eta$-reduction is not valid, as it would be changing the scope of $x$. In order to avoid this, an $\alpha$-reduction must be performed before the $\eta$-reduction.

Community
  • 1
Mauro ALLEGRANZA
  • 94,169