sp-eta-postpone/README.org
2025-02-03 22:47:49 -05:00

42 lines
2.2 KiB
Org Mode

* Church Rosser, surjective pairing, and strong normalization
This repository contains a mechanized proof that the lambda calculus
with beta and eta rules for functions and pairs is in fact confluent
for the subset of terms that are "strongly $\beta$-normalizing".
Our notion of $\beta$-strong normalization, based on Abel's POPLMark
Reloaded survey, is inductively defined
and captures a strict subset of the usual notion of strong
normalization in the sense that ill-formed terms such as $\pi_1
\lambda x . x$ are rejected.
* Joinability: A transitive equational relation
The joinability relation $a \Leftrightarrow b$, which is true exactly
when $\exists c, a \leadsto_{\beta\eta} c \wedge b
\leadsto_{\beta\eta} c$, is transitive on the set of strongly
normalizing terms thanks to the Church-Rosser theorem proven in this
repository.
** Joinability and logical predicates
When building a logical predicate where adequacy ensures beta strong
normalization, we can then show that two types $A \Leftrightarrow B$
share the same meaning if both are semantically well-formed. The
strong normalization constraint can be extracted from adequacy so we
have transitivity of $A \Leftrightarrow B$ available in the proof that
the logical predicate is functional over joinable terms.
** Joinability and typed equality
For systems with a type-directed equality, the same joinability
relation can be used to model the equality. The fundamental theorem
for the judgmental equality would take the following form: $\vdash a
\equiv b \in A$ implies $\vDash a, b \in A$ and $a \Leftrightarrow b$.
Note that such a result does not immediately give us the decidability
of type conversion because the algorithm of beta-eta normalization
may identify more terms when eta is not type-preserving. However, I
believe Coquand's algorithm can be proven correct using the method
described in [[https://www.researchgate.net/publication/226663076_Justifying_Algorithms_for_be-Conversion][Goguen 2005]]. We have all the ingredients needed:
- Strong normalization of beta (needed for the termination metric)
- Church-Rosser for $\beta$-SN terms (the disciplined expansion of
Coquand's algorithm preserves both SN and typing)