Towards Computational UIP in Cubical Agda

TL;DR

This paper explores the possibility of implementing a simplified version of Cubical Agda that retains key features like QITs and functional extensionality while removing univalence, by postulating UIP instead.

Contribution

It analyzes formulations of UIP in Cubical Agda, evaluates their computational rules, and implements a variant compatible with postulated UIP.

Findings

Analysis of UIP formulations and their computation rules.

Implementation of a Cubical Agda variant without Glue types.

Compatibility with postulated UIP demonstrated.

Abstract

Some advantages of Cubical Type Theory, as implemented by Cubical Agda, over intensional Martin-L\"of Type Theory include Quotient Inductive Types (QITs), which exist as instances of Higher Inductive Types, and functional extensionality, which is provable in Cubical Type Theory. However, HoTT features an infinite hierarchy of equalities that may become unwieldy in formalisations. Fortunately, QITs and functional extensionality are both preserved even if the equality levels of Cubical Type Theory are truncated to only homotopical Sets (h-Sets). In other words, removing the univalence axiom from Cubical Type Theory and instead postulating a conflicting axiom: the Uniqueness of Identity Proofs (UIP) postulate. Since univalence is proved in Cubical Type Theory from the so-called Glue Types, therefore, it is known that one can first remove the Glue Types (thus removing univalence) and then set-truncate all equalities (essentially assuming UIP), \`a la XTT. The result is a "h-Set Cubical Type Theory" that retains features such as functional extensionality and QITs. However, in Cubical Agda, there are currently only two unsatisfying ways to achieve h-Set Cubical Type Theory. The first is to give up on the canonicity of the theory and simply postulate the UIP axiom, while the second way is to use a standard result stating "type formers preserve h-levels" to manually prove UIP for every defined type. The latter is, however, laborious work best suited for an automatic implementation by the proof assistant. In this project, we analyse formulations of UIP and detail their computation rules for Cubical Agda, and evaluate their suitability for implementation. We also implement a variant of Cubical Agda without Glue, which is already compatible with postulated UIP, in anticipation of a future implementation of UIP in Cubical Agda.