History of Math Provers


LISP as Extraction Target

Originally, utyped lambda calculus was discovered as an inner language of the space (Curry, Church, 1932). This language was called LISP (McCarthy, 1958), and it was based on cons, nil, eq, atom, car, cdr, lambda, apply, and id. These are parts of inductive types lately known as inductive type constructors. Today untyped lambda calculus is still used as an extraction target for many provers (Idris, F*), and in different domain languages (JavaScript, Erlang).


The first computer-assisted prover AUTOMATH was devised by Nicolaas Govert de Bruijn starting in 1967. Type theory beneath AUTOMATH is typed lambda calculus. AUTOMATH was the first prover based on Curry–Howard correspondence.


Further teardown of inner space language was the ML programming language, founded merely on algebraic datatypes and algebra on higher terms rather than categorical semantic. Lately, it was fixed with categorical methods in CPL (Hagino, 1987) and Charity (Cockett, 1992). Assisted by Morris and Newey, Milner designed Meta Language for the purpose of building LCF in early 70-s. LCF was a predecessor family of automated math provers: HOL88, HOL90, HOL98, and HOL/Isabelle which is now built using Poly/ML.

Fully Automated Provers

Other automated math systems appeared during 80-90s: Mizar (Trybulec, 1989), PVS (Owre, Rushby, Shankar, 1995), ACL2 (Boyer, Kaufmann, Moore, 1996), and Otter (McCune, 1996).


Modern provers (based on consistent Martin-Löf Type Theory, 1972) like Agda, Coq, Lean, F*, and Idris are based on Barendregt and Coquand CoC with different flavors of infinity universe hierarchies and Calculus of Inductive Constructions. Some of them are automated and some are trying to also be general purpose programming languages with proving facilities, like Idris, Coq (, Agda (M-Alonso).

Cubical Type Theory

Further extensions of MLTT theory is Homotopy Types. It is necessary for reasoning about abstract shapes and multidimentional path types. The Cubical approach was chosen in favour of Globular and Simplical due to computability of Voevodsky's univalence axiom.

Infinity Language

From PTS to HTS. We want to have flexible and detachable layers on top of PTS core. Then Sigma for proving. Then well-founded trees or polynomial functors as known as data and record. Then higher path types, interval arithmetic, glue and comp for HIT. Layers are based on different math areas, the only common is the method — category theory.

Extensible Language Design. Encoding of inductive types is based on categorical semantic of compilation to CoC. All other syntax constructions are inductive definitions plugged into the stream parser. AST of the CoC language is also defined in terms of inductive constructions and thus allowed in the macros. The language of polynomial functors (data and record) and core language of the process calculus (spawn, receive, and send) are just a macrosystem over the CoC language, its syntax extensions.

Changable Encodings. In pure CoC we have only arrows, so all inductive type encodings would be Church-encoding variations. Most extended nowadays is Church-Boehm-Berrarducci encoding dedicated to inductive types. Another well known are Scott (laziness), Parigot (laziness and constant-time iterators), and CPS (continuations) encodings.

Proved Categorical Semantic. There was built a math model (using higher-order categorical logic) of encoding, which calculates (co)limits in a category of (co)algebras built with given set of (de)constructors. We call such encoding in honor of Lambek lemma that leads us to the equality of (co)initial object and (co)limit in the categories of (co)algebras. Such encoding works with dependent types and its consistency is proved in Lean model.