How proof assistants are implemented?

Question

What are the main blocks of a proof assistant?

I am just interested in knowing the internal logic of proof checking. For example, topics about graphical user interfaces of such assistants do not interest me.

A similar question to mine has been asked for compilers: https://softwareengineering.stackexchange.com/questions/165543/how-to-write-a-very-basic-compiler

My concern is the same but for proof checking systems.

Answer (Manuel Eberl)

I'm hardly an expert on the matter (I'm only a user of these systems; I don't worry too much about their internals) and this will probably only be a vague partial answer, but the two main approaches that I know of are:

• Dependently-typed systems (e.g. Coq, Lean, Agda) that use the Curry–Howard isomorphism. Statements are just types, and proofs are terms that have that type, so checking the validity of a proof is essentially just a special case of type checking a term. I don't want to say too much about this approach because I don't know too much about it and am afraid I'll get something wrong. Théo Winterhalter linked something in the comments above that may provide more context on this approach.

• LCF-style theorems provers (e.g. Isabelle, HOL Light, HOL 4). Here a theorem is (roughly speaking) an opaque value of type thm in the implementation language. Only the comparatively small 'proof kernel' can create these thm values and all other parts of the system interact with this proof kernel. The kernel offers an interface consisting of various small functions that implement small inference steps such as modus ponens (if you have a theorem A ⟹ B and a theorem A, you can get the theorem B) or ∀-introduction (if you have the theorem P x for a fixed variable x, you can get the theorem ∀x. P x) etc. The kernel also offers an interface for defining new constants. In principle, as long as you can trust that these functions faithfully implement the basic inference steps of the underlying logic, you can trust that any thm value you can produce really corresponds to a theorem in your logic. For LCF-style provers, the answer of what the actual proof is is a bit more difficult to answer because they usually don't build proof terms (e.g. Isabelle has them, but they are disabled by default and not widely used). I think one could say that the history of how the kernel primitives are called constitute the proof, and if one were to record it, it could – in principle – be replayed and checked in another system.

In both cases the idea is that you have a kernel (the type checker in the former case and the inference kernel in the latter) that you have to trust, and then you have a large ecosystem of additional procedures around this to provide more convenience layers. Since they have to interact with the kernel in order to actually produce theorems, however, you do not have to trust that code.

All these different systems have various trade-offs about what parts of the system are in the kernel and what parts are not. In general, I think it is fair to say that the dependently-typed systems tend to have considerably larger kernels than the LCF-based ones (e.g. HOL Light has a particularly small and simple kernel).

There are also other systems that I believe do not fit into these two categories (e.g. Mizar, ACL2, PVS, Metamath, NuPRL) but I don't know anything about how these are implemented.

Answer (Lawrence Paulson)

In the case of LCF, HOL and Isabelle, you'll find an extensive answer to your question in the journal article "From LCF to Isabelle/HOL". (It's open access.)

Most dependently typed systems, such as Coq, are also LCF-style theorem provers, as described in the article and in Eberl's answer. One significant difference is that such calculi incorporate full proof objects, so that one of the objectives of the LCF approach — to save space by not storing proofs — is abandoned. However, the objective of soundness is still met.