In recent years, the emergence of feature rich and mature interactive proof assistants has enabled large formalization efforts of high-profile conjectures and results previously established only by pen and paper. A medley of incompatible and philosophically diverging logics are at the core of all these proof assistants. Cousineau and Dowek (2007) have proposed the λΠ-calculus modulo as a universal target framework for other front-end proof languages and environments. We explain in this thesis how this particularly simple formalism allows for a small, modular and efficient proof checker upon which the consistency of entire systems can be made to rely upon. Proofs increasingly rely on computation both in the large, as exemplified by the proof of the four colour theorem by Gonthier (2007), and in the small following the SSReflect methodoly and supporting tools. Encoding proofs from other systems in the λΠ-calculus modulo bakes yet more computation into the proof terms. We show how to make the proof checking problem manageable by turning entire proof terms into functional programs and compiling them in one go using off-the-shelf compilers for standard programming languages. We use untyped normalization by evaluation (NbE) as an enabling technology and show how to optimize previous instances of it found in the literature. Through a single change to the interpretation of proof terms, we arrive at a representation of proof terms using higher order abstract syntax (HOAS) allowing for a proof checking algorithm devoid of any explicit typing context for all Pure Type Systems (PTS). We observe that this novel algorithm is a generalization to dependent types of a type checking algorithm found in the HOL proof assistants enabling on-the-fly checking of proofs. We thus arrive at a purely functional system with no explicit state, where all proofs are checked by construction. We formally verify in Coq the correspondence of the type system on higher order terms lying behind this algorithm with respect to the standard typing rules for PTS. This line of work can be seen as connecting two historic strands of proof assistants: LCF and its descendents, where proofs of untyped or simply typed formulae are checked by construction, versus Automath and its descendents, where proofs of dependently typed terms are checked a posteriori. The algorithms presented in this thesis are at the core of a new proof checker called Dedukti and in some cases have been transferred to the more mature platform that is Coq. In joint work with Denes, we show how to extend the untyped NbE algorithm to the syntax and reduction rules of the Calculus of Inductive Constructions (CIC). In joint work with Burel, we generalize previous work by Cousineau and Dowek (2007) on the embedding into the λΠ-calculus modulo of a large class of PTS to inductive types, pattern matching and fixpoint operators.
