A cornerstone of theoretical computer science is the Curry-Howard correspondence where formulas are types, proofs are programs, and proof normalization is computation. In this framework we introduce the atomic λμ-calculus, an interpretation of a classical deep inference proof system. It is based on two extensions of the λ-calculus, the λμ-calculus and the atomic λ-calculus. The former interprets classical logic, featuring continuation-like constructs, while the latter interprets intuitionistic deep inference, featuring explicit sharing operators. The main property of the atomic λ-calculus is reduction on individual constructors, derived from atomicity in deep inference. We thus work on open deduction, a deep inference formalism, allowing composition with connectives and with derivations, and using the medial rule to obtain atomicity. One challenge is to find a suitable formulation for deriving a computational interpretation of classical natural deduction. A second design challenge leads us to work on a variant of the λμ-calculus, the ΛμS-calculus, adding streams and dropping names. We show that our calculus has preservation of strong normalization (PSN), confluence, fully-lazy sharing, and subject reduction in the typed case. There are two challenges with PSN. First, we need to show that sharing reductions strongly normalize, underlining that only β, μ-reductions create divergence. Our proof is new and follows a graphical approach to terms close to the idea of sharing. Second, infinite reductions of the atomic calculus can appear in weakenings, creating infinite atomic paths corresponding to finite ΛμS-paths. Our solution is to separate the proof into two parts, isolating the problem of sharing from that of weakening. We first translate into anintermediate weakening calculus, which unfolds shared terms while keeping weakened ones, and preserves infinite reductions. We then design a reduction strategy preventing infinite paths from falling into weakenings.

Cette thèse étudie deux modèles de réalisabilité pour la logique classique construits sur la sémantique des jeux HO, interprétant la logique, l'arithmétique et l'analyse classiques directement par des programmes manipulant un espace de stockage d'ordre supérieur.La non-innocence en jeux HO autorise les références d'ordre supérieur, et le non parenthésage révèle la CPS des jeux HO et fournit une catégorie de continuations dans laquelle interpréter le lambda-mu calcul de Parigot. Deux modèles de réalisabilité sont construits sur cette interprétation calculatoire directe des preuves classiques.Le premier repose sur l'orthogonalité, comme celui de Krivine, mais il est simplement typé et au premier ordre. En l'absence de codage de l'absurdité au second ordre, une mu-variable libre dans les réaliseurs permet l'extraction. Nous définissons un bar-récurseur et prouvons qu'il réalise l'axiome du choix dépendant, utilisant deux conséquences de la structure de CPO du modèle de jeux: toute fonction sur les entiers (même non calculable) existe dans le modèle, et toute fonctionnelle sur des séquences est Scott-continue. La bar-récursion est habituellement utilisée pour réaliser intuitionnistiquement le « double negation shift » et en déduire la traduction négative de l'axiome du choix. Ici, nous réalisons directement l'axiome du choix dans un cadre classique.Le second, très spécifique au modèle de jeux, repose sur des conditions de gain: des ensembles de positions d'un jeu munis de propriétés de cohérence. Un réaliseur est alors une stratégie dont les positions sont toutes gagnantes. / This thesis investigates two realizability models for classical logic built on HO game semantics. The main motivation is to have a direct computational interpretation of classical logic, arithmetic and analysis with programs manipulating a higher-order store.Relaxing the innocence condition in HO games provides higher-order references, and dropping the well-bracketing of strategies reveals the CPS of HO games and gives a category of continuations in which we can interpret Parigot's lambda-mu calculus. This permits a direct computational interpretation of classical proofs from which we build two realizability models.The first model is orthogonality-based, as the one of Krivine. However, it is simply-typed and first-order. This means that we do not use a second-order coding of falsity, and extraction is handled by considering realizers with a free mu-variable. We provide a bar-recursor in this model and prove that it realizes the axiom of dependent choice, relying on two consequences of the CPO structure of the games model: every function on natural numbers (possibly non computable) exists in the model, and every functional on sequences is Scott-continuous. Usually, bar-recursion is used to intuitionistically realize the double negation shift and consequently the negative translation of the axiom of choice. Here, we directly realize the axiom of choice in a classical setting.The second model relies on winning conditions and is very specific to the games model. A winning condition is a set of positions in a game which satisfies some coherence properties, and a realizer of a formula is then a strategy which positions are all winning.