Fabio Zanasi

Relative entropy is a fundamental class of distances between probability distributions, with widespread applications in probability theory, statistics, and machine learning. In this work, we study relative entropy from a categorical perspective, viewing it as a quantitative enrichment of categories of stochastic matrices. We consider two natural monoidal structures on stochastic matrices, given by the Kronecker product and the direct sum. Our main results are complete axiomatisations of Kullback-Leibler divergence and, more generally, of Rényi divergences of arbitrary order, for each such structure. Our axiomatic theories are formulated within the framework of quantitative monoidal algebra, using a graphical language of string diagrams enriched with quantitative equations.

We develop layered monoidal theories – a generalisation of monoidal theories combining formal descriptions of a system at different levels of abstraction. Via their representation as string diagrams , monoidal theories provide a graphical formalism to reason algebraically about information flow in models across different fields of science. Layered monoidal theories allow mixing several monoidal theories (together with translations between them) within the same string diagram, while retaining mathematical precision and semantic interpretability. We develop the mathematical foundations of layered monoidal theories, as well as providing several instances of our approach, including digital and electrical circuits, quantum processes, chemical reactions, concurrent processes, and probability theory.

Interacting Hopf Algebras - la théorie des systèmes linéaires: Dans cette thèse, on présente la théorie algébrique IH par le biais de générateurs et d’équations.Le modèle libre de IH est la catégorie des sous-espaces linéaires sur un corps k. Les termes de IH sont des diagrammes de cordes, qui, selon le choix de k, peuvent exprimer différents types de réseaux et de formalismes graphiques, que l’on retrouve dans des domaines scientifiques divers, tels que les circuits quantiques, les circuits électriques et les réseaux de Petri. Les équations de IH sont obtenues via des lois distributives entre algèbres de Hopf – d’où le nom “Interacting Hopf algebras” (algèbres de Hopf interagissantes). La caractérisation via les sous-espaces permet de voir IH comme une syntaxe fondée sur les diagrammes de cordes pour l’algèbre linéaire: les applications linéaires, les espaces et leurs transformations ont chacun leur représentation fidèle dans le langage graphique. Cela aboutit à un point de vue alternatif, souvent fructueux, sur le domaine.On illustre cela en particulier en utilisant IH pour axiomatiser la sémantique formelle de circuits de calculs de signaux, pour lesquels on s’intéresse aux questions de la complète adéquation et de la réalisabilité. Notre analyse suggère un certain nombre d’enseignements au sujet du rôle de la causalité dans la sémantique des systèmes de calcul.

Modern deep learning models require immense computational resources, motivating research into low-precision training. Quantised training addresses this by representing training components in low-bit integers, but typically relies on discretising real-valued updates. We introduce an alternative approach where the update rule itself is discrete, avoiding the quantisation of continuous updates by design. We establish convergence guarantees for a general class of such discrete schemes, and present a multinomial update rule as a concrete example, supported by empirical evaluation. This perspective opens new avenues for efficient training, particularly for models with inherently discrete structure.

Post-hoc methods in Explainable AI (XAI) elucidate black-box models by identifying input features critical to the model's decision-making. Recent advancements in these methods have facilitated the generation of logic-based explanations that capture interactions among input features. However, these techniques often encounter critical limitations, notably the inability to ensure logical consistency and fidelity between generated explanations and the model's actual decision-making processes. Such inconsistencies jeopardize the reliability of explanations particularly in high-risk domains. To address this gap, we introduce a novel, theoretically rigorous approach rooted in category theory. Specifically, we propose the concept of an explaining functor, which preserves logical entailment structurally between the explanations and the decisions of black-box models. By establishing a categorical framework, our method guarantees the coherence and accuracy of extracted explanations, thus overcoming the common pitfalls associated with heuristic-based explanation methods. We demonstrate the practical efficacy of our theoretical contributions through two synthetic benchmarks that highlight significant reductions in contradictory and unfaithful explanations. Our experiments show how our framework can provide mathematically grounded, compositional, and coherent explanations.

String diagrams are pictorial representations for morphisms of symmetric monoidal categories. They constitute an intuitive and expressive graphical syntax, which has found application in a very diverse range of fields including concurrency theory, quantum computing, control theory, machine learning, linguistics, and digital circuits. Rewriting theory for string diagrams relies on a combinatorial interpretation as double-pushout rewriting of certain hypergraphs. As previously studied, there is a `tension' in this interpretation: in order to make it sound and complete, we either need to add structure on string diagrams (in particular, Frobenius algebra structure) or pose restrictions on double-pushout rewriting (resulting in 'convex' rewriting). From the string diagram viewpoint, imposing a full Frobenius structure may not always be natural or desirable in applications, which motivates our study of a weaker requirement: commutative monoid structure. In this work we characterise string diagram rewriting modulo commutative monoid equations, via a sound and complete interpretation in a suitable notion of double-pushout rewriting of hypergraphs.