In this thesis, I study fluctuations and rare events of time-additive observables of discrete-time Markov chains on finite state spaces. The observable of interest is the mean node connectivity visited by a random walk running on instances of an Erdős-Rényi (ER) random graph. I implement and analyze the Adaptive Power Method (APM) which converges to the driven process, a biased random walk defined through a control parameter that simulates trajectories corresponding to rare events of the observable in the original dynamics. The APM demonstrates good convergence and accurately produces the desired quantities from a single trajectory. Due to the bulk-dangling-chain structure in the ER graph, the driven process seems to undergo a dynamical phase transition (DPT) for infinitely large graphs, meaning the behavior of the trajectories changes abruptly as the control parameter is varied. Observations show that the random walk visits two distinct phases, being de-localized in the bulk or localized in the chain. Through two simpler models capturing the bulk-dangling-chain property of the ER graph I study how the DPT occurs as the graph size increases. I observe that the trajectories of the driven process near the transition show intermittent behavior between the two phases. The diverging time scale of the DPT is found to be the average time that the random walk spends in a phase before it transitions to the other one. On the ER graph the trajectories are also intermittent but the form of the time scaling remains open due to computational limits on the graph size.

De nombreux systèmes complexes soumis à un ajout continu d'énergie réagissent à cet ajout par une accumulation de tension au cours du temps, interrompue par de soudaines libérations d'énergie appelées avalanches. Récemment, il a été remarqué que plusieurs propriétés élémentaires de la dynamique d'avalanche sont issues de processus de relaxation ayant lieu à une échelle microscopique, processus qui sont négligés dans la plupart des modèles. Lors de ma thèse, j'ai étudié deux modèles classiques d'avalanches, modifiés par l'ajout d'une forme de relaxation la plus simple possible. Le premier système est une interface viscoélastique tirée à travers un milieu désordonné. En champ moyen, nous prouvons que l'interface a un comportement périodique caractérisé par une nouvelle échelle temporelle (émergente), avec des avalanches qui touchent l'ensemble du système. Le calcul semi-analytique de la force de friction agissant sur la surface donne des résultats compatibles avec les expériences de friction classique. En dimension finie (2D), les événements touchant l'ensemble du système (trouvés en champ moyen) deviennent localisés, et les simulations numériques donnent des résultats en bon accord avec plusieurs caractéristiques importantes des tremblements de terre, tant qualitativement que quantitativement. Le second système incluant également une forme très simple de relaxation est un modèle jouet d'avalanche : c'est la percolation dirigée. Dans notre étude d'une variante non-markovienne de la percolation dirigée, nous avons observé que la classe d'universalité était modifiée mais seulement partiellement. En particulier, un exposant change de valeur tandis que plusieurs relations d'échelle sont préservées. Cette idée d'une classe d'universalité étendue, obtenue par l'ajout d'une perturbation non-markovienne offre des perspectives prometteuses pour notre premier système. / Many complex systems respond to a continuous input of energy by an accumulation of stress over time, interrupted by sudden energy releases called avalanches. Recently, it has been pointed out that several basic features of avalanche dynamics are induced at the microscopic level by relaxation processes, which are neglected by most models. During my thesis, I studied two well-known models of avalanche dynamics, modified minimally by the inclusion of some forms of relaxation. The first system is that of a viscoelastic interface driven in a disordered medium. In mean-field, we prove that the interface has a periodic behaviour (with a new, emerging time scale), with avalanche events that span the whole system. We compute semi-analytically the friction force acting on this surface, and find that it is compatible with classical friction experiments. In finite dimensions (2D), the mean-field system-sized events become local, and numerical simulations give qualitative and quantitative results in good agreement with several important features of real earthquakes. The second system including a minimal form of relaxation consists in a toy model of avalanches: the Directed Percolation process. In our study of a non-Markovian variant of Directed Percolation, we observed that the universality class was modified but not completely. In particular, in the non-Markov case an exponent changes of value while several scaling relations still hold. This picture of an extended universality class obtained by the addition of a non-Markovian perturbation to the dynamics provides promising prospects for our first system.