We study an abstract model for the coevolution between mutating viruses and the adaptive immune system. In sequence space, these two populations are localized around transiently dominant strains. Delocalization or error thresholds exhibit a novel interdependence because immune response is conditional on the viral attack. An evolutionary chase is induced by stochastic fluctuations and can occur via periodic or intermittent cycles. Using simulations and stochastic analysis, we show how the transition between these two dynamic regimes depends on mutation rate, immune response, and population size.Evolution is commonly pictured as a dynamic process on a fitness landscape in sequence space. In general, this landscape depends not only on the genotype but varies dynamically as a function of the environment and coevolving interaction partners [1]. Prominent biological examples are the coevolutionary dynamics between the adaptive immune system and virus populations such as HIV [2, 3] or influenza [4], or between bacteria and their phages [5]. Continuous evolutionary innovations allow the virus to transiently escape immune suppression, triggering subsequent adaptations of the immune system. These dynamics can lead to coevolutionary cycles, which have been generally described in two different forms [4, 6]: either as an intermittent series of quasistationary states connected by stochastic jumps, or as periodic and largely deterministic oscillations. From a modeling perspective, this highly complex process is determined by three main features [7]. First, mutation rates are high and populations are large, which implies large genetic heterogeneity within the populations [8]. This has often been pictured in terms of broad quasispecies distributions around peaks in the fitness landscape [9, 10]. At the same time, continuous adaption and coevolutionary arms races are driven by strong ecological interactions [6, 11]. These modulate effective fitness landscapes [2,12,13] and lead to nontrivial nonlinear population dynamics. Finally, stochastic effects in finite populations become especially pronounced whenever the first two issues are relevant at the same time [11,[14][15][16].Here, we offer a synthetic perspective on these processes. In our model [see Fig. 1(a)], we consider a population of N viruses represented by their genotypes (binary sequences of length L and frequency x i ) and replication rates r i = 1. A small number n of these genotypes corresponding to particularly virulent strains have a fitness advantage α over the unit baseline, giving r i = 1 + α for i = p, q, . . .. Offspring sequences undergo mutations with per-base rate µ x . In the absence of immune suppression, and in the stationary state, the viral population localizes as so-called quasispecies around any of the fittest genotypes, provided the mutation rate is smaller than Eigen's error threshold µ c ≈ ln Fig. 1(a), right]. In the deterministic limit (N → ∞), our model is described by [20]where i and j run over all 2 L sequences. The fitness advantage α of ...