It is well known that, under some aperiodicity and irreducibility conditions, the number of occurrences of local patterns within a Markov chain (and, more generally, within the languages generated by weighted regular expressions/automata) follows a Gaussian distribution with both variance and mean in Θ(n). By contrast, when these conditions no longer hold, it has been observed that the limiting distribution may follow a whole diversity of distributions, including the uniform, powerlaw or even multimodal distribution, arising as tradeoffs between structural properties of the regular expression and the weight/probabilities associated with its transitions/letters. However these cases only partially cover the full diversity of behaviors induced within regular expressions, and a characterization of attainable distributions remained to be provided.In this article, we constructively show that the limiting distribution of the simplest foreseeable motif (a single letter!) may already follow an arbitrarily complex continuous distribution (or cadlag process). We also give applications in random generation (Boltzmann sampling) and bioinformatics (parsimonious segmentation of DNA).