Due to the representation of the continuous time series, there is no explicit machinery for addressing discrete events. Consider for example, a binary switching event such as the output of a classifier. This could be the output of a speech detection system which determines if the user speaking or not-speaking. In a time series, such a phenomenon will typically be represented as a step function. In the ARL, this could be potentially harmful due to the projection on the eigenspace. The onset of a step function is a highly non-linear and localized transition and eigenspace projection will unnaturally smooth it into a sigmoid like curve. Thus, one would require an alternative way of representing sudden changes which might involve fundamental changes to the time series representation.