<p>Extreme analysis, via e.g., GEV, was developed to deal with univariate time series, and is very difficult to extend beyond that dimension. Here we explore a different method, the archetypal analysis, which focuses on multivariate extremes. The method seeks to approximate the convex hull in high-dimensional state space, by identifying corners representing "pure" types, i.e. archetypes. The method, encompasses, in particular, the virtues of EOFs and clustering. The method is presented with a new manifold-based optimization algorithm, and applied to a number of atmospheric fields, including SST and SLP gridded data. The application to SST, in particular, reveals important features related to SST extremes. The strengths and weaknesses of the method and possible future perspectives will be discussed.</p>
High-dimensional data compression via PHLCT
