Weather characteristics, such as temperature, are often measured along time, leading to the analysis of a functional data set. Functional principal components is then a useful tool to reduce the dimension and to summarize the important aspects of the data. Interpretation of principal components is however not always obvious and alternative methods, such as varimax rotation, are often considered to provide a more informative summary of the data. In this talk, we introduce the concept of structural components which offers an alternative approach to define interpretable components. In particular, we distinguish between two types of components: block-components and difference-components. The former are positive on the whole time period considered and are therefore easier to interpret. In practice, our approach is shown to offer a useful compromise between functional principal components and varimax, as illustrated on weather data.

This is a joint work with Juhyun Park (University of Lancaster) and Theo Gasser (University of Zurich)