Information Geometry and Dimensional Reduction for Statistical Structural Features of Paper

Information geometry provides a metric on spaces of probability density functions. Here we apply it to the space of trivariate Gaussian distributions of joint variation among the areal density variables for pixels and their first and second neighbours, from radiographs and simulations. At a pixel scale of one millimetre these distributions can pick up essential structural features including flocculation intensity and scale. We do this by applying the technique of dimensionality reduction to large mixed data sets of samples and the results show promise for classification, including extraction of groupings that represent different former types. This kind of analysis could be valuable in evaluating trials, comparing different installations of similar formers and for identifying anomalous behaviour.