Skip to content

Features and data analysis

To compare many persistence results, we often turn each one into numbers indexed by common coordinates. The choice of coordinates, window and normalization determines what those numbers mean. Reading scalar fields in their original coordinates starts with that interpretation: inspect a landscape layer or a section of a higher-dimensional field, then assemble values whose coordinates remain explicit. A feature vector is useful because of the information it retains, which also makes its losses part of the explanation.

Repeated computations introduce a second question: how do their answers vary? Analyzing ensembles of persistence results aligns the coordinates and features before summarizing replicates. An ensemble here is a collection of repeated results with a stated relationship to the experiment. The guide separates descriptive variation from confidence claims, whose meaning depends on the sampling assumptions and whether a statement concerns one point or an entire domain.

Sometimes we want the locations at which we measure a feature to adapt to a task. For the generalized rank invariant landscape, or GRIL, a probe specifies where and how the landscape is queried. Region queries and GRIL features explain the quantities from which such features are built. Adapting probes to a task then introduces choices about the training objective, the parameters that may vary and which data remain separate for evaluation. A change in the underlying discrete structure can require different reasoning from a small change in a continuous parameter.

If your immediate task is to inspect a supplied feature table, distance table or embedding, comparing results through declared correspondences provides the adjoining route. It makes the sample identities and coordinate correspondence explicit before using shared scales. A visual grouping can then be traced to the values that produced it, without treating the drawing as a new mathematical guarantee.