AF: Small: Algorithms for Inference

California Institute Of Technology, Pasadena CA

Investigators

Abstract

The first focus of this award is the problem of inferring causal relationships. This problem is essential to many statistical applications. Generally speaking, causation can be inferred only by active intervention, through controlled experiment. However such experiments may be impossible or else practically or morally infeasible: for instance, in predicting potential effects regulations or laws on medical, educational and economic outcomes, or, in predicting whole-ecosystem effects of human activity. The starting point for Dr. Schulman's research in this area is Pearl's theory of Structured Causal Models (SCM) which, allows, in special circumstances, "identification" of a causal relationship (as opposed to a statistical correlation) from passive observation. The proposed research aims, in the first place, to expand the above special class of circumstances---and thereby make the theory more widely applicable---by using a relaxed but still useful notion of "weak identification" of causal relationships. The relaxed notion is more robust, and enables valid inference even if the posited SCM is slightly inaccurate. In the second place, the research aims to provide efficient and numerically stable algorithms for weak identification from empirical data. The second focus of this award, again in computational statistics, is the representation of a large data set (considered as an empirical measure) by a much smaller data set, in such a way that for a specific family of integrals, all integrals of the measure are approximately preserved. This work encompasses two separate application areas. The first concerns clustering and related high dimensional data analysis problems. Here the compressed data set is known as an epsilon-approximation or core-set of the input measure. A particular focus is on "underclustering", namely, preparation of core-sets for clustering in a normed space, before the norm has been specified. The technical tools needed in this application have to do with recently developed ideas about the "total sensitivity" of the family of integrals, as well as with, on the algorithmic side, bicriteria approximations. The second application area concerns signal processing (or approximation theory) on compact groups. Here the methods draw on representation theory, the classical theory of the moment problem, and convex geometry. This award will be used to train graduate students and postdoctoral fellows in research in algorithms, statistics, and underlying mathematical topics in algebra and geometry.