The most obvious Hopf algebras came from considering the structure that appears when one considers the real valued functions on a group, and the enveloping algebra spanned by the left invariant (differential) operators on the group. The group injects into its enveloping algebra of operators.
Within the framework an important subclass are the combinatorial algebras – these are graded – and provide natural locally finite linear bases for the functions on the associated group. A particularly important example is the group of paths (or streams). In this case the Enveloping algebra is the tensor series, and the commutative algebra is the shuffle algebra. It provides an alternative way of formally describing a path segment or stream – via its evaluation on these features, instead of giving its value at the ends of the segment. This top down approach is fundamentally different to the evaluation at points approach of Kolmogorov and Doob and opens up the way to describing much rougher paths, and generalizing the notion of smooth path in a way that allows calculus.
From this starting point Rough path theory aims to build an effective calculus that can model the interactions between complex oscillatory (rough) evolving systems. At its mathematical foundations, it is a combination of analysis blended with algebra that goes back to LC Young, and to KT Chen. Key to the theory is the essential need to incorporate additional non-commutative structure into areas of mathematics we thought were stable.
The approach connects to modern data science in two ways. First it models multimodal data streams very naturally, and secondly, the principle of mapping to a curve into a linear space where the dual is the space of functions on the curve is a fundamental idea providing a “perfect” kernel approach (because it linearises polynomial functions). It allows a direct connection with standard machine learning thinking.
The approach is fundamental, and descriptions via time series, and non-adaptive wavelet expansions are doomed to be relatively ineffective since they are linear features of the path. Classic results, by Clark, Cameron and Dickinson, demonstrate that a nonlinear approach to the data is often essential. Rough path theory lives up to this challenge and can be viewed as providing more efficient ways of approximately describing complex data; approaches that, after penetrating the basic ideas, are computationally tractable and lead to new scalable ways to regress, classify, and learn functional relationships from data.
One non-mathematical application that is already striking is the use of signatures on a daily basis (linked with deep learning) in the online recognition of Chinese Handwriting on mobile phones. Recognizing actions in video data is another highly competitive application.