Compression Is All You Need: Modeling Mathematics

The talk will exposit a recent eponymous arXiv posting with coauthors Vitaly Aksenov, Eve Bodnia, and Mike Mulligan. The approach is to think like a physicist and model a seemingly complex bit of reality: mathematics, by a simple toy model where exact computations can be carried out and then compared with observation. The models are finitely generated monoids and the data is derived from MathLib, a large Lean-based repository. The hierarchical nature of definitions and lemmas in math is modeled by adding redundant generators to the monoids – think of the powers of 10 within the natural numbers which support place notation. Place notation confers an exponential compression of how we describe numbers; exploration of MathLib shows that this theme persists to (human) mathematics writ large. We hope that the observables we describe will help our agents navigate to interesting mathematical destinations.

Zoom: https://harvard.zoom.us/j/99308274895?pwd=KhVOYBUfBvKQMuBkDWhe346Un2e7zv.1

Password: 850429