$L^p$-Extremal Teichmüller mappings between Riemann surfaces are diffeomorphisms

What the math gives to ML

The paper provides a principled distortion functional for orientation-preserving two-dimensional maps, rather than a generic penalty on Jacobian size. Its key transferable asset is the singular barrier at vanishing Jacobian: minimizing finite-p conformal distortion penalizes both anisotropic stretching and local fold formation, while the theorem states that the continuum minimizer is a unique diffeomorphism in a prescribed homotopy class. This can be transferred to neural coordinate maps, planar spatial transformers, and learned image-registration modules by replacing ad hoc determinant penalties with sampled quasiconformal distortion and continuation in p. The theorem does not automatically certify a finite neural model as globally invertible, so experiments must measure fold rate, minimum determinant, distortion, and task loss explicitly.