A new paper proves that EML trees, which are compositions of the EML (Exp-Minus-Log) elementary function, are universal approximators for continuous functions. This establishes EML trees as a theoretically grounded alternative to neural networks for function approximation, with potential implications for interpretable machine learning and symbolic computation. The proof constructs EML representations of binary operations, polynomials, hyperbolic tangent, and partitions of unity, using them as LEGO-like blocks. Technical challenges with the logarithm's domain are addressed via sign-based decompositions and affine maps.
Background
The EML function (Exp-Minus-Log) is a single binary operator that can express any standard real elementary function through repeated composition. Universal approximation means that a class of functions can approximate any continuous function on a compact set to arbitrary precision. This result parallels the universal approximation theorem for neural networks but uses a different primitive.