Sample Complexity of Scientific Discovery: PAC Learnability of Compositional Function Trees

arXiv:2606.29331v1 Announce Type: new Abstract: Scientific discovery via symbolic regression is often viewed as statistically and computationally intractable because the hypothesis space of expressions grows combinatorially with depth. This paper revisits the statistical side through the lens of PAC learning, focusing on compositional function trees built from a finite vocabulary of smooth operators (e.g., {+,imes,sin,exp} and affine maps). We prove that the relevant generalization quantity, Rademacher complexity, hence the excess risk, does not necessarily blow up exponentially with the number of distinct symbolic structures, but is controlled by (i) the depth d and (ii) the Lipschitz constants of the base operators along the composed computation graph. Concretely, under mild Lipschitz conditions on operators and bounded affine leaves, a finite-union bound over a vocabulary of size K=|H_{base}| together with Maurer-type vector contraction yields mathfrak{R}n(H{comp}^{d}) leq (Kbsqrt{2}L)^{d-1}mathfrak{R}n(H{comp}^{1}) with arity bound b; corresponding high-probability risk bounds scale as O(L^{d}/sqrt{n}) when K,b=O(1) and mathfrak{R}n(H{comp}^{1})=O(n^{-1/2}). We complement the theory with a modular codebase that trains differentiable operator trees (not MLPs) on synthetic "physics-like" targets of controlled depth and shows that the empirical generalization gap correlates positively with the predicted complexity term (widehat{L}^{d})/sqrt{n}.

Source: arXiv cs.LG | 2026-06-30