arXiv:2609.15592v2 Announce Type: replace Abstract: We study regression from noisy evaluations of uniformly sampled subsets containing exactly \(k\) items from a ground set of size \(d\), a setting motivated by combination screening. The fixed-cardinality constraint couples membership coordinates and prevents a direct use of product-domain notions of interaction and smoothness. Using the harmonic decomposition of the Johnson graph, we define an intrinsic Johnson--Sobolev scale and determine the exact dimensions of low-order interaction spaces. We establish finite-sample risk bounds for harmonic projection and a nonasymptotic minimax characterization when the Sobolev energy budget is bounded relative to the noise variance proxy. We also study sparsity when responses depend on few items, quantifying the reduced interaction dimension and the effects of unknown support and rare configurations. When the empirical Gram matrix is well conditioned, stable least squares removes the design fluctuation of the component already in the fitted space. A rank-deficiency analysis identifies the components left unresolved by random sampling. Combined with a centered completion estimator, it yields minimax bounds matching up to constants whenever \(\min(k,d-k)\) and the smoothness order are fixed, without restricting the signal-to-noise ratio. Numerical experiments distinguish observation noise, design fluctuation, and incomplete coverage; NCI-ALMANAC drug-pair scores provide a real-data illustration.