On a Joint Simultaneous Learning of Relevant Feature Subsets and Subspaces in Regression-like Problems

The paper extends Entropy-Optimal Manifold Clustering into Entropy-Optimal Manifold Regression (EOMR), jointly identifying relevant feature subsets and subspaces in nonlinear and nonstationary regression. The abstract reports linear-scaling iteration and memory complexity. On Lorenz-96 dynamics at forcing values F=8 and F=12, and on Hasegawa-Wakatani tokamak-plasma data, EOMR is compared with gradient-boosted random forests, deep neural networks, and TabPFN v0.3. It reportedly achieves substantially lower RMSE and lower model complexity. In one plasma example, it reduces the leading EOF dynamics to an eight-parameter linear causal autoregressive process.