Average Gradient Outer Product in kernel regression provably recovers the central subspace for multi-index models

The paper's core claim is sample efficiency: recovering low-dimensional structure requires fewer samples than predicting the full high-dimensional function. This is a gap the summary mentions but doesn't emphasize. The practical implication is that you can extract interpretable subspaces from models trained on limited data without retraining.

This connects directly to the refinery optimization work from earlier today. That piece highlighted the trust gap between mathematically sound algorithms and real-world deployment, where practitioners need to validate and contextualize model outputs before acting. AGOP offers a complementary tool: a way to extract and verify the low-dimensional structure a kernel model has learned, making the learned function more interpretable and auditable. If operators can see which input dimensions actually drive predictions, they gain confidence in the model's reasoning, not just its accuracy. It's part of the same pattern we've seen across recent coverage (the unlearning verification work, the failure prediction framework): practitioners need visibility into what models have actually learned before deployment.

If follow-up work applies AGOP to real high-dimensional datasets (materials science, genomics, finance) and shows that the recovered subspace matches domain expert intuition or known physical constraints, that confirms the method's practical value. If the sample complexity bounds hold empirically but the recovered subspace remains opaque or contradicts domain knowledge, the theoretical result remains isolated from deployment.