Certified Learning under Distribution Shift: Sound Verification and Identifiable Structure

Proposition. Let $f$ be a predictor trained on a distribution $P$ and evaluated on a shifted distribution $Q$. Under verifiable regularity and complexity constraints, the excess risk under shift admits an explicit upper bound determined by a computable shift metric and model parameters. We develop a unified framework in which (i) risk under distribution shift is certified by explicit inequalities, (ii) verification of learned models is sound for nontrivial sizes, and (iii) interpretability is enforced through identifiability conditions rather than post hoc explanations. All claims are stated with explicit assumptions. Failure modes are isolated. Non-certifiable regimes are characterized.

💡 Research Summary

**
The paper tackles the fundamental problem of guaranteeing the performance of a learned predictor when the test distribution differs from the training distribution, a scenario commonly referred to as distribution shift. The authors adopt a rigorous, certification‑oriented perspective: they require every term appearing in the risk bound to be either observable from data or computable from model parameters, thereby eliminating hidden or non‑verifiable quantities that often plague existing domain‑adaptation results.

Problem setting and assumptions

• A predictor f :X→ℝ^m is trained on a source distribution P over X×Y, and evaluated on a target distribution Q.
• The shift is modeled as a covariate shift: Q(dx,dy)=Q_X(dx)·P(dy|x).
• The magnitude of the shift is bounded by a Wasserstein‑1 distance ρ, i.e., W₁(P_X,Q_X) ≤ ρ. The radius ρ may be prescribed by design or estimated from data with a confidence guarantee.
• The loss ℓ(u,y) is L_ℓ‑Lipschitz in its prediction argument, uniformly over y.
• The predictor f is L_f‑Lipschitz on the input domain; L_f can be derived from network weights or from a sound bound‑propagation certificate.
• An interpretability constraint set c_S (e.g., sparsity, additivity, monotonicity, or a symbolic grammar) is imposed; feasibility of this constraint must be certifiable.

These four assumptions (A1‑A4) are deliberately strong, because they are the price paid for a fully verifiable guarantee.

Shift‑robust risk bound
Define the conditional risk function g_f(x)=E_{Y|X=x}