Robust Instrumental Variables: Sharp Rates and Inference under Adversarial Contamination

What the math gives to ML

The paper's transferable asset is a robust aggregation principle: replace ordinary sample averages with quantile-winsorized averages, obtaining error that scales as the adversarial fraction to the power 1-1/m under finite m-th moments. This suggests robust minibatch gradient aggregation, where poisoned or extreme examples cannot dominate an optimizer step. The most practical first test is coordinatewise winsorization of per-example gradients with a clipping quantile above the suspected contamination rate, comparing loss descent and clean-data accuracy under targeted gradient poisoning.