In multi-group learning, a single predictor must control its excess risk simultaneously over many overlapping subpopulations. Earlier work by Tosh and Hsu established slow 1/√n rates via a reduction to online learning with sleeping experts, and achieved 1/n rates only under realizability and with a different strategy.
The new paper, by Hanqi Sun, Evgenii Chzhen, Luca Ganassali, and Gilles Stoltz, identifies a group-wise Bernstein condition under which the same adaptive predictor attains the fast 1/n rate without knowing whether the condition holds. If the condition fails on some groups, the predictor retains the slow rate there, so the improvement is automatic and local.
The condition is implied by natural assumptions weaker than realizability. Lower bounds show that weakened versions of the condition leave every learner with slow rates. The authors also obtain intermediate rates under relaxed conditions and extend their results to VC classes.