Multicalibration is a fairness notion that demands a predictor be calibrated not only on average but also within every subpopulation from a specified class, called auditors. The paper examines how many samples are required to achieve this property when the learning algorithm has access to an empirical risk minimization (ERM) oracle.
The abstract defines the setting and frames the sample-complexity question, but the text provided stops before presenting any concrete results. This appears to be a theoretical contribution to the understanding of multicalibration in an oracle-based model, where the cost of obtaining calibrated predictions is measured in samples.
No other sources are available for comparison, so the summary relies solely on the abstract's opening description.