Tree boosting for learning EFT parameters

To construct the algorithm, we interpret the optimized loss function of a traditional decision tree as the maximal Fisher information in Poisson counting experiments. We promote the interpretation to general EFT predictions and develop a suitable boosting method.

The resulting “Boosted Information Tree” algorithm approximates the score, the derivative of the log-likelihood function with respect to the parameter. It thus provides a sufficient statistic in the vicinity of a reference point in parameter space where the estimator is trained. The training exploits per-event information of likelihood ratios for different theory parameter values available in the simulated EFT data sets.