Source of the excess committor discrepancy

Determine whether the small excess discrepancy between the unconditional EDM model's committor and the noise-aware classifier's class posterior near the decision regime arises from an error in the diffusion model or from errors in the classifier.

Background

For a perfect diffusion model, the probability that a stochastic continuation ends in a class should equal the corresponding posterior probability assigned by the data-trained classifier. The unconditional EDM model agrees closely with the classifier overall, but near noise level approximately 1.1 its total-variation discrepancy exceeds the classifier's low-noise error level by as much as about 0.06.

Adjusting for the model's class proportions does not remove this excess, while the classifier's own error varies with noise level by a comparable amount. Consequently, the study does not resolve whether the residual discrepancy is attributable to the model or to the classifier.