Feldman's memorization lower bound is a widely cited tool for analyzing when machine learning models memorize training data. It assigns a frequency coefficient to each data point based on how often that point appears in the training sample, using this coefficient to quantify the cost of not fitting the point exactly.
The new preprint presents an exact three-point counterexample that breaks both the conditional and averaged forms of this bound. The construction relies on independent fair random variables, showing that the frequency-based coefficient does not always hold as stated.
This is a correction rather than a complete dismissal: the bound may still be useful in specific settings, but its current formulation is not universally valid. The authors' counterexample gives researchers a concrete case to test and refine the theory.