Tang’s dequantisation programme established that several quantum machine learning algorithms admit efficient classical simulation when given ℓ2-norm sample-and-query (SQ) access to their input data. The specific choice of ℓ2-norm sampling—rather than uniform, ℓ1, or any other sampling rule—has hitherto been justified operationally by its polynomial-time implementability, but not axiomatically. We show that ℓ2-norm SQ sampling is the unique sampling functional satisfying five natural operational