On thePeakingof theHughes Mean Recognition Accuracy: The Resolutionofan ApparentParadox

Thepeaking phenomenon oftheBayesrecognition accuracy ofpattern classifiers with unknown underlying statistics is addressed. Itisshown that this effect, known astheHughes paradox, arises fromimproper comparisons ofstatistically incomparable models. Aformalization ofthenotion ofcomparability isintroduced, andsomeoftheresults obtained intheliterature arerevisited inthis context.