Issue No. 07 - July (1999 vol. 21)
ISSN: 0162-8828
pp: 643-645
ABSTRACT
<p><b>Abstract</b>—We give a short proof of the following result. Let <tmath>$(X,Y)$</tmath> be any distribution on <tmath>${\cal N} \times \{0,1\}$</tmath>, and let <tmath>$(X_1,Y_1),\ldots,(X_n,Y_n)$</tmath> be an i.i.d. sample drawn from this distribution. In discrimination, the Bayes error <tmath>$L^* = \inf_g {\bf P}\{g(X) \not= Y \}$</tmath> is of crucial importance. Here we show that without further conditions on the distribution of <tmath>$(X,Y)$</tmath>, no rate-of-convergence results can be obtained. Let <tmath>$\phi_n (X_1,Y_1,\ldots,X_n,Y_n)$</tmath> be an estimate of the Bayes error, and let <tmath>$\{ \phi_n(.) \}$</tmath> be a sequence of such estimates. For any sequence <tmath>$\{a_n\}$</tmath> of positive numbers converging to zero, a distribution of <tmath>$(X,Y)$</tmath> may be found such that <tmath>${\bf E} \left\{ | L^* - \phi_n (X_1,Y_1,\ldots,X_n,Y_n) | \right\} \ge a_n$</tmath> infinitely often.</p>
INDEX TERMS
Discrimination, statistical pattern recognition, nonparametric estimation, Bayes error, lower bounds, rates of convergence.
CITATION
András Antos, László Györfi, Luc Devroye, "Lower Bounds for Bayes Error Estimation", IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 21, no. , pp. 643-645, July 1999, doi:10.1109/34.777375