Sufficient And Necessary Condition For Convergence of Conditional Error Probability in NN-Pattern Discrimination

Abstract:

      Let (θ₁,X₁),…, (θ_n,X_n), (θ, X) be iid random vectors ,where θ∈{0,1},X∈R^d Denote by θ′_n the nearest neighbour discriminator of θ based on the training samples (θ₁,X₁),…, (θ_n,X_n) and the observed X; put(?). This paper gives a sufficient and necessary condition for (?) as n→∞, namely (P(θ=0, X=x)-P(θ=1, X=x))²·P(θ=0, X=x)·P(θ=1, X=x)=0 for every x∈R^d.This generalizes a previous result of the authors [5] and improves a result of Wagner, T.J. [2].