Pattern Recognition Letters 1 (1983) 291-292 July 1983
North-Holland
A note on nearest neighbour error rates
S.J. MAYBANK
Department of Computer Science, Birbeck College, University of London, U.K.
and Marconi Space and Defence Systems, Frimley, U.K.
Received 9 December 1982
Abstract: Let Ak, t, Ek,/denote the acceptance and error rates in the (k, l) case of the nearest neighbour decision rule. We show
that if q>k and Aq, m=Ak, I then Eq, m<-Ek, l.
Key words: Nearest neighbours, error rates.
1. Introduction
In Devijver and Kittler (1982, p. 105) the follow-
ing conjecture is made: if
Ak, l=Aq, m (k< q< oo),
then
Ek, l+l <_Eq, m<_Ek,t. (1)
In (1), Aab denotes the acceptance rate for the
nearest neighbour rule where a pattern, re-
presented by a point in Euclidean space, is ac-
cepted only if b (b > a/2) out of its first a nearest
neighbours in a large sample set of correctly
classified patterns are in the same pattern class.
The pattern is assigned to this class. Eab denotes
the corresponding error rate. The notation is fully
explained in Devijver and Kittler (1982).
In this note we prove the second inequality of
(1).
2. Proof of the inequality
First we note that in the case of M pattern
classes COl, . . . , (..OM,
l
,-~=~ (1-rli)Iqmp(x) dx, (2)
where rli = - I']i(X ) is the a posteriori probability of a
pattern x lying in class o9i, p(x) is the un-
conditional probability density function of x and
q q
Iffm=j~_m(j)rIJ(1--rli)q-J
is the incomplete Beta function (written as
I,,(m,q-m+l) in Zelen and Severo (1964, pp.
944, 945)). Now if we set
=I ,_Igm
then, since Ak, l=Aq, m,
t~J~iP(x) dx = 0.
i=l
From (2) we now have
Ek't-Eq'm =- i=1 ~ ! rliJ'7ip(x)dx"
By differentiating the integral representation of
the incomplete Beta function, we can show that
~te [0, 1] such that rli<_t implies J~,>>_O and rli>_t
implies J,i_<0. The result is also true if q=oo,
except at 11 i = t.
0167-8655/83/$3.00 © 1983, Elsevier Science Publishers B.V. (North-Holland) 291