
236 Supervised Classification Part 1
FIGURE 6.3
ROIs for supervised c lassification. The insert shows the training observations
projected onto the plane of the first two principal axes. (See color insert.)
cova riance matrices Σ
1
= Σ
2
, then it can be shown that the minimum in fact
occurs at s = 1/2. Approximating the minimum as s = 1/2 also for the cas e
where Σ
1
6= Σ
2
leads to the (somewhat les s tight) Bhattacharyya bound ǫ
B
,
ǫ ≤ ǫ
B
=
p
Pr(1)Pr(2)
Z
p
p(g | 1)p(g | 2) dg. (6.11)
This integral can be evaluated explicitly (Exercise 1). The result is
ǫ
B
=
p
Pr(1)Pr(2) e
−B
,