ISyE 7401, Spring - 2019Instructor : A.ShapiroHomework # 81. Let π1 and π2 be two populations of m × 1 random vectors X having normaldistributions with respective means µ1 and µ2 and the same covariance matrix Σ,i.e., X ∼ N(µi; Σ) conditional on X 2 πi, i = 1; 2. Let Pr(X 2 πi) = qi, i = 1; 2.Consider Fisher’s linear discriminant classifier: an observation x is assigned to π1 if(µ1
...[Show More]
ISyE 7401, Spring - 2019
Instructor : A.Shapiro
Homework # 8
1. Let π1 and π2 be two populations of m × 1 random vectors X having normal
distributions with respective means µ1 and µ2 and the same covariance matrix Σ,
i.e., X ∼ N(µi; Σ) conditional on X 2 πi, i = 1; 2. Let Pr(X 2 πi) = qi, i = 1; 2.
Consider Fisher’s linear discriminant classifier: an observation x is assigned to π1 if
(µ1 - µ2)0Σ-1x > c;
and x is assigned to π2 otherwise. Find the minimal total misclassification probability
q1Pr(X 2 R2jX 2 π1) + q2Pr(X 2 R1jX 2 π2)
of this classifier. For what value of the constant c this minimal value is attained?
We have that the minimal total misclassification probability is attained by the
classifier assigning π1 if f f1 1( (x x) ) > q q2 1, where
fi(x) = (2π)-m=2jΣj-1=2 expf-1 2(x - µi)0Σ-1(x - µi); i = 1; 2;
are the respective densities. By direct calculations
ln q q1 2f f1 1( (x x) ) = b0x + b0;
where b = Σ-1(µ1 - µ2) and
b0 = -12µ0 1Σ-1µ1 + 1 2µ0 2Σ-1µ2 + ln(q2=q1):
and hence
c = 1
2µ0 1Σ-1µ1 - 1 2µ0 2Σ-1µ2 - ln(q2=q1):
The total misclassification probability is
q1Pr(X 2 R2jX 2 π1) + q2Pr(X 2 R1jX 2 π2):
Now
Pr(X 2 R2jX 2 π1) = Pr(b0X + b0 < 0jX 2 π1):
Denote
∆ = q(µ1 - µ2)0Σ-1(µ1 - µ2)
the Mahalanobis distance between π1 and π2. Conditional on X 2 π1, b0X +b0
has normal distribution with mean
µ0 1Σ-1(µ1 - µ2) + b0 = 1 2∆2 + ln(q2=q1)
and variance b0Σb = ∆2: Therefore
Pr(X 2 R2jX 2 π1) = Φ -1 2∆ - ∆-1 ln(q2=q1) ;
where Φ(·) is the cdf of N(0; 1). Similarly
Pr(X 2 R1jX 2 π2) = Φ -12∆ + ∆-1 ln(q2=q1) :
1
[Show Less]