10/20/04 MWG 22.38 PROBABILITY AND ITS APPLICATIONS TO RELIABILITY, QUALITY CONTROL AND RISK ASSESSMENT Fall 2004 CONVERGENCE OF BINOMIAL AND NORMAL DISTRIBUTIONS FOR LARGE NUMBERS OF TRIALS We wish to show that the binomial distribution for m successes observed out of n trials can be approximated by the normal distribution when n and m are mapped into the form of the standard normal variable, h. (1) P(m,n ) ≅ Prob. (h) , where ↑ ↑ Binomial Normal Distribution Distribution Binomial Distribution: ⎛ n⎞ P ( m,n ) = ⎜ m⎟ p m q (n −m ) , ⎝ ⎠ (2) p + q = 1 , where (3) p = probability of success in a single trial, and q = probability of failure in a single trial. − h2 2 e ( ) , Prob. ( h) = 2π σ Normal Distribution: (4) ⎛ m − µ⎞ h ≡⎜ ⎟ , ⎝ σ ⎠ µ = np , (5) (Binomial Distribution Mean) σ = np q . (Binomial Distribution Standard Deviation) Recall Sterling Approximation: m! ≅ 2 π m m m e − m ⎞ n 1 ⎛⎜ ⎛ n⎞ n! ⎟ ≅ ⇒ ⎜⎝ m ⎟⎠ ≡ ⎜ 2 π ⎝ m(n − m )⎟⎠ m!(n − m )! P(m,n ) ≅ ⎞ 1 ⎛⎜ 1 ⎟ 2 π ⎜⎝ m(n − m)⎟⎠ 1 1 of 2 1 m m n⎞ ⎛ n ⎞ ⎟ . ⎜ ⎟ ⎜ ⎝ m⎠ ⎝ n − m⎠ 2⎛ m n− m) n p⎞ ⎛ n q ⎞ ( ⎜ ⎟ ⎜ ⎟ ⎝ m ⎠ ⎝ n − m⎠ 2⎛ (6) (7) ≅ ⎛ np ⎞ m ⎛ nq ⎞ ( n − m ) . ⎟ ⎜ ⎟ ⎜ 2 π (n pq ) ⎝ m ⎠ ⎝ n − m ⎠ 1 The result above uses the relationships: m = np+ h npq (n − m ) = nq − h n p q to obtain result ⎛ ⎛ m (n − m )⎞ pq ⎞ pq⎞⎛ ⎟⎟ ⎟⎟ ⎜⎜ q − h ⎜ ⎟ = n ⎜⎜ p + h ⎝ ⎠ n ⎠ n ⎠⎝ n ⎝ ≅ n pq . x2 Then, use expansion of ln(1+ x ) ≅ x − , about x=0 to evaluate Eq. 1, using 2 π n pq Prob. (h ) 2 as: −ln ( 2 π n pq Prob. (h)) ≅ −ln ( 2 π n pq P(m, n)) ⎡ ⎛ n p ⎞ m ⎛ n q ⎞ ( n − m) ⎤ ⎥ = ln ⎢ ⎜ ⎟ ⎜ ⎟ ⎢⎣ ⎝ m ⎠ ⎝ n − m ⎠ ⎥⎦ ⎛ ⎛ ⎞ q ⎞⎟ ⎜ 1− h p ⎟ = (n p + h n pq )ln ⎜⎜ 1+ h ln + n q − h n pq ( ) ⎜ n p ⎟⎠ nq ⎟⎠ ⎝ ⎝ ⎛ ⎛ q h2 q ⎟⎞ p h2 p ⎟⎞ ⎜ ≅ (n p + h n pq )⎜⎜ h − −h + nq − h n p q − ( ) ⎜⎝ n q 2n q ⎟⎠ n p 2n p ⎟⎠ ⎝ ⎛ ⎞ ⎛ ⎞ h2 h2 + q h2 ⎟⎟ + ⎜⎜ −h n p q − p = ⎜⎜ h n pq − q + ph 2 ⎟⎟ 2 2 ⎝ ⎠ ⎝ ⎠ h2 h2 = (p + q) = . 123 2 2 1 Thus, the result is obtained; verifying Eq. 4: e ( ) e ( ) Prob. (h) = = . QED! 2π n p q σ 2π − h2 2 2 of 2 − h2 2