Professional Documents
Culture Documents
1 Limit Theorems 49
A first step is to formalize the problem. Since we know process ’A’ exactly
we only need to concern ourselves with ’B’. We associate the random variable
Xi with wafer i. A reasonable (and somewhat simplifying) assumption is to
posit that all Xi are independent and identically distributed where all Xi
have the mean µB . Obviously we do not know µB — otherwise there would
be no reason for testing! We denote by X̄m the average of the yields of m
wafers using process ’B’. What we are interested in is the accuracy for
which the probability
δ = Pr(|X̄m − µB | > ) satisfies δ ≤ 0.05.
Let us now discuss how the various bounds behave. For the sake of the
argument assume that µB − µA = 20, i.e. the new process produces on
average 20 additional usable chips.
Hoeffding Since the yields are in the interval {0, . . . , 400} we have an ex-
plicit bound on the range of observations. Recall the inequality (2.16) which
bounds the failure probably δ = 0.05 by an exponential term. Solving this
for m yields
m ≥ 0.5|b − a|2 −2 log(2/δ) ≈ 737.8 (2.19)
In other words, we need at lest 738 wafers to determine whether process ’B’
is better. While this is a significant improvement of almost two orders of
magnitude, it still seems wasteful and we would like to do better.