MIT6_262S11_lec21

# Substituting x r0 for n r pr sn exp x 0 r0 x

This preview shows page 1. Sign up to view the full content.

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: µX (a) = γX (ro) − roa for a < X , r < 0 γX (ro) ￿ slope = γX (ro) = a Pr {Sn ≥ na} ≤ exp(nµX (a)) The Chernoﬀ b ound, optimized over r, is essentially ex­ ponentially tight; i.e., Pr {Sn ≥ na} ≥ exp(n(µX (a) − ￿)) for large enough n. 12 In looking at threshold problems, we want to ﬁnd the probability that Pr {Sn ≥ α} for any n. Thus we want a b ound that focuses on variable n for a ﬁxed α, i.e., on when the threshold is crossed if it is crossed. We want a b ound of the form Pr {Sn ≥ α} ≤ exp αf (n) Start with the b ound Pr {Sn ≥ na} ≤ exp(n[γX (r0) − r0a]), ￿ with α = an and r0 such that γX (r0) = α/n. Substituting ￿ α/γX (r0) for n, ￿￿ γ (r ) Pr {Sn ≥ α} ≤ exp α X 0 − r0 ￿ γX (r0) ￿￿ 13 ￿￿ γ (r ) Pr {Sn ≥ α} ≤ exp α X 0 − r0 ￿ γX (r0) 0 r ro ❅ ❅ ❅ ❅ ❅ ❅ slope = X r∗ ￿￿ ro − γ (ro)/γ ￿(r0) slope = γ ￿(ro) = α/n γ (ro) ￿ When n is very large, the slope γX (r0) is close to 0 and the horizontal intercept (the negative exponent) is very large. As n decreases, the intercept decreases to r∗ and then increases...
View Full Document

## This note was uploaded on 01/13/2012 for the course ELECTRICAL 6.262 taught by Professor Staff during the Fall '11 term at MIT.

Ask a homework question - tutors are online