Chernoff Bounds推荐.pdfVIP

CHAPTE R F OUR Chernoff Bounds Thi s chapter introduces what are common ly called Chernoff bounds. Chernoff bou nds are extremely powerful, giving exponential ly decreasing hounds on the tail di stri bu­ tion . These bounds are derived by Marko v s inequal ity on the moment gener­ ating fu nction of a random variabl e. We start this chapter by defining and discussing the properties of the moment generati ng fu nction . We then derive Chernoff bounds for the binomial di stribution and other related di stri but ions, using a set-bal anc i ng prob­ lem as an exampl e. To demonstrate the power of Chernoff bou nds. we a pply them to the anal ysis of randomized packet routing schemes on the hy percube and butterfl y networks. 4 . 1 . Moment Generating Functions Before developing Chernoff bounds, we di scuss the special rol e of the moment gener­ ating function E [ex ]. Definition 4.1: The moment generati ng function of a random variable X is We are mainly interested in the exi stence and properties of this fu nction in the neigh­ borhood of zero. The fu nction M x ( t) captures all of the moments of X. Theorem 4.1: Let X be a random variable with moment generating .fimction Mx U). Under the a:;;sumption that exchanging the expectation and d�fferentiation operands is legitimate, for all n 1 we then have E [X 11 ] M.�n (O) . l 1 t ) evaluated at t = 0. where M� (0) is the nth derivative of Mx ( 61 CHERNOFF BOUNDS Proof: Assuming that we can exchange the expectation and differentiation operands, then Computed at t = 0, thi s expression yields