Title:A new Probability inequality using typical moments and Concentration Results

Abstract: A new probability inequality strengthening Azuma's is presented. Instead of assuming an absolute bound on the Martingale differences as Azuma's does, we use information on conditional moments of each variable. Information on moments conditioned on "typical" as well as "worst-case" values are used. Also, a weaker assumption than Martingale differences suffices here. The inequality is strong enough to yield sub-Gaussian tail bounds in many situations. We present several applications. For bin-packing optimal concentration intervals are proved settling a question of Talagrand. For chromatic number of sparse random graphs, the first sub-Gaussian tail bounds are proved and the first treatment of the number of cliques of growing sizes in random graphs is given. For Longest Sub-sequences, a result matching Talagrand's is given. Brief comparisons to other recent inequalities are made.