IncrementalityBiddingviaReinforcementLearningunder MixedandDelayedRewards Appendix AFormalDefinitionofInhomogeneousPoissonProcess

We first introduce the main idea of the the PAMM algorithm. Therefore, E(Xφ) and E(Yφ) precisely locate the parameter of interest,θh. Therefore, each of these items is a sub-exponential random variable withkNthkψ CT. To prove Theorem 2, we need several auxiliary lemmas. The importance of this lemma is that we provide a reformulation of functionR(π;bθ, bF)by utilizing the fact that the conversion incrementality can only happen ifthe learner wins the opportunity to show the ad to the user. R(πt;θ,F) # | {z} (iv) (25) The first inequality above is based on the fact the reward (expected utility of the learner) at each roundisboundedby[0,1].