The choice probabilities , , and sum to one
ProvedDynamicAssortment.Separation.choice_probabilities_sum_to_onechoice-modeldynamic-assortmentrandom-utility
Consider the random-utility choice model of Sauré–Zeevi (§3.2): a customer facing an assortment has utilities for and for the no-purchase option (), where the shocks are i.i.d. with a density that is integrable, everywhere positive and integrates to . Writing , the probability that product is chosen is
and the no-purchase probability is
Claim. These numbers form a probability distribution on : for every finite and every ,
Probabilistically, is the probability that alternative has the largest utility, and ties have probability zero because the utilities have densities. Analytically, the integrand of the sum is the a.e. derivative of , which increases from to .
Preamble
import Mathlib import Definitions.Def_DynamicAssortment_Separation_ChoiceModel
Formal statement
namespace DynamicAssortment.Separation
/-- The choice probabilities of the random-utility model (p. 391, §3.2) form a probability
distribution on `S ∪ {0}`: `∑_{i ∈ S} p_i(S, μ) + p_0(S, μ) = 1` for every assortment `S` and
every mean-utility vector `μ`, when the shocks have a density. -/
theorem choice_probabilities_sum_to_one {N : ℕ} (f : ℝ → ℝ) (hf : IsNoiseDensity f)
(μ : Fin N → ℝ) (S : Finset (Fin N)) :
∑ i ∈ S, choiceProb f μ S i + noPurchaseProb f μ S = 1 := by sorry
end DynamicAssortment.SeparationSource
Sauré, Zeevi, Optimal Dynamic Assortment Planning with Demand Learning, Manuf. Serv. Oper. Manag. 15(3), 2013, p. 391, §3.2 (choice probabilities p_i(S, μ) of the random-utility model, a distribution over S ∪ {0}); used implicitly in the proof of Theorem 2 (App. A, pp. 402–403) whenever probabilities are bounded by one