Honesty is the Best Policy ... or is it? Demand Management for Probabilistic Service Options

For an order o arising at decision epoch k, define its time horizon in the system by h_o=t^A_o-t_k. Because (i) the points in time t_k are discrete, (ii) t^A_o is finite, and (iii) probability display and allocation occur at distinct decision epochs, it holds that h_o\in\mathbb{N}.

For every h\in\mathbb{N}, let \mathcal{U}(h) denote the following statement: For every probability display pre-decision state S_k\in\mathcal{S}^p and hypothetical order o_{kj} arising if request r_k chooses service option j\in\mathcal{J}^P with h_{o_{kj}}=h, the truthful allocation probabilities \psi^\pi_{o_{kj}}(S_{kj}^c) for o_{kj} exist and are uniquely determined. Consequently, under full transparency, the probability-display vector P^{\pi_\alpha}(S_k) exists and is unique. Note that the allocation time t_{r_k}^A is an attribute of the request such that all service options j\in\mathcal{J} lead to the same horizon h_{o_{kj}}. We prove \mathcal{U}(h) for every h\in\mathbb{N} by induction.

For the base case, consider h=1, meaning that the hypothetical order o_{kj} arising in k is allocated at the subsequent decision epoch k+1. By the prescribed ordering of events, no additional probability display decision occurs between the creation and allocation of o_{kj}. Hence, every order that can affect its allocation is already represented in post-choice state S_{kj}^c, and the exogenous-information distribution \mathbb{P}(W_{k+1}^s|S_{kj}^c) together with the transition function S_{k+1}=S^{MW^s}(S_{kj}^c,W^s_{k+1}) uniquely determine the distribution of the next pre-decision state S_{k+1}. Conditional on each realization of the latter, the fully specified allocation scheme \alpha determines a unique joint allocation outcome for all orders o\in\mathcal{O}_{k+1}^A, which includes o_{kj}. Therefore, the truthful allocation probabilities of o_{kj} for every j\in\mathcal{J}^P exist and are unique. Under full transparency, the displayed probabilities must coincide with the truthful probabilities, establishing \mathcal{U}(1).

For the induction step, fix an arbitrary h\geq2 and suppose that \mathcal{U}(h') holds for every h'<h. Consider an arbitrary pre-decision state S_k\in\mathcal{S}^p with hypothetical order o_{kj} satisfying t^A_{o_{kj}}-t_k=h. Additional requests may arise before t_{o_{kj}}^A and generate orders allocated before or jointly with o_{kj}. Consider any such order o' arising at a later decision epoch k' with t_k<t_{k'}<t^A_{o_{kj}}. There are two relevant cases to distinguish.

• Case 1: t^A_{o'}\leq t^A_{o_{kj}}. Because \alpha is non-anticipatory, o' can affect the allocation outcome of o_{kj} only if t^A_{o'}\leq t^A_{o_{kj}}. Its horizon h_{o'} satisfies

  \begin{aligned} h_{o'} = t^A_{o'}-t_{k'}\leq t^A_{o_{kj}}-t_{k'}< t^A_{o_{kj}}-t_k=h \end{aligned}

  Since probability display and allocation occur at distinct decision epochs, 1\leq h_{o'}<h. Hence, every future order capable of affecting the allocation of o_{kj} has a horizon in \{1,\ldots,h-1\} and is covered by the induction hypothesis. The truthful allocation probabilities, and therefore the fully transparent display vector at epoch k', are thus uniquely determined. Since the customer-choice model \mathbb{P}(W_{k'+1}^c|S_{k'},P^{\pi_\alpha}(S_{k'})) uniquely maps the state and probability display vector into a choice probability distribution, the conditional distribution of the customer’s choice is unique.

• Case 2: t^A_{o'}> t^A_{o_{kj}}. The order o' is allocated strictly after order o_{kj}. Following its creation, o' is included in the updated set of pending orders O_{k'+1}, but it does not occupy capacity before its allocation. Moreover, by non-anticipativity of \alpha, neither the existence nor the characteristics of o' affect any allocation decision made at or before t^A_{o_{kj}}. Additionally, the presence of o' does not influence the conditional distribution of the exogenous information relevant to the allocation of o_{kj}, including capacity availability, request arrivals, and customer choices that may affect allocations up to t^A_{o_{kj}}. Consequently, conditional on any realization of the process up to t^A_{o_{kj}}, the allocation outcome of o_{kj} is invariant to o'. Hence, when constructing the conditional distribution of the allocation state for o_{kj}, the customer choice and subsequent state information associated exclusively with o' can be marginalized out. Orders allocated strictly after o_{kj} therefore do not affect the truthful allocation probability of o_{kj}.

The finite composition of state transitions, which follow unique conditional distributions, together uniquely determines the probability distribution of the allocation state for order o_{kj} and the associated set of orders to be allocated. Conditional on each realization, scheme \alpha uniquely determines the joint allocation outcome or its probability distribution. Hence, the truthful allocation probability \psi_{o_{kj}}^{\pi_\alpha}(S_{kj}^c) exists and has a unique value. Therefore, the probability display vector in pre-decision state S_k is unique, and \mathcal{U}(h) holds.

By induction, \mathcal{U}(h) holds for every h\in\mathbb{N}. Since S_k and j were arbitrary, the truthful allocation probabilities exist and are uniquely determined in every post-choice state. Consequently, under \tau=0, the probability display decision is unique in every pre-decision state. Therefore, the full-transparency constraints admit a unique policy conditional on \alpha, which is the optimal policy \pi_\alpha^* given \alpha.