--- title: Fictitious Revelation Principle in Mechanism Design url: https://www.emergentmind.com/topics/fictitious-revelation-principle type: topic --- # Fictitious Revelation Principle in Mechanism Design As used across recent mechanism-design papers, the “fictitious revelation principle” denotes a family of departures from the classical revelation principle in which the relevant direct object is no longer an ordinary costless truthful direct mechanism. In the standard Bayesian-Nash formulation, if a mechanism \(T=(S_1,\dots,S_I,g(\cdot))\) implements a social choice function \(f(\cdot)\) in Bayesian Nash equilibrium, then \(f(\cdot)\) is truthfully implementable by the direct revelation mechanism \(T_{\text{direct}}=(\Theta_1,\dots,\Theta_I,f(\cdot))\) with equilibrium strategy \(s^*(\theta)=\theta\) [1008.3287]. The recent literature qualifies this benchmark in several distinct ways: direct revelation may become physically nonviable when strategies are actions rather than messages [1008.3287]; it may fail behaviorally, with outcome replication occurring only through “dishonest and disobedient” conduct [2011.07878]; it may fail under cumulative prospect theory and be recoverable only through mediation [2101.08722]; it may need to be replaced by a canonical operator in sequential-move games [2402.13580]; it may require filtered type reports that preserve informational content rather than full revelation [2603.12532]; or, in dynamic pricing, it may reappear as an exact optimal-regret equivalence between indirect learning and direct revelation [2604.24093]. ## 1. Classical benchmark and the direct-revelation template The classical benchmark is the Bayesian-Nash revelation principle. In the notation used in the literature under discussion, a mechanism is \( \Gamma=(S_1,\cdots,S_I,g(\cdot)) \), or equivalently \(T=(S_1,\dots,S_I,g(\cdot))\), where each \(S_i\) is a strategy set and \(g\) maps strategy profiles into outcomes. A direct mechanism sets the strategy sets equal to the type spaces, \( \bar{\Gamma}=(\Theta_1,\cdots,\Theta_I,\bar g(\cdot)) \), with \( \bar g(\theta)=f(\theta) \) for all \(\theta\in\Theta\), and truthful reporting \( \bar s_i^*(\theta_i)=\theta_i \) is required to be a Bayesian Nash equilibrium [1008.3287]. The corresponding incentive condition is the familiar truthfulness inequality \[ E_{\theta_{-i}}\!\big[u_i(f(\theta_i,\theta_{-i}),\theta_i)\mid \theta_i\big] \ge E_{\theta_{-i}}\!\big[u_i(f(\hat\theta_i,\theta_{-i}),\theta_i)\mid \theta_i\big] \quad \forall \hat\theta_i\in\Theta_i. \] The paper on action-format strategies cites the standard result as “Proposition 23.D.1 (Revelation Principle for Bayesian Nash Equilibrium): If some mechanism implements \(f\) in BNE, then \(f\) is truthfully implementable in BNE” [2011.07878]. This benchmark fixes the point of comparison for the later variants. In the classical template, the direct step is treated as informationally sufficient: once equilibrium play in an indirect mechanism is known, one can encode its outcome rule as truthful type reports. The fictitious variants arise precisely where that informational reduction is no longer adequate, no longer truthful, or no longer the correct object to preserve. ## 2. Action-format strategies, energy feasibility, and behavioral failure A central line of work distinguishes **message-format** and **action-format** strategies. In the message-format case, a strategy is an informational message \(m_i(\cdot):\Theta_i\to M_i\) in an indirect mechanism \( \Gamma_m=(M_1,\cdots,M_I,g_m(\cdot)) \). In the action-format case, a strategy is a realistic action \(a_i(\cdot):\Theta_i\to A_i\) in an indirect mechanism \( \Gamma_a=(A_1,\cdots,A_I,g_a(\cdot)) \). This distinction is decisive because message strategies are only information, whereas action strategies are materially executed conduct [2011.07878]. From the energy perspective, the key quantities are \(E_a\), the energy required for an agent to carry out an action; \(E_m\), the energy required for an agent to choose a message; \(E_{\text{send}}\), the energy to send out a message; and \(E_g\), the energy for the designer to execute the outcome function \(g(\cdot)\). The paper assumes \(E_a > E_m\), and this makes the direct-revelation step nontrivial when the original mechanism relies on actions rather than messages [1008.3287]. | Strategy format | Indirect mechanism \(T'\) | Direct mechanism \(I_{\text{direct}}\) | |---|---|---| | Message | agents: \(I(E_m+E_{\text{send}})\); designer: \(E_g\) | agents: \(I E_{\text{send}}\); designer: \(I E_m+E_g\) | | Action | agents: \(I E_a\); designer: \(E_g\) | agents: \(I E_{\text{send}}\); designer: \(I E_a+E_g\) | Neglecting \(E_m\), \(E_g\), and \(E_{\text{send}}\) relative to \(E_a\), the simplified matrix is \[ \begin{array}{c|cc} & \Gamma & I_{\text{direct}} \\ \hline \text{Message} & [0,0] & [0,0] \\ \text{Action} & [I E_a, 0] & [0, I E_a] \end{array} \] and the additional feasibility condition becomes \[ \text{Designer energy} \ge \sum_{i=1}^{I} E_{a,i}, \qquad \text{or in the homogeneous case, } \text{Designer energy} \ge I E_a. \] Without that condition, the paper argues that the revelation principle may fail as a physically meaningful statement because the direct mechanism is not actually implementable by the designer under realistic energy constraints [1008.3287]. A different criticism of action-format implementation is behavioral rather than physical. In the action case, the constructed direct mechanism is multistage: an agent reports a type \(\hat\theta_i\), receives a suggestion \(a_i^m(\hat\theta_i)\), chooses an actual action \(\hat a_i\), and the final outcome is determined by the actual performed actions \(g_a(\hat a_1,\dots,\hat a_I)\). The claim is that an agent can misreport at the report stage, ignore the suggestion, and still perform the same equilibrium action \( \hat a_i=a_i^*(\theta_i) \), so that the final outcome remains \( g_a(a^*(\theta))=f(\theta) \). On this view, “honest and obedient” is not the Bayesian Nash equilibrium of the direct mechanism; the social choice function is reproduced only “dishonestly and disobediently” [2011.07878]. Taken together, these two papers identify distinct limits of direct revelation in action environments. One limit concerns **feasibility**: the designer may not possess the energy needed to absorb the agents’ action costs. The other concerns **incentive form**: even when outcomes are reproduced, truthful reporting may no longer be the equilibrium behavior. Both arguments leave the classical message-based case intact, but they sharply restrict any unconditional reading of revelation as a universally free reduction. ## 3. Non-expected-utility preferences and mediated restoration A second route to a fictitious revelation principle arises when agents are modeled with cumulative prospect theory rather than expected utility. In that framework, each player \(i\) has a finite outcome set \(\Gamma_i\), a value function \(v_i:\Gamma_i\to\mathbb R\), and probability weighting functions \(w_i^+,w_i^-:[0,1]\to[0,1]\). Given a lottery \(L_i=\{(p_i^j,\gamma_i^j)\}_{1\le j\le k_i}\in\Delta(\Gamma_i)\), the CPT value is defined by rank-dependent decision weights: \[ V_i(L_i) := \sum_{j=1}^{j_r} \pi_i^+(j,p_i,\beta_i)\, v_i(\gamma_i^{\beta_i^j}) + \sum_{j=j_r+1}^{k_i} \pi_i^-(j,p_i,\beta_i)\, v_i(\gamma_i^{\beta_i^j}). \] Expected utility is recovered only in the special case \(w_i^+(p)=w_i^-(p)=p\), where \(V_i(L_i)=U_i(L_i)=\sum_{j=1}^{k_i} p_i^j v_i(\gamma_i^j)\) [2101.08722]. The failure of the classical revelation principle under CPT is traced to two structural facts. First, under CPT, utility over allocations does not summarize preferences in the same way as under EUT: an induced utility over lotteries on allocations is not generally representable as CPT over the allocation set itself. Second, mixtures of incentive-compatible rules need not remain incentive-compatible. The paper gives explicit examples where \[ V_i(L_i') = V_i(L_i'') > V_i(0.5L_i' + 0.5L_i''), \] which breaks the convexity and averaging logic used in standard revelation arguments. It then gives an example of an allocation choice function \(f^*\) that is implementable in Bayes-Nash equilibrium by a non-direct mechanism but cannot be truthfully implemented by any direct mechanism [2101.08722]. The proposed repair is **mediated mechanism design**. A mediated mechanism is \[ \mathcal{M}:=((\Phi_i)_{i\in[n]},D,(\Theta_i)_{i\in[n]},h), \] where a mediator first draws \(\phi\sim D\), player \(i\) privately observes \(\phi_i\), then reports \(\theta_i'\in\Theta_i\), and the operator chooses an allocation \(\alpha\sim h(\phi,\theta')\). The resulting revelation theorem states: 1. If an allocation choice function is implementable in Bayes-Nash equilibrium by a mediated mechanism, then it is also truthfully implementable in Bayes-Nash equilibrium by a direct mediated mechanism. 2. If an allocation choice function is implementable in dominant equilibrium by a publicly mediated mechanism, then it is truthfully implementable in dominant equilibrium by a direct publicly mediated mechanism. 3. If an allocation choice function is implementable in belief-dominant equilibrium by a mediated or publicly mediated mechanism, then it is truthfully implementable in the corresponding direct mediated or direct publicly mediated mechanism [2101.08722]. The constructive proof enlarges each message space to \[ \Phi_i' := \Phi_i \times (\Theta_i)^{\Theta_i}, \] with the mediator distribution pushing all equilibrium randomization into the message stage. The direct mediated allocation rule then reproduces the original lottery distribution exactly. This suggests that, under CPT, the classical direct mechanism is fictitious in a precise sense: truthful direct revelation is unavailable in the non-mediated model, but a richer direct object with mediator-generated signals restores the principle. ## 4. Sequential-move games and canonical operators In sequential-move mechanism design, the standard direct mechanism is no longer the obvious reduction target because information is revealed gradually along histories rather than all at once. The setting models a mechanism as an extensive-form game \[ G=\big[\mathcal I,\mathcal A,\mathcal H,\mathcal T,K,\phi,A,(\zeta_i)_{i\in\mathcal I},g\big], \] with type-contingent behavior strategies \(S_i:\Theta_i\to\mathcal B_i\), terminal history \(T^G(S(\theta))\), and implementation condition \[ g\!\left(T^G\!\left([S_i(\theta_i)]_{i\in\mathcal I}\right)\right)=f(\theta) \quad\forall \theta\in\Theta. \] The key issue is that distinct equilibrium paths can generate different intermediate information states and different type bundlings, so a single direct-report object does not capture the relevant disclosure process [2402.13580]. The proposed replacement is an **operator** \(\gamma\) that records how a current set of candidate states is refined by equilibrium actions. A semi-operator is \[ \gamma:\left(\prod_{i\in\mathcal I}2^{\Theta_i\setminus\{\varnothing\}}\right)\times \Theta \to \prod_{i\in\mathcal I}2^{\Theta_i\setminus\{\varnothing\}}, \] and an operator is a semi-operator whose slices partition the candidate set. The operator induced by a given mechanism-strategy pair is denoted \(\gamma^{[G,S]}\) [2402.13580]. Two revelation-style results follow. For **additive** solution concepts, sequential-move implementation is equivalent to simultaneous-move implementation. The paper lists \(\mu\)-PBE, weak-dominance\(^*\), and max-min equilibrium as examples, and proves, for instance, that in perfect-recall games: \[ f \text{ can be }\mu\text{-PBE implemented} \iff f \text{ satisfies Bayesian IC} \iff f \text{ is implementable by a simultaneous-move game.} \] For **monotonic** solution concepts, the paper defines a canonical operator \(\gamma^{(\rho,f)}\) and proves that, if \(\rho\) is dissectible, monotonic, and normal, \[ f \text{ is }\rho\text{-implementable} \iff \gamma^{(\rho,f)} \text{ is achievable}. \] Achievability means \[ \exists N\in\mathbb N \text{ such that } \gamma_{(N)}[\Theta,\theta]=\{\theta\} \quad \forall \theta\in\Theta. \] The paper gives explicit versions for obvious dominance and strong-obvious dominance: \[ f \text{ is OSP} \iff \gamma^{OD-f}\text{ is achievable}, \qquad f \text{ is SOSP} \iff \gamma^{SOD-f}\text{ is achievable}. \] [2402.13580] Here the “fictitious” direct object is neither a truthful direct mechanism nor a mediated direct mechanism, but a primitive-defined dynamic disclosure process. The operator plays the role that direct revelation plays in simultaneous games: it is the canonical representation to which implementation reduces. ## 5. Filtered reports, informational equivalence, and self-confirming mechanisms A more literal use of the phrase appears in work on self-confirming mechanisms. The environment is standard Bayesian mechanism design except that the designer does **not** know the prior distribution of types and instead learns from the behavior induced by the mechanism itself. A mechanism is \[ \mathcal{M}=(M,\omega), \] with message space \(M=\prod_i M^i\), outcome rule \(\omega:M\to\Delta(O)\), and strategy profile \[ \sigma=(\sigma^1,\dots,\sigma^n), \qquad \sigma^i:\Theta^i\to \Delta(M^i). \] For any prior \(\pi\), the induced message distribution is \(\sigma_\pi\), and the feasible-prior set generated by the observed mechanism data is \[ \Pi^{\mathcal{M}(\sigma,\sigma_{\pi_0})} := \left\{\pi \in \Delta(\Theta): \sigma_\pi=\sigma_{\pi_0}\right\}. \tag{1} \] The paper also defines the kernel informativeness preorder \[ g \succsim h \quad \Longleftrightarrow \quad \{\mu: g_\mu=g_{\mu_0}\}\subseteq \{\mu:h_\mu=h_{\mu_0}\} \ \ \text{for all }\mu_0, \tag{2} \] and \(g\sim h\) when each is at least as informative as the other [2603.12532]. The classical revelation principle is inadequate in this setting because truthful direct mechanisms generally reveal the full prior, whereas many indirect mechanisms reveal much less. To preserve informational content, the paper introduces a **fictitious direct mechanism**. A direct mechanism is \(\mathcal D=(\Theta,\delta)\), but the observed report is filtered through \[ \varphi=(\varphi^1,\dots,\varphi^n), \qquad \varphi^i:\Theta^i\to \Delta(\Theta^i). \] The feasible-prior set induced by filtered reports is \[ \Pi^{\mathcal{D}(\varphi,\varphi_{\pi_0})} := \left\{\pi\in \Delta(\Theta): \varphi_\pi=\varphi_{\pi_0}\right\}. \tag{3} \] An augmented mechanism \((\mathcal M,\sigma)\) and a fictitious direct mechanism \((\mathcal D,\varphi)\) are **equivalent** if \[ \omega\circ \sigma=\delta, \qquad \sigma\sim\varphi. \tag{4} \] The strong fictitious revelation principle then states that every fictitious direct mechanism admits an equivalent augmented mechanism, and conversely, under uncountable standard Borel message and type spaces, every augmented mechanism admits an equivalent fictitious direct mechanism. The weak finite version replaces filters by weak filters, and the equivalence target remains the same informational content in the kernel sense [2603.12532]. The same paper defines an incentive-compatible augmented mechanism \((\mathcal M,\sigma)\) to be \(\pi_0\)-self-confirming if there exists \[ \pi \in \Pi^{\mathcal{M}(\sigma,\sigma_{\pi_0})} \] such that \[ V^0(\mathcal{M},\sigma,\pi) \ge V^0(\mathcal{M}',\sigma',\pi) \] for all incentive-compatible augmented mechanisms \((\mathcal M',\sigma')\), where \[ V^0(\mathcal{M},\sigma,\pi) = \int_{\Theta} u^0((\omega\circ \sigma)(\theta),\theta)\,\pi(d\theta). \] The direct analogue replaces \((\mathcal M,\sigma)\) by \((\mathcal D,\varphi)\) and compares over incentive-compatible direct mechanisms \(\mathcal D'\). A robust refinement is then defined by perturbing the filter so that it reveals \(\varepsilon\)-truth on a small product set \(A=\prod_i A^i\), yielding the perturbed feasible-prior set \[ \Pi^{\mathcal{D}(\varphi^A,\varphi^A_{\pi_0})} = \Pi^{\mathcal{D}(\varphi,\varphi_{\pi_0})} \cap \left\{\pi \in \prod_{j\in N}\Delta(\Theta^j): \pi^i|_{A^i}=\pi_0^i|_{A^i},\ i\in N\right\}. \tag{6} \] A fictitious direct mechanism is robustly \(\pi_0\)-self-confirming if self-confirmation survives every sufficiently small truth-revealing perturbation [2603.12532]. In the monopoly application, a deterministic posted price \(p\) corresponds to \[ \delta^p(\theta)= \begin{cases} 1_{(1,p)}, & \theta\ge p,\\ 1_{(0,0)}, & \theta 0. \] In round \(t\), the mechanism chooses allocation probability \(a_t\in[0,1]\) and payment \(p_t\); the buyer’s discounted utility is \[ \sum_{t=1}^T \gamma_t(a_t v-p_t), \] and the effective discounted horizon is \[ T_\gamma := \sum_{t=1}^T \gamma_t. \] Regret is measured against the omniscient full-surplus benchmark: \[ Reg(M)=\sup_{v\in[0,1]} \bigl(Tv-Rev(v;M)\bigr), \qquad Rev(v;M)=\sum_{t=1}^T p_t. \] [2604.24093] The paper distinguishes **indirect learning mechanisms**, which adaptively post prices or menus \(M_t=\{(a,p(a)): a\in A\subseteq[0,1]\}\), from **direct revelation mechanisms**, in which the buyer first reports \(\hat v\) and the seller commits to a full sequence \((p_t(\hat v),a_t(\hat v))_{t=1}^T\) subject to IC and PIR. Its main conceptual result is an equivalence theorem: 1. For any indirect menu-based learning mechanism \(I\), there exists a direct mechanism \(D\) such that \[ Rev(v;D)=Rev(v;I)\qquad \forall v. \] 2. For any IC and PIR direct mechanism \(D\), there exists an indirect mechanism \(I\) such that \[ Rev(v;I)\ge Rev(v;D)\qquad \forall v. \] Hence the optimal regret achievable by indirect learning mechanisms equals the optimal regret achievable by direct revelation mechanisms [2604.24093]. The same paper shows that menu mechanisms can strictly improve regret relative to posted prices. With a quadratic payment menu satisfying \[ p_t'(\underline a_t)=\underline v_t,\qquad p_t'(1)=\overline v_t,\qquad p_t(\underline a_t)=\underline a_t\cdot \underline v_t, \] and delayed updates of length \[ D=T_\gamma \log(T_\gamma/\epsilon), \] the mechanism forces every non-myopic best response to be an \(\epsilon\)-myopic best response, \[ \left|v a_t-p_t(a_t)-\max_a\{va-p_t(a)\}\right|\le \epsilon, \] shrinks the candidate interval by \[ [\underline v_{t+D},\overline v_{t+D}] = \bigl[p_t'(a_t)-\sqrt{2\epsilon},\;p_t'(a_t)+\sqrt{2\epsilon}\bigr] \cap [\underline v_t,\overline v_t], \] and achieves total regret \[ O(T_\gamma\log T_\gamma). \] In the geometric discount case \(\gamma_t=\gamma^t\) for constant \(\gamma<1\), the paper notes that \(T_\gamma=O(1)\), so the regret becomes \(O(1)\) [2604.24093]. This is not a failure result. It is a revelation-principle-style equivalence in which adaptive, data-driven online learning and explicit type elicitation are “two languages towards solving the same problem.” Relative to the fictitious variants above, it shows that once the mechanism class is rich enough—here, menu mechanisms with fractional allocations—the indirect/direct divide can collapse again, and may do so at the level of optimal regret rather than only at the level of implementability. ## 7. Conceptual synthesis The supplied literature does not present a single universal theorem under the name “fictitious revelation principle.” It instead presents several mechanisms by which the classical reduction from indirect implementation to truthful direct revelation becomes qualified, redirected, or generalized. One recurring pattern is that the preservation target changes. In the classical benchmark, the target is **truthful implementability** of the same social choice function. In the action-energy critique, the issue is **physical feasibility** of the direct step and the redistribution of action costs to the designer [1008.3287]. In the action-format failure critique, the issue is **behavioral form**: the same outcome may be reproduced, but not through honest and obedient play [2011.07878]. Under CPT, the relevant target becomes **lottery-equivalent implementability under mediation**, because direct revelation without mediator-generated signals no longer works [2101.08722]. In sequential games, the reduction target becomes a **canonical disclosure operator** rather than a direct mechanism [2402.13580]. In self-confirming design, what must be preserved is the mechanism’s **informational content**, captured through feasible-prior sets and filters [2603.12532]. In dynamic pricing, the preserved object is **optimal regret**, and learning itself becomes revelation in disguise [2604.24093]. A plausible implication is that “fictitious revelation principle” is best understood as a representation program rather than a single doctrinal claim. The common question is not merely whether one can replace an indirect mechanism with a direct one, but **which features must survive the replacement**: truthfulness, implementability, energy feasibility, informational content, dynamic disclosure structure, or regret. On that reading, the recent literature does not abolish the revelation principle; it decomposes it into separable components and shows that different environments preserve different ones. Source: https://www.emergentmind.com/topics/fictitious-revelation-principle