Inferable Signaling Schemes

• Inferable signaling schemes are public signaling policies that enable agents to recover strategic posteriors from a committed disclosure rule, crucial in auction and advertising models.
• They guide optimal revenue and welfare design by balancing information disclosure, competition, and privacy constraints in both deterministic and randomized settings.
• Extensions include distributed signaling and identification in routing games and Bayesian persuasion, highlighting computational and design trade-offs.

Inferable signaling schemes are signaling policies whose informational consequences can be recovered from a committed disclosure rule and the realized signal, so that receivers, bidders, or other agents can compute the relevant posterior and act on it. In the foundational auction formulation of Emek, Feldman, Gamzu, Paes Leme, and Tennenholtz, the scheme is publicly committed to by the auctioneer, bidders infer posterior beliefs over goods, and second-price bids equal posterior expected values; later work extended the same inferability theme to privacy-constrained ad auctions, distributed mediation, prior identification in routing games, and repeated-interaction settings in which the hidden object to be inferred is itself a bias parameter, a prior, or the signaling scheme (Emek et al., 2012, Chen et al., 2013, Verbree et al., 2021, Chen et al., 2024, Probine et al., 1 Oct 2025).

1. Foundational formulation in auctions

The canonical inferable signaling model studies a probabilistic single-item auction with a finite state space of goods $\Omega=[m]$ and a prior $p \in \Delta(m)$. Nature draws $\omega \in \Omega$ according to $p$, and the auctioneer observes $\omega$ while bidders do not. In the known-valuations case, bidder $i$ has deterministic valuation $V(i,j)\ge 0$ for good $j$, and the normalized valuations are $\Psi(i,j)=p(j)\cdot V(i,j)$. In the Bayesian case, the auctioneer has probabilistic knowledge through valuation matrices $V_1,\ldots,V_k$ with $p \in \Delta(m)$0 over outcomes $p \in \Delta(m)$1, and $p \in \Delta(m)$2.

A signaling scheme consists of a finite set of signals $p \in \Delta(m)$3 and a signaling function $p \in \Delta(m)$4 satisfying

$p \in \Delta(m)$5

Given good $p \in \Delta(m)$6, the auctioneer sends signal $p \in \Delta(m)$7 with probability $p \in \Delta(m)$8. The rule is publicly committed to and hence inferable by bidders. The induced signal distribution and posterior are

$p \in \Delta(m)$9

with Bayes plausibility

$\omega \in \Omega$0

Upon observing $\omega \in \Omega$1, bidder $\omega \in \Omega$2's posterior expected value is $\omega \in \Omega$3, and in a second-price auction truth-telling remains a dominant strategy: for every $\omega \in \Omega$4 and $\omega \in \Omega$5, bidding $\omega \in \Omega$6 is a dominant strategy. Deterministic signaling imposes $\omega \in \Omega$7 and corresponds to clustering; randomized signaling allows fractional assignments $\omega \in \Omega$8, interpretable as fractional clustering (Emek et al., 2012).

This model was introduced as a formal representation of impression selling in display advertising. States correspond to impressions characterized by attributes such as time, cookies, demographics, and location; valuations reflect advertiser suitability to the realized impression; and inferability means that advertisers can compute posterior expected values from a public policy mapping impressions to coarse audience segments.

2. Revenue design, welfare, and the role of partial disclosure

In the known-valuations case, expected revenue under a signaling scheme $\omega \in \Omega$9 is

$p$0

and social welfare is

$p$1

In the Bayesian case, revenue averages the same second-highest quantity over the valuation matrices with weights $p$2. The central design problem is the revenue maximization signaling problem (RMS): choose the inferable signaling rule that maximizes second-price revenue while respecting Bayes plausibility.

Several structural properties are sharp. In the known-valuations setting, there exists an optimal signaling scheme with at most $p$3 signals, each uniquely identified by the ordered pair $p$4 of bidders achieving the highest and second-highest expected normalized valuation under $p$5. The resulting optimization is polynomially solvable by LP1, which has $p$6 variables and $p$7 constraints. The same analysis also establishes that the optimal revenue obtainable by a signaling scheme is at most twice the optimal revenue obtainable by a clustering scheme, and that this factor-$p$8 is tight. On the efficiency side, there exists a revenue-optimal signaling scheme $p$9 with

$\omega$0

where $\omega$1 and $\omega$2, so at least half of the maximum social welfare can be preserved within a revenue-optimal inferable scheme (Emek et al., 2012).

The canonical examples explain why inferable partial disclosure matters. With $\omega$3 bidders and $\omega$4 goods under a uniform prior, where bidder $\omega$5 values only good $\omega$6 at $\omega$7, no signaling yields second-price revenue $\omega$8, and full revelation yields revenue $\omega$9 because the second-highest realized value is always $i$0. Clustering goods into pairs yields revenue $i$1: when the signal is the pair containing the realized good, two bidders have expected value $i$2. This directly contradicts the naive view that more informative signals necessarily help the seller. In these models, full revelation can destroy competition, whereas an inferable but coarser public policy preserves competition by keeping at least two bidders with high posterior expected values.

A plausible implication is that inferability is not merely a transparency constraint. It is also a design primitive: public commitment lets receivers compute posteriors, but the seller can choose how much state information survives into those posteriors.

3. Structural bounds, signal complexity, and algorithmic frontiers

The Bayesian extension preserves some structure but sharply changes complexity. If the number of valuation outcomes $i$3 is fixed, there exists an optimal signaling scheme with only $i$4 relevant signal types, and the problem is solvable in polynomial time by LP2. If the number of goods $i$5 is fixed, the feasible signal vectors can be partitioned into regions of $i$6 cut out by hyperplanes enforcing a consistent top-two bidder pair across all $i$7, and the number of regions with non-empty interior is $i$8 by the Whitney number bound. In the general Bayesian case, however, the decision version of RMS is NP-complete already for $i$9 bidders. Independently of these computational frontiers, every probabilistic single-item auction admits an optimal signaling scheme with at most $V(i,j)\ge 0$0 signals, and this is tight: Example 5.1 shows instances with $V(i,j)\ge 0$1 requiring $V(i,j)\ge 0$2 signals, while Example 5.2 shows that if the signal budget is $V(i,j)\ge 0$3, achievable revenue can scale only as $V(i,j)\ge 0$4 although $V(i,j)\ge 0$5 signals achieve $V(i,j)\ge 0$6 (Emek et al., 2012).

A second line of work studies inferability under explicit communication or feasibility constraints. In constrained signaling, the auctioneer cannot necessarily describe the exact item and instead chooses a signaling map from a feasible family $V(i,j)\ge 0$7, such as a $V(i,j)\ge 0$8-bounded-length signal family with $V(i,j)\ge 0$9 or a bipartite feasibility system $j$0. In this setting, deterministic schemes suffice for welfare maximization, bipartite welfare with known valuations admits a randomized polynomial-time $j$1-approximation via monotone submodular maximization under a truncated partition matroid, and there is no polynomial-time $j$2 approximation for welfare or revenue in communication-constrained signaling with known valuations unless $j$3. The same literature gives two explicit inferable compression mechanisms: a Multiplicative Weights signaling scheme for inner-product valuations with $j$4 updates and bit-length $j$5, and a Johnson–Lindenstrauss signaling scheme for subspace valuations whose succinct encoding preserves enough geometry for bidders to reconstruct approximate valuations from the signal alone (Dughmi et al., 2013).

In large-state ad-auction models, signal complexity can itself be a barrier to inference. In the Bayesian-valuation setting with binary targeting indicators, any public signaling scheme using fewer than $j$6 signals cannot guarantee any constant multiplicative fraction of the optimal revenue. Yet under a tail-balanced prior and monotone hazard rate assumptions, a simple tail-pooling public scheme achieves revenue at least $j$7 while remaining operationally simple: reveal full information when there is enough competition, and pool exactly the one-high tail states to create competition (Badanidiyuru et al., 2017).

Taken together, these results show that inferability does not imply algorithmic simplicity. The number of posterior classes may be small because of structure, large because of approximation barriers, or deliberately compressed through geometric or learning-based encodings.

4. Privacy constraints and distributed information

One important modification weakens inferability on purpose. In the $j$8-anonymous signaling model for ad auctions, an impression belongs to one of $j$9 categories with prior probabilities $\Psi(i,j)=p(j)\cdot V(i,j)$0, bidder $\Psi(i,j)=p(j)\cdot V(i,j)$1 has value $\Psi(i,j)=p(j)\cdot V(i,j)$2, and the auctioneer uses deterministic signaling $\Psi(i,j)=p(j)\cdot V(i,j)$3. The $\Psi(i,j)=p(j)\cdot V(i,j)$4-anonymity constraint requires every signal support to have size at least $\Psi(i,j)=p(j)\cdot V(i,j)$5:

$\Psi(i,j)=p(j)\cdot V(i,j)$6

Equivalently, the posterior $\Psi(i,j)=p(j)\cdot V(i,j)$7 has support size at least $\Psi(i,j)=p(j)\cdot V(i,j)$8, so bidders cannot infer the exact category. Welfare and revenue reduce to bundling categories into bundles of size at least $\Psi(i,j)=p(j)\cdot V(i,j)$9 and running a second-price auction on each bundle. Under this constraint, there is no polynomial-time $V_1,\ldots,V_k$0-approximation for welfare maximization for any constant $V_1,\ldots,V_k$1 unless $V_1,\ldots,V_k$2, and revenue maximization is NP-hard even when the number of signals is constant. On the algorithmic side, there is a randomized polynomial-time $V_1,\ldots,V_k$3-approximation for welfare maximization and, via a general $V_1,\ldots,V_k$4 transfer from welfare to revenue, a randomized polynomial-time $V_1,\ldots,V_k$5-approximation for revenue maximization. The model makes the privacy–inferability trade-off explicit: increasing $V_1,\ldots,V_k$6 reduces posterior distinguishability and tends to reduce targeting precision, although broader bundles can also thicken the market (Chen et al., 2013).

A different departure distributes information across multiple sources rather than hiding it from bidders. In distributed signaling games, the state space is a set of items or contexts $V_1,\ldots,V_k$7, bidders bid in a second-price auction, and each mediator $V_1,\ldots,V_k$8 observes only a partition $V_1,\ldots,V_k$9 of the item space. A mediator may report a coarser partition $p \in \Delta(m)$00, and the public information seen by bidders is the joint partition

$p \in \Delta(m)$01

Because partitions are broadcast and the prior $p \in \Delta(m)$02 and valuation matrix $p \in \Delta(m)$03 are known, posteriors $p \in \Delta(m)$04 and counterfactual revenues are inferable from the reports. This allows the auctioneer to compute each mediator’s marginal contribution $p \in \Delta(m)$05 and pay mediators according to the Shapley value

$p \in \Delta(m)$06

Under this Shapley rule, a pure Nash equilibrium always exists and best-response dynamics converge to it; moreover, at every Nash equilibrium the auctioneer’s revenue is at least the silent baseline $p \in \Delta(m)$07. In the coordination version of the problem, general distributed approximation is hard, but if all mediators are local experts with partitions of the form $p \in \Delta(m)$08, there is an efficient $p \in \Delta(m)$09-approximation algorithm (Feldman et al., 2014).

These models enlarge the inferability agenda. Inferability can be restricted, as in $p \in \Delta(m)$10-anonymity, or mechanized, as in distributed mediation where counterfactual marginal contributions must themselves be computable from public disclosures.

5. Inferability as identification, detection, and learning

Later work repurposes signaling away from revenue extraction and toward identification. In bias detection via signaling, the agent and designer share a prior $p \in \Delta(m)$11 and utilities $p \in \Delta(m)$12, but the agent updates according to a convex-combination bias model

$p \in \Delta(m)$13

where $p \in \Delta(m)$14 measures prior-stickiness. A signaling scheme is $p \in \Delta(m)$15-detectable if some signal induces different optimal actions when $p \in \Delta(m)$16 versus $p \in \Delta(m)$17, and it is identifiable over $p \in \Delta(m)$18 if the map

$p \in \Delta(m)$19

is injective on $p \in \Delta(m)$20. The geometric characterization is exact: single-signal detectability at threshold $p \in \Delta(m)$21 is possible if and only if

$p \in \Delta(m)$22

and detectability at $p \in \Delta(m)$23 is feasible if and only if $p \in \Delta(m)$24 for at least one $p \in \Delta(m)$25. A revelation-principle reduction shows that one may restrict attention to direct schemes with $p \in \Delta(m)$26, and a polynomial-time LP maximizes the probability of boundary signals that place the $p \in \Delta(m)$27-biased posterior on the indifference boundary. The same paper proves that adaptive schemes do not improve worst-case sample complexity for threshold testing (Chen et al., 2024).

Inferability also appears as prior recovery from equilibrium behavior. In routing games with uncertain network state, a traffic information system commits to a public signaling scheme $p \in \Delta(m)$28, users form posteriors by Bayes’ rule, and route choices induce a Wardrop equilibrium flow $p \in \Delta(m)$29. A signaling scheme is inferable if the map

$p \in \Delta(m)$30

is injective. Under strictly increasing edge costs, continuity of equilibrium edge-flows, and the existence of at least two states with distinct pure-state Wardrop equilibrium sets, there exists a prior-identifying public signaling scheme with $p \in \Delta(m)$31 messages. The paper also gives an iterative finite-time algorithm that adjusts two-supported signal rows until each signal becomes informative, proves robustness under small perturbations when the induced linear system remains full rank, and shows that with multiple known priors $p \in \Delta(m)$32 the fractions $p \in \Delta(m)$33 are identifiable if and only if the stacked response matrix has rank $p \in \Delta(m)$34 (Verbree et al., 2021).

A third identification problem arises when the receiver does not initially know the signaling scheme. In Bayesian persuasion with repeated interaction, the sender commits to $p \in \Delta(m)$35, but the receiver estimates the posteriors from empirical counts of observed $p \in \Delta(m)$36 pairs. If $p \in \Delta(m)$37 is the true posterior for signal $p \in \Delta(m)$38, $p \in \Delta(m)$39 is its distance to the nearest receiver decision boundary, and $p \in \Delta(m)$40 measures posterior stochasticity, then the per-round performance gap satisfies

$p \in \Delta(m)$41

and also

$p \in \Delta(m)$42

The same work lower-bounds the number of samples needed to approach known-commitment performance, shows that persuasion requires more samples than the corresponding Stackelberg-leader problem in a family of flower games, and proposes two design methods for inferable schemes: stochastic gradient descent on the sender’s inference-setting utility and optimization with a boundedly-rational receiver model. In the reported safety-alert example, the SGD-derived scheme uses fewer effective signals and makes the receiver’s optimal actions more distinct (Probine et al., 1 Oct 2025).

This suggests a broadening of the term. In the early auction literature, inferability primarily means that receivers can compute posteriors from a public rule. In these later papers, inferability becomes an identifiability property of the induced action or equilibrium map.

6. Conceptual distinctions and adjacent frameworks

The meaning of inferability depends on the underlying model of signals. In causal interdependence models, generated signals are characterized by conditional distributions such as $p \in \Delta(m)$43 or $p \in \Delta(m)$44 and are conditionally independent given the underlying variable of interest, whereas interpreted signals are deterministic interpretations of an underlying attribute state and generally are not d-separated by the outcome variable. This distinction yields different inferability patterns: in generated models, inference about the state aggregates through conditional independence; in interpreted models, correctness variables can be dependent even when predictions appear independent, and conditional-independence tests become substantive evidence about the data-generating structure (Wellman et al., 2012).

A separate geometric literature studies signaling with commitment by letting the sender commit directly to a state-contingent distribution over sender actions. For each sender action $p \in \Delta(m)$45, the interim payoff graph is

$p \in \Delta(m)$46

and the sender’s attainable payoffs under action commitment are characterized by the topological join of the graphs $p \in \Delta(m)$47 subject to Bayes plausibility. When the sender can commit to both payoff-irrelevant messages and actions, the attainable set expands to the convex hull of the union of these graphs, and the value becomes the corresponding concave envelope. A necessary and sufficient condition for messages to be redundant is that the join envelope be concave. This makes ex post inferability especially transparent: under action commitment, the observable action itself indexes a unique posterior chosen ex ante by the sender’s committed rule (Boleslavsky et al., 2023).

In signaling Bayesian Stackelberg games with double-sided information asymmetry, the leader’s realized action is hidden, the follower has a private type and may misreport it, and the leader’s commitment is interpreted as a probability measure over the follower’s belief space. The feasible set is partitioned into reporting cells, the leader’s value is piecewise linear on these cells, and signaling can strictly improve the leader’s expected utility relative to standard Bayesian Stackelberg equilibrium by revealing partial information about the leader’s realized action in a type-contingent way (Li et al., 2022).

Finally, the power of inferable disclosure interacts with equilibrium inefficiency. In games with a random state of nature, the additional power of moving from full information to optimal public signaling, from optimal public to optimal private signaling, and from optimal private to optimal ex-ante private signaling is tightly bounded by the Price of Anarchy of the realized games. In cost-minimization games the multiplicative gain is at most $p \in \Delta(m)$48, while in payoff-maximization games the corresponding ratio is at least $p \in \Delta(m)$49, and these bounds are tight (Nachbar et al., 2020).

Across these strands, inferable signaling schemes are best understood not as a single canonical object but as a family of commitment devices whose common feature is recoverability of strategically relevant beliefs. What changes from one literature to another is the object to be inferred: a posterior over goods, a privacy-preserving indistinguishability class, a mediator’s marginal contribution, a prior, a bias parameter, a hidden action, or the signaling scheme itself.