Revelation Principle

The problem the principle solves

Mechanism design — sometimes called "reverse game theory" — flips the usual game-theoretic question on its head. Instead of taking a game as given and asking how rational players will behave, the designer specifies the desired equilibrium outcome (an allocation, a tax schedule, a public-goods provision rule) and asks: which game (which mechanism) implements that outcome when self-interested players follow their incentives? Auction design, optimal taxation, voting theory, matching markets, public-goods provision — all live in this space.

The catch is that "mechanism" is an extraordinarily wide concept. Formally, a mechanism is a tuple (M₁, …, Mₙ, g) — a message space for each of n agents and an outcome rule g mapping joint messages to outcomes. An agent's message space can be a real number (a bid), a ranking (a vote), a full strategy in a multi-round game (a signalling protocol), or anything else. The designer also gets to pick which equilibrium concept to deploy: dominant strategies, Bayes-Nash equilibrium, Nash, perfect Bayesian, the list goes on.

Without further structure, finding the "best" mechanism over all of this is a functional-analysis problem: optimise over functions g : M → A with no a-priori bound on the dimension of M, subject to equilibrium constraints that themselves depend on g. The space is non-convex, infinite-dimensional, and offers no obvious starting point. For a couple of decades after Hurwicz's 1960 paper opening the field, this was a real obstacle.

The principle, precisely

Let θᵢ ∈ Θᵢ be agent i's type — their private information (a valuation, a productivity, a preference ranking). A social choice function f : Θ₁ × ⋯ × Θₙ → A maps profiles of types to outcomes. A mechanism Γ = (M₁, …, Mₙ, g) implements f in some equilibrium concept E if there is a profile of strategies σᵢ : Θᵢ → Mᵢ that forms an E-equilibrium of Γ and satisfies g(σ₁(θ₁), …, σₙ(θₙ)) = f(θ) for all type profiles θ.

A direct revelation mechanism is the special case where Mᵢ = Θᵢ — agents simply report types — and the outcome rule is f itself. The direct mechanism is called incentive-compatible (IC) in equilibrium concept E if truth-telling — σᵢ(θᵢ) = θᵢ — is an E-equilibrium.

The revelation principle, in full generality, then states:

If a mechanism Γ implements f in equilibrium concept E,
then the direct revelation mechanism (Θ₁, …, Θₙ, f) is
incentive-compatible in equilibrium concept E.

The proof is one line: define σ*ᵢ(θᵢ) = σᵢ(θᵢ) implicitly in the direct mechanism. If misreporting were beneficial — if some type θᵢ preferred reporting θ̂ᵢ — then in the original Γ, the strategy "use σᵢ(θ̂ᵢ) instead of σᵢ(θᵢ) when your type is θᵢ" would have been a profitable deviation, contradicting that σ was an equilibrium of Γ.

Bayesian vs. dominant-strategy versions

The revelation principle splits naturally into two flavours, distinguished by which equilibrium concept E we plug in.

Version	Equilibrium	Credited to	Strength of IC	Information used
Bayesian	Bayes-Nash	Myerson 1979	Bayesian IC: u(θ, θ) ≥ E[u(θ, θ̂)]	Designer knows prior over types
Dominant-strategy	Dominant strategies	Gibbard 1973; Dasgupta-Hammond-Maskin 1979	Strategy-proof: u(θ, θ; θ₋ᵢ) ≥ u(θ, θ̂; θ₋ᵢ) for every θ₋ᵢ	Prior-free

The Bayesian version is the analytically more useful one in continuous-type, asymmetric-information settings (auctions with continuous valuations, taxation with continuous productivities). It requires the principal to know — or assume — the prior distribution over types, and the IC constraints are weaker because they only need to hold in expectation over other agents' types.

The dominant-strategy version is conceptually cleaner and prior-free: a mechanism is strategy-proof if telling the truth is optimal for every agent regardless of what others do. The Vickrey second-price auction is the classic example — bidding your own valuation is a weakly dominant strategy whatever your opponents do. Strategy-proofness is a strong property, however: combined with Gibbard-Satterthwaite, it severely restricts what can be implemented.

The Gibbard-Satterthwaite ceiling

The dominant-strategy revelation principle, paired with Gibbard (1973) and Satterthwaite (1975), produces a striking impossibility result. Suppose the outcome space has at least three alternatives, the preference domain is unrestricted, and the social choice function is onto. Then any strategy-proof direct mechanism is dictatorial — there exists a single agent whose top alternative is always chosen.

This is not a minor restriction. It means that any non-dictatorial voting rule for unrestricted preferences (Borda count, plurality, run-offs, instant-runoff voting, Condorcet methods) admits some preference profile at which some voter is better off lying. The pure-mechanism-design route to dominant-strategy implementation without dictatorship therefore requires escaping one of the hypotheses.

• Restrict preferences. If voters have single-peaked preferences over a one-dimensional alternative space, the median voter rule is strategy-proof. Black (1948) and Moulin (1980) developed the theory. This is why housing-location, public-spending-level, and many ideological-position rules are reasonably non-strategic empirically.
• Use payments. Vickrey (1961), Clarke (1971), and Groves (1973) showed that with transferable utility and quasi-linear preferences, the VCG mechanism — pay each agent the externality they impose on the others — is strategy-proof and selects the efficient outcome. The second-price auction is the single-good special case.
• Switch to Bayesian. Weaken the IC constraint to Bayes-Nash and impose distributional assumptions. The Myerson optimal auction, the Akerlof-Spence-Stiglitz screening models, the Mirrlees taxation model all live here.

Worked example: Myerson's optimal auction

The single application that most justifies the principle's prominence is Myerson's 1981 derivation of the revenue-maximising auction. Setting: a single seller, n risk-neutral bidders with private valuations vᵢ drawn independently from distributions Fᵢ with densities fᵢ. The seller wants the auction format maximising expected revenue.

Without the revelation principle, this is a search over arbitrary auction formats (English, Dutch, first-price, all-pay, multi-stage, …) — intractable. With the revelation principle, Myerson restricts to direct mechanisms specified by an allocation rule qᵢ(v) and an expected-payment rule tᵢ(v):

qᵢ(v) = probability bidder i wins given reports v = (v₁, …, vₙ)
tᵢ(v) = expected payment from bidder i given v
Σᵢ qᵢ(v) ≤ 1                                  (feasibility)

Incentive compatibility forces qᵢ to be monotone non-decreasing in vᵢ, and pins down the transfer up to a constant. Substituting back into the expected revenue and integrating by parts, the seller's problem reduces to maximising expected virtual surplus:

E[Σᵢ qᵢ(v) · ψᵢ(vᵢ)]
  where ψᵢ(vᵢ) = vᵢ − (1 − Fᵢ(vᵢ)) / fᵢ(vᵢ)        (virtual value)

Pointwise maximisation under feasibility gives the optimal rule: allocate the good to the bidder with the highest virtual value if that value is positive, otherwise keep it. When bidders are symmetric and regular (ψ strictly increasing), this is equivalent to a second-price auction with a reserve price r* set where ψ(r*) = 0.

The point is not the result itself — though it is elegant — but the route to it. Without the revelation principle, Myerson would have had to compute equilibria of arbitrary auction games. With it, he reduced the problem to optimising a function of (q, t) subject to one-dimensional monotonicity constraints. The same recipe drives nearly every modern optimal-mechanism paper.

Mirrlees and optimal income taxation

The other canonical application is James Mirrlees's 1971 paper, "An Exploration in the Theory of Optimum Income Taxation" (Nobel 1996). Setting: individuals differ in productivity w (a private type), choose labour supply y to produce income wy, and consume what's left after tax. The government picks a tax schedule T(income) to maximise a social welfare function over the utility distribution.

Without revelation, the government searches over all possible non-linear tax schedules — intractable. With the Bayesian revelation principle, Mirrlees reformulates the problem as a direct mechanism: ask each productivity type w to report w truthfully, give them consumption c(w) and require pre-tax income y(w). The IC constraint is that no type wants to mimic another:

u(c(w), y(w)/w) ≥ u(c(w′), y(w′)/w)   for all w, w′

Mirrlees showed this single-crossing condition reduces to a first-order envelope condition that can be integrated. The resulting optimal schedule has three famous qualitative properties: (1) the marginal tax rate at the very top type is zero (no need to distort the highest type), (2) marginal rates are bounded above by 1 throughout, (3) the schedule is generally not progressive in the textbook sense — middle-rate marginal taxes are often higher than top-rate ones. Diamond (1998) and Saez (2001) refined the empirical shape into the famous "Mirrlees-Diamond-Saez" formula relating optimal top marginal rates to the Pareto coefficient of the income distribution and the elasticity of taxable income. Every modern optimal-taxation paper inherits this revelation-principle scaffolding.

Matching markets and kidney exchange

Matching theory adopts revelation-principle thinking even without explicit payments. The deferred-acceptance algorithm of Gale-Shapley (1962) is, when run with one side as proposers, strategy-proof for the proposers — they have no incentive to misreport their preferences. The student-optimal Boston school-choice mechanism (replaced in 2005 specifically for being manipulable) and the doctor-side serial-dictatorship at the National Resident Matching Program are read through the lens of the revelation principle: the question "what does a stable, IC mechanism look like?" is meaningful only because the revelation principle tells us to look at direct, truth-eliciting rules.

Kidney exchange, designed by Roth, Sönmez, and Ünver from 2004 onward, takes this further. Patients and donors report compatibility profiles to a clearinghouse; the matching algorithm (Top Trading Cycles in its simplest form) constructs feasible chains and cycles of swaps. The system is strategy-proof for patient-donor pairs — they cannot do better by lying about compatibilities — and this property is what makes the U.S. multi-region kidney-exchange system viable. The revelation principle is the conceptual scaffolding that lets the designers prove it.

Why this isn't the end — the limits

The revelation principle is a powerful reduction, but it rests on assumptions worth respecting.

• Commitment. The principal must credibly commit to the mechanism in advance — including its losing branches and its payments — and not renegotiate after seeing reports. Without commitment, agents will rationally anticipate that the principal will redesign the rule upon receiving information; the result is the "ratchet effect" in dynamic principal-agent settings, and a substantially different theory (mechanism design without commitment — Bester-Strausz 2001) emerges.
• No collusion. The result speaks only to individual incentive compatibility, not coalitional. A second-price auction is individually IC but vulnerable to bidder-ring collusion (the ring submits one serious bid and the others stay out). Coalitional strengthenings — group strategy-proofness — are achievable in some settings (Top Trading Cycles, serial dictatorship) but generally restrict feasible f.
• Single-shot. In dynamic settings — repeated interactions, sequential arrivals, learning over time — the standard revelation principle becomes a "dynamic revelation principle" (Myerson 1986, Pavan-Segal-Toikka 2014): agents report types and updated beliefs at every history. The enlarged direct mechanism is far less analytically convenient, and existence results are correspondingly weaker.
• Bounded rationality. The theorem assumes agents can compute their equilibrium strategy in the indirect mechanism — and therefore the direct counterpart — perfectly. Empirically, agents in laboratory experiments often fail to play dominant strategies (overbidding in second-price auctions, misreporting in deferred-acceptance experiments). "Obviously strategy-proof" mechanisms (Li 2017) are a recent attempt to bound the cognitive load.

Variants and extensions

• Implementation in Nash equilibria. Maskin (1977/1999) studies what social choice functions can be implemented when Nash equilibrium (rather than dominant strategies or Bayes-Nash) is the solution concept. The revelation-principle analogue is weaker: a Nash-implementable f need not be implementable by a direct mechanism. The required condition is "Maskin monotonicity."
• Implementation in undominated strategies / iterated dominance. Refinements (Jackson 1992; Abreu-Matsushima 1992) provide intermediate equilibrium concepts where revelation-style reductions can still go through, often with arbitrarily small mechanisms.
• Robust mechanism design. Bergemann-Morris (2005) study mechanisms that work for all priors over types. The "robust revelation principle" reduces to ex-post implementation — truth-telling is optimal regardless of beliefs.
• Algorithmic / approximate mechanism design. Nisan-Ronen (1999) extended mechanism design to settings where the optimal allocation rule is computationally intractable and the designer must approximate. Approximate revelation principles trade off incentive compatibility against computational tractability.
• Without money. When transfers are forbidden (organ allocation, school choice, public-housing), the relevant question is which f are strategy-proof and Pareto-efficient. Hylland-Zeckhauser (1979) and Bogomolnaia-Moulin (2001) develop the random-assignment theory.

Where this shows up

• FCC spectrum auctions. The FCC has raised over $200 billion (cumulative through 2024) selling wireless spectrum. The auction designs — simultaneous multi-round (SMR) and the 2017 incentive auction — were drawn directly from mechanism-design theory in which the revelation principle underpins the optimality arguments. Milgrom, Wilson (Nobel 2020), and Roth contributed.
• Online ad auctions. Google AdWords pays out roughly $200 billion per year. The Generalised Second-Price (GSP) auction in use is not technically IC, but its design and the migration toward "VCG for sponsored search" trace through the revelation-principle literature (Edelman-Ostrovsky-Schwarz 2007).
• Kidney exchange. The UNOS Kidney Paired Donation programme uses a Top Trading Cycles variant, strategy-proof for participants. Over 1000 transplants per year as of 2023.
• School choice. Boston (2005), NYC (2003), New Orleans (2012), Newark, and Denver replaced strategy-vulnerable assignment systems with the Deferred Acceptance algorithm. Roth and Sönmez led the redesign.
• Tax-policy quantification. The Mirrlees-Diamond-Saez optimal-marginal-rate formula was used by Saez and others to argue for top US marginal income-tax rates in the 70-80% range in the 2010s. Every estimate in this literature begins with a revelation-principle reduction.

Common pitfalls

• Confusing "exists" with "unique." The revelation principle says a direct IC mechanism exists implementing f — not that the direct mechanism is the only or best one. The English open auction and the Vickrey sealed-bid auction implement the same outcome; in practice the English version has better robustness, transparency, and behavioural properties.
• Forgetting individual rationality. The revelation principle is about IC, but real designs also need individual rationality (IR): no agent is forced to participate in a mechanism that makes them worse off than their outside option. IR is a separate constraint and constrains the achievable f orthogonally.
• Treating Bayesian and dominant-strategy versions as interchangeable. They are not. Bayesian IC is strictly weaker; a Bayesian-IC direct mechanism may be vulnerable to misreporting once other agents' actual types are observable. The right version to use depends on whether the designer knows the prior and how robust the implementation needs to be.
• Ignoring commitment. Many real-world settings — particularly in dynamic regulation — lack the commitment the theorem assumes. Telling an agent "report your cost and I'll set the price" can fail if the agent expects the regulator to re-optimise upon observing the cost.
• Over-reading Gibbard-Satterthwaite. The result holds for unrestricted preferences. Empirically, preferences are restricted in many real settings (single-peaked, lexicographic, quasi-linear), and useful non-dictatorial strategy-proof mechanisms exist on those domains.