Bayesian Nash Equilibrium

Why Nash needed an upgrade

John Nash's 1950 equilibrium concept assumes complete information: every player knows the payoff matrix, every player knows every other player's preferences, and this is common knowledge. Most real strategic situations violate this assumption. You bid in an auction without knowing rivals' valuations. You negotiate a contract without knowing the other party's cost. You enter a market without knowing whether the incumbent is tough or weak.

John Harsanyi's contribution in three papers in Management Science (1967-68) was a technical and conceptual sleight of hand. Instead of trying to model the messy reality of "what I think about what you think I think...", he proposed: imagine that before the game begins, Nature draws each player's type θ_i from a publicly known prior distribution p(θ). Each player observes only their own type. From this point, the game proceeds with imperfect information about Nature's draw, but everyone knows the joint distribution and the structure.

This is the Harsanyi transformation. It converts every game of incomplete information into a Bayesian game (N, A, T, p, u) — players, actions, types, prior, payoffs — that fits inside the standard game-theoretic apparatus. Equilibrium of this Bayesian game is the Bayesian Nash equilibrium.

The formal definition

A Bayesian Nash equilibrium is a strategy profile σ* = (σ₁*, ..., σ_N*) where σ_i*(θ_i) is the action chosen by player i if their type is θ_i, satisfying for every player i and every type θ_i:

σ_i*(θ_i) ∈ arg max_ai Σ_θ−i p(θ_−i | θ_i) · u_i(a_i, σ_−i*(θ_−i), θ_i, θ_−i)

In words: knowing your own type θ_i, you choose the action that maximises your expected utility, taking the expectation over the other players' types using the posterior p(θ_−i | θ_i) — derived from the prior by Bayes' rule — and assuming the others play their equilibrium strategies. Every type's choice must satisfy this condition simultaneously. The fixed point is the BNE.

Worked example: first-price sealed-bid auction

The auction is the canonical BNE. Two bidders, valuations v₁, v₂ drawn independently from uniform [0, 1]. The seller awards the item to the highest bidder at their submitted bid. Each bidder knows their own valuation but not the rival's.

Look for a symmetric BNE in which each bidder's strategy is some increasing function b(v). Bidder 1 with valuation v₁ chooses bid b₁ to maximise expected payoff:

EU(b₁) = (v₁ − b₁) · Pr(b₁ > b(v₂))

If bidder 2 plays the symmetric strategy b(v), then Pr(b₁ > b(v₂)) = Pr(v₂ < b⁻¹(b₁)). Differentiating and setting equal to zero, and imposing symmetry b₁ = b(v₁), gives the differential equation

(v − b) · 1/b'(v) = b'(v) · b⁻¹(b(v)) − b(v)... → → → ... b(v) = v/2.

With two bidders and uniform values: each bids half their valuation. With N bidders: b(v) = v·(N−1)/N. With 10 bidders you bid 0.9v; with 100 bidders 0.99v. Bid shading shrinks as the field grows, reflecting fiercer competition.

The BNE is constructive: you have a closed-form strategy, plug in your private valuation, win some auctions, lose others, and earn a positive expected surplus equal to (1/N) of the value above the second-highest type. Compare to a complete-information setting where rivals know your valuation and could bid v − ε — leaving you no surplus.

Revenue equivalence theorem

Vickrey (1961) and Myerson (1981) proved a striking BNE result. In any standard auction format with independent private values, risk-neutral bidders, and zero reserve price, the seller's expected revenue is the same. First-price, second-price (Vickrey), all-pay, descending-Dutch — all yield identical expected revenue.

The intuition is the envelope theorem. Bidder utility in any standard auction is u(v) = (1/N!) ∫₀ᵛ N(N−1) y^N−2 dy, by Myerson's lemma. The total bidder rent is fixed by the value distribution and the allocation rule. Different formats redistribute that rent in different ways, but the seller's share is the same complement. Myerson shared the 2007 Nobel Prize for the broader mechanism-design framework that revenue equivalence belongs to.

BNE vs related solution concepts

	Bayesian Nash	Nash	Subgame Perfect	Perfect Bayesian	Sequential	Trembling-Hand
Information	Incomplete	Complete	Complete, dynamic	Incomplete, dynamic	Incomplete, dynamic	Any
Strategy is...	Type → action	Action	Action history → action	Type, history → action	Same as PBE plus consistency	Action with trembles
Belief consistency	Bayes on prior	—	—	Bayes where applicable	Bayes + Kreps-Wilson consistency	—
Eliminates...	Non-optimal types	Dominated strategies	Empty threats	Off-path madness	Same plus consistency	Knife-edge equilibria
Canonical paper	Harsanyi 1967-68	Nash 1950	Selten 1965	Fudenberg-Tirole 1991	Kreps-Wilson 1982	Selten 1975
Use when	Hidden types	All info public	Sequential, complete info	Sequential, hidden types	Same, with refinements	Multiple Nash equilibria

BNE is the workhorse. Perfect Bayesian and sequential equilibrium extend it to multi-stage games with private information; trembling-hand perfect equilibrium adds robustness against small mistakes.

Variants and extensions

• Independent private values (IPV). The cleanest setting: types are independent and only affect own payoff. First-price auction equilibrium derived above is IPV.
• Common values. Each bidder gets a noisy signal about a common true value (oil-lease auctions). Wilson (1969) and Milgrom-Weber (1982) generalised; introduces the winner's curse.
• Correlated types. Wilson (1985); Crémer-McLean (1985). Correlation generates linkage between types, and the seller can extract more rent than under independence.
• Mechanism design. Myerson (1981), Hurwicz, Maskin, Myerson (2007 Nobel). Design the rules of the Bayesian game to achieve a target outcome. Optimal auctions, monopoly screening, public-goods provision all live here.
• Global games. Carlsson-van Damme (1993); Morris-Shin (1998). Each player gets a noisy signal of a fundamental, slightly correlated across players. Resolves coordination-game multiplicity into a unique BNE.
• Bayesian persuasion. Kamenica-Gentzkow (2011). The informed party commits to an information disclosure rule, then plays a Bayesian game with the receiver. A rapidly growing literature.

Where BNE shows up in the real world

• FCC spectrum auctions. The 1994-present U.S. spectrum auctions (raised over $200B for the Treasury through 2023) are designed by Milgrom, Wilson, McAfee using BNE principles. The 2017-2020 incentive auction redistributed 84 MHz from broadcasters to wireless.
• Treasury bond auctions. Daily uniform-price multi-unit auctions of U.S. T-bills (≈ $5T annual volume). Each bidder submits a demand curve; uniform-price BNE differs from the discriminatory format the Treasury used pre-1992.
• Sponsored search auctions. Google AdWords, now Ads (≈ $300B annual revenue), runs a generalised second-price auction with millions of BNE-style bidders. Edelman-Ostrovsky-Schwarz (2007) and Varian (2007) analysed the equilibrium.
• Procurement. Government contracting uses first-price sealed-bid; Krasnokutskaya (2011) shows bid shading patterns consistent with BNE under heterogeneous bidder costs.
• Online advertising. Real-time bidding markets clear billions of ads per day in a BNE-style mechanism. RTB ad-tech is a BNE laboratory at industrial scale.

A brief history

John Harsanyi published his three foundational papers — "Games with incomplete information played by Bayesian players" — in Management Science in 1967 and 1968. Each paper was technical and densely formal; together they invented the discipline of Bayesian game theory. Harsanyi had survived a Nazi labour camp in Hungary, escaped to Australia in 1950, completed a PhD at Stanford, and was a UC Berkeley professor by the time of his Nobel.

The 1994 Nobel Prize citation read: "for their pioneering analysis of equilibria in the theory of non-cooperative games." It was shared by Nash (the 1950 equilibrium concept), Selten (subgame perfection 1965; trembling hand 1975), and Harsanyi (Bayesian games 1967-68). Selten, like Harsanyi, came from devastated wartime Europe — both built the theoretical framework for thinking about deception, signaling, and reputation that defines modern economics.

The 2007 Nobel to Hurwicz, Maskin, and Myerson cited mechanism design — the inverse problem to BNE. Where BNE asks "what equilibrium results from a given game?", mechanism design asks "what game gives the equilibrium I want?". Both directions of analysis use the same Harsanyi apparatus.

Common pitfalls

• Forgetting that strategies are functions, not actions. A BNE specifies what every type would do, not just what one type does. Confusion between the two is the most common student error.
• Assuming the prior is observed. Each player observes only their own type; the prior is common knowledge but other players' actual types are not.
• Computing one type's best response. The equilibrium condition is for every type. A single type's best response may not extend to a consistent strategy function.
• Ignoring the winner's curse. In common-value auctions, conditioning your bid on the event "I won" introduces selection — you systematically overestimate value. Naive BNE that ignores this lose money.
• Treating multiple BNEs as equivalent. Different equilibria give different distributions of welfare. Refinements (Cho-Kreps Intuitive Criterion, divinity, D1) pick among them.
• Conflating BNE with perfect Bayesian. BNE is for static (one-shot) games; PBE is for dynamic games with private information and adds belief-update structure at every information set.
• Assuming independence. The independent-private-values assumption is convenient but often violated empirically (collusion, correlated values, common shocks). Correlated-type BNEs are messier.