Arrow's impossibility theorem (1951) proved that no voting rule can satisfy all four conditions: unrestricted domain, unanimity, independence of irrelevant alternatives, and non-dictatorship. This is not a limitation of plurality voting or Borda counts—it is a fundamental constraint on preference aggregation itself: any system that aggregates multiple agents' preferences into a collective outcome must violate at least one axiom.
Mechanism design was hoped to be an escape: with transfers, information elicitation, and strategic incentives, surely we could do better than voting rules. We could not. The Gibbard-Satterthwaite theorem (1973) and Myerson-Satterthwaite theorem (1983) showed that the same impossibilities reappear at every level of sophistication. Budget-balanced auctions cannot be efficient and strategy-proof. Matching markets cannot be stable for both sides simultaneously. The constraints are fundamental, not contingent on rule design.
The Impossibility Cascade
What happened instead was a cascade of impossibilities, each one more specific than the last:
Arrow's Impossibility (1951): No social choice function satisfies unrestricted domain + unanimity + IIA + non-dictatorship.
Gibbard-Satterthwaite (1973-75): No mechanism can implement a non-dictatorial, onto social choice function while being strategy-proof (truth-telling is dominant).
Myerson-Satterthwaite (1983): In bilateral trade with asymmetric information, no mechanism achieves efficiency + incentive compatibility + individual rationality + budget balance.
Each theorem constrains a narrower problem. Arrow constrains what outcomes are possible. Gibbard-Satterthwaite constrains what mechanisms can implement those outcomes. Myerson-Satterthwaite constrains what's possible with real budget constraints.
The insight: social choice impossibilities are not artifacts of voting rules. They are fundamental constraints on any system that aggregates information and allocates resources.
How This Changes Everything
Here's what this means in practice:
Voting systems: No voting rule is "best." Any rule you choose sacrifices one of Arrow's axioms. Plurality voting sacrifices IIA (third-party candidates act as spoilers). Borda count sacrifices IIA. Approval voting sacrifices unanimity. The choice is not "which rule is best" but "which axiom are you willing to sacrifice?"
Auctions: The VCG (Vickrey-Clarke-Groves) mechanism achieves efficiency and truth-telling by having each agent pay the externality they impose on others. But it typically doesn't balance the budget — you need external subsidies. Myerson proved this is not an accident: efficiency + truth-telling + budget balance is impossible when valuations are private information. You must sacrifice one.
Market design: The 2012 Nobel Prize recognized matching market design (Alvin Roth) precisely because Arrow's constraints make certain matchings impossible. The deferred-acceptance algorithm achieves stability for one side of the market (proposers) at the cost of strategy-proofness for the other side (acceptors). Both sides can't be strategy-proof simultaneously.
AI alignment: Reward model design faces the same problem. You're trying to implement a principal's (your) preferences through an agent's (AI's) incentives. But if the AI has private information about its capabilities or goals, Gibbard-Satterthwaite applies: no mechanism can fully extract truth without either restricting the domain of possible goals, introducing randomness, or accepting that some goals can't be incentivized (dictatorship).
The Three Escapes (All Imperfect)
Mechanism designers discovered three ways to partially escape these impossibilities:
1. Introduce Money
If preferences are quasi-linear (u_i = v_i(x) + m_i, where m_i is money), then IIA no longer applies. The VCG mechanism achieves efficiency and truth-telling. But you sacrifice budget balance: payments may not sum to zero.
The catch: Money only works if utility is genuinely quasi-linear. If fairness matters more than payment, if wealth effects are large, if some agents refuse transfers — Arrow's constraints re-emerge.
2. Restrict the Domain
If preferences are single-peaked on a one-dimensional policy space, the median voter theorem applies: majority rule is strategy-proof and non-dictatorial. The problem disappears.
The catch: Most policy spaces are multidimensional. You can't just declare preferences one-dimensional if they aren't.
3. Use Randomization
Probabilistic voting rules (select a random dictator, implement their preference) are strategy-proof and non-dictatorial ex-ante, even without money.
The catch: Outcomes are uncertain. Sometimes the wrong person's preference gets implemented.
What's Actually Invariant
This is the profound lesson: You can't escape all four constraints simultaneously. You can only choose which one to sacrifice:
- Sacrifice domain restriction: Have general preferences → need money or dictatorship
- Sacrifice unanimity: Randomize outcomes → drop the guarantee that everyone prefers the chosen outcome
- Sacrifice IIA: Use money or richer preference information → rely on quasi-linear assumptions that may not hold
- Sacrifice non-dictatorship: Accept that one agent's preference sometimes determines outcomes → builds distrust
Real institutions accept this and design accordingly. Democratic legislatures use procedures that create single-peaked preferences (voting in sequence, agenda control). Auctions use subsidies or accept that not all efficient trades happen. Markets use price to aggregate information at the cost of excluding those who can't or won't pay.
The deepest insight: There is no neutral institutional design. Any mechanism embeds a choice about which axiom to violate. "Fairness" in institutional design is not the absence of tradeoffs — it's the transparency and justification of which tradeoffs you've chosen.
Arrow didn't prove that good institutions are impossible. He proved that perfect institutions are impossible. And the choice of which imperfection to accept is itself a design decision, not a discovery.