Chapter 11 asked: given preferences and endowments, do competitive markets produce efficient outcomes? The answer, yes under the welfare theorem conditions, takes the market mechanism as given. This chapter inverts the question: given a desired outcome, can we design a mechanism to achieve it?
Mechanism design is often called "reverse game theory." Instead of predicting the outcome of a game, we design the game to produce a desired outcome. Market design applies these ideas to real institutions: auctions, matching markets, spectrum allocation, kidney exchange.
Prerequisites: Chapters 7 (game theory basics, Nash equilibrium) and 10 (welfare theorems, general equilibrium).
The challenge: agents' types are private. How do we get them to reveal their types truthfully?
Figure 12.1. Mechanism design timeline.
The mechanism designer chooses the rules (message space and outcome function) to achieve a desired social choice function.
A direct mechanism asks each agent to simply report their type (their private information). It is incentive compatible (IC) if truthful reporting is an equilibrium strategy: no agent benefits from lying.
This is the most powerful simplification in mechanism design. In principle, the space of possible mechanisms is infinitely large. An auction could have any number of rounds, any bidding rules, any payment formula. A matching algorithm could work in any conceivable way. Searching over all possible mechanisms for the best one seems hopeless.
The revelation principle says: you don't have to search. Whatever outcome any mechanism can achieve, a direct mechanism (just ask everyone to report truthfully) can achieve the same outcome. So the mechanism design problem reduces to: find the best allocation rule and payment rule as functions of reported types, subject to the constraint that truth-telling is optimal. This transforms an impossibly broad search into a well-defined optimization problem.
What this says: If some elaborate game can reach an outcome, then a plain "just tell the truth" game can reach the same outcome — so there is never any need to study elaborate mechanisms. We only ever study truthful direct mechanisms, where each agent simply reports their private type. "Incentive compatible" then means exactly one thing: honesty is the agent's best move.
Why it matters: The two strengths differ in how robust that honesty is. Dominant-strategy (DSIC) honesty holds no matter what anyone else does — you never need to guess others' types. Bayesian (BIC) honesty holds only on average, assuming everyone else is also telling the truth; it is weaker but lets the designer implement more outcomes. The revelation principle is what turns an impossibly large search over all conceivable auctions, algorithms, and rules into one well-posed problem: find the allocation-and-payment rule that makes truth-telling optimal.
In Full Mode, Eq. 12.1 (DSIC) and Eq. 12.2 (BIC) state the two incentive-compatibility conditions formally.DSIC is stronger but harder to achieve. BIC is weaker but allows more mechanisms.
Auction theory and mechanism design were partly a Cold-War product — RAND, von Neumann, and the strategic-rationality program shaped the field.
This is the mechanism design analog of Arrow's impossibility theorem. It says that in general social choice settings, no non-dictatorial mechanism can elicit truthful preferences in dominant strategies.
The escape: restrict the domain. With quasi-linear preferences ($U_i = v_i(a) + t_i$, where $t_i$ is a monetary transfer), the Gibbard-Satterthwaite barrier falls. The VCG mechanism achieves efficiency and DSIC with transfers.
The Vickrey-Clarke-Groves (VCG) mechanism achieves efficient allocation with truth-telling as a dominant strategy, using monetary transfers.
Is competition an end-state to be engineered, or a discovery process that resists design? Hurwicz's incentive-compatibility program sits right on that fault line.
The efficient allocation maximizes total value: $a^*(\theta) = \arg\max_a \sum_i v_i(a, \theta_i)$.
Agent $i$ pays the externality she imposes on others: the difference between others' welfare with and without $i$.
Why is truth-telling dominant? Under truthful reporting, agent $i$'s payoff is:
$$v_i(a^*(\theta)) + t_i = v_i(a^*(\theta)) + \sum_{j \neq i} v_j(a^*(\theta_{-i})) - \sum_{j \neq i} v_j(a^*(\theta))$$
This simplifies to $\sum_j v_j(a^*(\theta)) - \sum_{j \neq i} v_j(a^*(\theta_{-i}))$. The second term doesn't depend on $i$'s report. So $i$ maximizes her payoff by choosing her report to maximize $\sum_j v_j(a^*(\theta))$, which happens when she reports truthfully, since $a^*$ already maximizes total value.
What this says: You pay the harm your presence imposes on everyone else — the difference between what the others could have achieved without you and what they achieve with you in the room. That bill depends only on the others' values, never on your own report, so you cannot shrink it by shading what you say.
Why it matters: Because your payment is locked by the others, your only remaining lever is to help the mechanism pick the outcome that maximizes total value — and that outcome is best for you precisely when you report your true value. Lying can only steer the allocation away from the efficient one, which can never make you better off. That is why truth-telling is a dominant strategy: it works regardless of what anyone else does. In the single-object case this rule is exactly the second-price (Vickrey) auction — the winner pays the runner-up's bid.
In Full Mode, the algebra shows the second term of the payoff is independent of agent $i$'s report.Enter agent values for a single indivisible object. The calculator computes VCG payments (equivalent to a second-price auction for a single item).
Figure 12.2. Agent values and VCG payments. Each agent pays the externality they impose on others. The winner pays the second-highest value (in a single-item auction, VCG reduces to the Vickrey auction).
Three citizens value a bridge at $v_1 = 30$, $v_2 = 25$, $v_3 = 15$. The cost is $C = 60$.
Build if $\sum v_i > C$: \$10 > 60$ → yes.
Clarke tax payments:
Total collected: \$10 + 15 + 5 = 40 < 60$. There's a budget deficit of 20; VCG does not generally achieve budget balance. Each agent pays their "pivotal" contribution.
| Format | Rules | Winner pays |
|---|---|---|
| English (ascending) | Bidders raise bids; last bidder wins | Second-highest value (approx.) |
| Dutch (descending) | Price drops until someone claims | Their bid |
| First-price sealed-bid | Highest bid wins | Their bid |
| Second-price sealed-bid (Vickrey) | Highest bid wins | Second-highest bid |
The Vickrey auction (second-price sealed-bid) is DSIC: each bidder's dominant strategy is to bid their true value $v_i$. Bidding above $v_i$ risks winning at a price above value; bidding below risks losing when the second-highest bid is below $v_i$.
The implication: under these conditions, the differences between auction formats (open vs sealed bid, ascending vs descending, first-price vs second-price) don't affect expected revenue.
Revenue equivalence breaks down in several common cases:
Set the number of bidders and their value distribution. Run single auctions to see individual outcomes, or run 100 rounds to observe revenue equivalence (average revenues converge across formats). Adjust the risk aversion slider to break equivalence.
Figure 12.3. Auction outcomes. In single runs, revenues differ across formats due to randomness. Over 100 runs, average revenues converge, demonstrating revenue equivalence. Increase risk aversion ($\rho > 0$) to break equivalence: first-price revenue rises above second-price.
When the seller wants to maximize revenue (not efficiency), Myerson showed the optimal mechanism uses the virtual value:
where $F$ is the CDF and $f$ is the PDF of the bidder's value distribution.
The optimal auction allocates to the bidder with the highest virtual value, provided it is positive. If all virtual values are negative, the seller retains the object. This implies a reserve price: the seller sets a minimum bid equal to $\psi^{-1}(0)$.
What this says: A revenue-maximizing seller does not treat a bid at face value. Each bid is mentally discounted by the "information rent" the seller must concede to keep bidders honest — the discounted figure is the bidder's virtual value. The seller awards the item to the highest virtual value, and only sells at all when even that discounted value clears zero.
Why it matters: That zero-crossing cutoff IS the optimal reserve price. Below it, holding the item beats selling, because the extra revenue squeezed from high-value bidders by keeping the reserve high outweighs the lost sales to low-value bidders. This is why even a seller who values the object at nothing should sometimes refuse to sell — the reserve is a strategic commitment, not a cost floor. The same logic reappears in optimal income taxation: the planner discounts each taxpayer by the incentive cost of taxing them, and only redistributes where the discounted gain stays positive.
In Full Mode, Eqs. 12.4–12.6 define the virtual value $\psi(\theta)$ and the reserve condition $\psi(r^*) = 0$.Values uniformly distributed on $[0, 1]$: $F(\theta) = \theta$, $f(\theta) = 1$.
$\psi(\theta) = \theta - (1-\theta)/1 = 2\theta - 1$
$\psi(\theta) = 0 \implies \theta = 1/2$. Optimal reserve price = $1/2$.
A second-price auction with reserve $1/2$ is optimal: the item is sold only if at least one bidder values it above $1/2$.
For values drawn from Uniform$[0, V_{\max}]$, the virtual value is $\psi(\theta) = 2\theta - V_{\max}$. Drag the reserve price slider. The revenue curve shows expected revenue as a function of the reserve. The optimal reserve (maximizing expected revenue) is highlighted.
Figure 12.4a. Virtual value function $\psi(\theta) = 2\theta - 1$ (for $U[0,1]$). The reserve price is set where $\psi(r) = 0$. Bidders with $\theta < r$ are excluded (shaded red).
Figure 12.4b. Expected revenue as a function of reserve price. The green dot marks the optimal reserve that maximizes expected revenue. Your chosen reserve is shown as a blue dot.
A government allocates a license to one of two firms. Firm $i$ has private value $\theta_i \in \{L, H\} = \{10, 50\}$, each equally likely.
Allocate to the firm reporting higher value; in case of a tie, allocate to firm 1. The winner pays 30.
Check IC for a high-value firm ($\theta = 50$):
Truthful is better. IC holds for type $H$.
Check IC for a low-value firm ($\theta = 10$):
Truthful is better. IC holds for type $L$. The mechanism is incentive compatible.
Two bidders with values drawn independently from $U[0, 100]$.
Second-price auction: Expected revenue = $E[\text{2nd highest value}] = 100/3 \approx 33.33$.
First-price auction: Optimal bid with 2 bidders: $b(\theta) = \theta/2$. Expected revenue = $E[\max(b_1, b_2)] = E[\max(\theta_1/2, \theta_2/2)] = E[\max(\theta_1, \theta_2)]/2 = (200/3)/2 = 100/3 \approx 33.33$.
Both formats yield expected revenue of \$100/3$, confirming revenue equivalence. The first-price auction generates less variable revenue (each winner pays exactly half their value) while the second-price auction has higher variance (the payment depends on the second-highest value, which can vary widely).
The seller wants to overstate her cost (to extract a higher price). The buyer wants to understate his value (to pay less). Incentive compatibility requires leaving "information rents" to both parties. These rents are costly, and with budget balance, there isn't enough surplus to pay both rents and ensure all efficient trades occur.
Real-world bargaining under private information always involves some inefficiency: salary negotiations, used car purchases, M&A deals. Institutions like posted prices, reputation systems, and standardized contracts mitigate the problem but cannot fully eliminate it.
Some goods cannot be allocated by prices: we don't (or shouldn't) sell school admissions, organ transplants, or residency positions. Matching markets use algorithms instead.
What this says: Watch the deferred-acceptance algorithm run: proposers always walk down their preference lists (each rejection sends them to a less-preferred choice), while responders always trade up (they only ever swap a tentative partner for a better one). Nobody is ever revisited after a rejection, so the process cannot loop forever — it must stop, and it does so quickly (in at most $n^2$ rounds).
Why it matters: When it stops, the matching is stable: there is no pair who both prefer each other to their assigned partners, so no one has an incentive to "elope." Stability is exactly what makes a match self-enforcing — it stays put without any prices or payments. That is why the same algorithm runs the medical-residency match, school choice, and kidney exchange: it manufactures a stable outcome in markets where money cannot do the allocating.
In Full Mode, Eq. 12.8 states the termination bound and the stable-matching guarantee.Theorem (Gale & Shapley, 1962). The algorithm terminates in at most $n^2$ rounds and produces a stable matching: no unmatched pair both prefer each other to their current match.
The deferred acceptance algorithm has four properties worth noting:
Enter preference lists for students and schools. The algorithm animates each round: proposals, tentative holds, and rejections. Enter preferences as comma-separated names (e.g., "W,X,Y,Z").
Four students (A, B, C, D) and four schools (W, X, Y, Z). Students propose.
| Student | Preferences | School | Preferences |
|---|---|---|---|
| A | W > X > Y > Z | W | B > A > D > C |
| B | X > W > Y > Z | X | A > B > C > D |
| C | W > Y > X > Z | Y | C > D > A > B |
| D | Y > W > X > Z | Z | D > C > B > A |
The final matching is A-W, B-X, C-Y, D-Z, which is stable: no pair wants to deviate. Use the interactive above to verify step by step.
Run Gale-Shapley with students proposing vs. schools proposing. Compare the two stable matchings. The proposing side always gets their best stable matching; the responding side gets their worst.
Alvin Roth (Nobel 2012, shared with Lloyd Shapley) describes this as the "economist as engineer" approach: using economic theory not just to explain the world but to design real institutions that improve people's lives.
Markets are not natural objects that arise spontaneously. They are designed institutions: rules, algorithms, and enforcement mechanisms that determine who gets what, at what price, and through what process. The design choices determine the outcomes.
The city decides to auction the exclusive right to operate a lemonade stand at the prime downtown corner. Three potential vendors: Maya ($v_M = 50$/day), Nate ($v_N = 35$/day), Olivia ($v_O = 20$/day). Values drawn from $U[0, 60]$.
Second-price auction (Vickrey): Dominant strategy is to bid truthfully. Maya bids 50, Nate bids 35, Olivia bids 20. Maya wins, pays 35.
Optimal auction (Myerson): Virtual values with $F(\theta) = \theta/60$, $f(\theta) = 1/60$:
$\psi(\theta) = \theta - (60 - \theta) = 2\theta - 60$
Reserve price: $\psi(\theta) = 0 \implies \theta = 30$.
Maya's virtual value: \$1(50) - 60 = 40$. Nate's: \$10$. Olivia's: $-20$ (excluded by optimal auction).
In a second-price auction with reserve 30: Maya wins, pays $\max(35, 30) = 35$.
Roth as "Economist as Engineer." Alvin Roth (Nobel Prize 2012) transformed mechanism design from pure theory into a practical discipline that redesigns real markets. His work demonstrates that markets are designed institutions, not natural phenomena.
The National Residency Matching Program (NRMP): Roth diagnosed why the original medical residency match was failing (instability, strategic manipulation) and redesigned it using deferred acceptance. The new system matches ~40,000 medical residents annually.
Kidney exchange: Roth, Sonmez, and Unver designed exchange protocols allowing incompatible donor-patient pairs to swap donors through chains of transplants, saving thousands of lives. This was pure market design: creating a market where none existed, without using prices.
School choice: Roth and colleagues replaced Boston's manipulable school assignment mechanism with a strategy-proof system. Under the old system, parents who reported their true preferences were punished; under the new system, honesty is always optimal.
Spectrum auctions: Milgrom and Wilson (Nobel Prize 2020) designed combinatorial auctions for the FCC, raising billions of dollars while efficiently allocating spectrum licenses. The 2017 incentive auction alone raised \$19.8 billion.
The common thread: economic theory provides the blueprint, but implementation requires understanding the specific institutional context, the "details" that pure theory abstracts away.
See mechanism design positioned in the wider genealogy of economic ideas — and the lineage from von Neumann's game theory to the design of real markets — in the intellectual-genealogy graph (History of Economic Thought timeline).
The intellectual lineage behind this apparatus — Hurwicz's incentive-compatibility program, Maskin and Myerson on implementation and optimal auctions, Roth on matching, Milgrom and Wilson on auction design (the 2007 and 2020 Nobel Prizes) — is the subject of the History of Economic Thought, Ch. 11 (Information economics and the game-theory revolution) (chapter forthcoming). This chapter teaches the modern apparatus; that chapter tells where it came from.
Hurwicz framed mechanism design as a response to a problem the Austrians posed first: how can decentralized agents, each holding private information, be induced to act on knowledge no central planner possesses? For that older debate — Hayek's "knowledge problem" and the idea that competition is a discovery process rather than a designable end-state — see the Austrian tradition (History of Economic Thought, Ch. 6).
| Label | Equation | Description |
|---|---|---|
| Eq. 12.1 | $U_i(\theta_i, \theta_i) \geq U_i(\hat{\theta}_i, \theta_i)$ for all $\hat{\theta}_i, \theta_{-i}$ | DSIC |
| Eq. 12.2 | $E[U_i(\theta_i, \theta_i)] \geq E[U_i(\hat{\theta}_i, \theta_i)]$ | BIC |
| Eq. 12.3 | $t_i = \sum_{j \neq i} v_j(a^*(\theta_{-i})) - \sum_{j \neq i} v_j(a^*(\theta))$ | VCG payment |
| Eq. 12.4 | $\psi(\theta) = \theta - (1-F(\theta))/f(\theta)$ | Myerson virtual value |
Myerson (1981); Vickrey (1961); Clarke (1971); Groves (1973); Gale & Shapley (1962); Roth (2002); Milgrom (2004).
Coming in Part V: graduate macro. The models get serious, and so do the policy debates.