算法博弈：机制设计

Lecturer: Constantinos Daskalakis

General Mechanism (Quasi-Linear Setting)

Def:

setting

1. players’ type spaces $T_1,…,T_n$
2. $F_i$ over $T_i$, Bayesian setting : $t_i$~$F_i$
3. a set of possible alternatives $A$
4. valuation $v_i: T_i\times A \rightarrow R$

Mechanism

1. players’ action space $X_1,…,X_n$

2. allocation function $x:X_1\times…\times X_n \rightarrow A$

3. price function $p_i: X_1\times…\times X_n \rightarrow R$

   “direct” : $X_i = T_i$,$\forall i$. e.g. English Auction

Incentive Compatibility of Direct Mechanisms

• Indirect mechanisms: properties of DS, ex-post Nash or Bayesian Nash equilibria of the incomplete information games

• direct mechanisms: whether truthtelling is equilibrium

• A direct mechanism $(x,p)$ is Dominant Strategy Incentive Compatible (DSIC) iff truth-telling is a dominant strategy
  

  $$\forall i, t_i, t_{-i}:v_i(t_i,x(t_i,t_{-i}))-p_i(t_i,t_{-i})\geq v_i(t_i,x(t_i',t_{-i}))-p_i(t_i',t_{-i})$$

• A direct mechanism $(x,p)$ is BIC iff truth-telling is a Bayesian Nash equilibrium
  

  $$\forall i, t_i, t_{-i}:E_{t_{-i}}[v_i(t_i,x(t_i,t_{-i}))-p_i(t_i,t_{-i})]\geq E_{t_{-i}}[v_i(t_i,x(t_i',t_{-i}))-p_i(t_i',t_{-i})]$$

Implementation: M(x,p) implements $f$ in dominant strategies if for some dominant strategy equilibrium $s_1,…,s_n$, we have for all $t_1,…,t_n$

$$x(s_1(t),...,s_n(t_n)\equiv f(t_1,...,t_n)$$

The Revelation Principle (DSE to DSIC)

Th. If there is an arbitrary mechanism that implements some allocation rule $f$ in dominant strategies, then there is also a direct ,DSIC mechanism that implements $f$. Moreover, the payments of the players in the direct mechanism are identical to those of the original mechanism, at equilibrium, point-wise with respect to $t_1,…,t_n$

Proof:

​ Let $s_1,…,s_n$ be a dominant strategy equilibrium of original mechanism such that $x(s_1(t_1),…,s_n(t_n))\equiv f(t_1,…,t_n)$, for all t.

​ Def new direct mechanism:

$$x'(t_1,...,t_n)=x(s_1(t_1),...,s_n(t_n))\\ p'(t_1,...,t_n)=p(s_1(t_1),...,s_n(t_n))$$
​ $s_i$ is DSE for i in original mechanism, we have:
$$v_i(t_i,x(s_i(t_i),x_{-i}))-p_i(s_i(t_i),x_{-i})\geq v_i(t_i,x(x_i',x_{-i}))-p_i(x_i',x_{-i})$$
​ $x_{-i}=s_{-i}(t_{-i})$ and $x_i'=s_i(t_i')$, so we get
$$v_i(t_i,x'(t_i,t_{-i}))-p_i(t_i,t_{-i})\geq v_i(t_i,x'(t_i',t_{-i}))-p_i(t_i',t_{-i})$$
​ i.e. $(x',p')$ is DSCI and $x'=f$

Technical Lemma (gedanken experiment): from ex-post to dominant strategy equilibria

​ Let $s_1,…,s_n$ be an ex-post NE of a game $(X_1,…,X_n: T_1,…,T_n:u_1,…,u_n)$. Define new action space $X_i'={s_i(t)|t_i\in T_u}$. Then $s_1,…,s_n$ is a DSE of new game.

Direct Auctions and DSIC

• Allocation rule $x$
• Payment rule $p$
• $(x,p)$ is DSIC iff for all $i$, $b_{-i}$ is optimal for bidder $i$ to its true value :
  
  • $zx_i(z,b_{-i})-p_i(z,b_{-i})\geq zx_i(z',b_{-i})-p_i(z',b_{-i}),\forall z,z'$

  Myerson’s Lemma

  Implementability

  An allocation rule $x$ for a single-dimensional environment is implementable if there is a payment rule $p$ s.t. the sealed-bid auction $(x,p)$ is DSIC (dominant strategy incentive compatible).

  Monotonicity

  An allocation rule x for a single-dimensional environment is monotone if for every bidder $i$ and bids $b_{−i}$ by the other bidders, the allocation $x_i(z,b_{−i})$ to $i$ is non-decreasing in $i$’s bid $z$.

  Myerson’s Lemma [Myerson’81]

  Fix a single-dimensional environment.

  (1) An allocation rule $x$ is implementable if and only if it is monotone.

  (2) If $x$ is implementable/monotone, there is an essentially unique payment rule such that the sealed-bid mechanism $(x,p)$ is DSIC, given by the formula:
  

  $$\forall i,b_{-i} : p_i(z, b_{-i}) = z\cdot x_i(z, b_{-i}) - \int_{0}^z x_i(t,b_{-i})dt + p_i(0,b_{-i})$$
  (3) In particular, there is a unique payment function such that the mechanism is DSIC and additionally IR with non-positive transfers (i.e. $b_i = 0$ implies $p_i(b) = 0, ∀b_{-i}$).

  ​

  Revenue = Virtual Welfare

  **[Myerson ’81] **Fix a Bayesian single-dimensional environment, where bidder distributions are F1,…,Fn, and F=F1x…xFn.

  Let also (x,p) be a BIC mechanism satisfying interim IR and NPT. The expected revenue of this mechanism under truth-telling is
  

  $$E_{v-F}[\sum_iP_i(v)] = E[\sum x_i(v)\phi_i(v_i)]\\ \phi_i(v_i) := v_i - (1-F_i(v_j))/f_i(v_i)$$
  One U[0,1] Bidder, One Item

  • Recall that posting a price of ½ achieves expected revenue ¼.

  • Is ¼ the optimal expected revenue of any IR auction?

    ​ Answer: yes, it is!
    Proof:
    The virtual transform for U[0,1] distribution is: $\phi(v) = 2v-1$
    By Myerson’s theorem, any BIC, IR, NPT mechanism (x,p) has revenue $E[x(v)(2v-1)]$
    Notice that for all possible x(): $E[x(v)()2v-1]\leq 1/4$
    Hence, posting a price of ½ is optimal as it achieves the best possible expected virtual welfare, and hence expected revenue.

    ​

  ​
  Two U[0,1] Bidders, One Item

  • Recall that Vickrey without reserve has expected revenue ⅓, while Vickrey with reserve ½ has expected revenue 5/12.

  • Is 5/12 the optimal expected revenue of any interim IR auction?

    ​ Answer: yes it is!

  • Recall that the virtual transform for v~U[0,1] is: φ(v)= 2v-1

  • By Myerson’s theorem,

    ​ Optimizing expected revenue = Optimizing expected virtual welfare
​ subject to interim monotonicity (needed for BIC)

  • Let’s try to optimize virtual welfare point-wise on every bid profile (forgetting about interim monotonicity temporarily)

  • On bid profile (v1,v2), the pair of virtual values are (φ(v1), φ(v2))=(2v1-1, 2v2-1). How should we allocate?

    • If max{v1,v2} ≥ ½, let’s give the item to the bidder with highest value/virtual value.
    • Otherwise, φ(v1), φ(v2) < 0. So let’s not allocate the item.

  • Note that the above allocation rule is monotone, so by Myerson’s lemma there is a price rule that makes it DSIC.

  • DSIC + pointwise optimal virtual welfare => DSIC, Revenue-optimal

  • Now, notice that the above auction is identical to Vickrey auction with reserve ½ (which has revenue 5/12 and is actually DSIC).

  ​
  ​

  Revenue-optimal Single-Dimensional Auction

  • By revelation principle, instead of optimizing over all (potentially indirect) interim IR auctions, can optimize over direct interim IR, BIC auctions.

  • In fact, can also assume interim NPT, as if interim NPT is not satisfied expected revenue can be increased by turning it interim NPT.

  • Hence, by Myerson’s theorem suffices to find the interim monotone allocation rule that optimizes expected virtual welfare.

  • Let’s ignore interim monotonicity temporarily. What allocation rule optimizes expected virtual welfare?

  • Answer: Optimize virtual welfare pointwise, i.e. on every bid profile v.

    ​ i.e. on bid profile v solve: max Σi xi φi (vi), s.t x ∈ X

  • Call this allocation rule the Virtual Welfare-Maximizing Rule.

  • Pertinent Question: Is the Virtual Welfare-Maximizing Rule monotone?

  • If yes, done! And, in fact, with a DSIC mechanism ☺

  • But answer depends on the distribution:

    ​ e.g. suppose one item is sold to one bidder whose value is ⅔U[0,1] + ⅓ U[1,3]; then virtual welfare-maximizing rule is not monotone

  • Distributions resulting in monotone allocation rules?

  Regular Distributions

  Definition 1 (Regular Distributions): A single-dimensional distribution F is regular if its virtual transform v- (1-F(v))/f(v) is non-decreasing.

  Definition 2 (Monotone Hazard Rate (MHR)): A single-dimensional distribution F has Monotone Hazard Rate, if (1-F(v))/f(v) is non-increasing.

  What distributions are in these classes?

  • MHR: uniform, exponential and Gaussian distributions, and many more.
  • Regular: MHR and Power-law
  • Irregular: Multi-modal or distributions with very heavy tails.

  When all the Fi’s are regular, the Virtual Welfare-Maximizing Rule is monotone.

  [Myerson ’81] Fix a Bayesian single-dimensional environment, where all bidders draw values from regular distributions.

  Then the auction whose allocation rule is the virtual welfare maximizing allocation rule (and whose price rule is uniquely determined by the allocation rule so that the resulting auction is interim IR, NPT) is DSIC and revenue-optimal (among all interim IR, potentially indirect auctions).