# Yang Cai Oct 01, 2014. An overview of todayâ€™s class Myersonâ€™s...

date post

14-Dec-2015Category

## Documents

view

215download

0

Embed Size (px)

### Transcript of Yang Cai Oct 01, 2014. An overview of todayâ€™s class Myersonâ€™s...

- Slide 1

Yang Cai Oct 01, 2014 Slide 2 An overview of todays class Myersons Auction Recap Challenge of Multi-Dimensional Settings Unit-Demand Pricing Slide 3 Myersons Auction Recap [Myerson 81 ] For any single-dimensional environment. Let F= F 1 F 2 ... F n be the joint value distribution, and (x,p) be a DSIC mechanism. The expected revenue of this mechanism E v~F [ i p i (v)]=E v~F [ i x i (v) i (v i )], where i (v i ) := v i - (1-F i (v i ))/f i (v i ) is called bidder is virtual value (f i is the density function for F i ). Slide 4 1 1 Bidders report their values; transformed The reported values are transformed into virtual - values; the virtual-welfare maximizing allocation is chosen. Charge the payments according to Myersons Lemma. Transformation = depends on the distributions; deterministic function (the virtual value function); Myersons Auction Recap Myersons auction looks like the following Slide 5 Nice Properties of Myersons Auction DSIC, but optimal among all Bayesian Incentive Compatible (BIC) mechanisms! Deterministic, but optimal among all possibly randomized mechanisms! Central open problem in Mathematical Economics: How can we extend Myersons result to Multi-Dimensional Settings? Important progress in the past a few years. See the Challenges first! Slide 6 Challenges in Multi-Dimensional Settings Slide 7 Example 1: A single buyer, 2 non-identical items Bidder is additive e.g. v({1,2}) = v 1 +v 2. Further simplify the setting, assume v 1 and v 2 are drawn i.i.d. from distribution F = U{1,2} (1 w.p. , and 2 w.p. ). Whats the optimal auction here? Natural attempt: How about sell both items using Myersons auction separately? Slide 8 Example 1: Selling each item separately with Myersons auction has expected revenue $2. Any other mechanism you might want to try? How about bundling the two items and offer it at $3? What is the expected revenue? Revenue = 3 Pr[v 1 +v 2 3] = 3 = 9/4 > 2! Lesson 1: Bundling Helps!!! Slide 9 Example 1: The effect of bundling becomes more obvious when the number of items is large. Since they are i.i.d., by the central limit theorem (or Chernoff bound) you know the bidders value for the grand bundle (contains everything) will be a Gaussian distribution. The variance of this distribution decreases quickly. If set the price slightly lower than the expected value, then the bidder will buy the grand bundle w.p. almost 1. Thus, revenue is almost the expected value! This is the best you could hope for. Slide 10 Example 2: Change F to be U{0,1,2}. Selling the items separately gives $4/3. The best way to sell the Grand bundle is set it at price $2, this again gives $4/3. Any other way to sell the items? Consider the following menu. The bidder picks the best for her. -Buy either of the two items for $2 -Buy both for $3 Slide 11 Example 2: Bidders choice: Expected Revenue = 3 3/9 + 2 2/9 =13/9 > 4/3! v 1 \v 2 012 0$0 $2 1$0 $3 2$2$3 Slide 12 Example 3: Change F 1 to be U{1,2}, F 2 to be U{1,3}. Consider the following menu. The bidder picks the best for her. -Buy both items with price $4. -A lottery: get the first item for sure, and get the second item with prob. . pay $2.50. The expected revenue is $2.65. Every deterministic auction where every outcome awards either nothing, the first item, the second item, or both items has strictly less expected revenue. Lesson 2: randomization could help! Slide 13 Unit-demand Bidder Pricing Problem Slide 14 Unit-Demand Bidder Pricing Problem (UPP) 1 i n A fundamental pricing problem v 1 ~ F 1 v i ~ F i v n ~ F n Bidder chooses the item that maximizes v i - p i, if any of them is positive. Revenue will be the corresponding p i. Focus on pricing only, not considering randomized ones. Its known randomized mechanism can only get a constant factor better than pricing. Slide 15 Our goal for UPP Goal: design a pricing scheme that achieves a constant fraction of the revenue that is achievable by the optimal pricing scheme. Assumption: F i s are regular. Theorem [CHK 07]: There exists a simple pricing scheme (poly-time computable), that achieves at least of the revenue of the optimal pricing scheme. Remark: the constant can be improved with a better analysis. Slide 16 What is the Benchmark??? When designing simple nearly-optimal auctions. The benchmark is clear. Myersons auction, or the miximum of the virtual welfare. In this setting we dont know what the optimal pricing scheme looks like. We want to compare to the optimal revenue, but we have no clue what the optimal revenue is? Any natural upper bound for the optimal revenue? Slide 17 (a) UPP One unit-demand bidder n items Bidders value for the i-th item v i is drawn independently from F i (b) Auction n bidders One item Bidder Is value for the item v i is drawn independently from F i Two Scenarios 1 i n v 1 ~ F 1 v i ~ F i v n ~ F n Item 1 i n Bidders v 1 ~ F 1 v i ~ F i v n ~ F n Slide 18 Benchmark Lemma 1: The optimal revenue achievable in scenario (a) is always less than the optimal revenue achievable in scenario (b). - Proof: See the board. - Remark: This gives a natural benchmark for the revenue in (a). Slide 19 An even simpler benchmark In a single-item auction, the optimal expected revenue E v~F [max i x i (v) i (v i )] = E v~F [max i i (v i ) + ] (the expected prize of the prophet) Remember the following mechanism RM we learned in Lecture 6. 1.Choose t such that Pr[max i i (v i ) + t] = . 2.Set a reserve price r i = i -1 (t) for each bidder i with the t defined above. 3.Give the item to the highest bidder that meets her reserve price (if any). 4.Charge the payments according to Myersons Lemma. By prophet inequality: ARev(RM) = E v~F [ i x i (v) i (v i )] E v~F [max i i (v i ) + ] = ARev(Myerson) Lets use the revenue of RM as the benchmark. Slide 20 Inherent loss of this approach Relaxing the benchmark to be Myersons revenue in (b) This step might lose a constant factor already. To get real optimal, a different approach is needed.

Recommended

*View more*