1 of 14

Multi-Unit Auctions:�Beyond Roberts

Shahar Dobzinski and Noam Nisan

“A 3 minutes video is worth a 20 minutes talk”

Noam Nisan’s Blog, March 5:

2 of 14

The main topic of this talk: a strange, unintuitive and useless mechanism for multi-unit auctions.

3 of 14

Multi-unit auctions: n bidders, m (identical) items.

vi(s) denotes the value of bidder i for a bundle�of s items. vi is represented by a black box.

Normalization: vi(0) = 0

Monotonicity: vi(s+1) ≥ vi(s)

Goal: find an allocation (s1,…,sn), Σsi≤m, that maximizes the welfare Σivi(si).

4 of 14

What is a Good Auction?I.e., what is a good algorithm/mechanism?

Fast

Good Approx

Truthful

VCG

1.01-approx

Allocate�nothing

?

Fast: running time is polynomial in�n and log m.

Truthful: truth telling maximizes the profit

profit=value-price

dominant strategies, private values

Good Approximation: find a solution that is within a factor of 1.01 of the value of the optimal solution.

5 of 14

Theorem: No scalable truthful mechanism can achieve an approximation ratio better than 2 in poly time.

Theorem: There exists a polynomial time truthful 2-approximation mechanism [Dobzinski-Nisan’07].

6 of 14

Fact: the truthful poly time truthful 2-approx mechanism is maximal in range.

Theorem [Dobzinski-Nisan’07]: every poly-time MIR algorithm for multi-unit auctions has an approx ratio ≥ 2.

Main idea: extend the impossibility result for MIR to hold for any truthful mechanism using the characterize and optimize approach.

7 of 14

Conjecture [characterize]: every truthful mechanism for multi-unit auctions with an approx ratio < 2 is weighted MIR.

Theorem [optimize]: every poly-time (weighted) MIR algorithm for multi-unit auctions has an approx ratio ≥ 2.

Characterize

Optimize

Impossibility

+

=

Success stories: [Lavi-Mu’alem-Nisan], [Papadimitriou-Schapira-Singer], [Dobzinski-Sundararajan],… but no auction domains!

1.99-approx

Conjectured world:

Truthful

MIR

FALSE

8 of 14

Theorem [characterize]: every scalable 2-bidder truthful algorithm with an approximation ratio better than 2 is (almost) a triage auction.

The real world:

(almost…)

Truthful

1.99-approx

MIR

Triage

New family! Triage auctions are truthful but not MIR and provide a 1.01-approx (not in poly time)

9 of 14

Theorem: No poly-time scalable mechanism provides an approx ratio better than 2.

First impossibility result for an auction domain.

Characterize

Optimize

Impossibility

+

=

Proof: Modify the optimization theorem for the maximal-in-range case.

10 of 14

The Main Highlights of the Characterization

Start with the 2-item characterization and use induced mechanisms to obtain a characterization for m>2 items.

Scalability is used only in the 2-item characterization.

Manipulating the payment functions, not the allocation functions.�Very different machinery than Roberts’ theorem machinery.

11 of 14

Triage Auctions

12 of 14

Warm up: VCG in our Notation�(truthfulness as a menu)

v2

v1

0

v2

v2 - v1

0

Values

Prices (for u)

u2

u1

0

Values

u2

u2 - u1

0

Prices (for v)

The taxation principle: truthful mechanism 🡸🡺 each bidder faces a menu where each bundle has a price (possibly infinity), and is assigned the most profitable bundle.

13 of 14

This talk: triage auction, symmetric version, with parameter θ > ½.

The triage auction is feasible, its approximation ratio is 1/θ.

v2

v1

0

v2

θv2

0

Values

Prices

v2

v1

0

v2

v2 - v1

0

Values

Prices

v2

v1

0

v1/θ

v1/θ - v1

0

Values

Prices

Low range: v1 < (1-θ)v2

Mid range: (1-θ)v2 < v1 < θv2

High range: v1 > θv2

θ=0.75

14 of 14

Summary

Lesson I: there are “useful” non-VCG mechanisms.

Introduced triage auctions, characterized mechanisms for multi-unit auctions, got a lower bound on the power of poly time truthful scalable mechanisms.

Lesson II: and we can characterize them!