app_rev_max_mul_items

Approximate Revenue Maximization
with Multiple
Items
Sergiu Hart
Noam Nisan
April 2012
Presented by: Nir Shabat
The Problem
β€’ How can a seller maximize its revenue when
selling multiple items to a single buyer?
Single Item
β€’ π‘₯ – Buyer’s item valuation
β€’ π‘₯ is unknown, but we know that π‘₯~𝐹
β€’ Selling at price 𝑝
– Buyer buys at probability 1 βˆ’ 𝐹 𝑝
– Revenue is 𝑝 β‹… 1 βˆ’ 𝐹 𝑝
β€’ Seller chooses price π‘βˆ— that maximize this
expression
Single Item
β€’ Can other mechanisms yield higher revenue?
– Indirect, different prices for different probabilities,
other?
β€’ No… Take-it-or-leave-it at price π‘βˆ— yields
optimal revenue among all mechanisms.
Myerson [1981]
Two Items
β€’
β€’
β€’
β€’
π‘₯, 𝑦 - Buyer’s valuation for item 1 and 2, resp.
π‘₯, 𝑦 i.i.d. according to 𝐹
Valuation for both items is π‘₯ + 𝑦 (additive)
Is selling each single item optimally
(separately) is the optimal way for selling both
of them?
β€’ Turns out it isn’t …
Example
β€’ Distribution takes values 1 and 2, each with
prob. ½
β€’ Selling a single item optimally:
– At price 1 (selling always) – revenue is 1
1
2
– At price 2 – revenue is β‹… 2 = 𝟏
Revenue from both items is 2
Example (Cont.)
β€’ Sell both as a bundle at price 3:
– Buyer buys with probability of
3
4
3
4
Revenue is 3 β‹… = 𝟐. πŸπŸ“
More than selling them separately…
Selling Separately vs. Bundling
β€’ Is bundling always better than selling
separately?
β€’ No.
β€’ Example:
– Distribution taking values 0 and 1, each with
probability ½
– Max. revenue with bundle is ¾
– Can sell each item for revenue of ½
In this case selling separately is better…
Other Mechanisms
β€’ Sometimes neither selling separately nor
bundling is optimal.
β€’ Example:
β€’ Distribution taking values 0, 1 and 2 each with
prob. 1/3
– Selling separately or as bundle – revenue is πŸ’/πŸ‘
– Buyer can buy each single item at price 2, or both
at price 3 – revenue is πŸπŸ‘/πŸ— (better)
Other Mechanisms (cont.)
β€’ Similar case for uniform distribution on
[0,1] (Manelli and Vincent [2006])
β€’ Sometimes not even deterministic (Hart and
Reny [2011]):
– Distribution takes value 1, 2 and 4, with
probabilities 1/6, 1/2, 1/3 resp.
– Buyer chooses between buying one item with
prob. ½ for a price of 1, and buying both items for
a price of 4 - Optimal
Result 1
β€’ For every one-dimensional distributions 𝐹1
and 𝐹2 :
𝑅𝐸𝑉1 𝐹1 + 𝑅𝐸𝑉1 𝐹2
1
β‰₯ β‹… 𝑅𝐸𝑉2 𝐹1 × πΉ2
2
Also generalizes for multi-dimensional
distributions.
Result 1 (cont.)
β€’ Generalizes for multiple buyers:
1
𝑛
𝑛
𝑅𝐸𝑉
𝑋 + 𝑅𝐸𝑉
π‘Œ β‰₯ β‹… 𝑅𝐸𝑉
2
𝑛
𝑋, π‘Œ
Where 𝑋 = 𝑋1 , … , 𝑋 𝑛 ∈ 𝑅 𝑛 , Y =
π‘Œ1 , … , π‘Œ 𝑛 ∈ 𝑅 𝑛 are the values of the first item
and the second item to the n buyers, resp.
Result 2
β€’ For equal distributions we get a tighter bound
)𝐹1 = 𝐹2 = 𝐹):
𝑒
𝑅𝐸𝑉1 𝐹 + 𝑅𝐸𝑉1 𝐹 β‰₯
β‹… 𝑅𝐸𝑉2 𝐹1 × πΉ2
𝑒+1
β€’
𝑒
~0.73
𝑒+1
β€’ Conjecture – 0.78 is the tight bound
Result 3
β€’ There exists a constant 𝑐 > 0 s.t. for every
integer π‘˜ β‰₯ 2, and every one-dimensional
distributions 𝐹1 , … , πΉπ‘˜ :
𝑅𝐸𝑉1 𝐹1 + β‹― + 𝑅𝐸𝑉1 πΉπ‘˜
𝑐
β‰₯
β‹… π‘…πΈπ‘‰π‘˜ 𝐹1 × β‹― × πΉπ‘˜
2
log π‘˜
Result 4
β€’ There exists a constant 𝑐 > 0 s.t. for every
integer π‘˜ β‰₯ 2, and every one-dimensional
distribution 𝐹:
𝑐
𝑅𝐸𝑉1 𝐹 βˆ— β‹― βˆ— 𝐹 β‰₯
β‹… π‘…πΈπ‘‰π‘˜ 𝐹 × β‹― × πΉ
log π‘˜
π‘˜
π‘˜
Open Problems
β€’ Characterizing optimal auctions:
– When is selling separately optimal?
– When is bundling optimal?
– When are deterministic auctions optimal?
Open Problems
β€’ Upper/Lower bound gaps:
– 2 items, i.i.d. F:
βˆ€πΉ
𝑆𝑅𝐸𝑉 𝐹 × πΉ
≀ 0.73 …
𝑅𝐸𝑉 𝐹 × πΉ
βˆƒπΉ
𝑆𝑅𝐸𝑉 𝐹 × πΉ
≀ 0.78 …
𝑅𝐸𝑉 𝐹 × πΉ
Open Problems
β€’ Upper/Lower bound gaps:
– π‘˜ items, i.i.d. 𝐹:
βˆ€πΉ
βˆƒπΉ
𝐡𝑅𝐸𝑉 𝐹 ×π‘˜
1
β‰₯
×π‘˜
𝑆𝑅𝐸𝑉 𝐹
4
𝐡𝑅𝐸𝑉 𝐹 ×π‘˜
≀ 0.57
×π‘˜
𝑆𝑅𝐸𝑉 𝐹
For all large enough π‘˜
Open Problems
β€’ Upper/Lower bound gaps:
– k items, i.i.d. F:
𝐡𝑅𝐸𝑉 𝐹 × πΉ
2
𝑒
βˆ€πΉ
β‰₯ β‹…
𝑆𝑅𝐸𝑉 𝐹 × πΉ
3 𝑒+1
βˆƒπΉ
𝐡𝑅𝐸𝑉 𝐹 × πΉ
2
≀
𝑆𝑅𝐸𝑉 𝐹 × πΉ
3
Some more gaps …