Transcript
  • CSC304 Lecture 19

    Fair Division 2: Cake-cutting, Indivisible goods

    CSC304 - Nisarg Shah 1

  • Recall: Cake-Cutting

    CSC304 - Nisarg Shah 2

    β€’ A heterogeneous, divisible good

    ➒ Represented as [0,1]

    β€’ Set of players 𝑁 = {1,… , 𝑛}

    ➒ Each player 𝑖 has valuation 𝑉𝑖

    β€’ Allocation 𝐴 = (𝐴1, … , 𝐴𝑛)

    ➒ Disjoint partition of the cake

  • Recall: Cake-Cutting

    CSC304 - Nisarg Shah 3

    β€’ We looked at two measures of fairness:

    β€’ Proportionality: βˆ€π‘– ∈ 𝑁: 𝑉𝑖 𝐴𝑖 β‰₯ Ξ€1

    𝑛

    ➒ β€œEvery agent should get her fair share.”

    β€’ Envy-freeness: βˆ€π‘–, 𝑗 ∈ 𝑁: 𝑉𝑖 𝐴𝑖 β‰₯ 𝑉𝑖 π΄π‘—βž’ β€œNo agent should prefer someone else’s allocation.”

  • Other Desiderata

    CSC304 - Nisarg Shah 4

    β€’ There are two more properties that we often desire from an allocation.

    β€’ Pareto optimality (PO)➒ Notion of efficiency

    ➒ Informally, it says that there should be no β€œobviously better” allocation

    β€’ Strategyproofness (SP)➒ No player should be able to gain by misreporting her

    valuation

  • Strategyproofness (SP)

    CSC304 - Nisarg Shah 5

    β€’ For deterministic mechanisms➒ β€œStrategyproof”: No player should be able to increase her

    utility by misreporting her valuation, irrespective of what other players report.

    β€’ For randomized mechanisms➒ β€œStrategyproof-in-expectation”: No player should be able

    to increase her expected utility by misreporting.

    ➒ For simplicity, we’ll call this strategyproofness, and assume we mean β€œin expectation” if the mechanism is randomized.

  • Strategyproofness (SP)

    CSC304 - Nisarg Shah 6

    β€’ Deterministic➒ Bad news!

    ➒ Theorem [Menon & Larson β€˜17] : No deterministic SP mechanism is (even approximately) proportional.

    β€’ Randomized➒ Good news!

    ➒ Theorem [Chen et al. β€˜13, Mossel & Tamuz β€˜10]: There is a randomized SP mechanism that always returns an envy-free allocation.

  • Perfect Partition

    CSC304 - Nisarg Shah 7

    β€’ Theorem [Lyapunov ’40]: ➒ There always exists a β€œperfect partition” (𝐡1, … , 𝐡𝑛) of

    the cake such that 𝑉𝑖 𝐡𝑗 = Ξ€1

    𝑛 for every 𝑖, 𝑗 ∈ [𝑛].

    ➒ Every agent values every bundle equally.

    β€’ Theorem [Alon β€˜87]: ➒ There exists a perfect partition that only cuts the cake at π‘π‘œπ‘™π‘¦(𝑛) points.

    ➒ In contrast, Lyapunov’s proof is non-constructive, and might need an unbounded number of cuts.

  • Perfect Partition

    CSC304 - Nisarg Shah 8

    β€’ Q: Can you use an algorithm for computing a perfect partition as a black-box to design a randomized SP+EF mechanism?

    ➒ Yes! Compute a perfect partition, and assign the 𝑛bundles to the 𝑛 players uniformly at random.

    ➒ Why is this EF? o Every agent values every bundle at Ξ€1 𝑛.

    ➒ Why is this SP-in-expectation?o Because an agent is assigned a random bundle, her expected

    utility is Ξ€1 𝑛, irrespective of what she reports.

  • Pareto Optimality (PO)

    CSC304 - Nisarg Shah 9

    β€’ Definition➒ We say that an allocation 𝐴 = (𝐴1, … , 𝐴𝑛) is PO if there is

    no alternative allocation 𝐡 = (𝐡1, … , 𝐡𝑛) such that

    1. Every agent is at least as happy: 𝑉𝑖 𝐡𝑖 β‰₯ 𝑉𝑖(𝐴𝑖), βˆ€π‘– ∈ 𝑁

    2. Some agent is strictly happier: 𝑉𝑖 𝐡𝑖 > 𝑉𝑖(𝐴𝑖), βˆƒπ‘– ∈ 𝑁

    ➒ I.e., an allocation is PO if there is no β€œbetter” allocation.

    β€’ Q: Is it PO to give the entire cake to player 1?

    β€’ A: Not necessarily. But yes if player 1 values β€œevery part of the cake positively”.

  • PO + EF

    CSC304 - Nisarg Shah 10

    β€’ Theorem [Weller β€˜85]:➒ There always exists an allocation of the cake that is both

    envy-free and Pareto optimal.

    β€’ One way to achieve PO+EF:➒ Nash-optimal allocation: argmax𝐴 Ο‚π‘–βˆˆπ‘π‘‰π‘– π΄π‘–βž’ Obviously, this is PO. The fact that it is EF is non-trivial.

    ➒ This is named after John Nash.o Nash social welfare = product of utilities

    o Different from utilitarian social welfare = sum of utilities

  • Nash-Optimal Allocation

    CSC304 - Nisarg Shah 11

    β€’ Example:➒ Green player has value 1 distributed over 0, Ξ€2 3➒ Blue player has value 1 distributed over [0,1]

    ➒ Without loss of generality (why?) suppose: o Green player gets π‘₯ fraction of [0, Ξ€2 3]

    o Blue player gets the remaining 1 βˆ’ π‘₯ fraction of [0, Ξ€2 3] AND all of [ Ξ€2

    3 , 1].

    ➒ Green’s utility = π‘₯, blue’s utility = 1 βˆ’ x β‹… 23+

    1

    3=

    3βˆ’2π‘₯

    3

    ➒ Maximize: π‘₯ β‹…3βˆ’2π‘₯

    3β‡’ π‘₯ = Ξ€3 4 ( Ξ€3 4 fraction of Ξ€2 3 is Ξ€1 2).

    0 1ΰ΅—2 3

    Allocation 0 1

    ΰ΅—1 2

    Each player’s utility = Ξ€3 4

  • Problem with Nash Solution

    CSC304 - Nisarg Shah 12

    β€’ Difficult to compute in general➒ I believe it should require an unbounded number of

    queries in the Robertson-Webb model. But I can’t find such a result in the literature.

    β€’ Theorem [Aziz & Ye β€˜14]:➒ For piecewise constant valuations, the Nash-optimal

    solution can be computed in polynomial time.

    0 1

    The density function of a piecewise constant valuation looks like this

  • Indivisible Goods

    CSC304 - Nisarg Shah 13

    β€’ Goods cannot be shared / divided among players➒ E.g., house, painting, car, jewelry, …

    β€’ Problem: Envy-free allocations may not exist!

  • Indivisible Goods: Setting

    CSC304 - Nisarg Shah 14

    8 7 20 5

    9 11 12 8

    9 10 18 3

    We assume additive values. So, e.g., 𝑉 , = 8 + 7 = 15

    Given such a matrix of numbers, assign each good to a player.

  • Indivisible Goods

    CSC304 - Nisarg Shah 15

    β€’ Envy-freeness up to one good (EF1):

    βˆ€π‘–, 𝑗 ∈ 𝑁, βˆƒπ‘” ∈ 𝐴𝑗 ∢ 𝑉𝑖 𝐴𝑖 β‰₯ 𝑉𝑖 𝐴𝑗\{𝑔}

    ➒ Technically, we need either this or 𝐴𝑗 = βˆ….

    ➒ β€œIf 𝑖 envies 𝑗, there must be some good in 𝑗’s bundle such that removing it would make 𝑖 envy-free of 𝑗.”

    β€’ Does there always exist an EF1 allocation?

  • EF1

    CSC304 - Nisarg Shah 16

    β€’ Yes! We can use Round Robin.

    ➒ Agents take turns in cyclic order: 1,2,… , 𝑛, 1,2,… , 𝑛, …

    ➒ In her turn, an agent picks the good she likes the most among the goods still not picked by anyone.

    β€’ Observation: This always yields an EF1 allocation.➒ Informal proof on the board.

    β€’ Sadly, on some instances, this returns an allocation that is not Pareto optimal.

  • EF1+PO?

    CSC304 - Nisarg Shah 17

    β€’ Nash welfare to rescue!

    β€’ Theorem [Caragiannis et al. β€˜16]:➒ The allocation argmax𝐴 Ο‚π‘–βˆˆπ‘π‘‰π‘– 𝐴𝑖 is EF1 + PO.

    ➒ Note: This maximization is over only β€œintegral” allocations that assign each good to some player in whole.

    ➒ Note: Subtle tie-breaking if all allocations have zero Nash welfare.o Step 1: Choose a subset of players 𝑆 βŠ† 𝑁 with largest |𝑆| such that

    it is possible to give a positive utility to every player in 𝑆simultaneously.

    o Step 2: Choose argmax𝐴 Ο‚π‘–βˆˆπ‘†π‘‰π‘– 𝐴𝑖

  • 8 7 20 5

    9 11 12 8

    9 10 18 3

    Integral Nash Allocation?

  • 8 7 20 5

    9 11 12 8

    9 10 18 3

    20 * (11+8) * 9 = 3420 is the maximum possible product

  • Computation

    CSC304 - Nisarg Shah 20

    β€’ For indivisible goods, Nash-optimal solution is strongly NP-hard to compute➒ That is, remains NP-hard even if all values in the matrix

    are bounded

    β€’ Open Question: If our goal is EF1+PO, is there a different polynomial time algorithm? ➒ Not sure. But a recent paper gives a pseudo-polynomial

    time algorithm for EF1+POo Time is polynomial in 𝑛, π‘š, and max

    𝑖,𝑔𝑉𝑖 𝑔 .

  • Stronger Fairness

    CSC304 - Nisarg Shah 21

    β€’ Open Question: Does there always exist an EFxallocation?

    β€’ EF1: βˆ€π‘–, 𝑗 ∈ 𝑁, βˆƒπ‘” ∈ 𝐴𝑗 ∢ 𝑉𝑖 𝐴𝑖 β‰₯ 𝑉𝑖 𝐴𝑗\{𝑔}➒ Note: Or 𝐴𝑗 = βˆ… also allowed.

    ➒ Intuitively, 𝑖 doesn’t envy 𝑗 if she gets to remove her most valued item from 𝑗’s bundle.

    β€’ EFx: βˆ€π‘–, 𝑗 ∈ 𝑁, βˆ€π‘” ∈ 𝐴𝑗 ∢ 𝑉𝑖 𝐴𝑖 β‰₯ 𝑉𝑖 𝐴𝑗\{𝑔}➒ Note: βˆ€π‘” ∈ 𝐴𝑗 such that 𝑉𝑖 𝑔 > 0.

    ➒ Intuitively, 𝑖 doesn’t envy 𝑗 even if she removes her least positively valued item from 𝑗’s bundle.

  • Stronger Fairness

    CSC304 - Nisarg Shah 22

    β€’ To clarify the difference between EF1 and EFx:➒ Suppose there are two players and three goods with

    values as follows.

    ➒ If you give {A} β†’ P1 and {B,C} β†’ P2, it’s EF1 but not EFx.o EF1 because if P1 removes C from P2’s bundle, all is fine.

    o Not EFx because removing B doesn’t eliminate envy.

    ➒ Instead, {A,B} β†’ P1 and {C} β†’ P2 would be EFx.

    A B C

    P1 5 1 10

    P2 0 1 10


Top Related