T RUTH J USTICE A LGOS Fair Division V: Indivisible Goods Teachers: Ariel Procaccia (this time) and Alex Psomas
INDIVISIBLE GOODS โข Set ๐ป of ๐ goods ๐ป โข Each good is indivisible โข Players ๐ = 1, โฆ , ๐ have valuations ๐ ๐ for bundles of goods โข Valuations are additive if for all ๐ โ ๐ป and ๐ โ ๐ ๐ = ฯ ๐โ๐ป ๐ ๐ , ๐ ๐ ๐ โข Assume additivity unless noted otherwise โข An allocation is a partition of the goods, denoted ๐ฉ = (๐ต 1 , โฆ , ๐ต ๐ ) โข Envy-freeness and proportionality are infeasible!
MAXIMIN SHARE GUARANTEE Total: Total: Total: $50 $30 $20 $3 $2 $50 $30 $5 $5 $5
MAXIMIN SHARE GUARANTEE Total: Total: Total: $50 $30 $20 $3 $2 $50 $30 $5 $5 $5 $3 $2 $5 $40 $10 $20 $20 Total: Total: Total: $30 $30 $40
MAXIMIN SHARE GUARANTEE โข Maximin share (MMS) guarantee [Budish 2011] of player ๐ : ๐ 1 ,โฆ,๐ ๐ min max ๐ ๐ (๐ ๐ ) ๐ โข An MMS allocation is such that ๐ ๐ (๐ต ๐ ) is at least ๐โs MMS guarantee for all ๐ โ ๐ โข For ๐ = 2 an MMS allocation always exists โข Theorem [Kurokawa et al. 2018]: โ๐ โฅ 3 there exist additive valuation functions that do not admit an MMS allocation
COUNTEREXAMPLE FOR ๐ = 3 17 25 12 1 17 25 12 1 17 25 12 1 17 17 25 25 12 12 1 1 2 22 3 28 2 22 3 28 2 22 3 28 2 22 3 28 2 22 3 28 11 0 21 23 11 0 21 23 11 0 21 11 0 21 23 11 0 21 23 23
COUNTEREXAMPLE FOR ๐ = 3 1 1 1 1 17 25 12 1 ร 10 6 ร 10 3 + + 1 1 1 1 2 22 3 28 1 1 1 1 11 0 21 23 3 -1 -1 -1 3 -1 0 0 3 0 -1 0 0 0 0 0 -1 0 0 0 0 0 -1 0 0 0 0 0 -1 0 0 0 0 0 0 -1 Player 1 Player 2 Player 3
APPROXIMATE ENVY-FREENESS โข Assume general monotonic valuations, i.e., for all ๐ โ ๐ โ ๐ป, ๐ ๐ ๐ โค ๐ ๐ (๐) โข An allocation ๐ต 1 , โฆ , ๐ต ๐ is envy free up to one good (EF1) if and only if โ๐, ๐ โ ๐, โ๐ โ ๐ต ๐ s.t. ๐ค ๐ ๐ต ๐ โฅ ๐ค ๐ ๐ต ๐ \{๐} โข Theorem [Lipton et al. 2004]: An EF1 allocation exists and can be found in polynomial time
PROOF OF THEOREM โข A partial allocation is an allocation of a subset of the goods โข Given a partial allocation ๐ฉ , we have an edge (๐, ๐) in its envy graph if ๐ envies ๐ โข Lemma: An EF1 partial allocation ๐ฉ can be transformed in polynomial time into an EF1 partial allocation ๐ช of the same goods with an acyclic envy graph
PROOF OF LEMMA โข If ๐ป has a cycle ๐ท , shift allocations along ๐ท to obtain ๐ 4 ๐ 1 ๐ฉโฒ ; clearly EF1 is maintained โข #edges in envy graph of ๐ฉโฒ decreased: ๐ 3 ๐ 2 โฆ Same edges between ๐ โ ๐ท โฆ Edges from ๐ โ ๐ท to ๐ท shifted ๐ 4 ๐ 2 โฆ Edges from ๐ท to ๐ โ ๐ท can only decrease โฆ Edges inside C decreased ๐ 1 ๐ 3 โข Iteratively remove cycles โ
PROOF OF THEOREM โข Maintain EF1 and acyclic envy graph โข In round 1, allocate good ๐ 1 to arbitrary agent โข ๐ 1 , โฆ , ๐ ๐โ1 are allocated in acyclic ๐ฉ โข Derive ๐ช by allocating ๐ ๐ to source ๐ โข ๐ ๐ ๐ถ ๐ = ๐ ๐ ๐ต ๐ โฅ ๐ ๐ ๐ต ๐ = ๐ ๐ ๐ถ ๐ โ ๐ ๐ โข Use lemma to eliminate cycles โ
ROUND ROBIN โข Let us return to additive valuations โข Now proving the existence of an EF1 allocation is trivial โข A round-robin allocation is EF1: Phase 1 Phase 2
IMPLICATIONS FOR CAKE CUTTING โข In cake cutting, we can define an allocation to be ๐ -envy free if for all ๐, ๐ โ ๐, ๐ ๐ ๐ต ๐ โฅ ๐ ๐ ๐ต ๐ โ ๐ โข The foregoing result has interesting implications for cake cutting! Poll 1 ? Complexity of ๐ -EF in the RW model? 1 ๐ โข ๐ โข ๐ ๐ 2 ๐ ๐ 2 1 โข ๐ โข ๐ ๐ 2 ๐
MAXIMUM NASH WELFARE โข An allocation ๐ฉ is Pareto efficient if there is no allocation ๐ฉโฒ such that โฒ โฅ ๐ ๐ ๐ ๐ต ๐ ๐ ๐ต ๐ for all ๐ โ ๐ , and โฒ > ๐ ๐ ๐ ๐ต ๐ ๐ ๐ต ๐ for some ๐ โ ๐ โข Round Robin is not efficient โข Is there a rule that guarantees both EF1 and efficiency?
MAXIMUM NASH WELFARE โข The Nash welfare of an allocation ๐ฉ is the product of values NW ๐ฉ = เท ๐ ๐ (๐ต ๐ ) ๐โ๐ โข The maximum Nash welfare (MNW) solution chooses an allocation that maximizes the Nash welfare โข For ease of exposition we ignore the case of NW ๐ฉ = 0 for all ๐ฉ โข Theorem [Caragiannis et al. 2016]: Assuming additive valuations, the MNW solution is EF1 and efficient
PROOF OF THEOREM โข Efficiency is obvious, so we focus on EF1 โข Assume for contradiction that ๐ envies ๐ by more than one good โข Let ๐ โ โ argmin ๐โ๐ต ๐ ,๐ ๐ ๐ >0 ๐ ๐ (๐)/๐ ๐ (๐) โข Move ๐ โ from ๐ to ๐ to obtain ๐ฉโฒ , we will show that NW ๐ฉ โฒ > NW(๐ฉ) โฒ ) for all ๐ โ ๐, ๐ , โข It holds that ๐ ๐ ๐ต ๐ = ๐ ๐ (๐ต ๐ โฒ = ๐ ๐ ๐ โ , and ๐ ๐ ๐ต ๐ ๐ ๐ต ๐ + ๐ โฒ = ๐ ๐ ๐ โ ๐ ๐ ๐ต ๐ ๐ ๐ต ๐ โ ๐
PROOF OF THEOREM โข NW ๐ต โฒ ๐ ๐ ๐ โ ๐ ๐ ๐ โ NW ๐ต > 1 โ 1 โ 1 + > 1 โ ๐ ๐ ๐ต ๐ ๐ ๐ ๐ต ๐ ๐ ๐ ๐ โ ๐ ๐ โ ๐ ๐ ๐ต ๐ + ๐ < ๐ ๐ ๐ต ๐ ๐ ๐ ๐ โ โข Due to our choice of ๐ โ , ฯ ๐โ๐ต ๐ ๐ ๐ ๐ โ ๐ ๐ ๐ ๐ ๐ = ๐ ๐ ๐ต ๐ ๐ ๐ โ โค ฯ ๐โ๐ต ๐ ๐ ๐ ๐ ๐ ๐ต ๐ โข Due to EF1 violation, we have ๐ ๐ โ < ๐ ๐ ๐ ๐ต ๐ + ๐ ๐ ๐ต ๐ โข Multiply the last two inequalities to get the first โ
TRACTABILITY OF MNW 30 25 20 Time (s) 15 10 5 0 5 10 15 20 25 30 35 40 45 50 Number of players [Caragiannis et al., 2016]
INTERFACE
AN OPEN PROBLEM โข An allocation ๐ต 1 , โฆ , ๐ต ๐ is envy free up to any good (EFX) if and only if โ๐, ๐ โ ๐, โ๐ โ ๐ต ๐ , ๐ค ๐ ๐ต ๐ โฅ ๐ค ๐ ๐ต ๐ \{๐} โข Strictly stronger than EF1, strictly weaker than EF โข An EFX allocation exists for two players with monotonic valuations โข Existence is an open problem for ๐ โฅ 3 players with additive valuations
Recommend
More recommend