indivisible goods
play

INDIVISIBLE GOODS Set of goods Each good is indivisible Players - PowerPoint PPT Presentation

T RUTH J USTICE A LGOS Fair Division V: Indivisible Goods Teachers: Ariel Procaccia (this time) and Alex Psomas INDIVISIBLE GOODS Set of goods Each good is indivisible Players = 1, , have valuations


  1. T RUTH J USTICE A LGOS Fair Division V: Indivisible Goods Teachers: Ariel Procaccia (this time) and Alex Psomas

  2. INDIVISIBLE GOODS โ€ข Set ๐ป of ๐‘› goods ๐ป โ€ข Each good is indivisible โ€ข Players ๐‘‚ = 1, โ€ฆ , ๐‘œ have valuations ๐‘Š ๐‘— for bundles of goods โ€ข Valuations are additive if for all ๐‘‡ โŠ† ๐ป and ๐‘— โˆˆ ๐‘— ๐‘‡ = ฯƒ ๐‘•โˆˆ๐ป ๐‘Š ๐‘‚ , ๐‘Š ๐‘— ๐‘• โ€ข Assume additivity unless noted otherwise โ€ข An allocation is a partition of the goods, denoted ๐‘ฉ = (๐ต 1 , โ€ฆ , ๐ต ๐‘œ ) โ€ข Envy-freeness and proportionality are infeasible!

  3. MAXIMIN SHARE GUARANTEE Total: Total: Total: $50 $30 $20 $3 $2 $50 $30 $5 $5 $5

  4. 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

  5. MAXIMIN SHARE GUARANTEE โ€ข Maximin share (MMS) guarantee [Budish 2011] of player ๐‘— : ๐‘Œ 1 ,โ€ฆ,๐‘Œ ๐‘œ min max ๐‘Š ๐‘— (๐‘Œ ๐‘˜ ) ๐‘˜ โ€ข An MMS allocation is such that ๐‘Š ๐‘— (๐ต ๐‘— ) is at least ๐‘—โ€™s MMS guarantee for all ๐‘— โˆˆ ๐‘‚ โ€ข For ๐‘œ = 2 an MMS allocation always exists โ€ข Theorem [Kurokawa et al. 2018]: โˆ€๐‘œ โ‰ฅ 3 there exist additive valuation functions that do not admit an MMS allocation

  6. COUNTEREXAMPLE FOR ๐‘œ = 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

  7. COUNTEREXAMPLE FOR ๐‘œ = 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

  8. APPROXIMATE ENVY-FREENESS โ€ข Assume general monotonic valuations, i.e., for all ๐‘‡ โŠ† ๐‘ˆ โŠ† ๐ป, ๐‘Š ๐‘— ๐‘‡ โ‰ค ๐‘Š ๐‘— (๐‘ˆ) โ€ข An allocation ๐ต 1 , โ€ฆ , ๐ต ๐‘œ is envy free up to one good (EF1) if and only if โˆ€๐‘—, ๐‘˜ โˆˆ ๐‘‚, โˆƒ๐‘• โˆˆ ๐ต ๐‘˜ s.t. ๐‘ค ๐‘— ๐ต ๐‘— โ‰ฅ ๐‘ค ๐‘— ๐ต ๐‘˜ \{๐‘•} โ€ข Theorem [Lipton et al. 2004]: An EF1 allocation exists and can be found in polynomial time

  9. PROOF OF THEOREM โ€ข A partial allocation is an allocation of a subset of the goods โ€ข Given a partial allocation ๐‘ฉ , we have an edge (๐‘—, ๐‘˜) in its envy graph if ๐‘— envies ๐‘˜ โ€ข Lemma: An EF1 partial allocation ๐‘ฉ can be transformed in polynomial time into an EF1 partial allocation ๐‘ช of the same goods with an acyclic envy graph

  10. PROOF OF LEMMA โ€ข If ๐ป has a cycle ๐ท , shift allocations along ๐ท to obtain ๐‘‡ 4 ๐‘‡ 1 ๐‘ฉโ€ฒ ; clearly EF1 is maintained โ€ข #edges in envy graph of ๐‘ฉโ€ฒ decreased: ๐‘‡ 3 ๐‘‡ 2 โ—ฆ Same edges between ๐‘‚ โˆ– ๐ท โ—ฆ Edges from ๐‘‚ โˆ– ๐ท to ๐ท shifted ๐‘‡ 4 ๐‘‡ 2 โ—ฆ Edges from ๐ท to ๐‘‚ โˆ– ๐ท can only decrease โ—ฆ Edges inside C decreased ๐‘‡ 1 ๐‘‡ 3 โ€ข Iteratively remove cycles โˆŽ

  11. PROOF OF THEOREM โ€ข Maintain EF1 and acyclic envy graph โ€ข In round 1, allocate good ๐‘• 1 to arbitrary agent โ€ข ๐‘• 1 , โ€ฆ , ๐‘• ๐‘™โˆ’1 are allocated in acyclic ๐‘ฉ โ€ข Derive ๐‘ช by allocating ๐‘• ๐‘™ to source ๐‘— โ€ข ๐‘Š ๐‘˜ ๐ถ ๐‘˜ = ๐‘Š ๐‘˜ ๐ต ๐‘˜ โ‰ฅ ๐‘Š ๐‘˜ ๐ต ๐‘— = ๐‘Š ๐‘˜ ๐ถ ๐‘— โˆ– ๐‘• ๐‘™ โ€ข Use lemma to eliminate cycles โˆŽ

  12. ROUND ROBIN โ€ข Let us return to additive valuations โ€ข Now proving the existence of an EF1 allocation is trivial โ€ข A round-robin allocation is EF1: Phase 1 Phase 2

  13. IMPLICATIONS FOR CAKE CUTTING โ€ข In cake cutting, we can define an allocation to be ๐œ— -envy free if for all ๐‘—, ๐‘˜ โˆˆ ๐‘‚, ๐‘Š ๐‘— ๐ต ๐‘— โ‰ฅ ๐‘Š ๐‘— ๐ต ๐‘˜ โˆ’ ๐œ— โ€ข The foregoing result has interesting implications for cake cutting! Poll 1 ? Complexity of ๐œ— -EF in the RW model? 1 ๐‘œ โ€ข ๐‘ƒ โ€ข ๐‘ƒ ๐œ— 2 ๐œ— ๐‘œ 2 1 โ€ข ๐‘ƒ โ€ข ๐‘ƒ ๐œ— 2 ๐œ—

  14. MAXIMUM NASH WELFARE โ€ข An allocation ๐‘ฉ is Pareto efficient if there is no allocation ๐‘ฉโ€ฒ such that โ€ฒ โ‰ฅ ๐‘Š ๐‘Š ๐‘— ๐ต ๐‘— ๐‘— ๐ต ๐‘— for all ๐‘— โˆˆ ๐‘‚ , and โ€ฒ > ๐‘Š ๐‘Š ๐‘˜ ๐ต ๐‘˜ ๐‘˜ ๐ต ๐‘˜ for some ๐‘˜ โˆˆ ๐‘‚ โ€ข Round Robin is not efficient โ€ข Is there a rule that guarantees both EF1 and efficiency?

  15. MAXIMUM NASH WELFARE โ€ข The Nash welfare of an allocation ๐‘ฉ is the product of values NW ๐‘ฉ = เท‘ ๐‘Š ๐‘— (๐ต ๐‘— ) ๐‘—โˆˆ๐‘‚ โ€ข The maximum Nash welfare (MNW) solution chooses an allocation that maximizes the Nash welfare โ€ข For ease of exposition we ignore the case of NW ๐‘ฉ = 0 for all ๐‘ฉ โ€ข Theorem [Caragiannis et al. 2016]: Assuming additive valuations, the MNW solution is EF1 and efficient

  16. PROOF OF THEOREM โ€ข Efficiency is obvious, so we focus on EF1 โ€ข Assume for contradiction that ๐‘— envies ๐‘˜ by more than one good โ€ข Let ๐‘• โ‹† โˆˆ argmin ๐‘•โˆˆ๐ต ๐‘˜ ,๐‘Š ๐‘— ๐‘• >0 ๐‘Š ๐‘˜ (๐‘•)/๐‘Š ๐‘— (๐‘•) โ€ข Move ๐‘• โ‹† from ๐‘˜ to ๐‘— to obtain ๐‘ฉโ€ฒ , we will show that NW ๐‘ฉ โ€ฒ > NW(๐‘ฉ) โ€ฒ ) for all ๐‘™ โ‰  ๐‘—, ๐‘˜ , โ€ข It holds that ๐‘Š ๐‘™ ๐ต ๐‘™ = ๐‘Š ๐‘™ (๐ต ๐‘™ โ€ฒ = ๐‘Š ๐‘— ๐‘• โ‹† , and ๐‘Š ๐‘— ๐ต ๐‘— ๐‘— ๐ต ๐‘— + ๐‘Š โ€ฒ = ๐‘Š ๐‘˜ ๐‘• โ‹† ๐‘Š ๐‘˜ ๐ต ๐‘˜ ๐‘˜ ๐ต ๐‘˜ โˆ’ ๐‘Š

  17. PROOF OF THEOREM โ€ข NW ๐ต โ€ฒ ๐‘Š ๐‘˜ ๐‘• โ‹† ๐‘Š ๐‘— ๐‘• โ‹† NW ๐ต > 1 โ‡” 1 โˆ’ 1 + > 1 โ‡” ๐‘Š ๐‘˜ ๐ต ๐‘˜ ๐‘Š ๐‘— ๐ต ๐‘— ๐‘Š ๐‘˜ ๐‘• โ‹† ๐‘— ๐‘• โ‹† ๐‘Š ๐‘— ๐ต ๐‘— + ๐‘Š < ๐‘Š ๐‘˜ ๐ต ๐‘˜ ๐‘Š ๐‘— ๐‘• โ‹† โ€ข Due to our choice of ๐‘• โ‹† , ฯƒ ๐‘•โˆˆ๐ต ๐‘˜ ๐‘Š ๐‘˜ ๐‘• โ‹† ๐‘˜ ๐‘• ๐‘Š ๐‘— ๐‘• = ๐‘Š ๐‘˜ ๐ต ๐‘˜ ๐‘— ๐‘• โ‹† โ‰ค ฯƒ ๐‘•โˆˆ๐ต ๐‘˜ ๐‘Š ๐‘Š ๐‘Š ๐‘— ๐ต ๐‘˜ โ€ข Due to EF1 violation, we have ๐‘— ๐‘• โ‹† < ๐‘Š ๐‘Š ๐‘— ๐ต ๐‘— + ๐‘Š ๐‘— ๐ต ๐‘˜ โ€ข Multiply the last two inequalities to get the first โˆŽ

  18. 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]

  19. INTERFACE

  20. AN OPEN PROBLEM โ€ข An allocation ๐ต 1 , โ€ฆ , ๐ต ๐‘œ is envy free up to any good (EFX) if and only if โˆ€๐‘—, ๐‘˜ โˆˆ ๐‘‚, โˆ€๐‘• โˆˆ ๐ต ๐‘˜ , ๐‘ค ๐‘— ๐ต ๐‘— โ‰ฅ ๐‘ค ๐‘— ๐ต ๐‘˜ \{๐‘•} โ€ข Strictly stronger than EF1, strictly weaker than EF โ€ข An EFX allocation exists for two players with monotonic valuations โ€ข Existence is an open problem for ๐‘œ โ‰ฅ 3 players with additive valuations

Recommend


More recommend