Truthful Mechanism for Facility Allocation: A Characterization and Improvement of Approximation Ratio Pinyan Lu, MSR Asia Yajun Wang, MSR Asia Yuan Zhou, Carnegie Mellon University 1 Problem discussed Design a mechanism for the following n-player game Players is located on a real line Each player report their location to the mechanism The mechanism decides a new location to build the facility mechanism g y x1 x2 2 Problem discussed (cont’d) Design a mechanism for the following n-player game Players is located on a real line Each player report their location to the mechanism The mechanism decides a new location to build the facility For example, the mean func., g = (x 1 + x 2 )=2 mechanism g = (x 1 + x 2 )=2 x1 y x2 3 Problem discussed (cont’d) Design a mechanism for the following n-player game Players is located on a real line Each player report their location to the mechanism The mechanism decides a new location to build the facility For example, the mean func., g = (x 1 + x 2 )=2 This encourages Player 1 to report x 01 = 2x 1 ¡ x 2 , then g(x 01 ; x 2 ) = x 1 becomes closer to Player 1’s real location. mechanism g = (x 1 + x 2 )=2 x 01 x1 y x2 4 Truthfulness Design a mechanism for the following n-player game Players is located on a real line Each player report their location to the mechanism The mechanism decides a new location to build the facility Truthful mechanism does not encourage player to report untruthful locations g(x 1 ; x 2 ) = 0 g(x 1 ; x 2 ) = x 1 g(x 1 ; x 2 ) = minf x 1 ; x 2 g mechanism g = x 1 x1 x2 5 Truthfulness of g(x 1 ; x 2 ) = minf x 1 ; x 2 g Suppose w.l.o.g. that x 1 · x 2 x 1 has no incentive to lie x 2 will not change the outcome of g if it misreports a value x 02 ¸ x 1 If x 2 misreports that x 02 < x 1 , then the decision of g will be even farther from x 2 g = minf x 1 ; x 2 g x1 x 02 6 Truthfulness of g(x 1 ; x 2 ) = minf x 1 ; x 2 g Suppose w.l.o.g. that x 1 · x 2 x 1 has no incentive to lie x 2 will not change the outcome of g if it misreports a value x 02 ¸ x 1 If x 2 misreports that x 02 < x 1 , then the decision of g will be even farther from x 2 Corollary: a mechanism which outputs the leftmost (rightmost) location among n players is truthful gg0== minf minf xx11;;xx202gg x1 x 02 7 A natural question Is there any other (non-trivial) truthful mechanisms? Can we fully characterize the set of truthful mechanisms? Gibbard-Satterthwaite Theorem. If players can give arbitrary preferences, then the only truthful mechanisms are dictatorships, i.e. g = f (x i ) for some i 2 [n] In our facility game, since players are not able to give arbitrary preferences, we have a set of richer truthful mechanisms, such as leftmost(rightmost), and … 8 Even more interesting truthful mechanisms Mechanism: gk (x 1 ; x 2 ; ¢¢¢; x n ) = k-t h left locat ion among input s Suppose w.l.o.g. that x 1 · x 2 · x 3 · ¢¢¢· x n x k has no incentive to lie x i (i 6 = k) can change the outcome only when it lies to be x 0i where x 0i and x i are on different sides of x k , but this makes the new outcome farther from x i Corollary: outputting the median ( g[n =2] ) is truthful gkk0 x1 x 0ii xk xn 9 Social cost and approximation ratio Good news! Median is truthful! Median also optimizes the social cost, i.e. the total distance from each player to the facility Xn scx 1 ;x 2 ;¢¢¢;x n (g) := jx i ¡ g(x 1 ; x 2 ; ¢¢¢; x n )j i= 1 Approximation ratio of½mechanism g ©(g) := max x 1 ;x 2 ;¢¢¢;x n scx 1 ;x 2 ;¢¢¢;x n (g) OPT(x 1 ; x 2 ; ¢¢¢; x n ) ¾ ©(median) = 1 10 Approximation ratio of other mechanisms ©(g ´ 0) = + 1 Gap instance: 0 x 1 ; x 2 ; ¢¢¢; x n ©(out put t ing t he left most player's locat ion) = n ¡ 1 Gap instance: x1 x 2 ; x 3 ; ¢¢¢; x n 11 Extend to two facility game Suppose we have more budget, and we can afford building two facilities Each player’s cost function: its distance to the closest facility Good truthful approximation? A simple try Mechanism: set facilities on the leftmost and rightmost player’s location 12 Extend to two facility game A simple try Mechanism: set facilities on the leftmost and rightmost player’s location ¾ Gap Instance: © ¸ n ¡ 2 ) ©= n¡ 2 ©· n¡ 2 sc(Mech.) = n ¡ 2 x1 x 2 ; x 3 ; ¢¢¢; x n ¡ ¡ 1 0 OPT = 1 1 xn 1 13 Randomized mechanisms The mechanism selects pair of locations according to some distribution Each player’s cost function is the expected distance to the closest facility Does randomness help approximation ratio? 14 Multiple locations per agent Agent i controls wi locations x i = (x i 1 ; x i 2 ; ¢¢¢; x i w i ) P wi Agent i ‘s cost function is j = 1 jg ¡ x i j j Social cost: P P wi n i= 1 j = 1 jg ¡ xi j j A randomized truthful mechanism Given f x 1 ; x 2 ; ¢¢¢; x n g, return med(x i ) with Pn probability wi = j = 1 wj Claim. The mechanism is truthful Theorem. The mechanism’s approximation ratio is 2 min1· j · n wj Pn 3¡ j = 1 wj 15 Summary of questions. Characterization Is there a full characterization for deterministic truthful mechanism in one-facility game? Approximation Upper/lower bound for two facility game in deterministic/randomized case? Lower bound for one facility game in randomized case when agents control multiple locations? 16 Our result and related work Give a full characterization of one-facility deterministic truthful mechanisms Similar result by [Moulin] and [Barbera-Jackson] Improve the bounds approximation ratio in several extended game settings Setting one facility two facilities deterministic deterministic Previous known* two facilities randomized one facility, randomized** 1 vs. 1 3/2 vs. n – 1 ? vs. n – 1 ? vs. ? Our result N/A 2 vs. n – 1 1.045 vs. n – 1 1.33 vs. 3 Follow-up result N/A Ω(n) vs. n – 1 1.045 vs. 4 N/A *: Most of previous results are due to [ProcacciaTennenholtz] **: In this setting, each player can control multiple locations 17 Outline Characterization of one-facility deterministic truthful mechanisms Lower bound for randomized two-facility games Lower bound for randomized one-facility games when agents control multiple locations Upper bound for randomized two-facility games 18 The characterization Generally speaking, the set of one-facility deterministic truthful mechanisms consists of min-max functions (and its variations) Actually we prove that all truthful mechanism can be written in a standard min-max form with 2n parameters (perhaps with some variation) max x1 standard form min min x2 max x3 med med x1 c1 med c1 x1 c2 x2 x3 med c3 x1 c4 med med x2 c5 x1 c6 med c7 x1 c8 19 More precise in the characterization The image set of the mechanism can be an arbitrary closed set U We restrict the min-max function onto U by finding the nearest point in U closed set U f : max min min x1 x2 max x3 mechanism g x1 c1 20 More precise in the characterization The image set of the mechanism can be an arbitrary closed set U We restrict the min-max function onto U by finding the nearest point in U closed set U f : max min min x1 x2 max x3 mechanism g x1 c1 21 More precise in the characterization The image set of the mechanism can be an arbitrary closed set U We restrict the min-max function onto U by finding the nearest point in U What about when there are 2 nearest points ? A tie-breaking gadget takes response of that ! closed set U f : max min min x1 mechanismtiegbreaker x2 max x3 x1 c1 22 The proof – warm-up part Image set of g Lemma. If g is a truthful mechanism, then g(x; x; ¢¢¢; x) goes to the closest point in I (g) from x , for all x Proof. For every y = g(x 1 ; x 2 ; ¢¢¢; x n ), jy ¡ xj = jg(x 1 ; x 2 ; ¢¢¢; x n ) ¡ xj ¸ jg(x; x 2 ; ¢¢¢; x n ) ¡ xj ¸ jg(x; x; ¢¢¢; x n ) ¡ xj ¢¢¢ ¸ jg(x; x; ¢¢¢; x) ¡ xj Corollary. U = I (g) is closed. Now, for simplicity, assume U = I (g) = (¡ 1 ; + 1 ) 23 Main lemma Lemma. For each truthful mechanism g, there exists a min-max function , such that g(x 1 ; x 2 ; ¢¢¢; x n ) is the closest point in I (g) from f (x 1 ; x 2 ; ¢¢¢; x n ) , for all inputs x 1 ; x 2 ; ¢¢¢; x n 2 R Proof (sketch). Prove by induction on n When n = 1, g(x) should output the closest point in I (g) from x : f (x) = x For n > 1 24 Main lemma 0 g n > 1 For , define x 1 ;x 2 ;¢¢¢;x n ¡ 1 (x n ) = g(x 1 ; x 2 ; ¢¢¢; x n ) Claim 1. g0 is truthful Claim 2. 9L = L(x 1 ; ¢¢¢; x n ¡ 1 ); R = R(x 1 ; ¢¢¢; x n ¡ 1 ); s:t: I (gx01 ;x 2 ;¢¢¢;x n ¡ 1 ) = I (g) \ [L ; R] Claim 3. L; R , as mechanisms for (n ¡ 1) -player game, are truthful Claim 4. 9L l ; L r : I (L ) = I (g) \ [L l ; L r ]; 9Rl ; Rr : I (R) = I (g) \ [Rl ; Rr ] I (g) I (gx01 ;x 2 ;¢¢¢;x n ¡ 1 ) L (x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ) I (g) I (L ) Ll R(x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ) I (R) Lr Rl Rr 25 Main lemma Thus, g(x 1 ; x 2 ; ¢¢¢; x n ) = gx01 ;x 2 ;¢¢¢;x n ¡ 1 = med(L ; x n ; R) L = med(L l ; g1 (x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ); L r ) R = med(Rl ; g2 (x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ); Rr ) g: I (g) med xn L I (gx01 ;x 2 ;¢¢¢;x n ¡ 1 ) L (x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ) I (g) I (L ) Ll R R(x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ) I (R) Lr Rl Rr 26 Main lemma Thus, g(x 1 ; x 2 ; ¢¢¢; x n ) = gx01 ;x 2 ;¢¢¢;x n ¡ 1 = med(L ; x n ; R) L = med(L l ; g1 (x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ); L r ) R = med(Rl ; g2 (x 1 ; x 2 ; ¢¢¢; x n ¡ 1 ); Rr ) g: xn med Ll g1 med Lr med R l g2 Rr 27 Main lemma g: med 1 player: xn med Ll g1 med R l g2 Lr g( 1) : x1 Rr 2 players: g( 2) : med x2 med c1 g( 1) 1 c2 med c3 g( 1) 2 c4 28 Main lemma g: med 1 player: xn med 3 players: Ll g1 med R l g2 Lr g( 1) : x1 Rr 2 players: g( 2) : med x2 med c1 x 1 c2 med c3 x 1 c4 29 Main lemma 2 players: 1 player: ( 1) g g( 2) : x1 : 3 players: ( 3) g c1 x 1 c9 med ( 2) c10 g1 med x 2 med c1 x 1 c2 c3 x 1 x3 x2 med : med med med c2 med c3 x 1 c4 med c11 med c12 ( 2) g2 med x 2 med c4 c5 x 1 c6 c7 x 1 c8 30 Main lemma 2 players: 1 player: ( 1) g g( 2) : x1 : 3 players: ( 3) g c1 x 1 c9 med c10 med x 2 med c1 x 1 c2 c3 x 1 x3 x2 med : med med med c2 med c3 x 1 c4 med c11 med c12 med x 2 med c4 c5 x 1 c6 c7 x 1 c8 31 The reverse direction Lemma. Every min-max function is truthful Observation. To prove a n -player mechanism is truthful, only need to prove the 1-player mechanisms gx01 ;¢¢¢;x i ¡ 1 ;x i + 1 ;¢¢¢;x n (x i ) are truthful for every i and x j (j 6= i ) Theorem. The characterization is full 32 Multiple locations per agent Theorem. Any randomized truthful mechanism of the one facility game has an approximation ration at least 1.33 in the setting that each agent controls multiple locations. Theorem (weaker). Any randomized truthful mechanism of the one facility game has an approximation ration at least 1.2 in the setting that each agent controls multiple locations. 33 Multiple locations per agent (cont’d) Player 1 Player 2 Proof. (weaker version) Instance 1 g = D1 Instance 2 g = D2 0 1 L 2 = Ex 2 D 2 [jx ¡ 0j ¢1x · 0 ] Instance 3 For For g = D3 R2 = Ex 2 D 2 [jx ¡ 1j ¢1x ¸ 1 ] 0 1 0 1 Player 1 at Instance 1 (compared to Instance 2) 2 ¢jx ¡ 0j x 2 D 1 + jx ¡ 1j x 2 D 1 · 2 ¢jx ¡ 0j x 2 D 2 + jx ¡ 1j x 2 D 2 ) jx ¡ 0j x 2 D 1 · jx ¡ 0j x 2 D 2 + 2(L 2 + R2 ) Player 2 at Instance 3 (compared to Instance 2) 2 ¢jx ¡ 1j x 2 D 3 + jx ¡ 0j x 2 D 3 · 2 ¢jx ¡ 1j x 2 D 2 + jx ¡ 0j x 2 D 2 ) jx ¡ 1j x 2 D 3 · jx ¡ 1j x 2 D 2 + 2(L 2 + R2 ) 34 Multiple locations per agent (cont’d) Player 1 Player 2 Proof. (weaker version) Instance 1 g = D1 Instance 2 g = D2 0 1 L 2 = Ex 2 D 2 [jx ¡ 0j ¢1x · 0 ] Instance 3 g = D3 R2 = Ex 2 D 2 [jx ¡ 1j ¢1x ¸ 1 ] 0 1 0 1 For Player 1 jx ¡ 0j x 2 D 1 · jx ¡ 0j x 2 D 2 + 2(L 2 + R2 ) For Player 2 jx ¡ 1j x 2 D 3 · jx ¡ 1j x 2 D 2 + 2(L 2 + R2 ) Assume <1.2 approx. For Inst. 1 2 ¢jx ¡ 0j x 2 D 1 + 4 ¢jx ¡ 1j x 2 D 1 < 1:2 ¢2 For Inst. 2 3 ¢jx ¡ 0j x 2 D 2 + 3 ¢jx ¡ 1j x 2 D 2 < 1:2 ¢3 For Inst. 3 4 ¢jx ¡ 0j x 2 D 3 + 2 ¢jx ¡ 1j x 2 D 3 < 1:2 ¢2 35 Multiple locations per agent (cont’d) Player 1 Player 2 Proof. (weaker version) Instance 1 g = D1 Instance 2 g = D2 0 1 L 2 = Ex 2 D 2 [jx ¡ 0j ¢1x · 0 ] Instance 3 g = D3 For Player 1 For Player 2 R2 = Ex 2 D 2 [jx ¡ 1j ¢1x ¸ 1 ] 0 1 0 1 jx ¡ 0j x 2 D 1 · jx ¡ 0j x 2 D 2 + 2(L 2 + R2 ) 1.6 < jx ¡ 1j x 2 D 3 · jx ¡ 1j x 2 D 2 + 2(L 2 + R2 ) Assume <1.2 approx. < 1.6 Contradiction For Inst. 1 jx ¡ 1j x 2 D 1 < 0:2 For Inst. 2 jx ¡ 0j x 2 D 2 + jx ¡ 1j x 2 D 2 < 1:2 For Inst. 3 jx ¡ 0j x 2 D 3 < 0:2 ) jx ¡ 0j x 2 D 1 > 0:8 ) L 2 + R2 < 0:1 ) jx ¡ 1j x 2 D 3 > 0:8 36 Multiple locations per agent (cont’d) Player 1 Player 2 Proof. (stronger version) Instance 1 Instance 2 0 1 Instance 3 0 1 Instance 4 0 1 Instance 5 0 1 0 1 37 Multiple locations per agent (cont’d) Player 1 Player 2 Proof. (stronger version) Instance i (1 · i · K ) £ (K + i ) Instance K + 1 0 £ (2K + 1) 0 Instance 2K + 2 ¡ i (K ¸ i ¸ 1) £ (K + 1 ¡ i ) £ (2K + 1) 0 £ (K + 1 ¡ i ) £ (2K + 1) 1 £ (2K + 1) 1 £ (K + i ) 1 38 Multiple locations per agent (cont’d) Linear Programming Take K = 500 : ® > 1:33 39 Lower bound for 2-facility randomized case Theorem. For any 2-facility randomized truthful mechanism, the approximation ratio is at least 1 a ¡ , where n ¸ 30 is the number of players 1:045 n¡ 3 Proof Consider instanceI :1 player at ¡ 1 , n ¡ 2 players at 0 , 1 player at 1 e1 + e2 + e3 ¸ 1 For mechanisms within 2-approx. : e2 · 2=(n ¡ 2) Assume w.l.o.g.: e3 ¸ 1=2 ¡ 1=(n ¡ 2) e1 x1 yl ¡ 1 e2 x 2 ; x 3 ; ¢¢¢; x n ¡ 0 yr e3 1 xn 1 40 Lower bound for 2-facility randomized case Theorem. For any 2-facility randomized truthful mechanism, the approximation ratio is at least 1 a ¡ , where n ¸ 30 is the number of players 1:045 n¡ 3 e3 ¸ 1=2 ¡ 1=(n ¡ 2) Proof Consider instanceI :1 player at ¡ 1 , n ¡ 2 players at 0 , 1 player at 1 Another instance I 0: 1 player at ¡ 1, n ¡ 2 players at 0 , 1 player at 1 + ® e1 x1 yl ¡ 1 e2 x 2 ; x 3 ; ¢¢¢; x n ¡ 0 yr e3 1 xx 0nn 1 +1 ® 41 Lower bound for 2-facility randomized case Theorem. For any 2-facility randomized truthful mechanism, the approximation ratio is at least 1 a ¡ , where n ¸ 30 is the number of players 1:045 n¡ 3 e3 ¸ 1=2 ¡ 1=(n ¡ 2) Proof Consider instanceI :1 player at ¡ 1 , n ¡ 2 players at 0 , 1 player at 1 Another instance I 0: 1 player at ¡ 1, n ¡ 2 players at 0 , 1 player at 1 + ® By truthfulness: e03 ¸ 1=2 ¡ 1=(n ¡ 2) ¡ ® e1 x1 yl ¡ 1 e2 x 2 ; x 3 ; ¢¢¢; x n ¡ 0 yr e3 1 e03 xn x 0n 1 1+ ® 42 Lower bound for 2-facility randomized case Theorem. For any 2-facility randomized truthful mechanism, the approximation ratio is at least 1 a ¡ , where n ¸ 30 is the number of players 1:045 n¡ 3 e03 ¸ 1=2 ¡ 1=(n ¡ 2) ¡ ® Proof sc(I 0) = ¸ 0 sc(I ) ¸ = 1 1 ] ¢1 + Pr[yr ¸ ] ¢1 + e03 n¡ 2 n¡ 2 1 1 1 1 1 ¡ Pr[¡ < yr < ]+ ¡ ¡ ® n¡ 2 n¡ 2 2 n¡ 2 e1 + (n ¡ 2)e2 + e03 ¸ Pr[yr · ¡ 1 1 1 1 Pr[yr · ¡ ] ¢1 + Pr[yr ¸ ] ¢1 + Pr[¡ < yr < ] ¢(1 + ®) n¡ 2 n¡ 2 n¡ 2 n¡ 2 1 1 1 + ® ¢Pr[¡ < yr < ] n¡ 2 n¡ 2 e1 x1 yl ¡ 1 e2 x 2 ; x 3 ; ¢¢¢; x n ¡ 0 yr e3 1 e03 xn x 0n 1 1+ ® 43 Lower bound for 2-facility randomized case Theorem. For any 2-facility randomized truthful mechanism, the approximation ratio is at least 1 a ¡ , where n ¸ 30 is the number of players 1:045 n¡ 3 Proof p 2¡ 1 1 1 0 p ¡ sc(I ) ¸ 1 + > 1:045 ¡ n¡ 2 12 ¡ 2 2 n ¡ 2 opt(I 0) = 1 Done. e1 x1 yl ¡ 1 e2 x 2 ; x 3 ; ¢¢¢; x n ¡ 0 yr e3 1 e03 xn x 0n 1 1+ ® 44 A 4-approx. randomized mechanism for 2-facility game Mechanism. Choose i 2 f 1; 2; ¢¢¢; ng by random, then choose j 2 f 1; 2; ¢¢¢; ngwith probability jx i ¡ x j j Pn j 0= 1 jx i ¡ x j 0 j set two facilities at x i ; x j Truthfulness: only need to prove the following 2facility mechanism is truthful Set one facility at c, and the other facility at x i with probability P jc ¡ x i j n j = 1 jc ¡ x j j 45 Proof of truthfulness Truthfulness: only need to prove the following 2facility mechanism is truthful Set one facility at c, and the other facility at x i with probability jc ¡ x i j P n j = 1 jc ¡ Proof. For player i , P j6 =i cost = 0 cost = j6 =i jx i ¡ x 0ij ¸ jb¡ b0j S minf jx j ¡ x i j; jc ¡ x i jgjx j ¡ cj S P = A+ b j6 = i jx j ¡ cj + jx i ¡ cj A when misreporting to x 0i, P S · Ab xj j b S b b’ minf jx j ¡ x i j; jc ¡ x i jgjx j ¡ cj + minf jx i ¡ cj; jx i ¡ x 0ijgjx 0i ¡ cj P 0 j6 = i jx j ¡ cj + jx i ¡ cj S + minf b; jx i ¡ = A + b0 A x 0ijgb0 0 b’ S + minf b; jb¡ b jgb0 ¸ A + b0 46 Proof of truthfulness (cont’d) Truthfulness: only need to prove the following 2facility mechanism is truthful Set one facility at c, and the other facility at x i with probability jc ¡ x i j P Proof. n j = 1 jc ¡ xj j S · Ab jx i ¡ x 0ij ¸ jb¡ b0j S + minf b; jb¡ b0jgb0 S cost ¡ cost ¸ ¡ A + b0 ³ A+ b ´ 1 0 0 0 = minf b; jb¡ b jgb (A + b) ¡ S(b ¡ b) (A + b0)(A + b) ³ ´ (assume b0 ¸ b) 1 0 0 0 ¸ minf b; jb¡ b jgb (A + b) ¡ Ab(b ¡ b) (A + b0)(A + b) ³ ´ 1 0 0 bb (A + b) ¡ Ab(b ¡ b) ¸ 0 ½ (when b < jb¡ b0j) = (A + b0)(A + b) ³ ´ 1 0 0 0 0 (when b · jb¡ b j) = (b ¡ b)b (A + b) ¡ Ab(b ¡ b) ¸ 0 (A + b0)(A + b) 0 47 Approximation ratio Claim. The mechanism approximates the optimal social cost within a factor of 4. Intuition When locations are “sparse”, opt is also bad x1 x2 ¢¢¢ ¢¢¢ ¢¢¢ xn When locations fall into two groups, opt is small, but Mechanism behaves very similar to opt x 1 ; x 2 ; ¢¢¢; x n =2 x n =2+ 1 ; x n =2+ 2 ; ¢¢¢; x n 48 Open problems Characterization Deterministic 2-facility game? Randomized 1-facility game? Approximation Still some gaps… Randomized 3-facility game? 49 Thank you! 50