EECS 495: Combinatorial Optimization Mechanism Design with Rounding Lecture Manolis, Nima Rn−1 : E[vi (f (vi , v−i )) − pi (vi , v−i )] ≥ E[vi (f (bi , v−i )) − pi (bi , v−i )] Motivation • Make a social choice that (approximately) maximizes the social welfare subject to the economic constraints of truthfulness Problem Formulation Multi Unit Combinatorial Auctions • When optimizing the social welfare is NP-hard and the standard solution, the VCG mechanism, cannot be computed efficiently find truthful approximation algorithms • A set [m] of items • A number B representing how many copies we have of each item • A set [n] of players • Player i has valuation function vi : 2[m] → ℜ+ , vi (∅) = 0, vi (A) ≤ vi (B) for A ⊆ B (monotone) Mechanism Design Basics • A set of possible outcomes A and a set of bidders N . • Feasible solution: allocation (S1 , . . . , Sn ) such that each item is allocated to at most B players (these sets must be disjoint for B = 1) • For each i ∈ N a private function vi : A → R+ 0 • A Mechanism elicits bid bi from each i ∈ N and outputs the set f (b1 , . . . , bn ) ∈ A and and charges player I ∈ N : pi (b1 , . . . , bn ) • i’s value for (S1 , . . . , Sn ) is vi (Si ) • Goal: Maximize P v (S ) i i i Social Welfare: • We relax f and p to be random variables over the possible outcomes and the Rn Lotteries and Value Oracles payment vectors • Value oracle takes set and returns value • The expected utility of a player i ∈ • Analogous oracle for lotteries N is given by E[vi (f (bi , . . . , bn )) − • For x ∈ [0, 1]m , Dx define pi (b1 , . . . , bn )] X Y Y Fv (x) = ES∼Dx [v(S)] = v(S) xj (1 − xj ) • A mechanism is truthful in expectaS j∈S j ∈S / tion if for every bi 6= vi and v−i ∈ 1 • That is the value of the lottery can be computed as the expected value of the outcome w.r.t. the distribution of the lottery following max X vi (S)xi,S i,S6=∅ subject to A general technique for constructing truthful approximation algorithms X xi,S ≤ 1 for each player i S6=∅ X X i xi,S ≤ 1 for each item j S:j∈S xi,S ≥ 0 for each i, S Deterministic Support Mechanism • Any randomized mechanism can be viewed as a deterministic mechanism with a different outcome space: the set of all lotteries over the original outcomes, the deterministic support mechanism Theorem 0.1 Given any α-approximation algorithm for a problem with the ”packing property” that proves also an integrality gap of at most α there is a randomized truthful in expectation α-approximation algorithm • Given any truthful in expectation mechThe mechanism is defined as follows: anism the corresponding deterministic support mechanism that charges the ex1. Compute the fractional optimal solution pected payment vector is truthful x∗ • The reverse is also true. A similar rule Clark-Groves could be used to ensure individual rationality in expectation. 2. Scale the solution by α: x′ = x∗ /α 3. Express x′ as a convex combination of integer solutions where there are polynomially many non-zero weights • These transformations preserve the social welfare at any input 4. Convert the deterministic support allocation into its corresponding randomized version. The new framework • The packing property of a problem is Claim: This algorithm returns a truthful in that its corresponding natural LP sat- expectation randomized mechanism isfies the following fact: if vector x is a solution to the problem then any 0 ≤ Proof: x′ ≤ x is also a solution • Consider the mechanism that did not scale by α and could probably express • The mutli-unit combinatorial auction the best fractional solution as a convex problem has the packing property: combination of exponentially many integral ones • The LP formulation we will use is the 2 • This coincides with the VCG mechanism on lotteries therefore its randomization is also truthful • If the optimal solution to the primal is 1 then the variables λ correspond to the weights of the convex combination • The randomization obtained in the case where we scale by α has simply scaled the utility of each agent at every input, hence, the truthfulness equation is satisfied • Primal P has exponentially many variables • It has polynomially many constrains; Since x∗ is an extreme point of the original problem, it has at most m + n non zero variables, i.e., |E| ≤ m + n. 2 Claim: We can express the scaled allocation as a convex combination of polynomially many integral solutions • Equivalently, dual D has exponentially many constraints but polynomially many variables Proof: • We will show how we can solve the Dual restricting the feasible space only to the optimal solutions since then we have a separation oracle • Let I be the set of integral solutions • Let x∗ be the fractional optimal solution • Let E = {(i, S) | x∗i,S > 0} + = max(wi,S , 0). Given any Claim: Let wi,S integral x to the original problem one can find • we define the following primal P and its integral solution xl such that dual D X X + xli,S wi,S = x̂i,S wi,S (i,S)∈E (i,S)∈E Primal min s.t. P P l P l∈I l∈I λl Proof: λl xli,S = x∗i,S /α for all (i, S) ∈ E 1. if wi,S > 0 set xli,S = xi,S λl ≥ 1 2. Otherwise set xli,S = 0. λl ≥ 0 for all l ∈ I Dual max α1 s.t. P P (i,S)∈E (i,S)∈E 2 Claim: For any weight vector w we can compute in polynomial time an integral solution such that X X 1 max xi,S wi,S xli,S wi,S ≥ α feasible x x∗i,S wi,S + z xli,S wi,S + z ≤ 1 for all l ∈ I (i,S)∈E z≥0 Proof: λl ≥ 0 for all l ∈ I 3 (i,S)∈E • If some components of w are negative take w+ • First prove that the optimum of the dual is 1: Set z = 1 and wi,S = 0 for all (i, S) ∈ E. • We can monotonize ′ wi,S = max T ⊆S:(i,T )∈E • Assume we can find something better using solution x∗ : Then by the previous claim we can find an integer solution that is α approximate ofx∗ which implies the that corresponding constrain is violated + wi,T • Apply approximation algorithm with valuations w′ and get solution y • Using this argument we add the inequality 1 X ∗ · xi,S wi,S + z ≥ 1 α ′ • By construction wi,S ≥ wi,S for all (i, S) ∈ E. • Note that we don’t care about as singing weigths for (i, S) ∈ / E such the respective wi,S are zero. (i,S)∈E • Run Ellipsoid method on D to identify a dual program with polynomial size set of inequalities that is equivalent to D (the violated inequalities that are returned by the separation oracle that are used to cut the ellipsoid) • We know that X ′ yi,S wi,S (i,S)∈E is greater than α fraction of the objective • The primal of this program has polynomial number of constraints and variables. • We need to find another y such that the same holds with weights w+ • If we put some weight to some (i, S) such + ′ that wi,S 6= wi,S , i.e., we increased this value due to monotonicity constraints we put all of this weight then we simply put weight 1 to (i, T ) where T ⊆ S and + ′ wi,T = wi,S • Therefore non zero variables to the integer solution are at most the number of variables plus the number of the constrains. • The separation oracle used is the following: If • Use the previous claim to conclude that there is an integral solution xl such that is greater than α fraction of the objective when we use values w 1 X ∗ xi,S wi,S + z > 1 α (i,S)∈E then we identify the violated constrain according to the previous Claim or otherwise we use the half space that is defined by this constraint to cut the ellipsoid 2 Claim: We can find a convex combination of x∗ with polynomially many non zero weights α in polynomial time Proof: 2 4 2 Applications of this technique Convex Rounding and MIDR • MIDR: Fix some algorithm. Let distribution Dw be outcome for objective function w. Let D be the set of all possible Dw . The algorithm is MIDR if ∀w, D ∈ D, Ex∼Dw [w(x)] ≥ Ex∼D [w(x)] These technique can be used on existing approximation algorithms for the MUCA problem and derive the following results: • For short valuation (each player is inter1 ested in one set and its subsets) O(m B+! for any B ≥ 1 and (1 + ǫ) for B = Ω(log m). • We can convert any MIDR algorithm to a truthful mechanism, with the same approximation guarantee, using VCG payments • For general valuation functions the same bound but with with ex post Nash equilibrium as a solution concept • Observation: Instead of solving the relaxation and then rounding, why not optimize over the outcome of the rounding scheme? • For the special case of multi unit auctions this technique can be proved to be 2 approximate maximize E[w(r(x))] subject to x ∈ R Convex Rounding Framework (2) • We don’t know if it is tractable Relaxations Lemma 0.2 Program 2 is MIDR • Π an optimization problem, ∀(S, w) ∈ Π maximize w(x) subject to x ∈ S (1) • For most typical roundings, 2 is hard to solve • e.g., if r(x) = x for x ∈ S, then it is more general than 1 • Π′ relaxation: ∀(S, w) ∈ Π defines convex and compact relaxed feasible set R ∈ ℜm and an extension wR : R → ℜ of the objective • So we probably should have the unusual property that r(x) 6= x for x ∈ S • So we have the following • Rounding scheme r is α-approximate if w(x) ≥ E[w(r(x))] ≥ αw(x), ∀x ∈ S maximize wR (x) subject to x ∈ R Lemma 0.3 If r is α-approximate, then 2 is an α-approximation to the original problem 1 • A rounding scheme r : R → S (possibly randomized) • Call r convex if E[w(r(x))] concave • If ∀x ∈ R, E[w(r(x))] ≥ αwR (x), then this is an α-approximation • For r convex 2 can be solve efficiently 5 • So the problem is to find α-approximate convex rounding Poisson Rounding Lemma 0.6 Consider f : 2V → ℜ monotone, submodular, and normalizes (f (∅) = 0). Consider set S ⊆ V and random set S ′ by choosing each element of S independently with prob p. Then E[f (S ′ )] ≥ p · f (S). Theorem 0.4 If ∃ polynomial, αapproximate, convex r for Π′ , then ∃ truthful-in-expectation, polynomial, α-approximate mechanism for Π Combinatorial Auctions Proof: Matroid Rank Sum – Fix an ordering on elements of S • Set function v : 2[m] → ℜ is an MRS function if ∃u1 . . . , uk (all matroid rank + functions), and weights Pk w 1 , . . . , w k ∈ ℜ such that v(S) = ℓ=1 wℓ uℓ (S) – Let Si be the first i elements of S (similarly for Si′ ) P – f (S) = 1≤i≤|S| f (Si ) − f (Si−1 ), where f (S0 ) = 0 • Includes weighted coverage functions, matroid weighted rank functions, and all convex combinations of them E[f (S ′ )] = E[ X ′ f (Si′ ) − f (Si−1 )] 1≤i≤|S ′ | • Negative result: no universally truthful, polynomial, VCG-based mechanism achieves constant factor assuming N P * P |P oly ≥ X p · (f (Si ) − f (Si−1 )) 1≤i≤|S| = p · f (S) 2 Results • Now we define the Poisson Rounding Theorem 0.5 ∃(1−1/e)-approximate mechanism for combinatorial auctions with MRS valuations • Given fractional solution x, independently for each item assign j to i with prob 1 − e−xij • Formulation P • Note 1 − e−xij ≤ xij maximize w(x) = i vi ({j : xij = 1}) P subject to ∀j i xij ≤ 1 Lemma 0.7 Poisson rounding is (1 − xij ∈ {0, 1} ∀i, j 1/e)-approximate for submodular valuations • R is relaxation to 0 ≤ xij ≤ 1 Proof: • For x ∈ [0, 1]m , Dx define the extension of the value function – Rounding applied to integer soluX Y Y Fv (x) = ES∼Dx [v(S)] = v(S) xj (1 − xj ) tion cancels each allocated item S j∈S j ∈S / with prob 1/e 6 – Consider (S1 , . . . , Sn ) integer and corresponding (random) ′ ′ (S1 , . . . , Sn ) – Si′ includes any j ∈ Si independently with prob 1 − 1/e – Then E[vi (Si′ )] ≥ (1 − 1/e)vi (Si ) 2 Lemma 0.8 Poisson rounding is concave for MRS valuations Proof: • Let (S1 , . . . , Sn ) = rpoiss (x) • Want P to prove E[w(rposs (x))] E[ vi (Si )] concave = • Show E[vi (Si )] concave • We prove concavity for a subclass: Coverage functions Definition 0.1 A function f : 2V → ℜ is a coverage function if ∃ a set A (different from V ) of “activities”, a value vi for each activity i ∈ A, and P a set Xj ⊆ A for eachSj, such that v(S) = i∈c(S) vi , where c(S) = j∈S vj . Proof: X E[v(S)] = E[ vi ] i∈c(S) = X vi P r[i ∈ c(S)] i∈A For each i ∈ A, define Yi = {j ∈ S|i ∈ Xj }. We have P r[i ∈ c(S)] = P r[S ∩ Xi 6= ∅] = 1 − e− P j∈Yi xj 2 2 7